Spaces:
Potre1qw
/
Running on Zero

flatcherlee commited on
Commit
932ae62
1 Parent(s): cd8605c

Upload 273 files

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. ComfyUI/.pylintrc +3 -0
  2. ComfyUI/0.4.12 +29 -0
  3. ComfyUI/LICENSE +674 -0
  4. ComfyUI/README.md +224 -0
  5. ComfyUI/app/__init__.py +0 -0
  6. ComfyUI/app/app_settings.py +54 -0
  7. ComfyUI/app/frontend_management.py +188 -0
  8. ComfyUI/app/user_manager.py +205 -0
  9. ComfyUI/comfy/checkpoint_pickle.py +13 -0
  10. ComfyUI/comfy/cldm/cldm.py +437 -0
  11. ComfyUI/comfy/cldm/control_types.py +10 -0
  12. ComfyUI/comfy/cldm/mmdit.py +77 -0
  13. ComfyUI/comfy/cli_args.py +180 -0
  14. ComfyUI/comfy/clip_config_bigg.json +23 -0
  15. ComfyUI/comfy/clip_model.py +196 -0
  16. ComfyUI/comfy/clip_vision.py +121 -0
  17. ComfyUI/comfy/clip_vision_config_g.json +18 -0
  18. ComfyUI/comfy/clip_vision_config_h.json +18 -0
  19. ComfyUI/comfy/clip_vision_config_vitl.json +18 -0
  20. ComfyUI/comfy/clip_vision_config_vitl_336.json +18 -0
  21. ComfyUI/comfy/conds.py +83 -0
  22. ComfyUI/comfy/controlnet.py +622 -0
  23. ComfyUI/comfy/diffusers_convert.py +281 -0
  24. ComfyUI/comfy/diffusers_load.py +36 -0
  25. ComfyUI/comfy/extra_samplers/uni_pc.py +875 -0
  26. ComfyUI/comfy/gligen.py +343 -0
  27. ComfyUI/comfy/k_diffusion/deis.py +121 -0
  28. ComfyUI/comfy/k_diffusion/sampling.py +1050 -0
  29. ComfyUI/comfy/k_diffusion/utils.py +313 -0
  30. ComfyUI/comfy/latent_formats.py +170 -0
  31. ComfyUI/comfy/ldm/audio/autoencoder.py +282 -0
  32. ComfyUI/comfy/ldm/audio/dit.py +891 -0
  33. ComfyUI/comfy/ldm/audio/embedders.py +108 -0
  34. ComfyUI/comfy/ldm/aura/mmdit.py +478 -0
  35. ComfyUI/comfy/ldm/cascade/common.py +154 -0
  36. ComfyUI/comfy/ldm/cascade/controlnet.py +93 -0
  37. ComfyUI/comfy/ldm/cascade/stage_a.py +255 -0
  38. ComfyUI/comfy/ldm/cascade/stage_b.py +256 -0
  39. ComfyUI/comfy/ldm/cascade/stage_c.py +273 -0
  40. ComfyUI/comfy/ldm/cascade/stage_c_coder.py +95 -0
  41. ComfyUI/comfy/ldm/common_dit.py +8 -0
  42. ComfyUI/comfy/ldm/flux/layers.py +263 -0
  43. ComfyUI/comfy/ldm/flux/math.py +35 -0
  44. ComfyUI/comfy/ldm/flux/model.py +142 -0
  45. ComfyUI/comfy/ldm/hydit/attn_layers.py +219 -0
  46. ComfyUI/comfy/ldm/hydit/models.py +405 -0
  47. ComfyUI/comfy/ldm/hydit/poolers.py +37 -0
  48. ComfyUI/comfy/ldm/hydit/posemb_layers.py +224 -0
  49. ComfyUI/comfy/ldm/models/autoencoder.py +226 -0
  50. ComfyUI/comfy/ldm/modules/attention.py +865 -0
ComfyUI/.pylintrc ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ [MESSAGES CONTROL]
2
+ disable=all
3
+ enable=eval-used
ComfyUI/0.4.12 ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Collecting timm
2
+ Downloading timm-1.0.9-py3-none-any.whl.metadata (42 kB)
3
+ Requirement already satisfied: torch in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from timm) (2.4.0)
4
+ Requirement already satisfied: torchvision in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from timm) (0.19.0)
5
+ Requirement already satisfied: pyyaml in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from timm) (6.0.2)
6
+ Requirement already satisfied: huggingface_hub in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from timm) (0.24.6)
7
+ Requirement already satisfied: safetensors in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from timm) (0.4.4)
8
+ Requirement already satisfied: filelock in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (3.15.4)
9
+ Requirement already satisfied: fsspec>=2023.5.0 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (2024.6.1)
10
+ Requirement already satisfied: packaging>=20.9 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (24.1)
11
+ Requirement already satisfied: requests in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (2.32.3)
12
+ Requirement already satisfied: tqdm>=4.42.1 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (4.66.5)
13
+ Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from huggingface_hub->timm) (4.12.2)
14
+ Requirement already satisfied: sympy in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from torch->timm) (1.13.2)
15
+ Requirement already satisfied: networkx in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from torch->timm) (3.3)
16
+ Requirement already satisfied: jinja2 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from torch->timm) (3.1.4)
17
+ Requirement already satisfied: numpy<2 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from torchvision->timm) (1.26.4)
18
+ Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from torchvision->timm) (9.5.0)
19
+ Requirement already satisfied: colorama in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from tqdm>=4.42.1->huggingface_hub->timm) (0.4.6)
20
+ Requirement already satisfied: MarkupSafe>=2.0 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from jinja2->torch->timm) (2.1.5)
21
+ Requirement already satisfied: charset-normalizer<4,>=2 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from requests->huggingface_hub->timm) (3.3.2)
22
+ Requirement already satisfied: idna<4,>=2.5 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from requests->huggingface_hub->timm) (3.8)
23
+ Requirement already satisfied: urllib3<3,>=1.21.1 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from requests->huggingface_hub->timm) (2.2.2)
24
+ Requirement already satisfied: certifi>=2017.4.17 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from requests->huggingface_hub->timm) (2024.7.4)
25
+ Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\users\ilya9\anaconda3\envs\comf_upscaler\lib\site-packages (from sympy->torch->timm) (1.3.0)
26
+ Downloading timm-1.0.9-py3-none-any.whl (2.3 MB)
27
+ ---------------------------------------- 2.3/2.3 MB 6.0 MB/s eta 0:00:00
28
+ Installing collected packages: timm
29
+ Successfully installed timm-1.0.9
ComfyUI/LICENSE ADDED
@@ -0,0 +1,674 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ GNU GENERAL PUBLIC LICENSE
2
+ Version 3, 29 June 2007
3
+
4
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
5
+ Everyone is permitted to copy and distribute verbatim copies
6
+ of this license document, but changing it is not allowed.
7
+
8
+ Preamble
9
+
10
+ The GNU General Public License is a free, copyleft license for
11
+ software and other kinds of works.
12
+
13
+ The licenses for most software and other practical works are designed
14
+ to take away your freedom to share and change the works. By contrast,
15
+ the GNU General Public License is intended to guarantee your freedom to
16
+ share and change all versions of a program--to make sure it remains free
17
+ software for all its users. We, the Free Software Foundation, use the
18
+ GNU General Public License for most of our software; it applies also to
19
+ any other work released this way by its authors. You can apply it to
20
+ your programs, too.
21
+
22
+ When we speak of free software, we are referring to freedom, not
23
+ price. Our General Public Licenses are designed to make sure that you
24
+ have the freedom to distribute copies of free software (and charge for
25
+ them if you wish), that you receive source code or can get it if you
26
+ want it, that you can change the software or use pieces of it in new
27
+ free programs, and that you know you can do these things.
28
+
29
+ To protect your rights, we need to prevent others from denying you
30
+ these rights or asking you to surrender the rights. Therefore, you have
31
+ certain responsibilities if you distribute copies of the software, or if
32
+ you modify it: responsibilities to respect the freedom of others.
33
+
34
+ For example, if you distribute copies of such a program, whether
35
+ gratis or for a fee, you must pass on to the recipients the same
36
+ freedoms that you received. You must make sure that they, too, receive
37
+ or can get the source code. And you must show them these terms so they
38
+ know their rights.
39
+
40
+ Developers that use the GNU GPL protect your rights with two steps:
41
+ (1) assert copyright on the software, and (2) offer you this License
42
+ giving you legal permission to copy, distribute and/or modify it.
43
+
44
+ For the developers' and authors' protection, the GPL clearly explains
45
+ that there is no warranty for this free software. For both users' and
46
+ authors' sake, the GPL requires that modified versions be marked as
47
+ changed, so that their problems will not be attributed erroneously to
48
+ authors of previous versions.
49
+
50
+ Some devices are designed to deny users access to install or run
51
+ modified versions of the software inside them, although the manufacturer
52
+ can do so. This is fundamentally incompatible with the aim of
53
+ protecting users' freedom to change the software. The systematic
54
+ pattern of such abuse occurs in the area of products for individuals to
55
+ use, which is precisely where it is most unacceptable. Therefore, we
56
+ have designed this version of the GPL to prohibit the practice for those
57
+ products. If such problems arise substantially in other domains, we
58
+ stand ready to extend this provision to those domains in future versions
59
+ of the GPL, as needed to protect the freedom of users.
60
+
61
+ Finally, every program is threatened constantly by software patents.
62
+ States should not allow patents to restrict development and use of
63
+ software on general-purpose computers, but in those that do, we wish to
64
+ avoid the special danger that patents applied to a free program could
65
+ make it effectively proprietary. To prevent this, the GPL assures that
66
+ patents cannot be used to render the program non-free.
67
+
68
+ The precise terms and conditions for copying, distribution and
69
+ modification follow.
70
+
71
+ TERMS AND CONDITIONS
72
+
73
+ 0. Definitions.
74
+
75
+ "This License" refers to version 3 of the GNU General Public License.
76
+
77
+ "Copyright" also means copyright-like laws that apply to other kinds of
78
+ works, such as semiconductor masks.
79
+
80
+ "The Program" refers to any copyrightable work licensed under this
81
+ License. Each licensee is addressed as "you". "Licensees" and
82
+ "recipients" may be individuals or organizations.
83
+
84
+ To "modify" a work means to copy from or adapt all or part of the work
85
+ in a fashion requiring copyright permission, other than the making of an
86
+ exact copy. The resulting work is called a "modified version" of the
87
+ earlier work or a work "based on" the earlier work.
88
+
89
+ A "covered work" means either the unmodified Program or a work based
90
+ on the Program.
91
+
92
+ To "propagate" a work means to do anything with it that, without
93
+ permission, would make you directly or secondarily liable for
94
+ infringement under applicable copyright law, except executing it on a
95
+ computer or modifying a private copy. Propagation includes copying,
96
+ distribution (with or without modification), making available to the
97
+ public, and in some countries other activities as well.
98
+
99
+ To "convey" a work means any kind of propagation that enables other
100
+ parties to make or receive copies. Mere interaction with a user through
101
+ a computer network, with no transfer of a copy, is not conveying.
102
+
103
+ An interactive user interface displays "Appropriate Legal Notices"
104
+ to the extent that it includes a convenient and prominently visible
105
+ feature that (1) displays an appropriate copyright notice, and (2)
106
+ tells the user that there is no warranty for the work (except to the
107
+ extent that warranties are provided), that licensees may convey the
108
+ work under this License, and how to view a copy of this License. If
109
+ the interface presents a list of user commands or options, such as a
110
+ menu, a prominent item in the list meets this criterion.
111
+
112
+ 1. Source Code.
113
+
114
+ The "source code" for a work means the preferred form of the work
115
+ for making modifications to it. "Object code" means any non-source
116
+ form of a work.
117
+
118
+ A "Standard Interface" means an interface that either is an official
119
+ standard defined by a recognized standards body, or, in the case of
120
+ interfaces specified for a particular programming language, one that
121
+ is widely used among developers working in that language.
122
+
123
+ The "System Libraries" of an executable work include anything, other
124
+ than the work as a whole, that (a) is included in the normal form of
125
+ packaging a Major Component, but which is not part of that Major
126
+ Component, and (b) serves only to enable use of the work with that
127
+ Major Component, or to implement a Standard Interface for which an
128
+ implementation is available to the public in source code form. A
129
+ "Major Component", in this context, means a major essential component
130
+ (kernel, window system, and so on) of the specific operating system
131
+ (if any) on which the executable work runs, or a compiler used to
132
+ produce the work, or an object code interpreter used to run it.
133
+
134
+ The "Corresponding Source" for a work in object code form means all
135
+ the source code needed to generate, install, and (for an executable
136
+ work) run the object code and to modify the work, including scripts to
137
+ control those activities. However, it does not include the work's
138
+ System Libraries, or general-purpose tools or generally available free
139
+ programs which are used unmodified in performing those activities but
140
+ which are not part of the work. For example, Corresponding Source
141
+ includes interface definition files associated with source files for
142
+ the work, and the source code for shared libraries and dynamically
143
+ linked subprograms that the work is specifically designed to require,
144
+ such as by intimate data communication or control flow between those
145
+ subprograms and other parts of the work.
146
+
147
+ The Corresponding Source need not include anything that users
148
+ can regenerate automatically from other parts of the Corresponding
149
+ Source.
150
+
151
+ The Corresponding Source for a work in source code form is that
152
+ same work.
153
+
154
+ 2. Basic Permissions.
155
+
156
+ All rights granted under this License are granted for the term of
157
+ copyright on the Program, and are irrevocable provided the stated
158
+ conditions are met. This License explicitly affirms your unlimited
159
+ permission to run the unmodified Program. The output from running a
160
+ covered work is covered by this License only if the output, given its
161
+ content, constitutes a covered work. This License acknowledges your
162
+ rights of fair use or other equivalent, as provided by copyright law.
163
+
164
+ You may make, run and propagate covered works that you do not
165
+ convey, without conditions so long as your license otherwise remains
166
+ in force. You may convey covered works to others for the sole purpose
167
+ of having them make modifications exclusively for you, or provide you
168
+ with facilities for running those works, provided that you comply with
169
+ the terms of this License in conveying all material for which you do
170
+ not control copyright. Those thus making or running the covered works
171
+ for you must do so exclusively on your behalf, under your direction
172
+ and control, on terms that prohibit them from making any copies of
173
+ your copyrighted material outside their relationship with you.
174
+
175
+ Conveying under any other circumstances is permitted solely under
176
+ the conditions stated below. Sublicensing is not allowed; section 10
177
+ makes it unnecessary.
178
+
179
+ 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180
+
181
+ No covered work shall be deemed part of an effective technological
182
+ measure under any applicable law fulfilling obligations under article
183
+ 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184
+ similar laws prohibiting or restricting circumvention of such
185
+ measures.
186
+
187
+ When you convey a covered work, you waive any legal power to forbid
188
+ circumvention of technological measures to the extent such circumvention
189
+ is effected by exercising rights under this License with respect to
190
+ the covered work, and you disclaim any intention to limit operation or
191
+ modification of the work as a means of enforcing, against the work's
192
+ users, your or third parties' legal rights to forbid circumvention of
193
+ technological measures.
194
+
195
+ 4. Conveying Verbatim Copies.
196
+
197
+ You may convey verbatim copies of the Program's source code as you
198
+ receive it, in any medium, provided that you conspicuously and
199
+ appropriately publish on each copy an appropriate copyright notice;
200
+ keep intact all notices stating that this License and any
201
+ non-permissive terms added in accord with section 7 apply to the code;
202
+ keep intact all notices of the absence of any warranty; and give all
203
+ recipients a copy of this License along with the Program.
204
+
205
+ You may charge any price or no price for each copy that you convey,
206
+ and you may offer support or warranty protection for a fee.
207
+
208
+ 5. Conveying Modified Source Versions.
209
+
210
+ You may convey a work based on the Program, or the modifications to
211
+ produce it from the Program, in the form of source code under the
212
+ terms of section 4, provided that you also meet all of these conditions:
213
+
214
+ a) The work must carry prominent notices stating that you modified
215
+ it, and giving a relevant date.
216
+
217
+ b) The work must carry prominent notices stating that it is
218
+ released under this License and any conditions added under section
219
+ 7. This requirement modifies the requirement in section 4 to
220
+ "keep intact all notices".
221
+
222
+ c) You must license the entire work, as a whole, under this
223
+ License to anyone who comes into possession of a copy. This
224
+ License will therefore apply, along with any applicable section 7
225
+ additional terms, to the whole of the work, and all its parts,
226
+ regardless of how they are packaged. This License gives no
227
+ permission to license the work in any other way, but it does not
228
+ invalidate such permission if you have separately received it.
229
+
230
+ d) If the work has interactive user interfaces, each must display
231
+ Appropriate Legal Notices; however, if the Program has interactive
232
+ interfaces that do not display Appropriate Legal Notices, your
233
+ work need not make them do so.
234
+
235
+ A compilation of a covered work with other separate and independent
236
+ works, which are not by their nature extensions of the covered work,
237
+ and which are not combined with it such as to form a larger program,
238
+ in or on a volume of a storage or distribution medium, is called an
239
+ "aggregate" if the compilation and its resulting copyright are not
240
+ used to limit the access or legal rights of the compilation's users
241
+ beyond what the individual works permit. Inclusion of a covered work
242
+ in an aggregate does not cause this License to apply to the other
243
+ parts of the aggregate.
244
+
245
+ 6. Conveying Non-Source Forms.
246
+
247
+ You may convey a covered work in object code form under the terms
248
+ of sections 4 and 5, provided that you also convey the
249
+ machine-readable Corresponding Source under the terms of this License,
250
+ in one of these ways:
251
+
252
+ a) Convey the object code in, or embodied in, a physical product
253
+ (including a physical distribution medium), accompanied by the
254
+ Corresponding Source fixed on a durable physical medium
255
+ customarily used for software interchange.
256
+
257
+ b) Convey the object code in, or embodied in, a physical product
258
+ (including a physical distribution medium), accompanied by a
259
+ written offer, valid for at least three years and valid for as
260
+ long as you offer spare parts or customer support for that product
261
+ model, to give anyone who possesses the object code either (1) a
262
+ copy of the Corresponding Source for all the software in the
263
+ product that is covered by this License, on a durable physical
264
+ medium customarily used for software interchange, for a price no
265
+ more than your reasonable cost of physically performing this
266
+ conveying of source, or (2) access to copy the
267
+ Corresponding Source from a network server at no charge.
268
+
269
+ c) Convey individual copies of the object code with a copy of the
270
+ written offer to provide the Corresponding Source. This
271
+ alternative is allowed only occasionally and noncommercially, and
272
+ only if you received the object code with such an offer, in accord
273
+ with subsection 6b.
274
+
275
+ d) Convey the object code by offering access from a designated
276
+ place (gratis or for a charge), and offer equivalent access to the
277
+ Corresponding Source in the same way through the same place at no
278
+ further charge. You need not require recipients to copy the
279
+ Corresponding Source along with the object code. If the place to
280
+ copy the object code is a network server, the Corresponding Source
281
+ may be on a different server (operated by you or a third party)
282
+ that supports equivalent copying facilities, provided you maintain
283
+ clear directions next to the object code saying where to find the
284
+ Corresponding Source. Regardless of what server hosts the
285
+ Corresponding Source, you remain obligated to ensure that it is
286
+ available for as long as needed to satisfy these requirements.
287
+
288
+ e) Convey the object code using peer-to-peer transmission, provided
289
+ you inform other peers where the object code and Corresponding
290
+ Source of the work are being offered to the general public at no
291
+ charge under subsection 6d.
292
+
293
+ A separable portion of the object code, whose source code is excluded
294
+ from the Corresponding Source as a System Library, need not be
295
+ included in conveying the object code work.
296
+
297
+ A "User Product" is either (1) a "consumer product", which means any
298
+ tangible personal property which is normally used for personal, family,
299
+ or household purposes, or (2) anything designed or sold for incorporation
300
+ into a dwelling. In determining whether a product is a consumer product,
301
+ doubtful cases shall be resolved in favor of coverage. For a particular
302
+ product received by a particular user, "normally used" refers to a
303
+ typical or common use of that class of product, regardless of the status
304
+ of the particular user or of the way in which the particular user
305
+ actually uses, or expects or is expected to use, the product. A product
306
+ is a consumer product regardless of whether the product has substantial
307
+ commercial, industrial or non-consumer uses, unless such uses represent
308
+ the only significant mode of use of the product.
309
+
310
+ "Installation Information" for a User Product means any methods,
311
+ procedures, authorization keys, or other information required to install
312
+ and execute modified versions of a covered work in that User Product from
313
+ a modified version of its Corresponding Source. The information must
314
+ suffice to ensure that the continued functioning of the modified object
315
+ code is in no case prevented or interfered with solely because
316
+ modification has been made.
317
+
318
+ If you convey an object code work under this section in, or with, or
319
+ specifically for use in, a User Product, and the conveying occurs as
320
+ part of a transaction in which the right of possession and use of the
321
+ User Product is transferred to the recipient in perpetuity or for a
322
+ fixed term (regardless of how the transaction is characterized), the
323
+ Corresponding Source conveyed under this section must be accompanied
324
+ by the Installation Information. But this requirement does not apply
325
+ if neither you nor any third party retains the ability to install
326
+ modified object code on the User Product (for example, the work has
327
+ been installed in ROM).
328
+
329
+ The requirement to provide Installation Information does not include a
330
+ requirement to continue to provide support service, warranty, or updates
331
+ for a work that has been modified or installed by the recipient, or for
332
+ the User Product in which it has been modified or installed. Access to a
333
+ network may be denied when the modification itself materially and
334
+ adversely affects the operation of the network or violates the rules and
335
+ protocols for communication across the network.
336
+
337
+ Corresponding Source conveyed, and Installation Information provided,
338
+ in accord with this section must be in a format that is publicly
339
+ documented (and with an implementation available to the public in
340
+ source code form), and must require no special password or key for
341
+ unpacking, reading or copying.
342
+
343
+ 7. Additional Terms.
344
+
345
+ "Additional permissions" are terms that supplement the terms of this
346
+ License by making exceptions from one or more of its conditions.
347
+ Additional permissions that are applicable to the entire Program shall
348
+ be treated as though they were included in this License, to the extent
349
+ that they are valid under applicable law. If additional permissions
350
+ apply only to part of the Program, that part may be used separately
351
+ under those permissions, but the entire Program remains governed by
352
+ this License without regard to the additional permissions.
353
+
354
+ When you convey a copy of a covered work, you may at your option
355
+ remove any additional permissions from that copy, or from any part of
356
+ it. (Additional permissions may be written to require their own
357
+ removal in certain cases when you modify the work.) You may place
358
+ additional permissions on material, added by you to a covered work,
359
+ for which you have or can give appropriate copyright permission.
360
+
361
+ Notwithstanding any other provision of this License, for material you
362
+ add to a covered work, you may (if authorized by the copyright holders of
363
+ that material) supplement the terms of this License with terms:
364
+
365
+ a) Disclaiming warranty or limiting liability differently from the
366
+ terms of sections 15 and 16 of this License; or
367
+
368
+ b) Requiring preservation of specified reasonable legal notices or
369
+ author attributions in that material or in the Appropriate Legal
370
+ Notices displayed by works containing it; or
371
+
372
+ c) Prohibiting misrepresentation of the origin of that material, or
373
+ requiring that modified versions of such material be marked in
374
+ reasonable ways as different from the original version; or
375
+
376
+ d) Limiting the use for publicity purposes of names of licensors or
377
+ authors of the material; or
378
+
379
+ e) Declining to grant rights under trademark law for use of some
380
+ trade names, trademarks, or service marks; or
381
+
382
+ f) Requiring indemnification of licensors and authors of that
383
+ material by anyone who conveys the material (or modified versions of
384
+ it) with contractual assumptions of liability to the recipient, for
385
+ any liability that these contractual assumptions directly impose on
386
+ those licensors and authors.
387
+
388
+ All other non-permissive additional terms are considered "further
389
+ restrictions" within the meaning of section 10. If the Program as you
390
+ received it, or any part of it, contains a notice stating that it is
391
+ governed by this License along with a term that is a further
392
+ restriction, you may remove that term. If a license document contains
393
+ a further restriction but permits relicensing or conveying under this
394
+ License, you may add to a covered work material governed by the terms
395
+ of that license document, provided that the further restriction does
396
+ not survive such relicensing or conveying.
397
+
398
+ If you add terms to a covered work in accord with this section, you
399
+ must place, in the relevant source files, a statement of the
400
+ additional terms that apply to those files, or a notice indicating
401
+ where to find the applicable terms.
402
+
403
+ Additional terms, permissive or non-permissive, may be stated in the
404
+ form of a separately written license, or stated as exceptions;
405
+ the above requirements apply either way.
406
+
407
+ 8. Termination.
408
+
409
+ You may not propagate or modify a covered work except as expressly
410
+ provided under this License. Any attempt otherwise to propagate or
411
+ modify it is void, and will automatically terminate your rights under
412
+ this License (including any patent licenses granted under the third
413
+ paragraph of section 11).
414
+
415
+ However, if you cease all violation of this License, then your
416
+ license from a particular copyright holder is reinstated (a)
417
+ provisionally, unless and until the copyright holder explicitly and
418
+ finally terminates your license, and (b) permanently, if the copyright
419
+ holder fails to notify you of the violation by some reasonable means
420
+ prior to 60 days after the cessation.
421
+
422
+ Moreover, your license from a particular copyright holder is
423
+ reinstated permanently if the copyright holder notifies you of the
424
+ violation by some reasonable means, this is the first time you have
425
+ received notice of violation of this License (for any work) from that
426
+ copyright holder, and you cure the violation prior to 30 days after
427
+ your receipt of the notice.
428
+
429
+ Termination of your rights under this section does not terminate the
430
+ licenses of parties who have received copies or rights from you under
431
+ this License. If your rights have been terminated and not permanently
432
+ reinstated, you do not qualify to receive new licenses for the same
433
+ material under section 10.
434
+
435
+ 9. Acceptance Not Required for Having Copies.
436
+
437
+ You are not required to accept this License in order to receive or
438
+ run a copy of the Program. Ancillary propagation of a covered work
439
+ occurring solely as a consequence of using peer-to-peer transmission
440
+ to receive a copy likewise does not require acceptance. However,
441
+ nothing other than this License grants you permission to propagate or
442
+ modify any covered work. These actions infringe copyright if you do
443
+ not accept this License. Therefore, by modifying or propagating a
444
+ covered work, you indicate your acceptance of this License to do so.
445
+
446
+ 10. Automatic Licensing of Downstream Recipients.
447
+
448
+ Each time you convey a covered work, the recipient automatically
449
+ receives a license from the original licensors, to run, modify and
450
+ propagate that work, subject to this License. You are not responsible
451
+ for enforcing compliance by third parties with this License.
452
+
453
+ An "entity transaction" is a transaction transferring control of an
454
+ organization, or substantially all assets of one, or subdividing an
455
+ organization, or merging organizations. If propagation of a covered
456
+ work results from an entity transaction, each party to that
457
+ transaction who receives a copy of the work also receives whatever
458
+ licenses to the work the party's predecessor in interest had or could
459
+ give under the previous paragraph, plus a right to possession of the
460
+ Corresponding Source of the work from the predecessor in interest, if
461
+ the predecessor has it or can get it with reasonable efforts.
462
+
463
+ You may not impose any further restrictions on the exercise of the
464
+ rights granted or affirmed under this License. For example, you may
465
+ not impose a license fee, royalty, or other charge for exercise of
466
+ rights granted under this License, and you may not initiate litigation
467
+ (including a cross-claim or counterclaim in a lawsuit) alleging that
468
+ any patent claim is infringed by making, using, selling, offering for
469
+ sale, or importing the Program or any portion of it.
470
+
471
+ 11. Patents.
472
+
473
+ A "contributor" is a copyright holder who authorizes use under this
474
+ License of the Program or a work on which the Program is based. The
475
+ work thus licensed is called the contributor's "contributor version".
476
+
477
+ A contributor's "essential patent claims" are all patent claims
478
+ owned or controlled by the contributor, whether already acquired or
479
+ hereafter acquired, that would be infringed by some manner, permitted
480
+ by this License, of making, using, or selling its contributor version,
481
+ but do not include claims that would be infringed only as a
482
+ consequence of further modification of the contributor version. For
483
+ purposes of this definition, "control" includes the right to grant
484
+ patent sublicenses in a manner consistent with the requirements of
485
+ this License.
486
+
487
+ Each contributor grants you a non-exclusive, worldwide, royalty-free
488
+ patent license under the contributor's essential patent claims, to
489
+ make, use, sell, offer for sale, import and otherwise run, modify and
490
+ propagate the contents of its contributor version.
491
+
492
+ In the following three paragraphs, a "patent license" is any express
493
+ agreement or commitment, however denominated, not to enforce a patent
494
+ (such as an express permission to practice a patent or covenant not to
495
+ sue for patent infringement). To "grant" such a patent license to a
496
+ party means to make such an agreement or commitment not to enforce a
497
+ patent against the party.
498
+
499
+ If you convey a covered work, knowingly relying on a patent license,
500
+ and the Corresponding Source of the work is not available for anyone
501
+ to copy, free of charge and under the terms of this License, through a
502
+ publicly available network server or other readily accessible means,
503
+ then you must either (1) cause the Corresponding Source to be so
504
+ available, or (2) arrange to deprive yourself of the benefit of the
505
+ patent license for this particular work, or (3) arrange, in a manner
506
+ consistent with the requirements of this License, to extend the patent
507
+ license to downstream recipients. "Knowingly relying" means you have
508
+ actual knowledge that, but for the patent license, your conveying the
509
+ covered work in a country, or your recipient's use of the covered work
510
+ in a country, would infringe one or more identifiable patents in that
511
+ country that you have reason to believe are valid.
512
+
513
+ If, pursuant to or in connection with a single transaction or
514
+ arrangement, you convey, or propagate by procuring conveyance of, a
515
+ covered work, and grant a patent license to some of the parties
516
+ receiving the covered work authorizing them to use, propagate, modify
517
+ or convey a specific copy of the covered work, then the patent license
518
+ you grant is automatically extended to all recipients of the covered
519
+ work and works based on it.
520
+
521
+ A patent license is "discriminatory" if it does not include within
522
+ the scope of its coverage, prohibits the exercise of, or is
523
+ conditioned on the non-exercise of one or more of the rights that are
524
+ specifically granted under this License. You may not convey a covered
525
+ work if you are a party to an arrangement with a third party that is
526
+ in the business of distributing software, under which you make payment
527
+ to the third party based on the extent of your activity of conveying
528
+ the work, and under which the third party grants, to any of the
529
+ parties who would receive the covered work from you, a discriminatory
530
+ patent license (a) in connection with copies of the covered work
531
+ conveyed by you (or copies made from those copies), or (b) primarily
532
+ for and in connection with specific products or compilations that
533
+ contain the covered work, unless you entered into that arrangement,
534
+ or that patent license was granted, prior to 28 March 2007.
535
+
536
+ Nothing in this License shall be construed as excluding or limiting
537
+ any implied license or other defenses to infringement that may
538
+ otherwise be available to you under applicable patent law.
539
+
540
+ 12. No Surrender of Others' Freedom.
541
+
542
+ If conditions are imposed on you (whether by court order, agreement or
543
+ otherwise) that contradict the conditions of this License, they do not
544
+ excuse you from the conditions of this License. If you cannot convey a
545
+ covered work so as to satisfy simultaneously your obligations under this
546
+ License and any other pertinent obligations, then as a consequence you may
547
+ not convey it at all. For example, if you agree to terms that obligate you
548
+ to collect a royalty for further conveying from those to whom you convey
549
+ the Program, the only way you could satisfy both those terms and this
550
+ License would be to refrain entirely from conveying the Program.
551
+
552
+ 13. Use with the GNU Affero General Public License.
553
+
554
+ Notwithstanding any other provision of this License, you have
555
+ permission to link or combine any covered work with a work licensed
556
+ under version 3 of the GNU Affero General Public License into a single
557
+ combined work, and to convey the resulting work. The terms of this
558
+ License will continue to apply to the part which is the covered work,
559
+ but the special requirements of the GNU Affero General Public License,
560
+ section 13, concerning interaction through a network will apply to the
561
+ combination as such.
562
+
563
+ 14. Revised Versions of this License.
564
+
565
+ The Free Software Foundation may publish revised and/or new versions of
566
+ the GNU General Public License from time to time. Such new versions will
567
+ be similar in spirit to the present version, but may differ in detail to
568
+ address new problems or concerns.
569
+
570
+ Each version is given a distinguishing version number. If the
571
+ Program specifies that a certain numbered version of the GNU General
572
+ Public License "or any later version" applies to it, you have the
573
+ option of following the terms and conditions either of that numbered
574
+ version or of any later version published by the Free Software
575
+ Foundation. If the Program does not specify a version number of the
576
+ GNU General Public License, you may choose any version ever published
577
+ by the Free Software Foundation.
578
+
579
+ If the Program specifies that a proxy can decide which future
580
+ versions of the GNU General Public License can be used, that proxy's
581
+ public statement of acceptance of a version permanently authorizes you
582
+ to choose that version for the Program.
583
+
584
+ Later license versions may give you additional or different
585
+ permissions. However, no additional obligations are imposed on any
586
+ author or copyright holder as a result of your choosing to follow a
587
+ later version.
588
+
589
+ 15. Disclaimer of Warranty.
590
+
591
+ THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592
+ APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593
+ HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594
+ OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595
+ THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596
+ PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597
+ IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598
+ ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599
+
600
+ 16. Limitation of Liability.
601
+
602
+ IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603
+ WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604
+ THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605
+ GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606
+ USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607
+ DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608
+ PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609
+ EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610
+ SUCH DAMAGES.
611
+
612
+ 17. Interpretation of Sections 15 and 16.
613
+
614
+ If the disclaimer of warranty and limitation of liability provided
615
+ above cannot be given local legal effect according to their terms,
616
+ reviewing courts shall apply local law that most closely approximates
617
+ an absolute waiver of all civil liability in connection with the
618
+ Program, unless a warranty or assumption of liability accompanies a
619
+ copy of the Program in return for a fee.
620
+
621
+ END OF TERMS AND CONDITIONS
622
+
623
+ How to Apply These Terms to Your New Programs
624
+
625
+ If you develop a new program, and you want it to be of the greatest
626
+ possible use to the public, the best way to achieve this is to make it
627
+ free software which everyone can redistribute and change under these terms.
628
+
629
+ To do so, attach the following notices to the program. It is safest
630
+ to attach them to the start of each source file to most effectively
631
+ state the exclusion of warranty; and each file should have at least
632
+ the "copyright" line and a pointer to where the full notice is found.
633
+
634
+ <one line to give the program's name and a brief idea of what it does.>
635
+ Copyright (C) <year> <name of author>
636
+
637
+ This program is free software: you can redistribute it and/or modify
638
+ it under the terms of the GNU General Public License as published by
639
+ the Free Software Foundation, either version 3 of the License, or
640
+ (at your option) any later version.
641
+
642
+ This program is distributed in the hope that it will be useful,
643
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
644
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645
+ GNU General Public License for more details.
646
+
647
+ You should have received a copy of the GNU General Public License
648
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
649
+
650
+ Also add information on how to contact you by electronic and paper mail.
651
+
652
+ If the program does terminal interaction, make it output a short
653
+ notice like this when it starts in an interactive mode:
654
+
655
+ <program> Copyright (C) <year> <name of author>
656
+ This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657
+ This is free software, and you are welcome to redistribute it
658
+ under certain conditions; type `show c' for details.
659
+
660
+ The hypothetical commands `show w' and `show c' should show the appropriate
661
+ parts of the General Public License. Of course, your program's commands
662
+ might be different; for a GUI interface, you would use an "about box".
663
+
664
+ You should also get your employer (if you work as a programmer) or school,
665
+ if any, to sign a "copyright disclaimer" for the program, if necessary.
666
+ For more information on this, and how to apply and follow the GNU GPL, see
667
+ <https://www.gnu.org/licenses/>.
668
+
669
+ The GNU General Public License does not permit incorporating your program
670
+ into proprietary programs. If your program is a subroutine library, you
671
+ may consider it more useful to permit linking proprietary applications with
672
+ the library. If this is what you want to do, use the GNU Lesser General
673
+ Public License instead of this License. But first, please read
674
+ <https://www.gnu.org/licenses/why-not-lgpl.html>.
ComfyUI/README.md ADDED
@@ -0,0 +1,224 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ComfyUI
2
+ =======
3
+ The most powerful and modular stable diffusion GUI and backend.
4
+ -----------
5
+ ![ComfyUI Screenshot](comfyui_screenshot.png)
6
+
7
+ This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
8
+ ### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
9
+
10
+ ### [Installing ComfyUI](#installing)
11
+
12
+ ## Features
13
+ - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
14
+ - Fully supports SD1.x, SD2.x, [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/), [Stable Cascade](https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/), [SD3](https://comfyanonymous.github.io/ComfyUI_examples/sd3/) and [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
15
+ - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
16
+ - Asynchronous Queue system
17
+ - Many optimizations: Only re-executes the parts of the workflow that changes between executions.
18
+ - Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
19
+ - Works even if you don't have a GPU with: ```--cpu``` (slow)
20
+ - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs and CLIP models.
21
+ - Embeddings/Textual inversion
22
+ - [Loras (regular, locon and loha)](https://comfyanonymous.github.io/ComfyUI_examples/lora/)
23
+ - [Hypernetworks](https://comfyanonymous.github.io/ComfyUI_examples/hypernetworks/)
24
+ - Loading full workflows (with seeds) from generated PNG, WebP and FLAC files.
25
+ - Saving/Loading workflows as Json files.
26
+ - Nodes interface can be used to create complex workflows like one for [Hires fix](https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/) or much more advanced ones.
27
+ - [Area Composition](https://comfyanonymous.github.io/ComfyUI_examples/area_composition/)
28
+ - [Inpainting](https://comfyanonymous.github.io/ComfyUI_examples/inpaint/) with both regular and inpainting models.
29
+ - [ControlNet and T2I-Adapter](https://comfyanonymous.github.io/ComfyUI_examples/controlnet/)
30
+ - [Upscale Models (ESRGAN, ESRGAN variants, SwinIR, Swin2SR, etc...)](https://comfyanonymous.github.io/ComfyUI_examples/upscale_models/)
31
+ - [unCLIP Models](https://comfyanonymous.github.io/ComfyUI_examples/unclip/)
32
+ - [GLIGEN](https://comfyanonymous.github.io/ComfyUI_examples/gligen/)
33
+ - [Model Merging](https://comfyanonymous.github.io/ComfyUI_examples/model_merging/)
34
+ - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
35
+ - [SDXL Turbo](https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/)
36
+ - [AuraFlow](https://comfyanonymous.github.io/ComfyUI_examples/aura_flow/)
37
+ - [HunyuanDiT](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_dit/)
38
+ - Latent previews with [TAESD](#how-to-show-high-quality-previews)
39
+ - Starts up very fast.
40
+ - Works fully offline: will never download anything.
41
+ - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
42
+
43
+ Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
44
+
45
+ ## Shortcuts
46
+
47
+ | Keybind | Explanation |
48
+ |------------------------------------|--------------------------------------------------------------------------------------------------------------------|
49
+ | Ctrl + Enter | Queue up current graph for generation |
50
+ | Ctrl + Shift + Enter | Queue up current graph as first for generation |
51
+ | Ctrl + Z/Ctrl + Y | Undo/Redo |
52
+ | Ctrl + S | Save workflow |
53
+ | Ctrl + O | Load workflow |
54
+ | Ctrl + A | Select all nodes |
55
+ | Alt + C | Collapse/uncollapse selected nodes |
56
+ | Ctrl + M | Mute/unmute selected nodes |
57
+ | Ctrl + B | Bypass selected nodes (acts like the node was removed from the graph and the wires reconnected through) |
58
+ | Delete/Backspace | Delete selected nodes |
59
+ | Ctrl + Backspace | Delete the current graph |
60
+ | Space | Move the canvas around when held and moving the cursor |
61
+ | Ctrl/Shift + Click | Add clicked node to selection |
62
+ | Ctrl + C/Ctrl + V | Copy and paste selected nodes (without maintaining connections to outputs of unselected nodes) |
63
+ | Ctrl + C/Ctrl + Shift + V | Copy and paste selected nodes (maintaining connections from outputs of unselected nodes to inputs of pasted nodes) |
64
+ | Shift + Drag | Move multiple selected nodes at the same time |
65
+ | Ctrl + D | Load default graph |
66
+ | Alt + `+` | Canvas Zoom in |
67
+ | Alt + `-` | Canvas Zoom out |
68
+ | Ctrl + Shift + LMB + Vertical drag | Canvas Zoom in/out |
69
+ | Q | Toggle visibility of the queue |
70
+ | H | Toggle visibility of history |
71
+ | R | Refresh graph |
72
+ | Double-Click LMB | Open node quick search palette |
73
+
74
+ Ctrl can also be replaced with Cmd instead for macOS users
75
+
76
+ # Installing
77
+
78
+ ## Windows
79
+
80
+ There is a portable standalone build for Windows that should work for running on Nvidia GPUs or for running on your CPU only on the [releases page](https://github.com/comfyanonymous/ComfyUI/releases).
81
+
82
+ ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z)
83
+
84
+ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints
85
+
86
+ If you have trouble extracting it, right click the file -> properties -> unblock
87
+
88
+ #### How do I share models between another UI and ComfyUI?
89
+
90
+ See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor.
91
+
92
+ ## Jupyter Notebook
93
+
94
+ To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
95
+
96
+ ## Manual Install (Windows, Linux)
97
+
98
+ Git clone this repo.
99
+
100
+ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints
101
+
102
+ Put your VAE in: models/vae
103
+
104
+
105
+ ### AMD GPUs (Linux only)
106
+ AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
107
+
108
+ ```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0```
109
+
110
+ This is the command to install the nightly with ROCm 6.0 which might have some performance improvements:
111
+
112
+ ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.1```
113
+
114
+ ### NVIDIA
115
+
116
+ Nvidia users should install stable pytorch using this command:
117
+
118
+ ```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121```
119
+
120
+ This is the command to install pytorch nightly instead which might have performance improvements:
121
+
122
+ ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124```
123
+
124
+ #### Troubleshooting
125
+
126
+ If you get the "Torch not compiled with CUDA enabled" error, uninstall torch with:
127
+
128
+ ```pip uninstall torch```
129
+
130
+ And install it again with the command above.
131
+
132
+ ### Dependencies
133
+
134
+ Install the dependencies by opening your terminal inside the ComfyUI folder and:
135
+
136
+ ```pip install -r requirements.txt```
137
+
138
+ After this you should have everything installed and can proceed to running ComfyUI.
139
+
140
+ ### Others:
141
+
142
+ #### Intel GPUs
143
+
144
+ Intel GPU support is available for all Intel GPUs supported by Intel's Extension for Pytorch (IPEX) with the support requirements listed in the [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) page. Choose your platform and method of install and follow the instructions. The steps are as follows:
145
+
146
+ 1. Start by installing the drivers or kernel listed or newer in the Installation page of IPEX linked above for Windows and Linux if needed.
147
+ 1. Follow the instructions to install [Intel's oneAPI Basekit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html) for your platform.
148
+ 1. Install the packages for IPEX using the instructions provided in the Installation page for your platform.
149
+ 1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux and run ComfyUI normally as described above after everything is installed.
150
+
151
+ Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476).
152
+
153
+ #### Apple Mac silicon
154
+
155
+ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS version.
156
+
157
+ 1. Install pytorch nightly. For instructions, read the [Accelerated PyTorch training on Mac](https://developer.apple.com/metal/pytorch/) Apple Developer guide (make sure to install the latest pytorch nightly).
158
+ 1. Follow the [ComfyUI manual installation](#manual-install-windows-linux) instructions for Windows and Linux.
159
+ 1. Install the ComfyUI [dependencies](#dependencies). If you have another Stable Diffusion UI [you might be able to reuse the dependencies](#i-already-have-another-ui-for-stable-diffusion-installed-do-i-really-have-to-install-all-of-these-dependencies).
160
+ 1. Launch ComfyUI by running `python main.py`
161
+
162
+ > **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux).
163
+
164
+ #### DirectML (AMD Cards on Windows)
165
+
166
+ ```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
167
+
168
+ # Running
169
+
170
+ ```python main.py```
171
+
172
+ ### For AMD cards not officially supported by ROCm
173
+
174
+ Try running it with this command if you have issues:
175
+
176
+ For 6700, 6600 and maybe other RDNA2 or older: ```HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py```
177
+
178
+ For AMD 7600 and maybe other RDNA3 cards: ```HSA_OVERRIDE_GFX_VERSION=11.0.0 python main.py```
179
+
180
+ # Notes
181
+
182
+ Only parts of the graph that have an output with all the correct inputs will be executed.
183
+
184
+ Only parts of the graph that change from each execution to the next will be executed, if you submit the same graph twice only the first will be executed. If you change the last part of the graph only the part you changed and the part that depends on it will be executed.
185
+
186
+ Dragging a generated png on the webpage or loading one will give you the full workflow including seeds that were used to create it.
187
+
188
+ You can use () to change emphasis of a word or phrase like: (good code:1.2) or (bad code:0.8). The default emphasis for () is 1.1. To use () characters in your actual prompt escape them like \\( or \\).
189
+
190
+ You can use {day|night}, for wildcard/dynamic prompts. With this syntax "{wild|card|test}" will be randomly replaced by either "wild", "card" or "test" by the frontend every time you queue the prompt. To use {} characters in your actual prompt escape them like: \\{ or \\}.
191
+
192
+ Dynamic prompts also support C-style comments, like `// comment` or `/* comment */`.
193
+
194
+ To use a textual inversion concepts/embeddings in a text prompt put them in the models/embeddings directory and use them in the CLIPTextEncode node like this (you can omit the .pt extension):
195
+
196
+ ```embedding:embedding_filename.pt```
197
+
198
+
199
+ ## How to show high-quality previews?
200
+
201
+ Use ```--preview-method auto``` to enable previews.
202
+
203
+ The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesd_decoder.pth) (for SD1.x and SD2.x) and [taesdxl_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesdxl_decoder.pth) (for SDXL) models and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI to enable high-quality previews.
204
+
205
+ ## How to use TLS/SSL?
206
+ Generate a self-signed certificate (not appropriate for shared/production use) and key by running the command: `openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -sha256 -days 3650 -nodes -subj "/C=XX/ST=StateName/L=CityName/O=CompanyName/OU=CompanySectionName/CN=CommonNameOrHostname"`
207
+
208
+ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`.
209
+
210
+ > Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
211
+ <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
212
+
213
+ ## Support and dev channel
214
+
215
+ [Matrix space: #comfyui_space:matrix.org](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) (it's like discord but open source).
216
+
217
+ See also: [https://www.comfy.org/](https://www.comfy.org/)
218
+
219
+ # QA
220
+
221
+ ### Which GPU should I buy for this?
222
+
223
+ [See this page for some recommendations](https://github.com/comfyanonymous/ComfyUI/wiki/Which-GPU-should-I-buy-for-ComfyUI)
224
+
ComfyUI/app/__init__.py ADDED
File without changes
ComfyUI/app/app_settings.py ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import json
3
+ from aiohttp import web
4
+
5
+
6
+ class AppSettings():
7
+ def __init__(self, user_manager):
8
+ self.user_manager = user_manager
9
+
10
+ def get_settings(self, request):
11
+ file = self.user_manager.get_request_user_filepath(
12
+ request, "comfy.settings.json")
13
+ if os.path.isfile(file):
14
+ with open(file) as f:
15
+ return json.load(f)
16
+ else:
17
+ return {}
18
+
19
+ def save_settings(self, request, settings):
20
+ file = self.user_manager.get_request_user_filepath(
21
+ request, "comfy.settings.json")
22
+ with open(file, "w") as f:
23
+ f.write(json.dumps(settings, indent=4))
24
+
25
+ def add_routes(self, routes):
26
+ @routes.get("/settings")
27
+ async def get_settings(request):
28
+ return web.json_response(self.get_settings(request))
29
+
30
+ @routes.get("/settings/{id}")
31
+ async def get_setting(request):
32
+ value = None
33
+ settings = self.get_settings(request)
34
+ setting_id = request.match_info.get("id", None)
35
+ if setting_id and setting_id in settings:
36
+ value = settings[setting_id]
37
+ return web.json_response(value)
38
+
39
+ @routes.post("/settings")
40
+ async def post_settings(request):
41
+ settings = self.get_settings(request)
42
+ new_settings = await request.json()
43
+ self.save_settings(request, {**settings, **new_settings})
44
+ return web.Response(status=200)
45
+
46
+ @routes.post("/settings/{id}")
47
+ async def post_setting(request):
48
+ setting_id = request.match_info.get("id", None)
49
+ if not setting_id:
50
+ return web.Response(status=400)
51
+ settings = self.get_settings(request)
52
+ settings[setting_id] = await request.json()
53
+ self.save_settings(request, settings)
54
+ return web.Response(status=200)
ComfyUI/app/frontend_management.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+ import argparse
3
+ import logging
4
+ import os
5
+ import re
6
+ import tempfile
7
+ import zipfile
8
+ from dataclasses import dataclass
9
+ from functools import cached_property
10
+ from pathlib import Path
11
+ from typing import TypedDict
12
+
13
+ import requests
14
+ from typing_extensions import NotRequired
15
+ from comfy.cli_args import DEFAULT_VERSION_STRING
16
+
17
+
18
+ REQUEST_TIMEOUT = 10 # seconds
19
+
20
+
21
+ class Asset(TypedDict):
22
+ url: str
23
+
24
+
25
+ class Release(TypedDict):
26
+ id: int
27
+ tag_name: str
28
+ name: str
29
+ prerelease: bool
30
+ created_at: str
31
+ published_at: str
32
+ body: str
33
+ assets: NotRequired[list[Asset]]
34
+
35
+
36
+ @dataclass
37
+ class FrontEndProvider:
38
+ owner: str
39
+ repo: str
40
+
41
+ @property
42
+ def folder_name(self) -> str:
43
+ return f"{self.owner}_{self.repo}"
44
+
45
+ @property
46
+ def release_url(self) -> str:
47
+ return f"https://api.github.com/repos/{self.owner}/{self.repo}/releases"
48
+
49
+ @cached_property
50
+ def all_releases(self) -> list[Release]:
51
+ releases = []
52
+ api_url = self.release_url
53
+ while api_url:
54
+ response = requests.get(api_url, timeout=REQUEST_TIMEOUT)
55
+ response.raise_for_status() # Raises an HTTPError if the response was an error
56
+ releases.extend(response.json())
57
+ # GitHub uses the Link header to provide pagination links. Check if it exists and update api_url accordingly.
58
+ if "next" in response.links:
59
+ api_url = response.links["next"]["url"]
60
+ else:
61
+ api_url = None
62
+ return releases
63
+
64
+ @cached_property
65
+ def latest_release(self) -> Release:
66
+ latest_release_url = f"{self.release_url}/latest"
67
+ response = requests.get(latest_release_url, timeout=REQUEST_TIMEOUT)
68
+ response.raise_for_status() # Raises an HTTPError if the response was an error
69
+ return response.json()
70
+
71
+ def get_release(self, version: str) -> Release:
72
+ if version == "latest":
73
+ return self.latest_release
74
+ else:
75
+ for release in self.all_releases:
76
+ if release["tag_name"] in [version, f"v{version}"]:
77
+ return release
78
+ raise ValueError(f"Version {version} not found in releases")
79
+
80
+
81
+ def download_release_asset_zip(release: Release, destination_path: str) -> None:
82
+ """Download dist.zip from github release."""
83
+ asset_url = None
84
+ for asset in release.get("assets", []):
85
+ if asset["name"] == "dist.zip":
86
+ asset_url = asset["url"]
87
+ break
88
+
89
+ if not asset_url:
90
+ raise ValueError("dist.zip not found in the release assets")
91
+
92
+ # Use a temporary file to download the zip content
93
+ with tempfile.TemporaryFile() as tmp_file:
94
+ headers = {"Accept": "application/octet-stream"}
95
+ response = requests.get(
96
+ asset_url, headers=headers, allow_redirects=True, timeout=REQUEST_TIMEOUT
97
+ )
98
+ response.raise_for_status() # Ensure we got a successful response
99
+
100
+ # Write the content to the temporary file
101
+ tmp_file.write(response.content)
102
+
103
+ # Go back to the beginning of the temporary file
104
+ tmp_file.seek(0)
105
+
106
+ # Extract the zip file content to the destination path
107
+ with zipfile.ZipFile(tmp_file, "r") as zip_ref:
108
+ zip_ref.extractall(destination_path)
109
+
110
+
111
+ class FrontendManager:
112
+ DEFAULT_FRONTEND_PATH = str(Path(__file__).parents[1] / "web")
113
+ CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
114
+
115
+ @classmethod
116
+ def parse_version_string(cls, value: str) -> tuple[str, str, str]:
117
+ """
118
+ Args:
119
+ value (str): The version string to parse.
120
+
121
+ Returns:
122
+ tuple[str, str]: A tuple containing provider name and version.
123
+
124
+ Raises:
125
+ argparse.ArgumentTypeError: If the version string is invalid.
126
+ """
127
+ VERSION_PATTERN = r"^([a-zA-Z0-9][a-zA-Z0-9-]{0,38})/([a-zA-Z0-9_.-]+)@(v?\d+\.\d+\.\d+|latest)$"
128
+ match_result = re.match(VERSION_PATTERN, value)
129
+ if match_result is None:
130
+ raise argparse.ArgumentTypeError(f"Invalid version string: {value}")
131
+
132
+ return match_result.group(1), match_result.group(2), match_result.group(3)
133
+
134
+ @classmethod
135
+ def init_frontend_unsafe(cls, version_string: str) -> str:
136
+ """
137
+ Initializes the frontend for the specified version.
138
+
139
+ Args:
140
+ version_string (str): The version string.
141
+
142
+ Returns:
143
+ str: The path to the initialized frontend.
144
+
145
+ Raises:
146
+ Exception: If there is an error during the initialization process.
147
+ main error source might be request timeout or invalid URL.
148
+ """
149
+ if version_string == DEFAULT_VERSION_STRING:
150
+ return cls.DEFAULT_FRONTEND_PATH
151
+
152
+ repo_owner, repo_name, version = cls.parse_version_string(version_string)
153
+ provider = FrontEndProvider(repo_owner, repo_name)
154
+ release = provider.get_release(version)
155
+
156
+ semantic_version = release["tag_name"].lstrip("v")
157
+ web_root = str(
158
+ Path(cls.CUSTOM_FRONTENDS_ROOT) / provider.folder_name / semantic_version
159
+ )
160
+ if not os.path.exists(web_root):
161
+ os.makedirs(web_root, exist_ok=True)
162
+ logging.info(
163
+ "Downloading frontend(%s) version(%s) to (%s)",
164
+ provider.folder_name,
165
+ semantic_version,
166
+ web_root,
167
+ )
168
+ logging.debug(release)
169
+ download_release_asset_zip(release, destination_path=web_root)
170
+ return web_root
171
+
172
+ @classmethod
173
+ def init_frontend(cls, version_string: str) -> str:
174
+ """
175
+ Initializes the frontend with the specified version string.
176
+
177
+ Args:
178
+ version_string (str): The version string to initialize the frontend with.
179
+
180
+ Returns:
181
+ str: The path of the initialized frontend.
182
+ """
183
+ try:
184
+ return cls.init_frontend_unsafe(version_string)
185
+ except Exception as e:
186
+ logging.error("Failed to initialize frontend: %s", e)
187
+ logging.info("Falling back to the default frontend.")
188
+ return cls.DEFAULT_FRONTEND_PATH
ComfyUI/app/user_manager.py ADDED
@@ -0,0 +1,205 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ import os
3
+ import re
4
+ import uuid
5
+ import glob
6
+ import shutil
7
+ from aiohttp import web
8
+ from comfy.cli_args import args
9
+ from folder_paths import user_directory
10
+ from .app_settings import AppSettings
11
+
12
+ default_user = "default"
13
+ users_file = os.path.join(user_directory, "users.json")
14
+
15
+
16
+ class UserManager():
17
+ def __init__(self):
18
+ global user_directory
19
+
20
+ self.settings = AppSettings(self)
21
+ if not os.path.exists(user_directory):
22
+ os.mkdir(user_directory)
23
+ if not args.multi_user:
24
+ print("****** User settings have been changed to be stored on the server instead of browser storage. ******")
25
+ print("****** For multi-user setups add the --multi-user CLI argument to enable multiple user profiles. ******")
26
+
27
+ if args.multi_user:
28
+ if os.path.isfile(users_file):
29
+ with open(users_file) as f:
30
+ self.users = json.load(f)
31
+ else:
32
+ self.users = {}
33
+ else:
34
+ self.users = {"default": "default"}
35
+
36
+ def get_request_user_id(self, request):
37
+ user = "default"
38
+ if args.multi_user and "comfy-user" in request.headers:
39
+ user = request.headers["comfy-user"]
40
+
41
+ if user not in self.users:
42
+ raise KeyError("Unknown user: " + user)
43
+
44
+ return user
45
+
46
+ def get_request_user_filepath(self, request, file, type="userdata", create_dir=True):
47
+ global user_directory
48
+
49
+ if type == "userdata":
50
+ root_dir = user_directory
51
+ else:
52
+ raise KeyError("Unknown filepath type:" + type)
53
+
54
+ user = self.get_request_user_id(request)
55
+ path = user_root = os.path.abspath(os.path.join(root_dir, user))
56
+
57
+ # prevent leaving /{type}
58
+ if os.path.commonpath((root_dir, user_root)) != root_dir:
59
+ return None
60
+
61
+ if file is not None:
62
+ # prevent leaving /{type}/{user}
63
+ path = os.path.abspath(os.path.join(user_root, file))
64
+ if os.path.commonpath((user_root, path)) != user_root:
65
+ return None
66
+
67
+ parent = os.path.split(path)[0]
68
+
69
+ if create_dir and not os.path.exists(parent):
70
+ os.makedirs(parent, exist_ok=True)
71
+
72
+ return path
73
+
74
+ def add_user(self, name):
75
+ name = name.strip()
76
+ if not name:
77
+ raise ValueError("username not provided")
78
+ user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name)
79
+ user_id = user_id + "_" + str(uuid.uuid4())
80
+
81
+ self.users[user_id] = name
82
+
83
+ global users_file
84
+ with open(users_file, "w") as f:
85
+ json.dump(self.users, f)
86
+
87
+ return user_id
88
+
89
+ def add_routes(self, routes):
90
+ self.settings.add_routes(routes)
91
+
92
+ @routes.get("/users")
93
+ async def get_users(request):
94
+ if args.multi_user:
95
+ return web.json_response({"storage": "server", "users": self.users})
96
+ else:
97
+ user_dir = self.get_request_user_filepath(request, None, create_dir=False)
98
+ return web.json_response({
99
+ "storage": "server",
100
+ "migrated": os.path.exists(user_dir)
101
+ })
102
+
103
+ @routes.post("/users")
104
+ async def post_users(request):
105
+ body = await request.json()
106
+ username = body["username"]
107
+ if username in self.users.values():
108
+ return web.json_response({"error": "Duplicate username."}, status=400)
109
+
110
+ user_id = self.add_user(username)
111
+ return web.json_response(user_id)
112
+
113
+ @routes.get("/userdata")
114
+ async def listuserdata(request):
115
+ directory = request.rel_url.query.get('dir', '')
116
+ if not directory:
117
+ return web.Response(status=400)
118
+
119
+ path = self.get_request_user_filepath(request, directory)
120
+ if not path:
121
+ return web.Response(status=403)
122
+
123
+ if not os.path.exists(path):
124
+ return web.Response(status=404)
125
+
126
+ recurse = request.rel_url.query.get('recurse', '').lower() == "true"
127
+ results = glob.glob(os.path.join(
128
+ glob.escape(path), '**/*'), recursive=recurse)
129
+ results = [os.path.relpath(x, path) for x in results if os.path.isfile(x)]
130
+
131
+ split_path = request.rel_url.query.get('split', '').lower() == "true"
132
+ if split_path:
133
+ results = [[x] + x.split(os.sep) for x in results]
134
+
135
+ return web.json_response(results)
136
+
137
+ def get_user_data_path(request, check_exists = False, param = "file"):
138
+ file = request.match_info.get(param, None)
139
+ if not file:
140
+ return web.Response(status=400)
141
+
142
+ path = self.get_request_user_filepath(request, file)
143
+ if not path:
144
+ return web.Response(status=403)
145
+
146
+ if check_exists and not os.path.exists(path):
147
+ return web.Response(status=404)
148
+
149
+ return path
150
+
151
+ @routes.get("/userdata/{file}")
152
+ async def getuserdata(request):
153
+ path = get_user_data_path(request, check_exists=True)
154
+ if not isinstance(path, str):
155
+ return path
156
+
157
+ return web.FileResponse(path)
158
+
159
+ @routes.post("/userdata/{file}")
160
+ async def post_userdata(request):
161
+ path = get_user_data_path(request)
162
+ if not isinstance(path, str):
163
+ return path
164
+
165
+ overwrite = request.query["overwrite"] != "false"
166
+ if not overwrite and os.path.exists(path):
167
+ return web.Response(status=409)
168
+
169
+ body = await request.read()
170
+
171
+ with open(path, "wb") as f:
172
+ f.write(body)
173
+
174
+ resp = os.path.relpath(path, self.get_request_user_filepath(request, None))
175
+ return web.json_response(resp)
176
+
177
+ @routes.delete("/userdata/{file}")
178
+ async def delete_userdata(request):
179
+ path = get_user_data_path(request, check_exists=True)
180
+ if not isinstance(path, str):
181
+ return path
182
+
183
+ os.remove(path)
184
+
185
+ return web.Response(status=204)
186
+
187
+ @routes.post("/userdata/{file}/move/{dest}")
188
+ async def move_userdata(request):
189
+ source = get_user_data_path(request, check_exists=True)
190
+ if not isinstance(source, str):
191
+ return source
192
+
193
+ dest = get_user_data_path(request, check_exists=False, param="dest")
194
+ if not isinstance(source, str):
195
+ return dest
196
+
197
+ overwrite = request.query["overwrite"] != "false"
198
+ if not overwrite and os.path.exists(dest):
199
+ return web.Response(status=409)
200
+
201
+ print(f"moving '{source}' -> '{dest}'")
202
+ shutil.move(source, dest)
203
+
204
+ resp = os.path.relpath(dest, self.get_request_user_filepath(request, None))
205
+ return web.json_response(resp)
ComfyUI/comfy/checkpoint_pickle.py ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import pickle
2
+
3
+ load = pickle.load
4
+
5
+ class Empty:
6
+ pass
7
+
8
+ class Unpickler(pickle.Unpickler):
9
+ def find_class(self, module, name):
10
+ #TODO: safe unpickle
11
+ if module.startswith("pytorch_lightning"):
12
+ return Empty
13
+ return super().find_class(module, name)
ComfyUI/comfy/cldm/cldm.py ADDED
@@ -0,0 +1,437 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #taken from: https://github.com/lllyasviel/ControlNet
2
+ #and modified
3
+
4
+ import torch
5
+ import torch as th
6
+ import torch.nn as nn
7
+
8
+ from ..ldm.modules.diffusionmodules.util import (
9
+ zero_module,
10
+ timestep_embedding,
11
+ )
12
+
13
+ from ..ldm.modules.attention import SpatialTransformer
14
+ from ..ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample
15
+ from ..ldm.util import exists
16
+ from .control_types import UNION_CONTROLNET_TYPES
17
+ from collections import OrderedDict
18
+ import comfy.ops
19
+ from comfy.ldm.modules.attention import optimized_attention
20
+
21
+ class OptimizedAttention(nn.Module):
22
+ def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None):
23
+ super().__init__()
24
+ self.heads = nhead
25
+ self.c = c
26
+
27
+ self.in_proj = operations.Linear(c, c * 3, bias=True, dtype=dtype, device=device)
28
+ self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
29
+
30
+ def forward(self, x):
31
+ x = self.in_proj(x)
32
+ q, k, v = x.split(self.c, dim=2)
33
+ out = optimized_attention(q, k, v, self.heads)
34
+ return self.out_proj(out)
35
+
36
+ class QuickGELU(nn.Module):
37
+ def forward(self, x: torch.Tensor):
38
+ return x * torch.sigmoid(1.702 * x)
39
+
40
+ class ResBlockUnionControlnet(nn.Module):
41
+ def __init__(self, dim, nhead, dtype=None, device=None, operations=None):
42
+ super().__init__()
43
+ self.attn = OptimizedAttention(dim, nhead, dtype=dtype, device=device, operations=operations)
44
+ self.ln_1 = operations.LayerNorm(dim, dtype=dtype, device=device)
45
+ self.mlp = nn.Sequential(
46
+ OrderedDict([("c_fc", operations.Linear(dim, dim * 4, dtype=dtype, device=device)), ("gelu", QuickGELU()),
47
+ ("c_proj", operations.Linear(dim * 4, dim, dtype=dtype, device=device))]))
48
+ self.ln_2 = operations.LayerNorm(dim, dtype=dtype, device=device)
49
+
50
+ def attention(self, x: torch.Tensor):
51
+ return self.attn(x)
52
+
53
+ def forward(self, x: torch.Tensor):
54
+ x = x + self.attention(self.ln_1(x))
55
+ x = x + self.mlp(self.ln_2(x))
56
+ return x
57
+
58
+ class ControlledUnetModel(UNetModel):
59
+ #implemented in the ldm unet
60
+ pass
61
+
62
+ class ControlNet(nn.Module):
63
+ def __init__(
64
+ self,
65
+ image_size,
66
+ in_channels,
67
+ model_channels,
68
+ hint_channels,
69
+ num_res_blocks,
70
+ dropout=0,
71
+ channel_mult=(1, 2, 4, 8),
72
+ conv_resample=True,
73
+ dims=2,
74
+ num_classes=None,
75
+ use_checkpoint=False,
76
+ dtype=torch.float32,
77
+ num_heads=-1,
78
+ num_head_channels=-1,
79
+ num_heads_upsample=-1,
80
+ use_scale_shift_norm=False,
81
+ resblock_updown=False,
82
+ use_new_attention_order=False,
83
+ use_spatial_transformer=False, # custom transformer support
84
+ transformer_depth=1, # custom transformer support
85
+ context_dim=None, # custom transformer support
86
+ n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
87
+ legacy=True,
88
+ disable_self_attentions=None,
89
+ num_attention_blocks=None,
90
+ disable_middle_self_attn=False,
91
+ use_linear_in_transformer=False,
92
+ adm_in_channels=None,
93
+ transformer_depth_middle=None,
94
+ transformer_depth_output=None,
95
+ attn_precision=None,
96
+ union_controlnet_num_control_type=None,
97
+ device=None,
98
+ operations=comfy.ops.disable_weight_init,
99
+ **kwargs,
100
+ ):
101
+ super().__init__()
102
+ assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
103
+ if use_spatial_transformer:
104
+ assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
105
+
106
+ if context_dim is not None:
107
+ assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
108
+ # from omegaconf.listconfig import ListConfig
109
+ # if type(context_dim) == ListConfig:
110
+ # context_dim = list(context_dim)
111
+
112
+ if num_heads_upsample == -1:
113
+ num_heads_upsample = num_heads
114
+
115
+ if num_heads == -1:
116
+ assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
117
+
118
+ if num_head_channels == -1:
119
+ assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
120
+
121
+ self.dims = dims
122
+ self.image_size = image_size
123
+ self.in_channels = in_channels
124
+ self.model_channels = model_channels
125
+
126
+ if isinstance(num_res_blocks, int):
127
+ self.num_res_blocks = len(channel_mult) * [num_res_blocks]
128
+ else:
129
+ if len(num_res_blocks) != len(channel_mult):
130
+ raise ValueError("provide num_res_blocks either as an int (globally constant) or "
131
+ "as a list/tuple (per-level) with the same length as channel_mult")
132
+ self.num_res_blocks = num_res_blocks
133
+
134
+ if disable_self_attentions is not None:
135
+ # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
136
+ assert len(disable_self_attentions) == len(channel_mult)
137
+ if num_attention_blocks is not None:
138
+ assert len(num_attention_blocks) == len(self.num_res_blocks)
139
+ assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
140
+
141
+ transformer_depth = transformer_depth[:]
142
+
143
+ self.dropout = dropout
144
+ self.channel_mult = channel_mult
145
+ self.conv_resample = conv_resample
146
+ self.num_classes = num_classes
147
+ self.use_checkpoint = use_checkpoint
148
+ self.dtype = dtype
149
+ self.num_heads = num_heads
150
+ self.num_head_channels = num_head_channels
151
+ self.num_heads_upsample = num_heads_upsample
152
+ self.predict_codebook_ids = n_embed is not None
153
+
154
+ time_embed_dim = model_channels * 4
155
+ self.time_embed = nn.Sequential(
156
+ operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
157
+ nn.SiLU(),
158
+ operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
159
+ )
160
+
161
+ if self.num_classes is not None:
162
+ if isinstance(self.num_classes, int):
163
+ self.label_emb = nn.Embedding(num_classes, time_embed_dim)
164
+ elif self.num_classes == "continuous":
165
+ print("setting up linear c_adm embedding layer")
166
+ self.label_emb = nn.Linear(1, time_embed_dim)
167
+ elif self.num_classes == "sequential":
168
+ assert adm_in_channels is not None
169
+ self.label_emb = nn.Sequential(
170
+ nn.Sequential(
171
+ operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
172
+ nn.SiLU(),
173
+ operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
174
+ )
175
+ )
176
+ else:
177
+ raise ValueError()
178
+
179
+ self.input_blocks = nn.ModuleList(
180
+ [
181
+ TimestepEmbedSequential(
182
+ operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
183
+ )
184
+ ]
185
+ )
186
+ self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations, dtype=self.dtype, device=device)])
187
+
188
+ self.input_hint_block = TimestepEmbedSequential(
189
+ operations.conv_nd(dims, hint_channels, 16, 3, padding=1, dtype=self.dtype, device=device),
190
+ nn.SiLU(),
191
+ operations.conv_nd(dims, 16, 16, 3, padding=1, dtype=self.dtype, device=device),
192
+ nn.SiLU(),
193
+ operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2, dtype=self.dtype, device=device),
194
+ nn.SiLU(),
195
+ operations.conv_nd(dims, 32, 32, 3, padding=1, dtype=self.dtype, device=device),
196
+ nn.SiLU(),
197
+ operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2, dtype=self.dtype, device=device),
198
+ nn.SiLU(),
199
+ operations.conv_nd(dims, 96, 96, 3, padding=1, dtype=self.dtype, device=device),
200
+ nn.SiLU(),
201
+ operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2, dtype=self.dtype, device=device),
202
+ nn.SiLU(),
203
+ operations.conv_nd(dims, 256, model_channels, 3, padding=1, dtype=self.dtype, device=device)
204
+ )
205
+
206
+ self._feature_size = model_channels
207
+ input_block_chans = [model_channels]
208
+ ch = model_channels
209
+ ds = 1
210
+ for level, mult in enumerate(channel_mult):
211
+ for nr in range(self.num_res_blocks[level]):
212
+ layers = [
213
+ ResBlock(
214
+ ch,
215
+ time_embed_dim,
216
+ dropout,
217
+ out_channels=mult * model_channels,
218
+ dims=dims,
219
+ use_checkpoint=use_checkpoint,
220
+ use_scale_shift_norm=use_scale_shift_norm,
221
+ dtype=self.dtype,
222
+ device=device,
223
+ operations=operations,
224
+ )
225
+ ]
226
+ ch = mult * model_channels
227
+ num_transformers = transformer_depth.pop(0)
228
+ if num_transformers > 0:
229
+ if num_head_channels == -1:
230
+ dim_head = ch // num_heads
231
+ else:
232
+ num_heads = ch // num_head_channels
233
+ dim_head = num_head_channels
234
+ if legacy:
235
+ #num_heads = 1
236
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
237
+ if exists(disable_self_attentions):
238
+ disabled_sa = disable_self_attentions[level]
239
+ else:
240
+ disabled_sa = False
241
+
242
+ if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
243
+ layers.append(
244
+ SpatialTransformer(
245
+ ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
246
+ disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
247
+ use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
248
+ )
249
+ )
250
+ self.input_blocks.append(TimestepEmbedSequential(*layers))
251
+ self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
252
+ self._feature_size += ch
253
+ input_block_chans.append(ch)
254
+ if level != len(channel_mult) - 1:
255
+ out_ch = ch
256
+ self.input_blocks.append(
257
+ TimestepEmbedSequential(
258
+ ResBlock(
259
+ ch,
260
+ time_embed_dim,
261
+ dropout,
262
+ out_channels=out_ch,
263
+ dims=dims,
264
+ use_checkpoint=use_checkpoint,
265
+ use_scale_shift_norm=use_scale_shift_norm,
266
+ down=True,
267
+ dtype=self.dtype,
268
+ device=device,
269
+ operations=operations
270
+ )
271
+ if resblock_updown
272
+ else Downsample(
273
+ ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
274
+ )
275
+ )
276
+ )
277
+ ch = out_ch
278
+ input_block_chans.append(ch)
279
+ self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
280
+ ds *= 2
281
+ self._feature_size += ch
282
+
283
+ if num_head_channels == -1:
284
+ dim_head = ch // num_heads
285
+ else:
286
+ num_heads = ch // num_head_channels
287
+ dim_head = num_head_channels
288
+ if legacy:
289
+ #num_heads = 1
290
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
291
+ mid_block = [
292
+ ResBlock(
293
+ ch,
294
+ time_embed_dim,
295
+ dropout,
296
+ dims=dims,
297
+ use_checkpoint=use_checkpoint,
298
+ use_scale_shift_norm=use_scale_shift_norm,
299
+ dtype=self.dtype,
300
+ device=device,
301
+ operations=operations
302
+ )]
303
+ if transformer_depth_middle >= 0:
304
+ mid_block += [SpatialTransformer( # always uses a self-attn
305
+ ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
306
+ disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
307
+ use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
308
+ ),
309
+ ResBlock(
310
+ ch,
311
+ time_embed_dim,
312
+ dropout,
313
+ dims=dims,
314
+ use_checkpoint=use_checkpoint,
315
+ use_scale_shift_norm=use_scale_shift_norm,
316
+ dtype=self.dtype,
317
+ device=device,
318
+ operations=operations
319
+ )]
320
+ self.middle_block = TimestepEmbedSequential(*mid_block)
321
+ self.middle_block_out = self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)
322
+ self._feature_size += ch
323
+
324
+ if union_controlnet_num_control_type is not None:
325
+ self.num_control_type = union_controlnet_num_control_type
326
+ num_trans_channel = 320
327
+ num_trans_head = 8
328
+ num_trans_layer = 1
329
+ num_proj_channel = 320
330
+ # task_scale_factor = num_trans_channel ** 0.5
331
+ self.task_embedding = nn.Parameter(torch.empty(self.num_control_type, num_trans_channel, dtype=self.dtype, device=device))
332
+
333
+ self.transformer_layes = nn.Sequential(*[ResBlockUnionControlnet(num_trans_channel, num_trans_head, dtype=self.dtype, device=device, operations=operations) for _ in range(num_trans_layer)])
334
+ self.spatial_ch_projs = operations.Linear(num_trans_channel, num_proj_channel, dtype=self.dtype, device=device)
335
+ #-----------------------------------------------------------------------------------------------------
336
+
337
+ control_add_embed_dim = 256
338
+ class ControlAddEmbedding(nn.Module):
339
+ def __init__(self, in_dim, out_dim, num_control_type, dtype=None, device=None, operations=None):
340
+ super().__init__()
341
+ self.num_control_type = num_control_type
342
+ self.in_dim = in_dim
343
+ self.linear_1 = operations.Linear(in_dim * num_control_type, out_dim, dtype=dtype, device=device)
344
+ self.linear_2 = operations.Linear(out_dim, out_dim, dtype=dtype, device=device)
345
+ def forward(self, control_type, dtype, device):
346
+ c_type = torch.zeros((self.num_control_type,), device=device)
347
+ c_type[control_type] = 1.0
348
+ c_type = timestep_embedding(c_type.flatten(), self.in_dim, repeat_only=False).to(dtype).reshape((-1, self.num_control_type * self.in_dim))
349
+ return self.linear_2(torch.nn.functional.silu(self.linear_1(c_type)))
350
+
351
+ self.control_add_embedding = ControlAddEmbedding(control_add_embed_dim, time_embed_dim, self.num_control_type, dtype=self.dtype, device=device, operations=operations)
352
+ else:
353
+ self.task_embedding = None
354
+ self.control_add_embedding = None
355
+
356
+ def union_controlnet_merge(self, hint, control_type, emb, context):
357
+ # Equivalent to: https://github.com/xinsir6/ControlNetPlus/tree/main
358
+ inputs = []
359
+ condition_list = []
360
+
361
+ for idx in range(min(1, len(control_type))):
362
+ controlnet_cond = self.input_hint_block(hint[idx], emb, context)
363
+ feat_seq = torch.mean(controlnet_cond, dim=(2, 3))
364
+ if idx < len(control_type):
365
+ feat_seq += self.task_embedding[control_type[idx]].to(dtype=feat_seq.dtype, device=feat_seq.device)
366
+
367
+ inputs.append(feat_seq.unsqueeze(1))
368
+ condition_list.append(controlnet_cond)
369
+
370
+ x = torch.cat(inputs, dim=1)
371
+ x = self.transformer_layes(x)
372
+ controlnet_cond_fuser = None
373
+ for idx in range(len(control_type)):
374
+ alpha = self.spatial_ch_projs(x[:, idx])
375
+ alpha = alpha.unsqueeze(-1).unsqueeze(-1)
376
+ o = condition_list[idx] + alpha
377
+ if controlnet_cond_fuser is None:
378
+ controlnet_cond_fuser = o
379
+ else:
380
+ controlnet_cond_fuser += o
381
+ return controlnet_cond_fuser
382
+
383
+ def make_zero_conv(self, channels, operations=None, dtype=None, device=None):
384
+ return TimestepEmbedSequential(operations.conv_nd(self.dims, channels, channels, 1, padding=0, dtype=dtype, device=device))
385
+
386
+ def forward(self, x, hint, timesteps, context, y=None, **kwargs):
387
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
388
+ emb = self.time_embed(t_emb)
389
+
390
+ guided_hint = None
391
+ if self.control_add_embedding is not None: #Union Controlnet
392
+ control_type = kwargs.get("control_type", [])
393
+
394
+ if any([c >= self.num_control_type for c in control_type]):
395
+ max_type = max(control_type)
396
+ max_type_name = {
397
+ v: k for k, v in UNION_CONTROLNET_TYPES.items()
398
+ }[max_type]
399
+ raise ValueError(
400
+ f"Control type {max_type_name}({max_type}) is out of range for the number of control types" +
401
+ f"({self.num_control_type}) supported.\n" +
402
+ "Please consider using the ProMax ControlNet Union model.\n" +
403
+ "https://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/main"
404
+ )
405
+
406
+ emb += self.control_add_embedding(control_type, emb.dtype, emb.device)
407
+ if len(control_type) > 0:
408
+ if len(hint.shape) < 5:
409
+ hint = hint.unsqueeze(dim=0)
410
+ guided_hint = self.union_controlnet_merge(hint, control_type, emb, context)
411
+
412
+ if guided_hint is None:
413
+ guided_hint = self.input_hint_block(hint, emb, context)
414
+
415
+ out_output = []
416
+ out_middle = []
417
+
418
+ hs = []
419
+ if self.num_classes is not None:
420
+ assert y.shape[0] == x.shape[0]
421
+ emb = emb + self.label_emb(y)
422
+
423
+ h = x
424
+ for module, zero_conv in zip(self.input_blocks, self.zero_convs):
425
+ if guided_hint is not None:
426
+ h = module(h, emb, context)
427
+ h += guided_hint
428
+ guided_hint = None
429
+ else:
430
+ h = module(h, emb, context)
431
+ out_output.append(zero_conv(h, emb, context))
432
+
433
+ h = self.middle_block(h, emb, context)
434
+ out_middle.append(self.middle_block_out(h, emb, context))
435
+
436
+ return {"middle": out_middle, "output": out_output}
437
+
ComfyUI/comfy/cldm/control_types.py ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ UNION_CONTROLNET_TYPES = {
2
+ "openpose": 0,
3
+ "depth": 1,
4
+ "hed/pidi/scribble/ted": 2,
5
+ "canny/lineart/anime_lineart/mlsd": 3,
6
+ "normal": 4,
7
+ "segment": 5,
8
+ "tile": 6,
9
+ "repaint": 7,
10
+ }
ComfyUI/comfy/cldm/mmdit.py ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from typing import Dict, Optional
3
+ import comfy.ldm.modules.diffusionmodules.mmdit
4
+
5
+ class ControlNet(comfy.ldm.modules.diffusionmodules.mmdit.MMDiT):
6
+ def __init__(
7
+ self,
8
+ num_blocks = None,
9
+ dtype = None,
10
+ device = None,
11
+ operations = None,
12
+ **kwargs,
13
+ ):
14
+ super().__init__(dtype=dtype, device=device, operations=operations, final_layer=False, num_blocks=num_blocks, **kwargs)
15
+ # controlnet_blocks
16
+ self.controlnet_blocks = torch.nn.ModuleList([])
17
+ for _ in range(len(self.joint_blocks)):
18
+ self.controlnet_blocks.append(operations.Linear(self.hidden_size, self.hidden_size, device=device, dtype=dtype))
19
+
20
+ self.pos_embed_input = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(
21
+ None,
22
+ self.patch_size,
23
+ self.in_channels,
24
+ self.hidden_size,
25
+ bias=True,
26
+ strict_img_size=False,
27
+ dtype=dtype,
28
+ device=device,
29
+ operations=operations
30
+ )
31
+
32
+ def forward(
33
+ self,
34
+ x: torch.Tensor,
35
+ timesteps: torch.Tensor,
36
+ y: Optional[torch.Tensor] = None,
37
+ context: Optional[torch.Tensor] = None,
38
+ hint = None,
39
+ ) -> torch.Tensor:
40
+
41
+ #weird sd3 controlnet specific stuff
42
+ y = torch.zeros_like(y)
43
+
44
+ if self.context_processor is not None:
45
+ context = self.context_processor(context)
46
+
47
+ hw = x.shape[-2:]
48
+ x = self.x_embedder(x) + self.cropped_pos_embed(hw, device=x.device).to(dtype=x.dtype, device=x.device)
49
+ x += self.pos_embed_input(hint)
50
+
51
+ c = self.t_embedder(timesteps, dtype=x.dtype)
52
+ if y is not None and self.y_embedder is not None:
53
+ y = self.y_embedder(y)
54
+ c = c + y
55
+
56
+ if context is not None:
57
+ context = self.context_embedder(context)
58
+
59
+ output = []
60
+
61
+ blocks = len(self.joint_blocks)
62
+ for i in range(blocks):
63
+ context, x = self.joint_blocks[i](
64
+ context,
65
+ x,
66
+ c=c,
67
+ use_checkpoint=self.use_checkpoint,
68
+ )
69
+
70
+ out = self.controlnet_blocks[i](x)
71
+ count = self.depth // blocks
72
+ if i == blocks - 1:
73
+ count -= 1
74
+ for j in range(count):
75
+ output.append(out)
76
+
77
+ return {"output": output}
ComfyUI/comfy/cli_args.py ADDED
@@ -0,0 +1,180 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import enum
3
+ import os
4
+ from typing import Optional
5
+ import comfy.options
6
+
7
+
8
+ class EnumAction(argparse.Action):
9
+ """
10
+ Argparse action for handling Enums
11
+ """
12
+ def __init__(self, **kwargs):
13
+ # Pop off the type value
14
+ enum_type = kwargs.pop("type", None)
15
+
16
+ # Ensure an Enum subclass is provided
17
+ if enum_type is None:
18
+ raise ValueError("type must be assigned an Enum when using EnumAction")
19
+ if not issubclass(enum_type, enum.Enum):
20
+ raise TypeError("type must be an Enum when using EnumAction")
21
+
22
+ # Generate choices from the Enum
23
+ choices = tuple(e.value for e in enum_type)
24
+ kwargs.setdefault("choices", choices)
25
+ kwargs.setdefault("metavar", f"[{','.join(list(choices))}]")
26
+
27
+ super(EnumAction, self).__init__(**kwargs)
28
+
29
+ self._enum = enum_type
30
+
31
+ def __call__(self, parser, namespace, values, option_string=None):
32
+ # Convert value back into an Enum
33
+ value = self._enum(values)
34
+ setattr(namespace, self.dest, value)
35
+
36
+
37
+ parser = argparse.ArgumentParser()
38
+
39
+ parser.add_argument("--listen", type=str, default="127.0.0.1", metavar="IP", nargs="?", const="0.0.0.0", help="Specify the IP address to listen on (default: 127.0.0.1). If --listen is provided without an argument, it defaults to 0.0.0.0. (listens on all)")
40
+ parser.add_argument("--port", type=int, default=8188, help="Set the listen port.")
41
+ parser.add_argument("--tls-keyfile", type=str, help="Path to TLS (SSL) key file. Enables TLS, makes app accessible at https://... requires --tls-certfile to function")
42
+ parser.add_argument("--tls-certfile", type=str, help="Path to TLS (SSL) certificate file. Enables TLS, makes app accessible at https://... requires --tls-keyfile to function")
43
+ parser.add_argument("--enable-cors-header", type=str, default=None, metavar="ORIGIN", nargs="?", const="*", help="Enable CORS (Cross-Origin Resource Sharing) with optional origin or allow all with default '*'.")
44
+ parser.add_argument("--max-upload-size", type=float, default=100, help="Set the maximum upload size in MB.")
45
+
46
+ parser.add_argument("--extra-model-paths-config", type=str, default=None, metavar="PATH", nargs='+', action='append', help="Load one or more extra_model_paths.yaml files.")
47
+ parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory.")
48
+ parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory).")
49
+ parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory.")
50
+ parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
51
+ parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
52
+ parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
53
+ cm_group = parser.add_mutually_exclusive_group()
54
+ cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
55
+ cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")
56
+
57
+
58
+ fp_group = parser.add_mutually_exclusive_group()
59
+ fp_group.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).")
60
+ fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.")
61
+
62
+ fpunet_group = parser.add_mutually_exclusive_group()
63
+ fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.")
64
+ fpunet_group.add_argument("--fp16-unet", action="store_true", help="Store unet weights in fp16.")
65
+ fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
66
+ fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
67
+
68
+ fpvae_group = parser.add_mutually_exclusive_group()
69
+ fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")
70
+ fpvae_group.add_argument("--fp32-vae", action="store_true", help="Run the VAE in full precision fp32.")
71
+ fpvae_group.add_argument("--bf16-vae", action="store_true", help="Run the VAE in bf16.")
72
+
73
+ parser.add_argument("--cpu-vae", action="store_true", help="Run the VAE on the CPU.")
74
+
75
+ fpte_group = parser.add_mutually_exclusive_group()
76
+ fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Store text encoder weights in fp8 (e4m3fn variant).")
77
+ fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
78
+ fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
79
+ fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
80
+
81
+ parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
82
+
83
+ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
84
+
85
+ parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize when loading models with Intel GPUs.")
86
+
87
+ class LatentPreviewMethod(enum.Enum):
88
+ NoPreviews = "none"
89
+ Auto = "auto"
90
+ Latent2RGB = "latent2rgb"
91
+ TAESD = "taesd"
92
+
93
+ parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
94
+
95
+ attn_group = parser.add_mutually_exclusive_group()
96
+ attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
97
+ attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
98
+ attn_group.add_argument("--use-pytorch-cross-attention", action="store_true", help="Use the new pytorch 2.0 cross attention function.")
99
+
100
+ parser.add_argument("--disable-xformers", action="store_true", help="Disable xformers.")
101
+
102
+ upcast = parser.add_mutually_exclusive_group()
103
+ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force enable attention upcasting, please report if it fixes black images.")
104
+ upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.")
105
+
106
+
107
+ vram_group = parser.add_mutually_exclusive_group()
108
+ vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
109
+ vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
110
+ vram_group.add_argument("--normalvram", action="store_true", help="Used to force normal vram use if lowvram gets automatically enabled.")
111
+ vram_group.add_argument("--lowvram", action="store_true", help="Split the unet in parts to use less vram.")
112
+ vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
113
+ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
114
+
115
+ parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
116
+
117
+ parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
118
+ parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
119
+
120
+ parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
121
+ parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
122
+ parser.add_argument("--windows-standalone-build", action="store_true", help="Windows standalone build: Enable convenient things that most people using the standalone windows build will probably enjoy (like auto opening the page on startup).")
123
+
124
+ parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
125
+ parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
126
+
127
+ parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
128
+
129
+ parser.add_argument("--verbose", action="store_true", help="Enables more debug prints.")
130
+
131
+ # The default built-in provider hosted under web/
132
+ DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest"
133
+
134
+ parser.add_argument(
135
+ "--front-end-version",
136
+ type=str,
137
+ default=DEFAULT_VERSION_STRING,
138
+ help="""
139
+ Specifies the version of the frontend to be used. This command needs internet connectivity to query and
140
+ download available frontend implementations from GitHub releases.
141
+
142
+ The version string should be in the format of:
143
+ [repoOwner]/[repoName]@[version]
144
+ where version is one of: "latest" or a valid version number (e.g. "1.0.0")
145
+ """,
146
+ )
147
+
148
+ def is_valid_directory(path: Optional[str]) -> Optional[str]:
149
+ """Validate if the given path is a directory."""
150
+ if path is None:
151
+ return None
152
+
153
+ if not os.path.isdir(path):
154
+ raise argparse.ArgumentTypeError(f"{path} is not a valid directory.")
155
+ return path
156
+
157
+ parser.add_argument(
158
+ "--front-end-root",
159
+ type=is_valid_directory,
160
+ default=None,
161
+ help="The local filesystem path to the directory where the frontend is located. Overrides --front-end-version.",
162
+ )
163
+
164
+ if comfy.options.args_parsing:
165
+ args = parser.parse_args()
166
+ else:
167
+ args = parser.parse_args([])
168
+
169
+ if args.windows_standalone_build:
170
+ args.auto_launch = True
171
+
172
+ if args.disable_auto_launch:
173
+ args.auto_launch = False
174
+
175
+ import logging
176
+ logging_level = logging.INFO
177
+ if args.verbose:
178
+ logging_level = logging.DEBUG
179
+
180
+ logging.basicConfig(format="%(message)s", level=logging_level)
ComfyUI/comfy/clip_config_bigg.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "CLIPTextModel"
4
+ ],
5
+ "attention_dropout": 0.0,
6
+ "bos_token_id": 0,
7
+ "dropout": 0.0,
8
+ "eos_token_id": 49407,
9
+ "hidden_act": "gelu",
10
+ "hidden_size": 1280,
11
+ "initializer_factor": 1.0,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 5120,
14
+ "layer_norm_eps": 1e-05,
15
+ "max_position_embeddings": 77,
16
+ "model_type": "clip_text_model",
17
+ "num_attention_heads": 20,
18
+ "num_hidden_layers": 32,
19
+ "pad_token_id": 1,
20
+ "projection_dim": 1280,
21
+ "torch_dtype": "float32",
22
+ "vocab_size": 49408
23
+ }
ComfyUI/comfy/clip_model.py ADDED
@@ -0,0 +1,196 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from comfy.ldm.modules.attention import optimized_attention_for_device
3
+ import comfy.ops
4
+
5
+ class CLIPAttention(torch.nn.Module):
6
+ def __init__(self, embed_dim, heads, dtype, device, operations):
7
+ super().__init__()
8
+
9
+ self.heads = heads
10
+ self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
11
+ self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
12
+ self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
13
+
14
+ self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
15
+
16
+ def forward(self, x, mask=None, optimized_attention=None):
17
+ q = self.q_proj(x)
18
+ k = self.k_proj(x)
19
+ v = self.v_proj(x)
20
+
21
+ out = optimized_attention(q, k, v, self.heads, mask)
22
+ return self.out_proj(out)
23
+
24
+ ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a),
25
+ "gelu": torch.nn.functional.gelu,
26
+ }
27
+
28
+ class CLIPMLP(torch.nn.Module):
29
+ def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations):
30
+ super().__init__()
31
+ self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device)
32
+ self.activation = ACTIVATIONS[activation]
33
+ self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device)
34
+
35
+ def forward(self, x):
36
+ x = self.fc1(x)
37
+ x = self.activation(x)
38
+ x = self.fc2(x)
39
+ return x
40
+
41
+ class CLIPLayer(torch.nn.Module):
42
+ def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
43
+ super().__init__()
44
+ self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
45
+ self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations)
46
+ self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
47
+ self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations)
48
+
49
+ def forward(self, x, mask=None, optimized_attention=None):
50
+ x += self.self_attn(self.layer_norm1(x), mask, optimized_attention)
51
+ x += self.mlp(self.layer_norm2(x))
52
+ return x
53
+
54
+
55
+ class CLIPEncoder(torch.nn.Module):
56
+ def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
57
+ super().__init__()
58
+ self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)])
59
+
60
+ def forward(self, x, mask=None, intermediate_output=None):
61
+ optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)
62
+
63
+ if intermediate_output is not None:
64
+ if intermediate_output < 0:
65
+ intermediate_output = len(self.layers) + intermediate_output
66
+
67
+ intermediate = None
68
+ for i, l in enumerate(self.layers):
69
+ x = l(x, mask, optimized_attention)
70
+ if i == intermediate_output:
71
+ intermediate = x.clone()
72
+ return x, intermediate
73
+
74
+ class CLIPEmbeddings(torch.nn.Module):
75
+ def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None, operations=None):
76
+ super().__init__()
77
+ self.token_embedding = operations.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
78
+ self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
79
+
80
+ def forward(self, input_tokens, dtype=torch.float32):
81
+ return self.token_embedding(input_tokens, out_dtype=dtype) + comfy.ops.cast_to(self.position_embedding.weight, dtype=dtype, device=input_tokens.device)
82
+
83
+
84
+ class CLIPTextModel_(torch.nn.Module):
85
+ def __init__(self, config_dict, dtype, device, operations):
86
+ num_layers = config_dict["num_hidden_layers"]
87
+ embed_dim = config_dict["hidden_size"]
88
+ heads = config_dict["num_attention_heads"]
89
+ intermediate_size = config_dict["intermediate_size"]
90
+ intermediate_activation = config_dict["hidden_act"]
91
+ self.eos_token_id = config_dict["eos_token_id"]
92
+
93
+ super().__init__()
94
+ self.embeddings = CLIPEmbeddings(embed_dim, dtype=dtype, device=device, operations=operations)
95
+ self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
96
+ self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
97
+
98
+ def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
99
+ x = self.embeddings(input_tokens, dtype=dtype)
100
+ mask = None
101
+ if attention_mask is not None:
102
+ mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
103
+ mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
104
+
105
+ causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
106
+ if mask is not None:
107
+ mask += causal_mask
108
+ else:
109
+ mask = causal_mask
110
+
111
+ x, i = self.encoder(x, mask=mask, intermediate_output=intermediate_output)
112
+ x = self.final_layer_norm(x)
113
+ if i is not None and final_layer_norm_intermediate:
114
+ i = self.final_layer_norm(i)
115
+
116
+ pooled_output = x[torch.arange(x.shape[0], device=x.device), (torch.round(input_tokens).to(dtype=torch.int, device=x.device) == self.eos_token_id).int().argmax(dim=-1),]
117
+ return x, i, pooled_output
118
+
119
+ class CLIPTextModel(torch.nn.Module):
120
+ def __init__(self, config_dict, dtype, device, operations):
121
+ super().__init__()
122
+ self.num_layers = config_dict["num_hidden_layers"]
123
+ self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
124
+ embed_dim = config_dict["hidden_size"]
125
+ self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
126
+ self.text_projection.weight.copy_(torch.eye(embed_dim))
127
+ self.dtype = dtype
128
+
129
+ def get_input_embeddings(self):
130
+ return self.text_model.embeddings.token_embedding
131
+
132
+ def set_input_embeddings(self, embeddings):
133
+ self.text_model.embeddings.token_embedding = embeddings
134
+
135
+ def forward(self, *args, **kwargs):
136
+ x = self.text_model(*args, **kwargs)
137
+ out = self.text_projection(x[2])
138
+ return (x[0], x[1], out, x[2])
139
+
140
+
141
+ class CLIPVisionEmbeddings(torch.nn.Module):
142
+ def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dtype=None, device=None, operations=None):
143
+ super().__init__()
144
+ self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device))
145
+
146
+ self.patch_embedding = operations.Conv2d(
147
+ in_channels=num_channels,
148
+ out_channels=embed_dim,
149
+ kernel_size=patch_size,
150
+ stride=patch_size,
151
+ bias=False,
152
+ dtype=dtype,
153
+ device=device
154
+ )
155
+
156
+ num_patches = (image_size // patch_size) ** 2
157
+ num_positions = num_patches + 1
158
+ self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
159
+
160
+ def forward(self, pixel_values):
161
+ embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
162
+ return torch.cat([comfy.ops.cast_to_input(self.class_embedding, embeds).expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + comfy.ops.cast_to_input(self.position_embedding.weight, embeds)
163
+
164
+
165
+ class CLIPVision(torch.nn.Module):
166
+ def __init__(self, config_dict, dtype, device, operations):
167
+ super().__init__()
168
+ num_layers = config_dict["num_hidden_layers"]
169
+ embed_dim = config_dict["hidden_size"]
170
+ heads = config_dict["num_attention_heads"]
171
+ intermediate_size = config_dict["intermediate_size"]
172
+ intermediate_activation = config_dict["hidden_act"]
173
+
174
+ self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=dtype, device=device, operations=operations)
175
+ self.pre_layrnorm = operations.LayerNorm(embed_dim)
176
+ self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
177
+ self.post_layernorm = operations.LayerNorm(embed_dim)
178
+
179
+ def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
180
+ x = self.embeddings(pixel_values)
181
+ x = self.pre_layrnorm(x)
182
+ #TODO: attention_mask?
183
+ x, i = self.encoder(x, mask=None, intermediate_output=intermediate_output)
184
+ pooled_output = self.post_layernorm(x[:, 0, :])
185
+ return x, i, pooled_output
186
+
187
+ class CLIPVisionModelProjection(torch.nn.Module):
188
+ def __init__(self, config_dict, dtype, device, operations):
189
+ super().__init__()
190
+ self.vision_model = CLIPVision(config_dict, dtype, device, operations)
191
+ self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)
192
+
193
+ def forward(self, *args, **kwargs):
194
+ x = self.vision_model(*args, **kwargs)
195
+ out = self.visual_projection(x[2])
196
+ return (x[0], x[1], out)
ComfyUI/comfy/clip_vision.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
2
+ import os
3
+ import torch
4
+ import json
5
+ import logging
6
+
7
+ import comfy.ops
8
+ import comfy.model_patcher
9
+ import comfy.model_management
10
+ import comfy.utils
11
+ import comfy.clip_model
12
+
13
+ class Output:
14
+ def __getitem__(self, key):
15
+ return getattr(self, key)
16
+ def __setitem__(self, key, item):
17
+ setattr(self, key, item)
18
+
19
+ def clip_preprocess(image, size=224):
20
+ mean = torch.tensor([ 0.48145466,0.4578275,0.40821073], device=image.device, dtype=image.dtype)
21
+ std = torch.tensor([0.26862954,0.26130258,0.27577711], device=image.device, dtype=image.dtype)
22
+ image = image.movedim(-1, 1)
23
+ if not (image.shape[2] == size and image.shape[3] == size):
24
+ scale = (size / min(image.shape[2], image.shape[3]))
25
+ image = torch.nn.functional.interpolate(image, size=(round(scale * image.shape[2]), round(scale * image.shape[3])), mode="bicubic", antialias=True)
26
+ h = (image.shape[2] - size)//2
27
+ w = (image.shape[3] - size)//2
28
+ image = image[:,:,h:h+size,w:w+size]
29
+ image = torch.clip((255. * image), 0, 255).round() / 255.0
30
+ return (image - mean.view([3,1,1])) / std.view([3,1,1])
31
+
32
+ class ClipVisionModel():
33
+ def __init__(self, json_config):
34
+ with open(json_config) as f:
35
+ config = json.load(f)
36
+
37
+ self.image_size = config.get("image_size", 224)
38
+ self.load_device = comfy.model_management.text_encoder_device()
39
+ offload_device = comfy.model_management.text_encoder_offload_device()
40
+ self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
41
+ self.model = comfy.clip_model.CLIPVisionModelProjection(config, self.dtype, offload_device, comfy.ops.manual_cast)
42
+ self.model.eval()
43
+
44
+ self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
45
+
46
+ def load_sd(self, sd):
47
+ return self.model.load_state_dict(sd, strict=False)
48
+
49
+ def get_sd(self):
50
+ return self.model.state_dict()
51
+
52
+ def encode_image(self, image):
53
+ comfy.model_management.load_model_gpu(self.patcher)
54
+ pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size).float()
55
+ out = self.model(pixel_values=pixel_values, intermediate_output=-2)
56
+
57
+ outputs = Output()
58
+ outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
59
+ outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
60
+ outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
61
+ return outputs
62
+
63
+ def convert_to_transformers(sd, prefix):
64
+ sd_k = sd.keys()
65
+ if "{}transformer.resblocks.0.attn.in_proj_weight".format(prefix) in sd_k:
66
+ keys_to_replace = {
67
+ "{}class_embedding".format(prefix): "vision_model.embeddings.class_embedding",
68
+ "{}conv1.weight".format(prefix): "vision_model.embeddings.patch_embedding.weight",
69
+ "{}positional_embedding".format(prefix): "vision_model.embeddings.position_embedding.weight",
70
+ "{}ln_post.bias".format(prefix): "vision_model.post_layernorm.bias",
71
+ "{}ln_post.weight".format(prefix): "vision_model.post_layernorm.weight",
72
+ "{}ln_pre.bias".format(prefix): "vision_model.pre_layrnorm.bias",
73
+ "{}ln_pre.weight".format(prefix): "vision_model.pre_layrnorm.weight",
74
+ }
75
+
76
+ for x in keys_to_replace:
77
+ if x in sd_k:
78
+ sd[keys_to_replace[x]] = sd.pop(x)
79
+
80
+ if "{}proj".format(prefix) in sd_k:
81
+ sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1)
82
+
83
+ sd = transformers_convert(sd, prefix, "vision_model.", 48)
84
+ else:
85
+ replace_prefix = {prefix: ""}
86
+ sd = state_dict_prefix_replace(sd, replace_prefix)
87
+ return sd
88
+
89
+ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
90
+ if convert_keys:
91
+ sd = convert_to_transformers(sd, prefix)
92
+ if "vision_model.encoder.layers.47.layer_norm1.weight" in sd:
93
+ json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_g.json")
94
+ elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
95
+ json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
96
+ elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
97
+ if sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577:
98
+ json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
99
+ else:
100
+ json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
101
+ else:
102
+ return None
103
+
104
+ clip = ClipVisionModel(json_config)
105
+ m, u = clip.load_sd(sd)
106
+ if len(m) > 0:
107
+ logging.warning("missing clip vision: {}".format(m))
108
+ u = set(u)
109
+ keys = list(sd.keys())
110
+ for k in keys:
111
+ if k not in u:
112
+ t = sd.pop(k)
113
+ del t
114
+ return clip
115
+
116
+ def load(ckpt_path):
117
+ sd = load_torch_file(ckpt_path)
118
+ if "visual.transformer.resblocks.0.attn.in_proj_weight" in sd:
119
+ return load_clipvision_from_sd(sd, prefix="visual.", convert_keys=True)
120
+ else:
121
+ return load_clipvision_from_sd(sd)
ComfyUI/comfy/clip_vision_config_g.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attention_dropout": 0.0,
3
+ "dropout": 0.0,
4
+ "hidden_act": "gelu",
5
+ "hidden_size": 1664,
6
+ "image_size": 224,
7
+ "initializer_factor": 1.0,
8
+ "initializer_range": 0.02,
9
+ "intermediate_size": 8192,
10
+ "layer_norm_eps": 1e-05,
11
+ "model_type": "clip_vision_model",
12
+ "num_attention_heads": 16,
13
+ "num_channels": 3,
14
+ "num_hidden_layers": 48,
15
+ "patch_size": 14,
16
+ "projection_dim": 1280,
17
+ "torch_dtype": "float32"
18
+ }
ComfyUI/comfy/clip_vision_config_h.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attention_dropout": 0.0,
3
+ "dropout": 0.0,
4
+ "hidden_act": "gelu",
5
+ "hidden_size": 1280,
6
+ "image_size": 224,
7
+ "initializer_factor": 1.0,
8
+ "initializer_range": 0.02,
9
+ "intermediate_size": 5120,
10
+ "layer_norm_eps": 1e-05,
11
+ "model_type": "clip_vision_model",
12
+ "num_attention_heads": 16,
13
+ "num_channels": 3,
14
+ "num_hidden_layers": 32,
15
+ "patch_size": 14,
16
+ "projection_dim": 1024,
17
+ "torch_dtype": "float32"
18
+ }
ComfyUI/comfy/clip_vision_config_vitl.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attention_dropout": 0.0,
3
+ "dropout": 0.0,
4
+ "hidden_act": "quick_gelu",
5
+ "hidden_size": 1024,
6
+ "image_size": 224,
7
+ "initializer_factor": 1.0,
8
+ "initializer_range": 0.02,
9
+ "intermediate_size": 4096,
10
+ "layer_norm_eps": 1e-05,
11
+ "model_type": "clip_vision_model",
12
+ "num_attention_heads": 16,
13
+ "num_channels": 3,
14
+ "num_hidden_layers": 24,
15
+ "patch_size": 14,
16
+ "projection_dim": 768,
17
+ "torch_dtype": "float32"
18
+ }
ComfyUI/comfy/clip_vision_config_vitl_336.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attention_dropout": 0.0,
3
+ "dropout": 0.0,
4
+ "hidden_act": "quick_gelu",
5
+ "hidden_size": 1024,
6
+ "image_size": 336,
7
+ "initializer_factor": 1.0,
8
+ "initializer_range": 0.02,
9
+ "intermediate_size": 4096,
10
+ "layer_norm_eps": 1e-5,
11
+ "model_type": "clip_vision_model",
12
+ "num_attention_heads": 16,
13
+ "num_channels": 3,
14
+ "num_hidden_layers": 24,
15
+ "patch_size": 14,
16
+ "projection_dim": 768,
17
+ "torch_dtype": "float32"
18
+ }
ComfyUI/comfy/conds.py ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import math
3
+ import comfy.utils
4
+
5
+
6
+ def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
7
+ return abs(a*b) // math.gcd(a, b)
8
+
9
+ class CONDRegular:
10
+ def __init__(self, cond):
11
+ self.cond = cond
12
+
13
+ def _copy_with(self, cond):
14
+ return self.__class__(cond)
15
+
16
+ def process_cond(self, batch_size, device, **kwargs):
17
+ return self._copy_with(comfy.utils.repeat_to_batch_size(self.cond, batch_size).to(device))
18
+
19
+ def can_concat(self, other):
20
+ if self.cond.shape != other.cond.shape:
21
+ return False
22
+ return True
23
+
24
+ def concat(self, others):
25
+ conds = [self.cond]
26
+ for x in others:
27
+ conds.append(x.cond)
28
+ return torch.cat(conds)
29
+
30
+ class CONDNoiseShape(CONDRegular):
31
+ def process_cond(self, batch_size, device, area, **kwargs):
32
+ data = self.cond
33
+ if area is not None:
34
+ dims = len(area) // 2
35
+ for i in range(dims):
36
+ data = data.narrow(i + 2, area[i + dims], area[i])
37
+
38
+ return self._copy_with(comfy.utils.repeat_to_batch_size(data, batch_size).to(device))
39
+
40
+
41
+ class CONDCrossAttn(CONDRegular):
42
+ def can_concat(self, other):
43
+ s1 = self.cond.shape
44
+ s2 = other.cond.shape
45
+ if s1 != s2:
46
+ if s1[0] != s2[0] or s1[2] != s2[2]: #these 2 cases should not happen
47
+ return False
48
+
49
+ mult_min = lcm(s1[1], s2[1])
50
+ diff = mult_min // min(s1[1], s2[1])
51
+ if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
52
+ return False
53
+ return True
54
+
55
+ def concat(self, others):
56
+ conds = [self.cond]
57
+ crossattn_max_len = self.cond.shape[1]
58
+ for x in others:
59
+ c = x.cond
60
+ crossattn_max_len = lcm(crossattn_max_len, c.shape[1])
61
+ conds.append(c)
62
+
63
+ out = []
64
+ for c in conds:
65
+ if c.shape[1] < crossattn_max_len:
66
+ c = c.repeat(1, crossattn_max_len // c.shape[1], 1) #padding with repeat doesn't change result
67
+ out.append(c)
68
+ return torch.cat(out)
69
+
70
+ class CONDConstant(CONDRegular):
71
+ def __init__(self, cond):
72
+ self.cond = cond
73
+
74
+ def process_cond(self, batch_size, device, **kwargs):
75
+ return self._copy_with(self.cond)
76
+
77
+ def can_concat(self, other):
78
+ if self.cond != other.cond:
79
+ return False
80
+ return True
81
+
82
+ def concat(self, others):
83
+ return self.cond
ComfyUI/comfy/controlnet.py ADDED
@@ -0,0 +1,622 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import math
3
+ import os
4
+ import logging
5
+ import comfy.utils
6
+ import comfy.model_management
7
+ import comfy.model_detection
8
+ import comfy.model_patcher
9
+ import comfy.ops
10
+ import comfy.latent_formats
11
+
12
+ import comfy.cldm.cldm
13
+ import comfy.t2i_adapter.adapter
14
+ import comfy.ldm.cascade.controlnet
15
+ import comfy.cldm.mmdit
16
+
17
+
18
+ def broadcast_image_to(tensor, target_batch_size, batched_number):
19
+ current_batch_size = tensor.shape[0]
20
+ #print(current_batch_size, target_batch_size)
21
+ if current_batch_size == 1:
22
+ return tensor
23
+
24
+ per_batch = target_batch_size // batched_number
25
+ tensor = tensor[:per_batch]
26
+
27
+ if per_batch > tensor.shape[0]:
28
+ tensor = torch.cat([tensor] * (per_batch // tensor.shape[0]) + [tensor[:(per_batch % tensor.shape[0])]], dim=0)
29
+
30
+ current_batch_size = tensor.shape[0]
31
+ if current_batch_size == target_batch_size:
32
+ return tensor
33
+ else:
34
+ return torch.cat([tensor] * batched_number, dim=0)
35
+
36
+ class ControlBase:
37
+ def __init__(self, device=None):
38
+ self.cond_hint_original = None
39
+ self.cond_hint = None
40
+ self.strength = 1.0
41
+ self.timestep_percent_range = (0.0, 1.0)
42
+ self.latent_format = None
43
+ self.vae = None
44
+ self.global_average_pooling = False
45
+ self.timestep_range = None
46
+ self.compression_ratio = 8
47
+ self.upscale_algorithm = 'nearest-exact'
48
+ self.extra_args = {}
49
+
50
+ if device is None:
51
+ device = comfy.model_management.get_torch_device()
52
+ self.device = device
53
+ self.previous_controlnet = None
54
+
55
+ def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None):
56
+ self.cond_hint_original = cond_hint
57
+ self.strength = strength
58
+ self.timestep_percent_range = timestep_percent_range
59
+ if self.latent_format is not None:
60
+ self.vae = vae
61
+ return self
62
+
63
+ def pre_run(self, model, percent_to_timestep_function):
64
+ self.timestep_range = (percent_to_timestep_function(self.timestep_percent_range[0]), percent_to_timestep_function(self.timestep_percent_range[1]))
65
+ if self.previous_controlnet is not None:
66
+ self.previous_controlnet.pre_run(model, percent_to_timestep_function)
67
+
68
+ def set_previous_controlnet(self, controlnet):
69
+ self.previous_controlnet = controlnet
70
+ return self
71
+
72
+ def cleanup(self):
73
+ if self.previous_controlnet is not None:
74
+ self.previous_controlnet.cleanup()
75
+ if self.cond_hint is not None:
76
+ del self.cond_hint
77
+ self.cond_hint = None
78
+ self.timestep_range = None
79
+
80
+ def get_models(self):
81
+ out = []
82
+ if self.previous_controlnet is not None:
83
+ out += self.previous_controlnet.get_models()
84
+ return out
85
+
86
+ def copy_to(self, c):
87
+ c.cond_hint_original = self.cond_hint_original
88
+ c.strength = self.strength
89
+ c.timestep_percent_range = self.timestep_percent_range
90
+ c.global_average_pooling = self.global_average_pooling
91
+ c.compression_ratio = self.compression_ratio
92
+ c.upscale_algorithm = self.upscale_algorithm
93
+ c.latent_format = self.latent_format
94
+ c.extra_args = self.extra_args.copy()
95
+ c.vae = self.vae
96
+
97
+ def inference_memory_requirements(self, dtype):
98
+ if self.previous_controlnet is not None:
99
+ return self.previous_controlnet.inference_memory_requirements(dtype)
100
+ return 0
101
+
102
+ def control_merge(self, control, control_prev, output_dtype):
103
+ out = {'input':[], 'middle':[], 'output': []}
104
+
105
+ for key in control:
106
+ control_output = control[key]
107
+ applied_to = set()
108
+ for i in range(len(control_output)):
109
+ x = control_output[i]
110
+ if x is not None:
111
+ if self.global_average_pooling:
112
+ x = torch.mean(x, dim=(2, 3), keepdim=True).repeat(1, 1, x.shape[2], x.shape[3])
113
+
114
+ if x not in applied_to: #memory saving strategy, allow shared tensors and only apply strength to shared tensors once
115
+ applied_to.add(x)
116
+ x *= self.strength
117
+
118
+ if x.dtype != output_dtype:
119
+ x = x.to(output_dtype)
120
+
121
+ out[key].append(x)
122
+
123
+ if control_prev is not None:
124
+ for x in ['input', 'middle', 'output']:
125
+ o = out[x]
126
+ for i in range(len(control_prev[x])):
127
+ prev_val = control_prev[x][i]
128
+ if i >= len(o):
129
+ o.append(prev_val)
130
+ elif prev_val is not None:
131
+ if o[i] is None:
132
+ o[i] = prev_val
133
+ else:
134
+ if o[i].shape[0] < prev_val.shape[0]:
135
+ o[i] = prev_val + o[i]
136
+ else:
137
+ o[i] = prev_val + o[i] #TODO: change back to inplace add if shared tensors stop being an issue
138
+ return out
139
+
140
+ def set_extra_arg(self, argument, value=None):
141
+ self.extra_args[argument] = value
142
+
143
+
144
+ class ControlNet(ControlBase):
145
+ def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None):
146
+ super().__init__(device)
147
+ self.control_model = control_model
148
+ self.load_device = load_device
149
+ if control_model is not None:
150
+ self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
151
+
152
+ self.compression_ratio = compression_ratio
153
+ self.global_average_pooling = global_average_pooling
154
+ self.model_sampling_current = None
155
+ self.manual_cast_dtype = manual_cast_dtype
156
+ self.latent_format = latent_format
157
+
158
+ def get_control(self, x_noisy, t, cond, batched_number):
159
+ control_prev = None
160
+ if self.previous_controlnet is not None:
161
+ control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
162
+
163
+ if self.timestep_range is not None:
164
+ if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
165
+ if control_prev is not None:
166
+ return control_prev
167
+ else:
168
+ return None
169
+
170
+ dtype = self.control_model.dtype
171
+ if self.manual_cast_dtype is not None:
172
+ dtype = self.manual_cast_dtype
173
+
174
+ output_dtype = x_noisy.dtype
175
+ if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]:
176
+ if self.cond_hint is not None:
177
+ del self.cond_hint
178
+ self.cond_hint = None
179
+ compression_ratio = self.compression_ratio
180
+ if self.vae is not None:
181
+ compression_ratio *= self.vae.downscale_ratio
182
+ self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * compression_ratio, x_noisy.shape[2] * compression_ratio, self.upscale_algorithm, "center")
183
+ if self.vae is not None:
184
+ loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
185
+ self.cond_hint = self.vae.encode(self.cond_hint.movedim(1, -1))
186
+ comfy.model_management.load_models_gpu(loaded_models)
187
+ if self.latent_format is not None:
188
+ self.cond_hint = self.latent_format.process_in(self.cond_hint)
189
+ self.cond_hint = self.cond_hint.to(device=self.device, dtype=dtype)
190
+ if x_noisy.shape[0] != self.cond_hint.shape[0]:
191
+ self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
192
+
193
+ context = cond.get('crossattn_controlnet', cond['c_crossattn'])
194
+ extra = self.extra_args.copy()
195
+ for c in ["y", "guidance"]: #TODO
196
+ temp = cond.get(c, None)
197
+ if temp is not None:
198
+ extra[c] = temp.to(dtype)
199
+
200
+ timestep = self.model_sampling_current.timestep(t)
201
+ x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
202
+
203
+ control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
204
+ return self.control_merge(control, control_prev, output_dtype)
205
+
206
+ def copy(self):
207
+ c = ControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
208
+ c.control_model = self.control_model
209
+ c.control_model_wrapped = self.control_model_wrapped
210
+ self.copy_to(c)
211
+ return c
212
+
213
+ def get_models(self):
214
+ out = super().get_models()
215
+ out.append(self.control_model_wrapped)
216
+ return out
217
+
218
+ def pre_run(self, model, percent_to_timestep_function):
219
+ super().pre_run(model, percent_to_timestep_function)
220
+ self.model_sampling_current = model.model_sampling
221
+
222
+ def cleanup(self):
223
+ self.model_sampling_current = None
224
+ super().cleanup()
225
+
226
+ class ControlLoraOps:
227
+ class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp):
228
+ def __init__(self, in_features: int, out_features: int, bias: bool = True,
229
+ device=None, dtype=None) -> None:
230
+ factory_kwargs = {'device': device, 'dtype': dtype}
231
+ super().__init__()
232
+ self.in_features = in_features
233
+ self.out_features = out_features
234
+ self.weight = None
235
+ self.up = None
236
+ self.down = None
237
+ self.bias = None
238
+
239
+ def forward(self, input):
240
+ weight, bias = comfy.ops.cast_bias_weight(self, input)
241
+ if self.up is not None:
242
+ return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
243
+ else:
244
+ return torch.nn.functional.linear(input, weight, bias)
245
+
246
+ class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
247
+ def __init__(
248
+ self,
249
+ in_channels,
250
+ out_channels,
251
+ kernel_size,
252
+ stride=1,
253
+ padding=0,
254
+ dilation=1,
255
+ groups=1,
256
+ bias=True,
257
+ padding_mode='zeros',
258
+ device=None,
259
+ dtype=None
260
+ ):
261
+ super().__init__()
262
+ self.in_channels = in_channels
263
+ self.out_channels = out_channels
264
+ self.kernel_size = kernel_size
265
+ self.stride = stride
266
+ self.padding = padding
267
+ self.dilation = dilation
268
+ self.transposed = False
269
+ self.output_padding = 0
270
+ self.groups = groups
271
+ self.padding_mode = padding_mode
272
+
273
+ self.weight = None
274
+ self.bias = None
275
+ self.up = None
276
+ self.down = None
277
+
278
+
279
+ def forward(self, input):
280
+ weight, bias = comfy.ops.cast_bias_weight(self, input)
281
+ if self.up is not None:
282
+ return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
283
+ else:
284
+ return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
285
+
286
+
287
+ class ControlLora(ControlNet):
288
+ def __init__(self, control_weights, global_average_pooling=False, device=None):
289
+ ControlBase.__init__(self, device)
290
+ self.control_weights = control_weights
291
+ self.global_average_pooling = global_average_pooling
292
+
293
+ def pre_run(self, model, percent_to_timestep_function):
294
+ super().pre_run(model, percent_to_timestep_function)
295
+ controlnet_config = model.model_config.unet_config.copy()
296
+ controlnet_config.pop("out_channels")
297
+ controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1]
298
+ self.manual_cast_dtype = model.manual_cast_dtype
299
+ dtype = model.get_dtype()
300
+ if self.manual_cast_dtype is None:
301
+ class control_lora_ops(ControlLoraOps, comfy.ops.disable_weight_init):
302
+ pass
303
+ else:
304
+ class control_lora_ops(ControlLoraOps, comfy.ops.manual_cast):
305
+ pass
306
+ dtype = self.manual_cast_dtype
307
+
308
+ controlnet_config["operations"] = control_lora_ops
309
+ controlnet_config["dtype"] = dtype
310
+ self.control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
311
+ self.control_model.to(comfy.model_management.get_torch_device())
312
+ diffusion_model = model.diffusion_model
313
+ sd = diffusion_model.state_dict()
314
+ cm = self.control_model.state_dict()
315
+
316
+ for k in sd:
317
+ weight = sd[k]
318
+ try:
319
+ comfy.utils.set_attr_param(self.control_model, k, weight)
320
+ except:
321
+ pass
322
+
323
+ for k in self.control_weights:
324
+ if k not in {"lora_controlnet"}:
325
+ comfy.utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device()))
326
+
327
+ def copy(self):
328
+ c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
329
+ self.copy_to(c)
330
+ return c
331
+
332
+ def cleanup(self):
333
+ del self.control_model
334
+ self.control_model = None
335
+ super().cleanup()
336
+
337
+ def get_models(self):
338
+ out = ControlBase.get_models(self)
339
+ return out
340
+
341
+ def inference_memory_requirements(self, dtype):
342
+ return comfy.utils.calculate_parameters(self.control_weights) * comfy.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
343
+
344
+ def controlnet_config(sd):
345
+ model_config = comfy.model_detection.model_config_from_unet(sd, "", True)
346
+
347
+ supported_inference_dtypes = model_config.supported_inference_dtypes
348
+
349
+ controlnet_config = model_config.unet_config
350
+ unet_dtype = comfy.model_management.unet_dtype(supported_dtypes=supported_inference_dtypes)
351
+ load_device = comfy.model_management.get_torch_device()
352
+ manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
353
+ if manual_cast_dtype is not None:
354
+ operations = comfy.ops.manual_cast
355
+ else:
356
+ operations = comfy.ops.disable_weight_init
357
+
358
+ return model_config, operations, load_device, unet_dtype, manual_cast_dtype
359
+
360
+ def controlnet_load_state_dict(control_model, sd):
361
+ missing, unexpected = control_model.load_state_dict(sd, strict=False)
362
+
363
+ if len(missing) > 0:
364
+ logging.warning("missing controlnet keys: {}".format(missing))
365
+
366
+ if len(unexpected) > 0:
367
+ logging.debug("unexpected controlnet keys: {}".format(unexpected))
368
+ return control_model
369
+
370
+ def load_controlnet_mmdit(sd):
371
+ new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
372
+ model_config, operations, load_device, unet_dtype, manual_cast_dtype = controlnet_config(new_sd)
373
+ num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
374
+ for k in sd:
375
+ new_sd[k] = sd[k]
376
+
377
+ control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=load_device, dtype=unet_dtype, **model_config.unet_config)
378
+ control_model = controlnet_load_state_dict(control_model, new_sd)
379
+
380
+ latent_format = comfy.latent_formats.SD3()
381
+ latent_format.shift_factor = 0 #SD3 controlnet weirdness
382
+ control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
383
+ return control
384
+
385
+
386
+ def load_controlnet(ckpt_path, model=None):
387
+ controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
388
+ if "lora_controlnet" in controlnet_data:
389
+ return ControlLora(controlnet_data)
390
+
391
+ controlnet_config = None
392
+ supported_inference_dtypes = None
393
+
394
+ if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
395
+ controlnet_config = comfy.model_detection.unet_config_from_diffusers_unet(controlnet_data)
396
+ diffusers_keys = comfy.utils.unet_to_diffusers(controlnet_config)
397
+ diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
398
+ diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
399
+
400
+ count = 0
401
+ loop = True
402
+ while loop:
403
+ suffix = [".weight", ".bias"]
404
+ for s in suffix:
405
+ k_in = "controlnet_down_blocks.{}{}".format(count, s)
406
+ k_out = "zero_convs.{}.0{}".format(count, s)
407
+ if k_in not in controlnet_data:
408
+ loop = False
409
+ break
410
+ diffusers_keys[k_in] = k_out
411
+ count += 1
412
+
413
+ count = 0
414
+ loop = True
415
+ while loop:
416
+ suffix = [".weight", ".bias"]
417
+ for s in suffix:
418
+ if count == 0:
419
+ k_in = "controlnet_cond_embedding.conv_in{}".format(s)
420
+ else:
421
+ k_in = "controlnet_cond_embedding.blocks.{}{}".format(count - 1, s)
422
+ k_out = "input_hint_block.{}{}".format(count * 2, s)
423
+ if k_in not in controlnet_data:
424
+ k_in = "controlnet_cond_embedding.conv_out{}".format(s)
425
+ loop = False
426
+ diffusers_keys[k_in] = k_out
427
+ count += 1
428
+
429
+ new_sd = {}
430
+ for k in diffusers_keys:
431
+ if k in controlnet_data:
432
+ new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
433
+
434
+ if "control_add_embedding.linear_1.bias" in controlnet_data: #Union Controlnet
435
+ controlnet_config["union_controlnet_num_control_type"] = controlnet_data["task_embedding"].shape[0]
436
+ for k in list(controlnet_data.keys()):
437
+ new_k = k.replace('.attn.in_proj_', '.attn.in_proj.')
438
+ new_sd[new_k] = controlnet_data.pop(k)
439
+
440
+ leftover_keys = controlnet_data.keys()
441
+ if len(leftover_keys) > 0:
442
+ logging.warning("leftover keys: {}".format(leftover_keys))
443
+ controlnet_data = new_sd
444
+ elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
445
+ return load_controlnet_mmdit(controlnet_data)
446
+
447
+ pth_key = 'control_model.zero_convs.0.0.weight'
448
+ pth = False
449
+ key = 'zero_convs.0.0.weight'
450
+ if pth_key in controlnet_data:
451
+ pth = True
452
+ key = pth_key
453
+ prefix = "control_model."
454
+ elif key in controlnet_data:
455
+ prefix = ""
456
+ else:
457
+ net = load_t2i_adapter(controlnet_data)
458
+ if net is None:
459
+ logging.error("error checkpoint does not contain controlnet or t2i adapter data {}".format(ckpt_path))
460
+ return net
461
+
462
+ if controlnet_config is None:
463
+ model_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, True)
464
+ supported_inference_dtypes = model_config.supported_inference_dtypes
465
+ controlnet_config = model_config.unet_config
466
+
467
+ load_device = comfy.model_management.get_torch_device()
468
+ if supported_inference_dtypes is None:
469
+ unet_dtype = comfy.model_management.unet_dtype()
470
+ else:
471
+ unet_dtype = comfy.model_management.unet_dtype(supported_dtypes=supported_inference_dtypes)
472
+
473
+ manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
474
+ if manual_cast_dtype is not None:
475
+ controlnet_config["operations"] = comfy.ops.manual_cast
476
+ controlnet_config["dtype"] = unet_dtype
477
+ controlnet_config.pop("out_channels")
478
+ controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
479
+ control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
480
+
481
+ if pth:
482
+ if 'difference' in controlnet_data:
483
+ if model is not None:
484
+ comfy.model_management.load_models_gpu([model])
485
+ model_sd = model.model_state_dict()
486
+ for x in controlnet_data:
487
+ c_m = "control_model."
488
+ if x.startswith(c_m):
489
+ sd_key = "diffusion_model.{}".format(x[len(c_m):])
490
+ if sd_key in model_sd:
491
+ cd = controlnet_data[x]
492
+ cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
493
+ else:
494
+ logging.warning("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
495
+
496
+ class WeightsLoader(torch.nn.Module):
497
+ pass
498
+ w = WeightsLoader()
499
+ w.control_model = control_model
500
+ missing, unexpected = w.load_state_dict(controlnet_data, strict=False)
501
+ else:
502
+ missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False)
503
+
504
+ if len(missing) > 0:
505
+ logging.warning("missing controlnet keys: {}".format(missing))
506
+
507
+ if len(unexpected) > 0:
508
+ logging.debug("unexpected controlnet keys: {}".format(unexpected))
509
+
510
+ global_average_pooling = False
511
+ filename = os.path.splitext(ckpt_path)[0]
512
+ if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling
513
+ global_average_pooling = True
514
+
515
+ control = ControlNet(control_model, global_average_pooling=global_average_pooling, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
516
+ return control
517
+
518
+ class T2IAdapter(ControlBase):
519
+ def __init__(self, t2i_model, channels_in, compression_ratio, upscale_algorithm, device=None):
520
+ super().__init__(device)
521
+ self.t2i_model = t2i_model
522
+ self.channels_in = channels_in
523
+ self.control_input = None
524
+ self.compression_ratio = compression_ratio
525
+ self.upscale_algorithm = upscale_algorithm
526
+
527
+ def scale_image_to(self, width, height):
528
+ unshuffle_amount = self.t2i_model.unshuffle_amount
529
+ width = math.ceil(width / unshuffle_amount) * unshuffle_amount
530
+ height = math.ceil(height / unshuffle_amount) * unshuffle_amount
531
+ return width, height
532
+
533
+ def get_control(self, x_noisy, t, cond, batched_number):
534
+ control_prev = None
535
+ if self.previous_controlnet is not None:
536
+ control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
537
+
538
+ if self.timestep_range is not None:
539
+ if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
540
+ if control_prev is not None:
541
+ return control_prev
542
+ else:
543
+ return None
544
+
545
+ if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]:
546
+ if self.cond_hint is not None:
547
+ del self.cond_hint
548
+ self.control_input = None
549
+ self.cond_hint = None
550
+ width, height = self.scale_image_to(x_noisy.shape[3] * self.compression_ratio, x_noisy.shape[2] * self.compression_ratio)
551
+ self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, width, height, self.upscale_algorithm, "center").float().to(self.device)
552
+ if self.channels_in == 1 and self.cond_hint.shape[1] > 1:
553
+ self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True)
554
+ if x_noisy.shape[0] != self.cond_hint.shape[0]:
555
+ self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
556
+ if self.control_input is None:
557
+ self.t2i_model.to(x_noisy.dtype)
558
+ self.t2i_model.to(self.device)
559
+ self.control_input = self.t2i_model(self.cond_hint.to(x_noisy.dtype))
560
+ self.t2i_model.cpu()
561
+
562
+ control_input = {}
563
+ for k in self.control_input:
564
+ control_input[k] = list(map(lambda a: None if a is None else a.clone(), self.control_input[k]))
565
+
566
+ return self.control_merge(control_input, control_prev, x_noisy.dtype)
567
+
568
+ def copy(self):
569
+ c = T2IAdapter(self.t2i_model, self.channels_in, self.compression_ratio, self.upscale_algorithm)
570
+ self.copy_to(c)
571
+ return c
572
+
573
+ def load_t2i_adapter(t2i_data):
574
+ compression_ratio = 8
575
+ upscale_algorithm = 'nearest-exact'
576
+
577
+ if 'adapter' in t2i_data:
578
+ t2i_data = t2i_data['adapter']
579
+ if 'adapter.body.0.resnets.0.block1.weight' in t2i_data: #diffusers format
580
+ prefix_replace = {}
581
+ for i in range(4):
582
+ for j in range(2):
583
+ prefix_replace["adapter.body.{}.resnets.{}.".format(i, j)] = "body.{}.".format(i * 2 + j)
584
+ prefix_replace["adapter.body.{}.".format(i, j)] = "body.{}.".format(i * 2)
585
+ prefix_replace["adapter."] = ""
586
+ t2i_data = comfy.utils.state_dict_prefix_replace(t2i_data, prefix_replace)
587
+ keys = t2i_data.keys()
588
+
589
+ if "body.0.in_conv.weight" in keys:
590
+ cin = t2i_data['body.0.in_conv.weight'].shape[1]
591
+ model_ad = comfy.t2i_adapter.adapter.Adapter_light(cin=cin, channels=[320, 640, 1280, 1280], nums_rb=4)
592
+ elif 'conv_in.weight' in keys:
593
+ cin = t2i_data['conv_in.weight'].shape[1]
594
+ channel = t2i_data['conv_in.weight'].shape[0]
595
+ ksize = t2i_data['body.0.block2.weight'].shape[2]
596
+ use_conv = False
597
+ down_opts = list(filter(lambda a: a.endswith("down_opt.op.weight"), keys))
598
+ if len(down_opts) > 0:
599
+ use_conv = True
600
+ xl = False
601
+ if cin == 256 or cin == 768:
602
+ xl = True
603
+ model_ad = comfy.t2i_adapter.adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl)
604
+ elif "backbone.0.0.weight" in keys:
605
+ model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.0.weight'].shape[1], proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63])
606
+ compression_ratio = 32
607
+ upscale_algorithm = 'bilinear'
608
+ elif "backbone.10.blocks.0.weight" in keys:
609
+ model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.weight'].shape[1], bottleneck_mode="large", proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63])
610
+ compression_ratio = 1
611
+ upscale_algorithm = 'nearest-exact'
612
+ else:
613
+ return None
614
+
615
+ missing, unexpected = model_ad.load_state_dict(t2i_data)
616
+ if len(missing) > 0:
617
+ logging.warning("t2i missing {}".format(missing))
618
+
619
+ if len(unexpected) > 0:
620
+ logging.debug("t2i unexpected {}".format(unexpected))
621
+
622
+ return T2IAdapter(model_ad, model_ad.input_channels, compression_ratio, upscale_algorithm)
ComfyUI/comfy/diffusers_convert.py ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ import torch
3
+ import logging
4
+
5
+ # conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py
6
+
7
+ # =================#
8
+ # UNet Conversion #
9
+ # =================#
10
+
11
+ unet_conversion_map = [
12
+ # (stable-diffusion, HF Diffusers)
13
+ ("time_embed.0.weight", "time_embedding.linear_1.weight"),
14
+ ("time_embed.0.bias", "time_embedding.linear_1.bias"),
15
+ ("time_embed.2.weight", "time_embedding.linear_2.weight"),
16
+ ("time_embed.2.bias", "time_embedding.linear_2.bias"),
17
+ ("input_blocks.0.0.weight", "conv_in.weight"),
18
+ ("input_blocks.0.0.bias", "conv_in.bias"),
19
+ ("out.0.weight", "conv_norm_out.weight"),
20
+ ("out.0.bias", "conv_norm_out.bias"),
21
+ ("out.2.weight", "conv_out.weight"),
22
+ ("out.2.bias", "conv_out.bias"),
23
+ ]
24
+
25
+ unet_conversion_map_resnet = [
26
+ # (stable-diffusion, HF Diffusers)
27
+ ("in_layers.0", "norm1"),
28
+ ("in_layers.2", "conv1"),
29
+ ("out_layers.0", "norm2"),
30
+ ("out_layers.3", "conv2"),
31
+ ("emb_layers.1", "time_emb_proj"),
32
+ ("skip_connection", "conv_shortcut"),
33
+ ]
34
+
35
+ unet_conversion_map_layer = []
36
+ # hardcoded number of downblocks and resnets/attentions...
37
+ # would need smarter logic for other networks.
38
+ for i in range(4):
39
+ # loop over downblocks/upblocks
40
+
41
+ for j in range(2):
42
+ # loop over resnets/attentions for downblocks
43
+ hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
44
+ sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
45
+ unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
46
+
47
+ if i < 3:
48
+ # no attention layers in down_blocks.3
49
+ hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
50
+ sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
51
+ unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
52
+
53
+ for j in range(3):
54
+ # loop over resnets/attentions for upblocks
55
+ hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
56
+ sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
57
+ unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
58
+
59
+ if i > 0:
60
+ # no attention layers in up_blocks.0
61
+ hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
62
+ sd_up_atn_prefix = f"output_blocks.{3 * i + j}.1."
63
+ unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
64
+
65
+ if i < 3:
66
+ # no downsample in down_blocks.3
67
+ hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
68
+ sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
69
+ unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
70
+
71
+ # no upsample in up_blocks.3
72
+ hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
73
+ sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
74
+ unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
75
+
76
+ hf_mid_atn_prefix = "mid_block.attentions.0."
77
+ sd_mid_atn_prefix = "middle_block.1."
78
+ unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
79
+
80
+ for j in range(2):
81
+ hf_mid_res_prefix = f"mid_block.resnets.{j}."
82
+ sd_mid_res_prefix = f"middle_block.{2 * j}."
83
+ unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
84
+
85
+
86
+ def convert_unet_state_dict(unet_state_dict):
87
+ # buyer beware: this is a *brittle* function,
88
+ # and correct output requires that all of these pieces interact in
89
+ # the exact order in which I have arranged them.
90
+ mapping = {k: k for k in unet_state_dict.keys()}
91
+ for sd_name, hf_name in unet_conversion_map:
92
+ mapping[hf_name] = sd_name
93
+ for k, v in mapping.items():
94
+ if "resnets" in k:
95
+ for sd_part, hf_part in unet_conversion_map_resnet:
96
+ v = v.replace(hf_part, sd_part)
97
+ mapping[k] = v
98
+ for k, v in mapping.items():
99
+ for sd_part, hf_part in unet_conversion_map_layer:
100
+ v = v.replace(hf_part, sd_part)
101
+ mapping[k] = v
102
+ new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
103
+ return new_state_dict
104
+
105
+
106
+ # ================#
107
+ # VAE Conversion #
108
+ # ================#
109
+
110
+ vae_conversion_map = [
111
+ # (stable-diffusion, HF Diffusers)
112
+ ("nin_shortcut", "conv_shortcut"),
113
+ ("norm_out", "conv_norm_out"),
114
+ ("mid.attn_1.", "mid_block.attentions.0."),
115
+ ]
116
+
117
+ for i in range(4):
118
+ # down_blocks have two resnets
119
+ for j in range(2):
120
+ hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
121
+ sd_down_prefix = f"encoder.down.{i}.block.{j}."
122
+ vae_conversion_map.append((sd_down_prefix, hf_down_prefix))
123
+
124
+ if i < 3:
125
+ hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
126
+ sd_downsample_prefix = f"down.{i}.downsample."
127
+ vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
128
+
129
+ hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
130
+ sd_upsample_prefix = f"up.{3 - i}.upsample."
131
+ vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
132
+
133
+ # up_blocks have three resnets
134
+ # also, up blocks in hf are numbered in reverse from sd
135
+ for j in range(3):
136
+ hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
137
+ sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
138
+ vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
139
+
140
+ # this part accounts for mid blocks in both the encoder and the decoder
141
+ for i in range(2):
142
+ hf_mid_res_prefix = f"mid_block.resnets.{i}."
143
+ sd_mid_res_prefix = f"mid.block_{i + 1}."
144
+ vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
145
+
146
+ vae_conversion_map_attn = [
147
+ # (stable-diffusion, HF Diffusers)
148
+ ("norm.", "group_norm."),
149
+ ("q.", "query."),
150
+ ("k.", "key."),
151
+ ("v.", "value."),
152
+ ("q.", "to_q."),
153
+ ("k.", "to_k."),
154
+ ("v.", "to_v."),
155
+ ("proj_out.", "to_out.0."),
156
+ ("proj_out.", "proj_attn."),
157
+ ]
158
+
159
+
160
+ def reshape_weight_for_sd(w):
161
+ # convert HF linear weights to SD conv2d weights
162
+ return w.reshape(*w.shape, 1, 1)
163
+
164
+
165
+ def convert_vae_state_dict(vae_state_dict):
166
+ mapping = {k: k for k in vae_state_dict.keys()}
167
+ for k, v in mapping.items():
168
+ for sd_part, hf_part in vae_conversion_map:
169
+ v = v.replace(hf_part, sd_part)
170
+ mapping[k] = v
171
+ for k, v in mapping.items():
172
+ if "attentions" in k:
173
+ for sd_part, hf_part in vae_conversion_map_attn:
174
+ v = v.replace(hf_part, sd_part)
175
+ mapping[k] = v
176
+ new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
177
+ weights_to_convert = ["q", "k", "v", "proj_out"]
178
+ for k, v in new_state_dict.items():
179
+ for weight_name in weights_to_convert:
180
+ if f"mid.attn_1.{weight_name}.weight" in k:
181
+ logging.debug(f"Reshaping {k} for SD format")
182
+ new_state_dict[k] = reshape_weight_for_sd(v)
183
+ return new_state_dict
184
+
185
+
186
+ # =========================#
187
+ # Text Encoder Conversion #
188
+ # =========================#
189
+
190
+
191
+ textenc_conversion_lst = [
192
+ # (stable-diffusion, HF Diffusers)
193
+ ("resblocks.", "text_model.encoder.layers."),
194
+ ("ln_1", "layer_norm1"),
195
+ ("ln_2", "layer_norm2"),
196
+ (".c_fc.", ".fc1."),
197
+ (".c_proj.", ".fc2."),
198
+ (".attn", ".self_attn"),
199
+ ("ln_final.", "transformer.text_model.final_layer_norm."),
200
+ ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
201
+ ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
202
+ ]
203
+ protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
204
+ textenc_pattern = re.compile("|".join(protected.keys()))
205
+
206
+ # Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
207
+ code2idx = {"q": 0, "k": 1, "v": 2}
208
+
209
+ # This function exists because at the time of writing torch.cat can't do fp8 with cuda
210
+ def cat_tensors(tensors):
211
+ x = 0
212
+ for t in tensors:
213
+ x += t.shape[0]
214
+
215
+ shape = [x] + list(tensors[0].shape)[1:]
216
+ out = torch.empty(shape, device=tensors[0].device, dtype=tensors[0].dtype)
217
+
218
+ x = 0
219
+ for t in tensors:
220
+ out[x:x + t.shape[0]] = t
221
+ x += t.shape[0]
222
+
223
+ return out
224
+
225
+ def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
226
+ new_state_dict = {}
227
+ capture_qkv_weight = {}
228
+ capture_qkv_bias = {}
229
+ for k, v in text_enc_dict.items():
230
+ if not k.startswith(prefix):
231
+ continue
232
+ if (
233
+ k.endswith(".self_attn.q_proj.weight")
234
+ or k.endswith(".self_attn.k_proj.weight")
235
+ or k.endswith(".self_attn.v_proj.weight")
236
+ ):
237
+ k_pre = k[: -len(".q_proj.weight")]
238
+ k_code = k[-len("q_proj.weight")]
239
+ if k_pre not in capture_qkv_weight:
240
+ capture_qkv_weight[k_pre] = [None, None, None]
241
+ capture_qkv_weight[k_pre][code2idx[k_code]] = v
242
+ continue
243
+
244
+ if (
245
+ k.endswith(".self_attn.q_proj.bias")
246
+ or k.endswith(".self_attn.k_proj.bias")
247
+ or k.endswith(".self_attn.v_proj.bias")
248
+ ):
249
+ k_pre = k[: -len(".q_proj.bias")]
250
+ k_code = k[-len("q_proj.bias")]
251
+ if k_pre not in capture_qkv_bias:
252
+ capture_qkv_bias[k_pre] = [None, None, None]
253
+ capture_qkv_bias[k_pre][code2idx[k_code]] = v
254
+ continue
255
+
256
+ text_proj = "transformer.text_projection.weight"
257
+ if k.endswith(text_proj):
258
+ new_state_dict[k.replace(text_proj, "text_projection")] = v.transpose(0, 1).contiguous()
259
+ else:
260
+ relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
261
+ new_state_dict[relabelled_key] = v
262
+
263
+ for k_pre, tensors in capture_qkv_weight.items():
264
+ if None in tensors:
265
+ raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
266
+ relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
267
+ new_state_dict[relabelled_key + ".in_proj_weight"] = cat_tensors(tensors)
268
+
269
+ for k_pre, tensors in capture_qkv_bias.items():
270
+ if None in tensors:
271
+ raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
272
+ relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
273
+ new_state_dict[relabelled_key + ".in_proj_bias"] = cat_tensors(tensors)
274
+
275
+ return new_state_dict
276
+
277
+
278
+ def convert_text_enc_state_dict(text_enc_dict):
279
+ return text_enc_dict
280
+
281
+
ComfyUI/comfy/diffusers_load.py ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import comfy.sd
4
+
5
+ def first_file(path, filenames):
6
+ for f in filenames:
7
+ p = os.path.join(path, f)
8
+ if os.path.exists(p):
9
+ return p
10
+ return None
11
+
12
+ def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_directory=None):
13
+ diffusion_model_names = ["diffusion_pytorch_model.fp16.safetensors", "diffusion_pytorch_model.safetensors", "diffusion_pytorch_model.fp16.bin", "diffusion_pytorch_model.bin"]
14
+ unet_path = first_file(os.path.join(model_path, "unet"), diffusion_model_names)
15
+ vae_path = first_file(os.path.join(model_path, "vae"), diffusion_model_names)
16
+
17
+ text_encoder_model_names = ["model.fp16.safetensors", "model.safetensors", "pytorch_model.fp16.bin", "pytorch_model.bin"]
18
+ text_encoder1_path = first_file(os.path.join(model_path, "text_encoder"), text_encoder_model_names)
19
+ text_encoder2_path = first_file(os.path.join(model_path, "text_encoder_2"), text_encoder_model_names)
20
+
21
+ text_encoder_paths = [text_encoder1_path]
22
+ if text_encoder2_path is not None:
23
+ text_encoder_paths.append(text_encoder2_path)
24
+
25
+ unet = comfy.sd.load_unet(unet_path)
26
+
27
+ clip = None
28
+ if output_clip:
29
+ clip = comfy.sd.load_clip(text_encoder_paths, embedding_directory=embedding_directory)
30
+
31
+ vae = None
32
+ if output_vae:
33
+ sd = comfy.utils.load_torch_file(vae_path)
34
+ vae = comfy.sd.VAE(sd=sd)
35
+
36
+ return (unet, clip, vae)
ComfyUI/comfy/extra_samplers/uni_pc.py ADDED
@@ -0,0 +1,875 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #code taken from: https://github.com/wl-zhao/UniPC and modified
2
+
3
+ import torch
4
+ import torch.nn.functional as F
5
+ import math
6
+
7
+ from tqdm.auto import trange, tqdm
8
+
9
+
10
+ class NoiseScheduleVP:
11
+ def __init__(
12
+ self,
13
+ schedule='discrete',
14
+ betas=None,
15
+ alphas_cumprod=None,
16
+ continuous_beta_0=0.1,
17
+ continuous_beta_1=20.,
18
+ ):
19
+ """Create a wrapper class for the forward SDE (VP type).
20
+
21
+ ***
22
+ Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
23
+ We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
24
+ ***
25
+
26
+ The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
27
+ We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
28
+ Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
29
+
30
+ log_alpha_t = self.marginal_log_mean_coeff(t)
31
+ sigma_t = self.marginal_std(t)
32
+ lambda_t = self.marginal_lambda(t)
33
+
34
+ Moreover, as lambda(t) is an invertible function, we also support its inverse function:
35
+
36
+ t = self.inverse_lambda(lambda_t)
37
+
38
+ ===============================================================
39
+
40
+ We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
41
+
42
+ 1. For discrete-time DPMs:
43
+
44
+ For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
45
+ t_i = (i + 1) / N
46
+ e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
47
+ We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
48
+
49
+ Args:
50
+ betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
51
+ alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
52
+
53
+ Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
54
+
55
+ **Important**: Please pay special attention for the args for `alphas_cumprod`:
56
+ The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
57
+ q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
58
+ Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
59
+ alpha_{t_n} = \sqrt{\hat{alpha_n}},
60
+ and
61
+ log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
62
+
63
+
64
+ 2. For continuous-time DPMs:
65
+
66
+ We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
67
+ schedule are the default settings in DDPM and improved-DDPM:
68
+
69
+ Args:
70
+ beta_min: A `float` number. The smallest beta for the linear schedule.
71
+ beta_max: A `float` number. The largest beta for the linear schedule.
72
+ cosine_s: A `float` number. The hyperparameter in the cosine schedule.
73
+ cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
74
+ T: A `float` number. The ending time of the forward process.
75
+
76
+ ===============================================================
77
+
78
+ Args:
79
+ schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
80
+ 'linear' or 'cosine' for continuous-time DPMs.
81
+ Returns:
82
+ A wrapper object of the forward SDE (VP type).
83
+
84
+ ===============================================================
85
+
86
+ Example:
87
+
88
+ # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
89
+ >>> ns = NoiseScheduleVP('discrete', betas=betas)
90
+
91
+ # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
92
+ >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
93
+
94
+ # For continuous-time DPMs (VPSDE), linear schedule:
95
+ >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
96
+
97
+ """
98
+
99
+ if schedule not in ['discrete', 'linear', 'cosine']:
100
+ raise ValueError("Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(schedule))
101
+
102
+ self.schedule = schedule
103
+ if schedule == 'discrete':
104
+ if betas is not None:
105
+ log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
106
+ else:
107
+ assert alphas_cumprod is not None
108
+ log_alphas = 0.5 * torch.log(alphas_cumprod)
109
+ self.total_N = len(log_alphas)
110
+ self.T = 1.
111
+ self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
112
+ self.log_alpha_array = log_alphas.reshape((1, -1,))
113
+ else:
114
+ self.total_N = 1000
115
+ self.beta_0 = continuous_beta_0
116
+ self.beta_1 = continuous_beta_1
117
+ self.cosine_s = 0.008
118
+ self.cosine_beta_max = 999.
119
+ self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
120
+ self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
121
+ self.schedule = schedule
122
+ if schedule == 'cosine':
123
+ # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
124
+ # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
125
+ self.T = 0.9946
126
+ else:
127
+ self.T = 1.
128
+
129
+ def marginal_log_mean_coeff(self, t):
130
+ """
131
+ Compute log(alpha_t) of a given continuous-time label t in [0, T].
132
+ """
133
+ if self.schedule == 'discrete':
134
+ return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), self.log_alpha_array.to(t.device)).reshape((-1))
135
+ elif self.schedule == 'linear':
136
+ return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
137
+ elif self.schedule == 'cosine':
138
+ log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
139
+ log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
140
+ return log_alpha_t
141
+
142
+ def marginal_alpha(self, t):
143
+ """
144
+ Compute alpha_t of a given continuous-time label t in [0, T].
145
+ """
146
+ return torch.exp(self.marginal_log_mean_coeff(t))
147
+
148
+ def marginal_std(self, t):
149
+ """
150
+ Compute sigma_t of a given continuous-time label t in [0, T].
151
+ """
152
+ return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
153
+
154
+ def marginal_lambda(self, t):
155
+ """
156
+ Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
157
+ """
158
+ log_mean_coeff = self.marginal_log_mean_coeff(t)
159
+ log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
160
+ return log_mean_coeff - log_std
161
+
162
+ def inverse_lambda(self, lamb):
163
+ """
164
+ Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
165
+ """
166
+ if self.schedule == 'linear':
167
+ tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
168
+ Delta = self.beta_0**2 + tmp
169
+ return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
170
+ elif self.schedule == 'discrete':
171
+ log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
172
+ t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), torch.flip(self.t_array.to(lamb.device), [1]))
173
+ return t.reshape((-1,))
174
+ else:
175
+ log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
176
+ t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
177
+ t = t_fn(log_alpha)
178
+ return t
179
+
180
+
181
+ def model_wrapper(
182
+ model,
183
+ noise_schedule,
184
+ model_type="noise",
185
+ model_kwargs={},
186
+ guidance_type="uncond",
187
+ condition=None,
188
+ unconditional_condition=None,
189
+ guidance_scale=1.,
190
+ classifier_fn=None,
191
+ classifier_kwargs={},
192
+ ):
193
+ """Create a wrapper function for the noise prediction model.
194
+
195
+ DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
196
+ firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
197
+
198
+ We support four types of the diffusion model by setting `model_type`:
199
+
200
+ 1. "noise": noise prediction model. (Trained by predicting noise).
201
+
202
+ 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
203
+
204
+ 3. "v": velocity prediction model. (Trained by predicting the velocity).
205
+ The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
206
+
207
+ [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
208
+ arXiv preprint arXiv:2202.00512 (2022).
209
+ [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
210
+ arXiv preprint arXiv:2210.02303 (2022).
211
+
212
+ 4. "score": marginal score function. (Trained by denoising score matching).
213
+ Note that the score function and the noise prediction model follows a simple relationship:
214
+ ```
215
+ noise(x_t, t) = -sigma_t * score(x_t, t)
216
+ ```
217
+
218
+ We support three types of guided sampling by DPMs by setting `guidance_type`:
219
+ 1. "uncond": unconditional sampling by DPMs.
220
+ The input `model` has the following format:
221
+ ``
222
+ model(x, t_input, **model_kwargs) -> noise | x_start | v | score
223
+ ``
224
+
225
+ 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
226
+ The input `model` has the following format:
227
+ ``
228
+ model(x, t_input, **model_kwargs) -> noise | x_start | v | score
229
+ ``
230
+
231
+ The input `classifier_fn` has the following format:
232
+ ``
233
+ classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
234
+ ``
235
+
236
+ [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
237
+ in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
238
+
239
+ 3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
240
+ The input `model` has the following format:
241
+ ``
242
+ model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
243
+ ``
244
+ And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
245
+
246
+ [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
247
+ arXiv preprint arXiv:2207.12598 (2022).
248
+
249
+
250
+ The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
251
+ or continuous-time labels (i.e. epsilon to T).
252
+
253
+ We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
254
+ ``
255
+ def model_fn(x, t_continuous) -> noise:
256
+ t_input = get_model_input_time(t_continuous)
257
+ return noise_pred(model, x, t_input, **model_kwargs)
258
+ ``
259
+ where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
260
+
261
+ ===============================================================
262
+
263
+ Args:
264
+ model: A diffusion model with the corresponding format described above.
265
+ noise_schedule: A noise schedule object, such as NoiseScheduleVP.
266
+ model_type: A `str`. The parameterization type of the diffusion model.
267
+ "noise" or "x_start" or "v" or "score".
268
+ model_kwargs: A `dict`. A dict for the other inputs of the model function.
269
+ guidance_type: A `str`. The type of the guidance for sampling.
270
+ "uncond" or "classifier" or "classifier-free".
271
+ condition: A pytorch tensor. The condition for the guided sampling.
272
+ Only used for "classifier" or "classifier-free" guidance type.
273
+ unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
274
+ Only used for "classifier-free" guidance type.
275
+ guidance_scale: A `float`. The scale for the guided sampling.
276
+ classifier_fn: A classifier function. Only used for the classifier guidance.
277
+ classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
278
+ Returns:
279
+ A noise prediction model that accepts the noised data and the continuous time as the inputs.
280
+ """
281
+
282
+ def get_model_input_time(t_continuous):
283
+ """
284
+ Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
285
+ For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
286
+ For continuous-time DPMs, we just use `t_continuous`.
287
+ """
288
+ if noise_schedule.schedule == 'discrete':
289
+ return (t_continuous - 1. / noise_schedule.total_N) * 1000.
290
+ else:
291
+ return t_continuous
292
+
293
+ def noise_pred_fn(x, t_continuous, cond=None):
294
+ if t_continuous.reshape((-1,)).shape[0] == 1:
295
+ t_continuous = t_continuous.expand((x.shape[0]))
296
+ t_input = get_model_input_time(t_continuous)
297
+ output = model(x, t_input, **model_kwargs)
298
+ if model_type == "noise":
299
+ return output
300
+ elif model_type == "x_start":
301
+ alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
302
+ dims = x.dim()
303
+ return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
304
+ elif model_type == "v":
305
+ alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
306
+ dims = x.dim()
307
+ return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
308
+ elif model_type == "score":
309
+ sigma_t = noise_schedule.marginal_std(t_continuous)
310
+ dims = x.dim()
311
+ return -expand_dims(sigma_t, dims) * output
312
+
313
+ def cond_grad_fn(x, t_input):
314
+ """
315
+ Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
316
+ """
317
+ with torch.enable_grad():
318
+ x_in = x.detach().requires_grad_(True)
319
+ log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
320
+ return torch.autograd.grad(log_prob.sum(), x_in)[0]
321
+
322
+ def model_fn(x, t_continuous):
323
+ """
324
+ The noise predicition model function that is used for DPM-Solver.
325
+ """
326
+ if t_continuous.reshape((-1,)).shape[0] == 1:
327
+ t_continuous = t_continuous.expand((x.shape[0]))
328
+ if guidance_type == "uncond":
329
+ return noise_pred_fn(x, t_continuous)
330
+ elif guidance_type == "classifier":
331
+ assert classifier_fn is not None
332
+ t_input = get_model_input_time(t_continuous)
333
+ cond_grad = cond_grad_fn(x, t_input)
334
+ sigma_t = noise_schedule.marginal_std(t_continuous)
335
+ noise = noise_pred_fn(x, t_continuous)
336
+ return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
337
+ elif guidance_type == "classifier-free":
338
+ if guidance_scale == 1. or unconditional_condition is None:
339
+ return noise_pred_fn(x, t_continuous, cond=condition)
340
+ else:
341
+ x_in = torch.cat([x] * 2)
342
+ t_in = torch.cat([t_continuous] * 2)
343
+ c_in = torch.cat([unconditional_condition, condition])
344
+ noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
345
+ return noise_uncond + guidance_scale * (noise - noise_uncond)
346
+
347
+ assert model_type in ["noise", "x_start", "v"]
348
+ assert guidance_type in ["uncond", "classifier", "classifier-free"]
349
+ return model_fn
350
+
351
+
352
+ class UniPC:
353
+ def __init__(
354
+ self,
355
+ model_fn,
356
+ noise_schedule,
357
+ predict_x0=True,
358
+ thresholding=False,
359
+ max_val=1.,
360
+ variant='bh1',
361
+ ):
362
+ """Construct a UniPC.
363
+
364
+ We support both data_prediction and noise_prediction.
365
+ """
366
+ self.model = model_fn
367
+ self.noise_schedule = noise_schedule
368
+ self.variant = variant
369
+ self.predict_x0 = predict_x0
370
+ self.thresholding = thresholding
371
+ self.max_val = max_val
372
+
373
+ def dynamic_thresholding_fn(self, x0, t=None):
374
+ """
375
+ The dynamic thresholding method.
376
+ """
377
+ dims = x0.dim()
378
+ p = self.dynamic_thresholding_ratio
379
+ s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
380
+ s = expand_dims(torch.maximum(s, self.thresholding_max_val * torch.ones_like(s).to(s.device)), dims)
381
+ x0 = torch.clamp(x0, -s, s) / s
382
+ return x0
383
+
384
+ def noise_prediction_fn(self, x, t):
385
+ """
386
+ Return the noise prediction model.
387
+ """
388
+ return self.model(x, t)
389
+
390
+ def data_prediction_fn(self, x, t):
391
+ """
392
+ Return the data prediction model (with thresholding).
393
+ """
394
+ noise = self.noise_prediction_fn(x, t)
395
+ dims = x.dim()
396
+ alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
397
+ x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
398
+ if self.thresholding:
399
+ p = 0.995 # A hyperparameter in the paper of "Imagen" [1].
400
+ s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
401
+ s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
402
+ x0 = torch.clamp(x0, -s, s) / s
403
+ return x0
404
+
405
+ def model_fn(self, x, t):
406
+ """
407
+ Convert the model to the noise prediction model or the data prediction model.
408
+ """
409
+ if self.predict_x0:
410
+ return self.data_prediction_fn(x, t)
411
+ else:
412
+ return self.noise_prediction_fn(x, t)
413
+
414
+ def get_time_steps(self, skip_type, t_T, t_0, N, device):
415
+ """Compute the intermediate time steps for sampling.
416
+ """
417
+ if skip_type == 'logSNR':
418
+ lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
419
+ lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
420
+ logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
421
+ return self.noise_schedule.inverse_lambda(logSNR_steps)
422
+ elif skip_type == 'time_uniform':
423
+ return torch.linspace(t_T, t_0, N + 1).to(device)
424
+ elif skip_type == 'time_quadratic':
425
+ t_order = 2
426
+ t = torch.linspace(t_T**(1. / t_order), t_0**(1. / t_order), N + 1).pow(t_order).to(device)
427
+ return t
428
+ else:
429
+ raise ValueError("Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
430
+
431
+ def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
432
+ """
433
+ Get the order of each step for sampling by the singlestep DPM-Solver.
434
+ """
435
+ if order == 3:
436
+ K = steps // 3 + 1
437
+ if steps % 3 == 0:
438
+ orders = [3,] * (K - 2) + [2, 1]
439
+ elif steps % 3 == 1:
440
+ orders = [3,] * (K - 1) + [1]
441
+ else:
442
+ orders = [3,] * (K - 1) + [2]
443
+ elif order == 2:
444
+ if steps % 2 == 0:
445
+ K = steps // 2
446
+ orders = [2,] * K
447
+ else:
448
+ K = steps // 2 + 1
449
+ orders = [2,] * (K - 1) + [1]
450
+ elif order == 1:
451
+ K = steps
452
+ orders = [1,] * steps
453
+ else:
454
+ raise ValueError("'order' must be '1' or '2' or '3'.")
455
+ if skip_type == 'logSNR':
456
+ # To reproduce the results in DPM-Solver paper
457
+ timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
458
+ else:
459
+ timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[torch.cumsum(torch.tensor([0,] + orders), 0).to(device)]
460
+ return timesteps_outer, orders
461
+
462
+ def denoise_to_zero_fn(self, x, s):
463
+ """
464
+ Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
465
+ """
466
+ return self.data_prediction_fn(x, s)
467
+
468
+ def multistep_uni_pc_update(self, x, model_prev_list, t_prev_list, t, order, **kwargs):
469
+ if len(t.shape) == 0:
470
+ t = t.view(-1)
471
+ if 'bh' in self.variant:
472
+ return self.multistep_uni_pc_bh_update(x, model_prev_list, t_prev_list, t, order, **kwargs)
473
+ else:
474
+ assert self.variant == 'vary_coeff'
475
+ return self.multistep_uni_pc_vary_update(x, model_prev_list, t_prev_list, t, order, **kwargs)
476
+
477
+ def multistep_uni_pc_vary_update(self, x, model_prev_list, t_prev_list, t, order, use_corrector=True):
478
+ print(f'using unified predictor-corrector with order {order} (solver type: vary coeff)')
479
+ ns = self.noise_schedule
480
+ assert order <= len(model_prev_list)
481
+
482
+ # first compute rks
483
+ t_prev_0 = t_prev_list[-1]
484
+ lambda_prev_0 = ns.marginal_lambda(t_prev_0)
485
+ lambda_t = ns.marginal_lambda(t)
486
+ model_prev_0 = model_prev_list[-1]
487
+ sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
488
+ log_alpha_t = ns.marginal_log_mean_coeff(t)
489
+ alpha_t = torch.exp(log_alpha_t)
490
+
491
+ h = lambda_t - lambda_prev_0
492
+
493
+ rks = []
494
+ D1s = []
495
+ for i in range(1, order):
496
+ t_prev_i = t_prev_list[-(i + 1)]
497
+ model_prev_i = model_prev_list[-(i + 1)]
498
+ lambda_prev_i = ns.marginal_lambda(t_prev_i)
499
+ rk = (lambda_prev_i - lambda_prev_0) / h
500
+ rks.append(rk)
501
+ D1s.append((model_prev_i - model_prev_0) / rk)
502
+
503
+ rks.append(1.)
504
+ rks = torch.tensor(rks, device=x.device)
505
+
506
+ K = len(rks)
507
+ # build C matrix
508
+ C = []
509
+
510
+ col = torch.ones_like(rks)
511
+ for k in range(1, K + 1):
512
+ C.append(col)
513
+ col = col * rks / (k + 1)
514
+ C = torch.stack(C, dim=1)
515
+
516
+ if len(D1s) > 0:
517
+ D1s = torch.stack(D1s, dim=1) # (B, K)
518
+ C_inv_p = torch.linalg.inv(C[:-1, :-1])
519
+ A_p = C_inv_p
520
+
521
+ if use_corrector:
522
+ print('using corrector')
523
+ C_inv = torch.linalg.inv(C)
524
+ A_c = C_inv
525
+
526
+ hh = -h if self.predict_x0 else h
527
+ h_phi_1 = torch.expm1(hh)
528
+ h_phi_ks = []
529
+ factorial_k = 1
530
+ h_phi_k = h_phi_1
531
+ for k in range(1, K + 2):
532
+ h_phi_ks.append(h_phi_k)
533
+ h_phi_k = h_phi_k / hh - 1 / factorial_k
534
+ factorial_k *= (k + 1)
535
+
536
+ model_t = None
537
+ if self.predict_x0:
538
+ x_t_ = (
539
+ sigma_t / sigma_prev_0 * x
540
+ - alpha_t * h_phi_1 * model_prev_0
541
+ )
542
+ # now predictor
543
+ x_t = x_t_
544
+ if len(D1s) > 0:
545
+ # compute the residuals for predictor
546
+ for k in range(K - 1):
547
+ x_t = x_t - alpha_t * h_phi_ks[k + 1] * torch.einsum('bkchw,k->bchw', D1s, A_p[k])
548
+ # now corrector
549
+ if use_corrector:
550
+ model_t = self.model_fn(x_t, t)
551
+ D1_t = (model_t - model_prev_0)
552
+ x_t = x_t_
553
+ k = 0
554
+ for k in range(K - 1):
555
+ x_t = x_t - alpha_t * h_phi_ks[k + 1] * torch.einsum('bkchw,k->bchw', D1s, A_c[k][:-1])
556
+ x_t = x_t - alpha_t * h_phi_ks[K] * (D1_t * A_c[k][-1])
557
+ else:
558
+ log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
559
+ x_t_ = (
560
+ (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
561
+ - (sigma_t * h_phi_1) * model_prev_0
562
+ )
563
+ # now predictor
564
+ x_t = x_t_
565
+ if len(D1s) > 0:
566
+ # compute the residuals for predictor
567
+ for k in range(K - 1):
568
+ x_t = x_t - sigma_t * h_phi_ks[k + 1] * torch.einsum('bkchw,k->bchw', D1s, A_p[k])
569
+ # now corrector
570
+ if use_corrector:
571
+ model_t = self.model_fn(x_t, t)
572
+ D1_t = (model_t - model_prev_0)
573
+ x_t = x_t_
574
+ k = 0
575
+ for k in range(K - 1):
576
+ x_t = x_t - sigma_t * h_phi_ks[k + 1] * torch.einsum('bkchw,k->bchw', D1s, A_c[k][:-1])
577
+ x_t = x_t - sigma_t * h_phi_ks[K] * (D1_t * A_c[k][-1])
578
+ return x_t, model_t
579
+
580
+ def multistep_uni_pc_bh_update(self, x, model_prev_list, t_prev_list, t, order, x_t=None, use_corrector=True):
581
+ # print(f'using unified predictor-corrector with order {order} (solver type: B(h))')
582
+ ns = self.noise_schedule
583
+ assert order <= len(model_prev_list)
584
+ dims = x.dim()
585
+
586
+ # first compute rks
587
+ t_prev_0 = t_prev_list[-1]
588
+ lambda_prev_0 = ns.marginal_lambda(t_prev_0)
589
+ lambda_t = ns.marginal_lambda(t)
590
+ model_prev_0 = model_prev_list[-1]
591
+ sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
592
+ log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
593
+ alpha_t = torch.exp(log_alpha_t)
594
+
595
+ h = lambda_t - lambda_prev_0
596
+
597
+ rks = []
598
+ D1s = []
599
+ for i in range(1, order):
600
+ t_prev_i = t_prev_list[-(i + 1)]
601
+ model_prev_i = model_prev_list[-(i + 1)]
602
+ lambda_prev_i = ns.marginal_lambda(t_prev_i)
603
+ rk = ((lambda_prev_i - lambda_prev_0) / h)[0]
604
+ rks.append(rk)
605
+ D1s.append((model_prev_i - model_prev_0) / rk)
606
+
607
+ rks.append(1.)
608
+ rks = torch.tensor(rks, device=x.device)
609
+
610
+ R = []
611
+ b = []
612
+
613
+ hh = -h[0] if self.predict_x0 else h[0]
614
+ h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1
615
+ h_phi_k = h_phi_1 / hh - 1
616
+
617
+ factorial_i = 1
618
+
619
+ if self.variant == 'bh1':
620
+ B_h = hh
621
+ elif self.variant == 'bh2':
622
+ B_h = torch.expm1(hh)
623
+ else:
624
+ raise NotImplementedError()
625
+
626
+ for i in range(1, order + 1):
627
+ R.append(torch.pow(rks, i - 1))
628
+ b.append(h_phi_k * factorial_i / B_h)
629
+ factorial_i *= (i + 1)
630
+ h_phi_k = h_phi_k / hh - 1 / factorial_i
631
+
632
+ R = torch.stack(R)
633
+ b = torch.tensor(b, device=x.device)
634
+
635
+ # now predictor
636
+ use_predictor = len(D1s) > 0 and x_t is None
637
+ if len(D1s) > 0:
638
+ D1s = torch.stack(D1s, dim=1) # (B, K)
639
+ if x_t is None:
640
+ # for order 2, we use a simplified version
641
+ if order == 2:
642
+ rhos_p = torch.tensor([0.5], device=b.device)
643
+ else:
644
+ rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1])
645
+ else:
646
+ D1s = None
647
+
648
+ if use_corrector:
649
+ # print('using corrector')
650
+ # for order 1, we use a simplified version
651
+ if order == 1:
652
+ rhos_c = torch.tensor([0.5], device=b.device)
653
+ else:
654
+ rhos_c = torch.linalg.solve(R, b)
655
+
656
+ model_t = None
657
+ if self.predict_x0:
658
+ x_t_ = (
659
+ expand_dims(sigma_t / sigma_prev_0, dims) * x
660
+ - expand_dims(alpha_t * h_phi_1, dims)* model_prev_0
661
+ )
662
+
663
+ if x_t is None:
664
+ if use_predictor:
665
+ pred_res = torch.einsum('k,bkchw->bchw', rhos_p, D1s)
666
+ else:
667
+ pred_res = 0
668
+ x_t = x_t_ - expand_dims(alpha_t * B_h, dims) * pred_res
669
+
670
+ if use_corrector:
671
+ model_t = self.model_fn(x_t, t)
672
+ if D1s is not None:
673
+ corr_res = torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s)
674
+ else:
675
+ corr_res = 0
676
+ D1_t = (model_t - model_prev_0)
677
+ x_t = x_t_ - expand_dims(alpha_t * B_h, dims) * (corr_res + rhos_c[-1] * D1_t)
678
+ else:
679
+ x_t_ = (
680
+ expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
681
+ - expand_dims(sigma_t * h_phi_1, dims) * model_prev_0
682
+ )
683
+ if x_t is None:
684
+ if use_predictor:
685
+ pred_res = torch.einsum('k,bkchw->bchw', rhos_p, D1s)
686
+ else:
687
+ pred_res = 0
688
+ x_t = x_t_ - expand_dims(sigma_t * B_h, dims) * pred_res
689
+
690
+ if use_corrector:
691
+ model_t = self.model_fn(x_t, t)
692
+ if D1s is not None:
693
+ corr_res = torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s)
694
+ else:
695
+ corr_res = 0
696
+ D1_t = (model_t - model_prev_0)
697
+ x_t = x_t_ - expand_dims(sigma_t * B_h, dims) * (corr_res + rhos_c[-1] * D1_t)
698
+ return x_t, model_t
699
+
700
+
701
+ def sample(self, x, timesteps, t_start=None, t_end=None, order=3, skip_type='time_uniform',
702
+ method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver',
703
+ atol=0.0078, rtol=0.05, corrector=False, callback=None, disable_pbar=False
704
+ ):
705
+ # t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
706
+ # t_T = self.noise_schedule.T if t_start is None else t_start
707
+ device = x.device
708
+ steps = len(timesteps) - 1
709
+ if method == 'multistep':
710
+ assert steps >= order
711
+ # timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
712
+ assert timesteps.shape[0] - 1 == steps
713
+ # with torch.no_grad():
714
+ for step_index in trange(steps, disable=disable_pbar):
715
+ if step_index == 0:
716
+ vec_t = timesteps[0].expand((x.shape[0]))
717
+ model_prev_list = [self.model_fn(x, vec_t)]
718
+ t_prev_list = [vec_t]
719
+ elif step_index < order:
720
+ init_order = step_index
721
+ # Init the first `order` values by lower order multistep DPM-Solver.
722
+ # for init_order in range(1, order):
723
+ vec_t = timesteps[init_order].expand(x.shape[0])
724
+ x, model_x = self.multistep_uni_pc_update(x, model_prev_list, t_prev_list, vec_t, init_order, use_corrector=True)
725
+ if model_x is None:
726
+ model_x = self.model_fn(x, vec_t)
727
+ model_prev_list.append(model_x)
728
+ t_prev_list.append(vec_t)
729
+ else:
730
+ extra_final_step = 0
731
+ if step_index == (steps - 1):
732
+ extra_final_step = 1
733
+ for step in range(step_index, step_index + 1 + extra_final_step):
734
+ vec_t = timesteps[step].expand(x.shape[0])
735
+ if lower_order_final:
736
+ step_order = min(order, steps + 1 - step)
737
+ else:
738
+ step_order = order
739
+ # print('this step order:', step_order)
740
+ if step == steps:
741
+ # print('do not run corrector at the last step')
742
+ use_corrector = False
743
+ else:
744
+ use_corrector = True
745
+ x, model_x = self.multistep_uni_pc_update(x, model_prev_list, t_prev_list, vec_t, step_order, use_corrector=use_corrector)
746
+ for i in range(order - 1):
747
+ t_prev_list[i] = t_prev_list[i + 1]
748
+ model_prev_list[i] = model_prev_list[i + 1]
749
+ t_prev_list[-1] = vec_t
750
+ # We do not need to evaluate the final model value.
751
+ if step < steps:
752
+ if model_x is None:
753
+ model_x = self.model_fn(x, vec_t)
754
+ model_prev_list[-1] = model_x
755
+ if callback is not None:
756
+ callback({'x': x, 'i': step_index, 'denoised': model_prev_list[-1]})
757
+ else:
758
+ raise NotImplementedError()
759
+ # if denoise_to_zero:
760
+ # x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
761
+ return x
762
+
763
+
764
+ #############################################################
765
+ # other utility functions
766
+ #############################################################
767
+
768
+ def interpolate_fn(x, xp, yp):
769
+ """
770
+ A piecewise linear function y = f(x), using xp and yp as keypoints.
771
+ We implement f(x) in a differentiable way (i.e. applicable for autograd).
772
+ The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
773
+
774
+ Args:
775
+ x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
776
+ xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
777
+ yp: PyTorch tensor with shape [C, K].
778
+ Returns:
779
+ The function values f(x), with shape [N, C].
780
+ """
781
+ N, K = x.shape[0], xp.shape[1]
782
+ all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
783
+ sorted_all_x, x_indices = torch.sort(all_x, dim=2)
784
+ x_idx = torch.argmin(x_indices, dim=2)
785
+ cand_start_idx = x_idx - 1
786
+ start_idx = torch.where(
787
+ torch.eq(x_idx, 0),
788
+ torch.tensor(1, device=x.device),
789
+ torch.where(
790
+ torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
791
+ ),
792
+ )
793
+ end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
794
+ start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
795
+ end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
796
+ start_idx2 = torch.where(
797
+ torch.eq(x_idx, 0),
798
+ torch.tensor(0, device=x.device),
799
+ torch.where(
800
+ torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
801
+ ),
802
+ )
803
+ y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
804
+ start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
805
+ end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
806
+ cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
807
+ return cand
808
+
809
+
810
+ def expand_dims(v, dims):
811
+ """
812
+ Expand the tensor `v` to the dim `dims`.
813
+
814
+ Args:
815
+ `v`: a PyTorch tensor with shape [N].
816
+ `dim`: a `int`.
817
+ Returns:
818
+ a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
819
+ """
820
+ return v[(...,) + (None,)*(dims - 1)]
821
+
822
+
823
+ class SigmaConvert:
824
+ schedule = ""
825
+ def marginal_log_mean_coeff(self, sigma):
826
+ return 0.5 * torch.log(1 / ((sigma * sigma) + 1))
827
+
828
+ def marginal_alpha(self, t):
829
+ return torch.exp(self.marginal_log_mean_coeff(t))
830
+
831
+ def marginal_std(self, t):
832
+ return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
833
+
834
+ def marginal_lambda(self, t):
835
+ """
836
+ Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
837
+ """
838
+ log_mean_coeff = self.marginal_log_mean_coeff(t)
839
+ log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
840
+ return log_mean_coeff - log_std
841
+
842
+ def predict_eps_sigma(model, input, sigma_in, **kwargs):
843
+ sigma = sigma_in.view(sigma_in.shape[:1] + (1,) * (input.ndim - 1))
844
+ input = input * ((sigma ** 2 + 1.0) ** 0.5)
845
+ return (input - model(input, sigma_in, **kwargs)) / sigma
846
+
847
+
848
+ def sample_unipc(model, noise, sigmas, extra_args=None, callback=None, disable=False, variant='bh1'):
849
+ timesteps = sigmas.clone()
850
+ if sigmas[-1] == 0:
851
+ timesteps = sigmas[:]
852
+ timesteps[-1] = 0.001
853
+ else:
854
+ timesteps = sigmas.clone()
855
+ ns = SigmaConvert()
856
+
857
+ noise = noise / torch.sqrt(1.0 + timesteps[0] ** 2.0)
858
+ model_type = "noise"
859
+
860
+ model_fn = model_wrapper(
861
+ lambda input, sigma, **kwargs: predict_eps_sigma(model, input, sigma, **kwargs),
862
+ ns,
863
+ model_type=model_type,
864
+ guidance_type="uncond",
865
+ model_kwargs=extra_args,
866
+ )
867
+
868
+ order = min(3, len(timesteps) - 2)
869
+ uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, variant=variant)
870
+ x = uni_pc.sample(noise, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=order, lower_order_final=True, callback=callback, disable_pbar=disable)
871
+ x /= ns.marginal_alpha(timesteps[-1])
872
+ return x
873
+
874
+ def sample_unipc_bh2(model, noise, sigmas, extra_args=None, callback=None, disable=False):
875
+ return sample_unipc(model, noise, sigmas, extra_args, callback, disable, variant='bh2')
ComfyUI/comfy/gligen.py ADDED
@@ -0,0 +1,343 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+ from .ldm.modules.attention import CrossAttention
4
+ from inspect import isfunction
5
+ import comfy.ops
6
+ ops = comfy.ops.manual_cast
7
+
8
+ def exists(val):
9
+ return val is not None
10
+
11
+
12
+ def uniq(arr):
13
+ return{el: True for el in arr}.keys()
14
+
15
+
16
+ def default(val, d):
17
+ if exists(val):
18
+ return val
19
+ return d() if isfunction(d) else d
20
+
21
+
22
+ # feedforward
23
+ class GEGLU(nn.Module):
24
+ def __init__(self, dim_in, dim_out):
25
+ super().__init__()
26
+ self.proj = ops.Linear(dim_in, dim_out * 2)
27
+
28
+ def forward(self, x):
29
+ x, gate = self.proj(x).chunk(2, dim=-1)
30
+ return x * torch.nn.functional.gelu(gate)
31
+
32
+
33
+ class FeedForward(nn.Module):
34
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
35
+ super().__init__()
36
+ inner_dim = int(dim * mult)
37
+ dim_out = default(dim_out, dim)
38
+ project_in = nn.Sequential(
39
+ ops.Linear(dim, inner_dim),
40
+ nn.GELU()
41
+ ) if not glu else GEGLU(dim, inner_dim)
42
+
43
+ self.net = nn.Sequential(
44
+ project_in,
45
+ nn.Dropout(dropout),
46
+ ops.Linear(inner_dim, dim_out)
47
+ )
48
+
49
+ def forward(self, x):
50
+ return self.net(x)
51
+
52
+
53
+ class GatedCrossAttentionDense(nn.Module):
54
+ def __init__(self, query_dim, context_dim, n_heads, d_head):
55
+ super().__init__()
56
+
57
+ self.attn = CrossAttention(
58
+ query_dim=query_dim,
59
+ context_dim=context_dim,
60
+ heads=n_heads,
61
+ dim_head=d_head,
62
+ operations=ops)
63
+ self.ff = FeedForward(query_dim, glu=True)
64
+
65
+ self.norm1 = ops.LayerNorm(query_dim)
66
+ self.norm2 = ops.LayerNorm(query_dim)
67
+
68
+ self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
69
+ self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
70
+
71
+ # this can be useful: we can externally change magnitude of tanh(alpha)
72
+ # for example, when it is set to 0, then the entire model is same as
73
+ # original one
74
+ self.scale = 1
75
+
76
+ def forward(self, x, objs):
77
+
78
+ x = x + self.scale * \
79
+ torch.tanh(self.alpha_attn) * self.attn(self.norm1(x), objs, objs)
80
+ x = x + self.scale * \
81
+ torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
82
+
83
+ return x
84
+
85
+
86
+ class GatedSelfAttentionDense(nn.Module):
87
+ def __init__(self, query_dim, context_dim, n_heads, d_head):
88
+ super().__init__()
89
+
90
+ # we need a linear projection since we need cat visual feature and obj
91
+ # feature
92
+ self.linear = ops.Linear(context_dim, query_dim)
93
+
94
+ self.attn = CrossAttention(
95
+ query_dim=query_dim,
96
+ context_dim=query_dim,
97
+ heads=n_heads,
98
+ dim_head=d_head,
99
+ operations=ops)
100
+ self.ff = FeedForward(query_dim, glu=True)
101
+
102
+ self.norm1 = ops.LayerNorm(query_dim)
103
+ self.norm2 = ops.LayerNorm(query_dim)
104
+
105
+ self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
106
+ self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
107
+
108
+ # this can be useful: we can externally change magnitude of tanh(alpha)
109
+ # for example, when it is set to 0, then the entire model is same as
110
+ # original one
111
+ self.scale = 1
112
+
113
+ def forward(self, x, objs):
114
+
115
+ N_visual = x.shape[1]
116
+ objs = self.linear(objs)
117
+
118
+ x = x + self.scale * torch.tanh(self.alpha_attn) * self.attn(
119
+ self.norm1(torch.cat([x, objs], dim=1)))[:, 0:N_visual, :]
120
+ x = x + self.scale * \
121
+ torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
122
+
123
+ return x
124
+
125
+
126
+ class GatedSelfAttentionDense2(nn.Module):
127
+ def __init__(self, query_dim, context_dim, n_heads, d_head):
128
+ super().__init__()
129
+
130
+ # we need a linear projection since we need cat visual feature and obj
131
+ # feature
132
+ self.linear = ops.Linear(context_dim, query_dim)
133
+
134
+ self.attn = CrossAttention(
135
+ query_dim=query_dim, context_dim=query_dim, dim_head=d_head, operations=ops)
136
+ self.ff = FeedForward(query_dim, glu=True)
137
+
138
+ self.norm1 = ops.LayerNorm(query_dim)
139
+ self.norm2 = ops.LayerNorm(query_dim)
140
+
141
+ self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
142
+ self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
143
+
144
+ # this can be useful: we can externally change magnitude of tanh(alpha)
145
+ # for example, when it is set to 0, then the entire model is same as
146
+ # original one
147
+ self.scale = 1
148
+
149
+ def forward(self, x, objs):
150
+
151
+ B, N_visual, _ = x.shape
152
+ B, N_ground, _ = objs.shape
153
+
154
+ objs = self.linear(objs)
155
+
156
+ # sanity check
157
+ size_v = math.sqrt(N_visual)
158
+ size_g = math.sqrt(N_ground)
159
+ assert int(size_v) == size_v, "Visual tokens must be square rootable"
160
+ assert int(size_g) == size_g, "Grounding tokens must be square rootable"
161
+ size_v = int(size_v)
162
+ size_g = int(size_g)
163
+
164
+ # select grounding token and resize it to visual token size as residual
165
+ out = self.attn(self.norm1(torch.cat([x, objs], dim=1)))[
166
+ :, N_visual:, :]
167
+ out = out.permute(0, 2, 1).reshape(B, -1, size_g, size_g)
168
+ out = torch.nn.functional.interpolate(
169
+ out, (size_v, size_v), mode='bicubic')
170
+ residual = out.reshape(B, -1, N_visual).permute(0, 2, 1)
171
+
172
+ # add residual to visual feature
173
+ x = x + self.scale * torch.tanh(self.alpha_attn) * residual
174
+ x = x + self.scale * \
175
+ torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
176
+
177
+ return x
178
+
179
+
180
+ class FourierEmbedder():
181
+ def __init__(self, num_freqs=64, temperature=100):
182
+
183
+ self.num_freqs = num_freqs
184
+ self.temperature = temperature
185
+ self.freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs)
186
+
187
+ @torch.no_grad()
188
+ def __call__(self, x, cat_dim=-1):
189
+ "x: arbitrary shape of tensor. dim: cat dim"
190
+ out = []
191
+ for freq in self.freq_bands:
192
+ out.append(torch.sin(freq * x))
193
+ out.append(torch.cos(freq * x))
194
+ return torch.cat(out, cat_dim)
195
+
196
+
197
+ class PositionNet(nn.Module):
198
+ def __init__(self, in_dim, out_dim, fourier_freqs=8):
199
+ super().__init__()
200
+ self.in_dim = in_dim
201
+ self.out_dim = out_dim
202
+
203
+ self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs)
204
+ self.position_dim = fourier_freqs * 2 * 4 # 2 is sin&cos, 4 is xyxy
205
+
206
+ self.linears = nn.Sequential(
207
+ ops.Linear(self.in_dim + self.position_dim, 512),
208
+ nn.SiLU(),
209
+ ops.Linear(512, 512),
210
+ nn.SiLU(),
211
+ ops.Linear(512, out_dim),
212
+ )
213
+
214
+ self.null_positive_feature = torch.nn.Parameter(
215
+ torch.zeros([self.in_dim]))
216
+ self.null_position_feature = torch.nn.Parameter(
217
+ torch.zeros([self.position_dim]))
218
+
219
+ def forward(self, boxes, masks, positive_embeddings):
220
+ B, N, _ = boxes.shape
221
+ masks = masks.unsqueeze(-1)
222
+ positive_embeddings = positive_embeddings
223
+
224
+ # embedding position (it may includes padding as placeholder)
225
+ xyxy_embedding = self.fourier_embedder(boxes) # B*N*4 --> B*N*C
226
+
227
+ # learnable null embedding
228
+ positive_null = self.null_positive_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1)
229
+ xyxy_null = self.null_position_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1)
230
+
231
+ # replace padding with learnable null embedding
232
+ positive_embeddings = positive_embeddings * \
233
+ masks + (1 - masks) * positive_null
234
+ xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
235
+
236
+ objs = self.linears(
237
+ torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
238
+ assert objs.shape == torch.Size([B, N, self.out_dim])
239
+ return objs
240
+
241
+
242
+ class Gligen(nn.Module):
243
+ def __init__(self, modules, position_net, key_dim):
244
+ super().__init__()
245
+ self.module_list = nn.ModuleList(modules)
246
+ self.position_net = position_net
247
+ self.key_dim = key_dim
248
+ self.max_objs = 30
249
+ self.current_device = torch.device("cpu")
250
+
251
+ def _set_position(self, boxes, masks, positive_embeddings):
252
+ objs = self.position_net(boxes, masks, positive_embeddings)
253
+ def func(x, extra_options):
254
+ key = extra_options["transformer_index"]
255
+ module = self.module_list[key]
256
+ return module(x, objs.to(device=x.device, dtype=x.dtype))
257
+ return func
258
+
259
+ def set_position(self, latent_image_shape, position_params, device):
260
+ batch, c, h, w = latent_image_shape
261
+ masks = torch.zeros([self.max_objs], device="cpu")
262
+ boxes = []
263
+ positive_embeddings = []
264
+ for p in position_params:
265
+ x1 = (p[4]) / w
266
+ y1 = (p[3]) / h
267
+ x2 = (p[4] + p[2]) / w
268
+ y2 = (p[3] + p[1]) / h
269
+ masks[len(boxes)] = 1.0
270
+ boxes += [torch.tensor((x1, y1, x2, y2)).unsqueeze(0)]
271
+ positive_embeddings += [p[0]]
272
+ append_boxes = []
273
+ append_conds = []
274
+ if len(boxes) < self.max_objs:
275
+ append_boxes = [torch.zeros(
276
+ [self.max_objs - len(boxes), 4], device="cpu")]
277
+ append_conds = [torch.zeros(
278
+ [self.max_objs - len(boxes), self.key_dim], device="cpu")]
279
+
280
+ box_out = torch.cat(
281
+ boxes + append_boxes).unsqueeze(0).repeat(batch, 1, 1)
282
+ masks = masks.unsqueeze(0).repeat(batch, 1)
283
+ conds = torch.cat(positive_embeddings +
284
+ append_conds).unsqueeze(0).repeat(batch, 1, 1)
285
+ return self._set_position(
286
+ box_out.to(device),
287
+ masks.to(device),
288
+ conds.to(device))
289
+
290
+ def set_empty(self, latent_image_shape, device):
291
+ batch, c, h, w = latent_image_shape
292
+ masks = torch.zeros([self.max_objs], device="cpu").repeat(batch, 1)
293
+ box_out = torch.zeros([self.max_objs, 4],
294
+ device="cpu").repeat(batch, 1, 1)
295
+ conds = torch.zeros([self.max_objs, self.key_dim],
296
+ device="cpu").repeat(batch, 1, 1)
297
+ return self._set_position(
298
+ box_out.to(device),
299
+ masks.to(device),
300
+ conds.to(device))
301
+
302
+
303
+ def load_gligen(sd):
304
+ sd_k = sd.keys()
305
+ output_list = []
306
+ key_dim = 768
307
+ for a in ["input_blocks", "middle_block", "output_blocks"]:
308
+ for b in range(20):
309
+ k_temp = filter(lambda k: "{}.{}.".format(a, b)
310
+ in k and ".fuser." in k, sd_k)
311
+ k_temp = map(lambda k: (k, k.split(".fuser.")[-1]), k_temp)
312
+
313
+ n_sd = {}
314
+ for k in k_temp:
315
+ n_sd[k[1]] = sd[k[0]]
316
+ if len(n_sd) > 0:
317
+ query_dim = n_sd["linear.weight"].shape[0]
318
+ key_dim = n_sd["linear.weight"].shape[1]
319
+
320
+ if key_dim == 768: # SD1.x
321
+ n_heads = 8
322
+ d_head = query_dim // n_heads
323
+ else:
324
+ d_head = 64
325
+ n_heads = query_dim // d_head
326
+
327
+ gated = GatedSelfAttentionDense(
328
+ query_dim, key_dim, n_heads, d_head)
329
+ gated.load_state_dict(n_sd, strict=False)
330
+ output_list.append(gated)
331
+
332
+ if "position_net.null_positive_feature" in sd_k:
333
+ in_dim = sd["position_net.null_positive_feature"].shape[0]
334
+ out_dim = sd["position_net.linears.4.weight"].shape[0]
335
+
336
+ class WeightsLoader(torch.nn.Module):
337
+ pass
338
+ w = WeightsLoader()
339
+ w.position_net = PositionNet(in_dim, out_dim)
340
+ w.load_state_dict(sd, strict=False)
341
+
342
+ gligen = Gligen(output_list, w.position_net, key_dim)
343
+ return gligen
ComfyUI/comfy/k_diffusion/deis.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #Taken from: https://github.com/zju-pi/diff-sampler/blob/main/gits-main/solver_utils.py
2
+ #under Apache 2 license
3
+ import torch
4
+ import numpy as np
5
+
6
+ # A pytorch reimplementation of DEIS (https://github.com/qsh-zh/deis).
7
+ #############################
8
+ ### Utils for DEIS solver ###
9
+ #############################
10
+ #----------------------------------------------------------------------------
11
+ # Transfer from the input time (sigma) used in EDM to that (t) used in DEIS.
12
+
13
+ def edm2t(edm_steps, epsilon_s=1e-3, sigma_min=0.002, sigma_max=80):
14
+ vp_sigma = lambda beta_d, beta_min: lambda t: (np.e ** (0.5 * beta_d * (t ** 2) + beta_min * t) - 1) ** 0.5
15
+ vp_sigma_inv = lambda beta_d, beta_min: lambda sigma: ((beta_min ** 2 + 2 * beta_d * (sigma ** 2 + 1).log()).sqrt() - beta_min) / beta_d
16
+ vp_beta_d = 2 * (np.log(torch.tensor(sigma_min).cpu() ** 2 + 1) / epsilon_s - np.log(torch.tensor(sigma_max).cpu() ** 2 + 1)) / (epsilon_s - 1)
17
+ vp_beta_min = np.log(torch.tensor(sigma_max).cpu() ** 2 + 1) - 0.5 * vp_beta_d
18
+ t_steps = vp_sigma_inv(vp_beta_d.clone().detach().cpu(), vp_beta_min.clone().detach().cpu())(edm_steps.clone().detach().cpu())
19
+ return t_steps, vp_beta_min, vp_beta_d + vp_beta_min
20
+
21
+ #----------------------------------------------------------------------------
22
+
23
+ def cal_poly(prev_t, j, taus):
24
+ poly = 1
25
+ for k in range(prev_t.shape[0]):
26
+ if k == j:
27
+ continue
28
+ poly *= (taus - prev_t[k]) / (prev_t[j] - prev_t[k])
29
+ return poly
30
+
31
+ #----------------------------------------------------------------------------
32
+ # Transfer from t to alpha_t.
33
+
34
+ def t2alpha_fn(beta_0, beta_1, t):
35
+ return torch.exp(-0.5 * t ** 2 * (beta_1 - beta_0) - t * beta_0)
36
+
37
+ #----------------------------------------------------------------------------
38
+
39
+ def cal_intergrand(beta_0, beta_1, taus):
40
+ with torch.inference_mode(mode=False):
41
+ taus = taus.clone()
42
+ beta_0 = beta_0.clone()
43
+ beta_1 = beta_1.clone()
44
+ with torch.enable_grad():
45
+ taus.requires_grad_(True)
46
+ alpha = t2alpha_fn(beta_0, beta_1, taus)
47
+ log_alpha = alpha.log()
48
+ log_alpha.sum().backward()
49
+ d_log_alpha_dtau = taus.grad
50
+ integrand = -0.5 * d_log_alpha_dtau / torch.sqrt(alpha * (1 - alpha))
51
+ return integrand
52
+
53
+ #----------------------------------------------------------------------------
54
+
55
+ def get_deis_coeff_list(t_steps, max_order, N=10000, deis_mode='tab'):
56
+ """
57
+ Get the coefficient list for DEIS sampling.
58
+
59
+ Args:
60
+ t_steps: A pytorch tensor. The time steps for sampling.
61
+ max_order: A `int`. Maximum order of the solver. 1 <= max_order <= 4
62
+ N: A `int`. Use how many points to perform the numerical integration when deis_mode=='tab'.
63
+ deis_mode: A `str`. Select between 'tab' and 'rhoab'. Type of DEIS.
64
+ Returns:
65
+ A pytorch tensor. A batch of generated samples or sampling trajectories if return_inters=True.
66
+ """
67
+ if deis_mode == 'tab':
68
+ t_steps, beta_0, beta_1 = edm2t(t_steps)
69
+ C = []
70
+ for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):
71
+ order = min(i+1, max_order)
72
+ if order == 1:
73
+ C.append([])
74
+ else:
75
+ taus = torch.linspace(t_cur, t_next, N) # split the interval for integral appximation
76
+ dtau = (t_next - t_cur) / N
77
+ prev_t = t_steps[[i - k for k in range(order)]]
78
+ coeff_temp = []
79
+ integrand = cal_intergrand(beta_0, beta_1, taus)
80
+ for j in range(order):
81
+ poly = cal_poly(prev_t, j, taus)
82
+ coeff_temp.append(torch.sum(integrand * poly) * dtau)
83
+ C.append(coeff_temp)
84
+
85
+ elif deis_mode == 'rhoab':
86
+ # Analytical solution, second order
87
+ def get_def_intergral_2(a, b, start, end, c):
88
+ coeff = (end**3 - start**3) / 3 - (end**2 - start**2) * (a + b) / 2 + (end - start) * a * b
89
+ return coeff / ((c - a) * (c - b))
90
+
91
+ # Analytical solution, third order
92
+ def get_def_intergral_3(a, b, c, start, end, d):
93
+ coeff = (end**4 - start**4) / 4 - (end**3 - start**3) * (a + b + c) / 3 \
94
+ + (end**2 - start**2) * (a*b + a*c + b*c) / 2 - (end - start) * a * b * c
95
+ return coeff / ((d - a) * (d - b) * (d - c))
96
+
97
+ C = []
98
+ for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):
99
+ order = min(i, max_order)
100
+ if order == 0:
101
+ C.append([])
102
+ else:
103
+ prev_t = t_steps[[i - k for k in range(order+1)]]
104
+ if order == 1:
105
+ coeff_cur = ((t_next - prev_t[1])**2 - (t_cur - prev_t[1])**2) / (2 * (t_cur - prev_t[1]))
106
+ coeff_prev1 = (t_next - t_cur)**2 / (2 * (prev_t[1] - t_cur))
107
+ coeff_temp = [coeff_cur, coeff_prev1]
108
+ elif order == 2:
109
+ coeff_cur = get_def_intergral_2(prev_t[1], prev_t[2], t_cur, t_next, t_cur)
110
+ coeff_prev1 = get_def_intergral_2(t_cur, prev_t[2], t_cur, t_next, prev_t[1])
111
+ coeff_prev2 = get_def_intergral_2(t_cur, prev_t[1], t_cur, t_next, prev_t[2])
112
+ coeff_temp = [coeff_cur, coeff_prev1, coeff_prev2]
113
+ elif order == 3:
114
+ coeff_cur = get_def_intergral_3(prev_t[1], prev_t[2], prev_t[3], t_cur, t_next, t_cur)
115
+ coeff_prev1 = get_def_intergral_3(t_cur, prev_t[2], prev_t[3], t_cur, t_next, prev_t[1])
116
+ coeff_prev2 = get_def_intergral_3(t_cur, prev_t[1], prev_t[3], t_cur, t_next, prev_t[2])
117
+ coeff_prev3 = get_def_intergral_3(t_cur, prev_t[1], prev_t[2], t_cur, t_next, prev_t[3])
118
+ coeff_temp = [coeff_cur, coeff_prev1, coeff_prev2, coeff_prev3]
119
+ C.append(coeff_temp)
120
+ return C
121
+
ComfyUI/comfy/k_diffusion/sampling.py ADDED
@@ -0,0 +1,1050 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+
3
+ from scipy import integrate
4
+ import torch
5
+ from torch import nn
6
+ import torchsde
7
+ from tqdm.auto import trange, tqdm
8
+
9
+ from . import utils
10
+ from . import deis
11
+ import comfy.model_patcher
12
+
13
+ def append_zero(x):
14
+ return torch.cat([x, x.new_zeros([1])])
15
+
16
+
17
+ def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'):
18
+ """Constructs the noise schedule of Karras et al. (2022)."""
19
+ ramp = torch.linspace(0, 1, n, device=device)
20
+ min_inv_rho = sigma_min ** (1 / rho)
21
+ max_inv_rho = sigma_max ** (1 / rho)
22
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
23
+ return append_zero(sigmas).to(device)
24
+
25
+
26
+ def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'):
27
+ """Constructs an exponential noise schedule."""
28
+ sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp()
29
+ return append_zero(sigmas)
30
+
31
+
32
+ def get_sigmas_polyexponential(n, sigma_min, sigma_max, rho=1., device='cpu'):
33
+ """Constructs an polynomial in log sigma noise schedule."""
34
+ ramp = torch.linspace(1, 0, n, device=device) ** rho
35
+ sigmas = torch.exp(ramp * (math.log(sigma_max) - math.log(sigma_min)) + math.log(sigma_min))
36
+ return append_zero(sigmas)
37
+
38
+
39
+ def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
40
+ """Constructs a continuous VP noise schedule."""
41
+ t = torch.linspace(1, eps_s, n, device=device)
42
+ sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
43
+ return append_zero(sigmas)
44
+
45
+
46
+ def to_d(x, sigma, denoised):
47
+ """Converts a denoiser output to a Karras ODE derivative."""
48
+ return (x - denoised) / utils.append_dims(sigma, x.ndim)
49
+
50
+
51
+ def get_ancestral_step(sigma_from, sigma_to, eta=1.):
52
+ """Calculates the noise level (sigma_down) to step down to and the amount
53
+ of noise to add (sigma_up) when doing an ancestral sampling step."""
54
+ if not eta:
55
+ return sigma_to, 0.
56
+ sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5)
57
+ sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5
58
+ return sigma_down, sigma_up
59
+
60
+
61
+ def default_noise_sampler(x):
62
+ return lambda sigma, sigma_next: torch.randn_like(x)
63
+
64
+
65
+ class BatchedBrownianTree:
66
+ """A wrapper around torchsde.BrownianTree that enables batches of entropy."""
67
+
68
+ def __init__(self, x, t0, t1, seed=None, **kwargs):
69
+ self.cpu_tree = True
70
+ if "cpu" in kwargs:
71
+ self.cpu_tree = kwargs.pop("cpu")
72
+ t0, t1, self.sign = self.sort(t0, t1)
73
+ w0 = kwargs.get('w0', torch.zeros_like(x))
74
+ if seed is None:
75
+ seed = torch.randint(0, 2 ** 63 - 1, []).item()
76
+ self.batched = True
77
+ try:
78
+ assert len(seed) == x.shape[0]
79
+ w0 = w0[0]
80
+ except TypeError:
81
+ seed = [seed]
82
+ self.batched = False
83
+ if self.cpu_tree:
84
+ self.trees = [torchsde.BrownianTree(t0.cpu(), w0.cpu(), t1.cpu(), entropy=s, **kwargs) for s in seed]
85
+ else:
86
+ self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
87
+
88
+ @staticmethod
89
+ def sort(a, b):
90
+ return (a, b, 1) if a < b else (b, a, -1)
91
+
92
+ def __call__(self, t0, t1):
93
+ t0, t1, sign = self.sort(t0, t1)
94
+ if self.cpu_tree:
95
+ w = torch.stack([tree(t0.cpu().float(), t1.cpu().float()).to(t0.dtype).to(t0.device) for tree in self.trees]) * (self.sign * sign)
96
+ else:
97
+ w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
98
+
99
+ return w if self.batched else w[0]
100
+
101
+
102
+ class BrownianTreeNoiseSampler:
103
+ """A noise sampler backed by a torchsde.BrownianTree.
104
+
105
+ Args:
106
+ x (Tensor): The tensor whose shape, device and dtype to use to generate
107
+ random samples.
108
+ sigma_min (float): The low end of the valid interval.
109
+ sigma_max (float): The high end of the valid interval.
110
+ seed (int or List[int]): The random seed. If a list of seeds is
111
+ supplied instead of a single integer, then the noise sampler will
112
+ use one BrownianTree per batch item, each with its own seed.
113
+ transform (callable): A function that maps sigma to the sampler's
114
+ internal timestep.
115
+ """
116
+
117
+ def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x, cpu=False):
118
+ self.transform = transform
119
+ t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
120
+ self.tree = BatchedBrownianTree(x, t0, t1, seed, cpu=cpu)
121
+
122
+ def __call__(self, sigma, sigma_next):
123
+ t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
124
+ return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
125
+
126
+
127
+ @torch.no_grad()
128
+ def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
129
+ """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
130
+ extra_args = {} if extra_args is None else extra_args
131
+ s_in = x.new_ones([x.shape[0]])
132
+ for i in trange(len(sigmas) - 1, disable=disable):
133
+ if s_churn > 0:
134
+ gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
135
+ sigma_hat = sigmas[i] * (gamma + 1)
136
+ else:
137
+ gamma = 0
138
+ sigma_hat = sigmas[i]
139
+
140
+ if gamma > 0:
141
+ eps = torch.randn_like(x) * s_noise
142
+ x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
143
+ denoised = model(x, sigma_hat * s_in, **extra_args)
144
+ d = to_d(x, sigma_hat, denoised)
145
+ if callback is not None:
146
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
147
+ dt = sigmas[i + 1] - sigma_hat
148
+ # Euler method
149
+ x = x + d * dt
150
+ return x
151
+
152
+
153
+ @torch.no_grad()
154
+ def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
155
+ """Ancestral sampling with Euler method steps."""
156
+ extra_args = {} if extra_args is None else extra_args
157
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
158
+ s_in = x.new_ones([x.shape[0]])
159
+ for i in trange(len(sigmas) - 1, disable=disable):
160
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
161
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
162
+ if callback is not None:
163
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
164
+ d = to_d(x, sigmas[i], denoised)
165
+ # Euler method
166
+ dt = sigma_down - sigmas[i]
167
+ x = x + d * dt
168
+ if sigmas[i + 1] > 0:
169
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
170
+ return x
171
+
172
+
173
+ @torch.no_grad()
174
+ def sample_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
175
+ """Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
176
+ extra_args = {} if extra_args is None else extra_args
177
+ s_in = x.new_ones([x.shape[0]])
178
+ for i in trange(len(sigmas) - 1, disable=disable):
179
+ if s_churn > 0:
180
+ gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
181
+ sigma_hat = sigmas[i] * (gamma + 1)
182
+ else:
183
+ gamma = 0
184
+ sigma_hat = sigmas[i]
185
+
186
+ sigma_hat = sigmas[i] * (gamma + 1)
187
+ if gamma > 0:
188
+ eps = torch.randn_like(x) * s_noise
189
+ x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
190
+ denoised = model(x, sigma_hat * s_in, **extra_args)
191
+ d = to_d(x, sigma_hat, denoised)
192
+ if callback is not None:
193
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
194
+ dt = sigmas[i + 1] - sigma_hat
195
+ if sigmas[i + 1] == 0:
196
+ # Euler method
197
+ x = x + d * dt
198
+ else:
199
+ # Heun's method
200
+ x_2 = x + d * dt
201
+ denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
202
+ d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
203
+ d_prime = (d + d_2) / 2
204
+ x = x + d_prime * dt
205
+ return x
206
+
207
+
208
+ @torch.no_grad()
209
+ def sample_dpm_2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
210
+ """A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022)."""
211
+ extra_args = {} if extra_args is None else extra_args
212
+ s_in = x.new_ones([x.shape[0]])
213
+ for i in trange(len(sigmas) - 1, disable=disable):
214
+ if s_churn > 0:
215
+ gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
216
+ sigma_hat = sigmas[i] * (gamma + 1)
217
+ else:
218
+ gamma = 0
219
+ sigma_hat = sigmas[i]
220
+
221
+ if gamma > 0:
222
+ eps = torch.randn_like(x) * s_noise
223
+ x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
224
+ denoised = model(x, sigma_hat * s_in, **extra_args)
225
+ d = to_d(x, sigma_hat, denoised)
226
+ if callback is not None:
227
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
228
+ if sigmas[i + 1] == 0:
229
+ # Euler method
230
+ dt = sigmas[i + 1] - sigma_hat
231
+ x = x + d * dt
232
+ else:
233
+ # DPM-Solver-2
234
+ sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp()
235
+ dt_1 = sigma_mid - sigma_hat
236
+ dt_2 = sigmas[i + 1] - sigma_hat
237
+ x_2 = x + d * dt_1
238
+ denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
239
+ d_2 = to_d(x_2, sigma_mid, denoised_2)
240
+ x = x + d_2 * dt_2
241
+ return x
242
+
243
+
244
+ @torch.no_grad()
245
+ def sample_dpm_2_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
246
+ """Ancestral sampling with DPM-Solver second-order steps."""
247
+ extra_args = {} if extra_args is None else extra_args
248
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
249
+ s_in = x.new_ones([x.shape[0]])
250
+ for i in trange(len(sigmas) - 1, disable=disable):
251
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
252
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
253
+ if callback is not None:
254
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
255
+ d = to_d(x, sigmas[i], denoised)
256
+ if sigma_down == 0:
257
+ # Euler method
258
+ dt = sigma_down - sigmas[i]
259
+ x = x + d * dt
260
+ else:
261
+ # DPM-Solver-2
262
+ sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
263
+ dt_1 = sigma_mid - sigmas[i]
264
+ dt_2 = sigma_down - sigmas[i]
265
+ x_2 = x + d * dt_1
266
+ denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
267
+ d_2 = to_d(x_2, sigma_mid, denoised_2)
268
+ x = x + d_2 * dt_2
269
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
270
+ return x
271
+
272
+
273
+ def linear_multistep_coeff(order, t, i, j):
274
+ if order - 1 > i:
275
+ raise ValueError(f'Order {order} too high for step {i}')
276
+ def fn(tau):
277
+ prod = 1.
278
+ for k in range(order):
279
+ if j == k:
280
+ continue
281
+ prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
282
+ return prod
283
+ return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0]
284
+
285
+
286
+ @torch.no_grad()
287
+ def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, order=4):
288
+ extra_args = {} if extra_args is None else extra_args
289
+ s_in = x.new_ones([x.shape[0]])
290
+ sigmas_cpu = sigmas.detach().cpu().numpy()
291
+ ds = []
292
+ for i in trange(len(sigmas) - 1, disable=disable):
293
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
294
+ d = to_d(x, sigmas[i], denoised)
295
+ ds.append(d)
296
+ if len(ds) > order:
297
+ ds.pop(0)
298
+ if callback is not None:
299
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
300
+ cur_order = min(i + 1, order)
301
+ coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
302
+ x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
303
+ return x
304
+
305
+
306
+ class PIDStepSizeController:
307
+ """A PID controller for ODE adaptive step size control."""
308
+ def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8):
309
+ self.h = h
310
+ self.b1 = (pcoeff + icoeff + dcoeff) / order
311
+ self.b2 = -(pcoeff + 2 * dcoeff) / order
312
+ self.b3 = dcoeff / order
313
+ self.accept_safety = accept_safety
314
+ self.eps = eps
315
+ self.errs = []
316
+
317
+ def limiter(self, x):
318
+ return 1 + math.atan(x - 1)
319
+
320
+ def propose_step(self, error):
321
+ inv_error = 1 / (float(error) + self.eps)
322
+ if not self.errs:
323
+ self.errs = [inv_error, inv_error, inv_error]
324
+ self.errs[0] = inv_error
325
+ factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3
326
+ factor = self.limiter(factor)
327
+ accept = factor >= self.accept_safety
328
+ if accept:
329
+ self.errs[2] = self.errs[1]
330
+ self.errs[1] = self.errs[0]
331
+ self.h *= factor
332
+ return accept
333
+
334
+
335
+ class DPMSolver(nn.Module):
336
+ """DPM-Solver. See https://arxiv.org/abs/2206.00927."""
337
+
338
+ def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None):
339
+ super().__init__()
340
+ self.model = model
341
+ self.extra_args = {} if extra_args is None else extra_args
342
+ self.eps_callback = eps_callback
343
+ self.info_callback = info_callback
344
+
345
+ def t(self, sigma):
346
+ return -sigma.log()
347
+
348
+ def sigma(self, t):
349
+ return t.neg().exp()
350
+
351
+ def eps(self, eps_cache, key, x, t, *args, **kwargs):
352
+ if key in eps_cache:
353
+ return eps_cache[key], eps_cache
354
+ sigma = self.sigma(t) * x.new_ones([x.shape[0]])
355
+ eps = (x - self.model(x, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t)
356
+ if self.eps_callback is not None:
357
+ self.eps_callback()
358
+ return eps, {key: eps, **eps_cache}
359
+
360
+ def dpm_solver_1_step(self, x, t, t_next, eps_cache=None):
361
+ eps_cache = {} if eps_cache is None else eps_cache
362
+ h = t_next - t
363
+ eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
364
+ x_1 = x - self.sigma(t_next) * h.expm1() * eps
365
+ return x_1, eps_cache
366
+
367
+ def dpm_solver_2_step(self, x, t, t_next, r1=1 / 2, eps_cache=None):
368
+ eps_cache = {} if eps_cache is None else eps_cache
369
+ h = t_next - t
370
+ eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
371
+ s1 = t + r1 * h
372
+ u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
373
+ eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
374
+ x_2 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps)
375
+ return x_2, eps_cache
376
+
377
+ def dpm_solver_3_step(self, x, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None):
378
+ eps_cache = {} if eps_cache is None else eps_cache
379
+ h = t_next - t
380
+ eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
381
+ s1 = t + r1 * h
382
+ s2 = t + r2 * h
383
+ u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
384
+ eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
385
+ u2 = x - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps)
386
+ eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', u2, s2)
387
+ x_3 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps)
388
+ return x_3, eps_cache
389
+
390
+ def dpm_solver_fast(self, x, t_start, t_end, nfe, eta=0., s_noise=1., noise_sampler=None):
391
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
392
+ if not t_end > t_start and eta:
393
+ raise ValueError('eta must be 0 for reverse sampling')
394
+
395
+ m = math.floor(nfe / 3) + 1
396
+ ts = torch.linspace(t_start, t_end, m + 1, device=x.device)
397
+
398
+ if nfe % 3 == 0:
399
+ orders = [3] * (m - 2) + [2, 1]
400
+ else:
401
+ orders = [3] * (m - 1) + [nfe % 3]
402
+
403
+ for i in range(len(orders)):
404
+ eps_cache = {}
405
+ t, t_next = ts[i], ts[i + 1]
406
+ if eta:
407
+ sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta)
408
+ t_next_ = torch.minimum(t_end, self.t(sd))
409
+ su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5
410
+ else:
411
+ t_next_, su = t_next, 0.
412
+
413
+ eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
414
+ denoised = x - self.sigma(t) * eps
415
+ if self.info_callback is not None:
416
+ self.info_callback({'x': x, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised})
417
+
418
+ if orders[i] == 1:
419
+ x, eps_cache = self.dpm_solver_1_step(x, t, t_next_, eps_cache=eps_cache)
420
+ elif orders[i] == 2:
421
+ x, eps_cache = self.dpm_solver_2_step(x, t, t_next_, eps_cache=eps_cache)
422
+ else:
423
+ x, eps_cache = self.dpm_solver_3_step(x, t, t_next_, eps_cache=eps_cache)
424
+
425
+ x = x + su * s_noise * noise_sampler(self.sigma(t), self.sigma(t_next))
426
+
427
+ return x
428
+
429
+ def dpm_solver_adaptive(self, x, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None):
430
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
431
+ if order not in {2, 3}:
432
+ raise ValueError('order should be 2 or 3')
433
+ forward = t_end > t_start
434
+ if not forward and eta:
435
+ raise ValueError('eta must be 0 for reverse sampling')
436
+ h_init = abs(h_init) * (1 if forward else -1)
437
+ atol = torch.tensor(atol)
438
+ rtol = torch.tensor(rtol)
439
+ s = t_start
440
+ x_prev = x
441
+ accept = True
442
+ pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety)
443
+ info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0}
444
+
445
+ while s < t_end - 1e-5 if forward else s > t_end + 1e-5:
446
+ eps_cache = {}
447
+ t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h)
448
+ if eta:
449
+ sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta)
450
+ t_ = torch.minimum(t_end, self.t(sd))
451
+ su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5
452
+ else:
453
+ t_, su = t, 0.
454
+
455
+ eps, eps_cache = self.eps(eps_cache, 'eps', x, s)
456
+ denoised = x - self.sigma(s) * eps
457
+
458
+ if order == 2:
459
+ x_low, eps_cache = self.dpm_solver_1_step(x, s, t_, eps_cache=eps_cache)
460
+ x_high, eps_cache = self.dpm_solver_2_step(x, s, t_, eps_cache=eps_cache)
461
+ else:
462
+ x_low, eps_cache = self.dpm_solver_2_step(x, s, t_, r1=1 / 3, eps_cache=eps_cache)
463
+ x_high, eps_cache = self.dpm_solver_3_step(x, s, t_, eps_cache=eps_cache)
464
+ delta = torch.maximum(atol, rtol * torch.maximum(x_low.abs(), x_prev.abs()))
465
+ error = torch.linalg.norm((x_low - x_high) / delta) / x.numel() ** 0.5
466
+ accept = pid.propose_step(error)
467
+ if accept:
468
+ x_prev = x_low
469
+ x = x_high + su * s_noise * noise_sampler(self.sigma(s), self.sigma(t))
470
+ s = t
471
+ info['n_accept'] += 1
472
+ else:
473
+ info['n_reject'] += 1
474
+ info['nfe'] += order
475
+ info['steps'] += 1
476
+
477
+ if self.info_callback is not None:
478
+ self.info_callback({'x': x, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info})
479
+
480
+ return x, info
481
+
482
+
483
+ @torch.no_grad()
484
+ def sample_dpm_fast(model, x, sigma_min, sigma_max, n, extra_args=None, callback=None, disable=None, eta=0., s_noise=1., noise_sampler=None):
485
+ """DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927."""
486
+ if sigma_min <= 0 or sigma_max <= 0:
487
+ raise ValueError('sigma_min and sigma_max must not be 0')
488
+ with tqdm(total=n, disable=disable) as pbar:
489
+ dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
490
+ if callback is not None:
491
+ dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
492
+ return dpm_solver.dpm_solver_fast(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise, noise_sampler)
493
+
494
+
495
+ @torch.no_grad()
496
+ def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callback=None, disable=None, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None, return_info=False):
497
+ """DPM-Solver-12 and 23 (adaptive step size). See https://arxiv.org/abs/2206.00927."""
498
+ if sigma_min <= 0 or sigma_max <= 0:
499
+ raise ValueError('sigma_min and sigma_max must not be 0')
500
+ with tqdm(disable=disable) as pbar:
501
+ dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
502
+ if callback is not None:
503
+ dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
504
+ x, info = dpm_solver.dpm_solver_adaptive(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise, noise_sampler)
505
+ if return_info:
506
+ return x, info
507
+ return x
508
+
509
+
510
+ @torch.no_grad()
511
+ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
512
+ """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
513
+ extra_args = {} if extra_args is None else extra_args
514
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
515
+ s_in = x.new_ones([x.shape[0]])
516
+ sigma_fn = lambda t: t.neg().exp()
517
+ t_fn = lambda sigma: sigma.log().neg()
518
+
519
+ for i in trange(len(sigmas) - 1, disable=disable):
520
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
521
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
522
+ if callback is not None:
523
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
524
+ if sigma_down == 0:
525
+ # Euler method
526
+ d = to_d(x, sigmas[i], denoised)
527
+ dt = sigma_down - sigmas[i]
528
+ x = x + d * dt
529
+ else:
530
+ # DPM-Solver++(2S)
531
+ t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
532
+ r = 1 / 2
533
+ h = t_next - t
534
+ s = t + r * h
535
+ x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * denoised
536
+ denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
537
+ x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_2
538
+ # Noise addition
539
+ if sigmas[i + 1] > 0:
540
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
541
+ return x
542
+
543
+
544
+ @torch.no_grad()
545
+ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
546
+ """DPM-Solver++ (stochastic)."""
547
+ if len(sigmas) <= 1:
548
+ return x
549
+
550
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
551
+ seed = extra_args.get("seed", None)
552
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
553
+ extra_args = {} if extra_args is None else extra_args
554
+ s_in = x.new_ones([x.shape[0]])
555
+ sigma_fn = lambda t: t.neg().exp()
556
+ t_fn = lambda sigma: sigma.log().neg()
557
+
558
+ for i in trange(len(sigmas) - 1, disable=disable):
559
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
560
+ if callback is not None:
561
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
562
+ if sigmas[i + 1] == 0:
563
+ # Euler method
564
+ d = to_d(x, sigmas[i], denoised)
565
+ dt = sigmas[i + 1] - sigmas[i]
566
+ x = x + d * dt
567
+ else:
568
+ # DPM-Solver++
569
+ t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
570
+ h = t_next - t
571
+ s = t + h * r
572
+ fac = 1 / (2 * r)
573
+
574
+ # Step 1
575
+ sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
576
+ s_ = t_fn(sd)
577
+ x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
578
+ x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
579
+ denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
580
+
581
+ # Step 2
582
+ sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
583
+ t_next_ = t_fn(sd)
584
+ denoised_d = (1 - fac) * denoised + fac * denoised_2
585
+ x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
586
+ x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
587
+ return x
588
+
589
+
590
+ @torch.no_grad()
591
+ def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
592
+ """DPM-Solver++(2M)."""
593
+ extra_args = {} if extra_args is None else extra_args
594
+ s_in = x.new_ones([x.shape[0]])
595
+ sigma_fn = lambda t: t.neg().exp()
596
+ t_fn = lambda sigma: sigma.log().neg()
597
+ old_denoised = None
598
+
599
+ for i in trange(len(sigmas) - 1, disable=disable):
600
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
601
+ if callback is not None:
602
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
603
+ t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
604
+ h = t_next - t
605
+ if old_denoised is None or sigmas[i + 1] == 0:
606
+ x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
607
+ else:
608
+ h_last = t - t_fn(sigmas[i - 1])
609
+ r = h_last / h
610
+ denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
611
+ x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
612
+ old_denoised = denoised
613
+ return x
614
+
615
+ @torch.no_grad()
616
+ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
617
+ """DPM-Solver++(2M) SDE."""
618
+ if len(sigmas) <= 1:
619
+ return x
620
+
621
+ if solver_type not in {'heun', 'midpoint'}:
622
+ raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
623
+
624
+ seed = extra_args.get("seed", None)
625
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
626
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
627
+ extra_args = {} if extra_args is None else extra_args
628
+ s_in = x.new_ones([x.shape[0]])
629
+
630
+ old_denoised = None
631
+ h_last = None
632
+ h = None
633
+
634
+ for i in trange(len(sigmas) - 1, disable=disable):
635
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
636
+ if callback is not None:
637
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
638
+ if sigmas[i + 1] == 0:
639
+ # Denoising step
640
+ x = denoised
641
+ else:
642
+ # DPM-Solver++(2M) SDE
643
+ t, s = -sigmas[i].log(), -sigmas[i + 1].log()
644
+ h = s - t
645
+ eta_h = eta * h
646
+
647
+ x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
648
+
649
+ if old_denoised is not None:
650
+ r = h_last / h
651
+ if solver_type == 'heun':
652
+ x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
653
+ elif solver_type == 'midpoint':
654
+ x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
655
+
656
+ if eta:
657
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
658
+
659
+ old_denoised = denoised
660
+ h_last = h
661
+ return x
662
+
663
+ @torch.no_grad()
664
+ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
665
+ """DPM-Solver++(3M) SDE."""
666
+
667
+ if len(sigmas) <= 1:
668
+ return x
669
+
670
+ seed = extra_args.get("seed", None)
671
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
672
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
673
+ extra_args = {} if extra_args is None else extra_args
674
+ s_in = x.new_ones([x.shape[0]])
675
+
676
+ denoised_1, denoised_2 = None, None
677
+ h, h_1, h_2 = None, None, None
678
+
679
+ for i in trange(len(sigmas) - 1, disable=disable):
680
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
681
+ if callback is not None:
682
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
683
+ if sigmas[i + 1] == 0:
684
+ # Denoising step
685
+ x = denoised
686
+ else:
687
+ t, s = -sigmas[i].log(), -sigmas[i + 1].log()
688
+ h = s - t
689
+ h_eta = h * (eta + 1)
690
+
691
+ x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised
692
+
693
+ if h_2 is not None:
694
+ r0 = h_1 / h
695
+ r1 = h_2 / h
696
+ d1_0 = (denoised - denoised_1) / r0
697
+ d1_1 = (denoised_1 - denoised_2) / r1
698
+ d1 = d1_0 + (d1_0 - d1_1) * r0 / (r0 + r1)
699
+ d2 = (d1_0 - d1_1) / (r0 + r1)
700
+ phi_2 = h_eta.neg().expm1() / h_eta + 1
701
+ phi_3 = phi_2 / h_eta - 0.5
702
+ x = x + phi_2 * d1 - phi_3 * d2
703
+ elif h_1 is not None:
704
+ r = h_1 / h
705
+ d = (denoised - denoised_1) / r
706
+ phi_2 = h_eta.neg().expm1() / h_eta + 1
707
+ x = x + phi_2 * d
708
+
709
+ if eta:
710
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
711
+
712
+ denoised_1, denoised_2 = denoised, denoised_1
713
+ h_1, h_2 = h, h_1
714
+ return x
715
+
716
+ @torch.no_grad()
717
+ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
718
+ if len(sigmas) <= 1:
719
+ return x
720
+
721
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
722
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
723
+ return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
724
+
725
+ @torch.no_grad()
726
+ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
727
+ if len(sigmas) <= 1:
728
+ return x
729
+
730
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
731
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
732
+ return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
733
+
734
+ @torch.no_grad()
735
+ def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
736
+ if len(sigmas) <= 1:
737
+ return x
738
+
739
+ sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
740
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
741
+ return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r)
742
+
743
+
744
+ def DDPMSampler_step(x, sigma, sigma_prev, noise, noise_sampler):
745
+ alpha_cumprod = 1 / ((sigma * sigma) + 1)
746
+ alpha_cumprod_prev = 1 / ((sigma_prev * sigma_prev) + 1)
747
+ alpha = (alpha_cumprod / alpha_cumprod_prev)
748
+
749
+ mu = (1.0 / alpha).sqrt() * (x - (1 - alpha) * noise / (1 - alpha_cumprod).sqrt())
750
+ if sigma_prev > 0:
751
+ mu += ((1 - alpha) * (1. - alpha_cumprod_prev) / (1. - alpha_cumprod)).sqrt() * noise_sampler(sigma, sigma_prev)
752
+ return mu
753
+
754
+ def generic_step_sampler(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None, step_function=None):
755
+ extra_args = {} if extra_args is None else extra_args
756
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
757
+ s_in = x.new_ones([x.shape[0]])
758
+
759
+ for i in trange(len(sigmas) - 1, disable=disable):
760
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
761
+ if callback is not None:
762
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
763
+ x = step_function(x / torch.sqrt(1.0 + sigmas[i] ** 2.0), sigmas[i], sigmas[i + 1], (x - denoised) / sigmas[i], noise_sampler)
764
+ if sigmas[i + 1] != 0:
765
+ x *= torch.sqrt(1.0 + sigmas[i + 1] ** 2.0)
766
+ return x
767
+
768
+
769
+ @torch.no_grad()
770
+ def sample_ddpm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
771
+ return generic_step_sampler(model, x, sigmas, extra_args, callback, disable, noise_sampler, DDPMSampler_step)
772
+
773
+ @torch.no_grad()
774
+ def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
775
+ extra_args = {} if extra_args is None else extra_args
776
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
777
+ s_in = x.new_ones([x.shape[0]])
778
+ for i in trange(len(sigmas) - 1, disable=disable):
779
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
780
+ if callback is not None:
781
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
782
+
783
+ x = denoised
784
+ if sigmas[i + 1] > 0:
785
+ x = model.inner_model.inner_model.model_sampling.noise_scaling(sigmas[i + 1], noise_sampler(sigmas[i], sigmas[i + 1]), x)
786
+ return x
787
+
788
+
789
+
790
+ @torch.no_grad()
791
+ def sample_heunpp2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
792
+ # From MIT licensed: https://github.com/Carzit/sd-webui-samplers-scheduler/
793
+ extra_args = {} if extra_args is None else extra_args
794
+ s_in = x.new_ones([x.shape[0]])
795
+ s_end = sigmas[-1]
796
+ for i in trange(len(sigmas) - 1, disable=disable):
797
+ gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
798
+ eps = torch.randn_like(x) * s_noise
799
+ sigma_hat = sigmas[i] * (gamma + 1)
800
+ if gamma > 0:
801
+ x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
802
+ denoised = model(x, sigma_hat * s_in, **extra_args)
803
+ d = to_d(x, sigma_hat, denoised)
804
+ if callback is not None:
805
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
806
+ dt = sigmas[i + 1] - sigma_hat
807
+ if sigmas[i + 1] == s_end:
808
+ # Euler method
809
+ x = x + d * dt
810
+ elif sigmas[i + 2] == s_end:
811
+
812
+ # Heun's method
813
+ x_2 = x + d * dt
814
+ denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
815
+ d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
816
+
817
+ w = 2 * sigmas[0]
818
+ w2 = sigmas[i+1]/w
819
+ w1 = 1 - w2
820
+
821
+ d_prime = d * w1 + d_2 * w2
822
+
823
+
824
+ x = x + d_prime * dt
825
+
826
+ else:
827
+ # Heun++
828
+ x_2 = x + d * dt
829
+ denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
830
+ d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
831
+ dt_2 = sigmas[i + 2] - sigmas[i + 1]
832
+
833
+ x_3 = x_2 + d_2 * dt_2
834
+ denoised_3 = model(x_3, sigmas[i + 2] * s_in, **extra_args)
835
+ d_3 = to_d(x_3, sigmas[i + 2], denoised_3)
836
+
837
+ w = 3 * sigmas[0]
838
+ w2 = sigmas[i + 1] / w
839
+ w3 = sigmas[i + 2] / w
840
+ w1 = 1 - w2 - w3
841
+
842
+ d_prime = w1 * d + w2 * d_2 + w3 * d_3
843
+ x = x + d_prime * dt
844
+ return x
845
+
846
+
847
+ #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
848
+ #under Apache 2 license
849
+ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
850
+ extra_args = {} if extra_args is None else extra_args
851
+ s_in = x.new_ones([x.shape[0]])
852
+
853
+ x_next = x
854
+
855
+ buffer_model = []
856
+ for i in trange(len(sigmas) - 1, disable=disable):
857
+ t_cur = sigmas[i]
858
+ t_next = sigmas[i + 1]
859
+
860
+ x_cur = x_next
861
+
862
+ denoised = model(x_cur, t_cur * s_in, **extra_args)
863
+ if callback is not None:
864
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
865
+
866
+ d_cur = (x_cur - denoised) / t_cur
867
+
868
+ order = min(max_order, i+1)
869
+ if order == 1: # First Euler step.
870
+ x_next = x_cur + (t_next - t_cur) * d_cur
871
+ elif order == 2: # Use one history point.
872
+ x_next = x_cur + (t_next - t_cur) * (3 * d_cur - buffer_model[-1]) / 2
873
+ elif order == 3: # Use two history points.
874
+ x_next = x_cur + (t_next - t_cur) * (23 * d_cur - 16 * buffer_model[-1] + 5 * buffer_model[-2]) / 12
875
+ elif order == 4: # Use three history points.
876
+ x_next = x_cur + (t_next - t_cur) * (55 * d_cur - 59 * buffer_model[-1] + 37 * buffer_model[-2] - 9 * buffer_model[-3]) / 24
877
+
878
+ if len(buffer_model) == max_order - 1:
879
+ for k in range(max_order - 2):
880
+ buffer_model[k] = buffer_model[k+1]
881
+ buffer_model[-1] = d_cur
882
+ else:
883
+ buffer_model.append(d_cur)
884
+
885
+ return x_next
886
+
887
+ #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
888
+ #under Apache 2 license
889
+ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
890
+ extra_args = {} if extra_args is None else extra_args
891
+ s_in = x.new_ones([x.shape[0]])
892
+
893
+ x_next = x
894
+ t_steps = sigmas
895
+
896
+ buffer_model = []
897
+ for i in trange(len(sigmas) - 1, disable=disable):
898
+ t_cur = sigmas[i]
899
+ t_next = sigmas[i + 1]
900
+
901
+ x_cur = x_next
902
+
903
+ denoised = model(x_cur, t_cur * s_in, **extra_args)
904
+ if callback is not None:
905
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
906
+
907
+ d_cur = (x_cur - denoised) / t_cur
908
+
909
+ order = min(max_order, i+1)
910
+ if order == 1: # First Euler step.
911
+ x_next = x_cur + (t_next - t_cur) * d_cur
912
+ elif order == 2: # Use one history point.
913
+ h_n = (t_next - t_cur)
914
+ h_n_1 = (t_cur - t_steps[i-1])
915
+ coeff1 = (2 + (h_n / h_n_1)) / 2
916
+ coeff2 = -(h_n / h_n_1) / 2
917
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1])
918
+ elif order == 3: # Use two history points.
919
+ h_n = (t_next - t_cur)
920
+ h_n_1 = (t_cur - t_steps[i-1])
921
+ h_n_2 = (t_steps[i-1] - t_steps[i-2])
922
+ temp = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
923
+ coeff1 = (2 + (h_n / h_n_1)) / 2 + temp
924
+ coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp
925
+ coeff3 = temp * h_n_1 / h_n_2
926
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2])
927
+ elif order == 4: # Use three history points.
928
+ h_n = (t_next - t_cur)
929
+ h_n_1 = (t_cur - t_steps[i-1])
930
+ h_n_2 = (t_steps[i-1] - t_steps[i-2])
931
+ h_n_3 = (t_steps[i-2] - t_steps[i-3])
932
+ temp1 = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
933
+ temp2 = ((1 - h_n / (3 * (h_n + h_n_1))) / 2 + (1 - h_n / (2 * (h_n + h_n_1))) * h_n / (6 * (h_n + h_n_1 + h_n_2))) \
934
+ * (h_n * (h_n + h_n_1) * (h_n + h_n_1 + h_n_2)) / (h_n_1 * (h_n_1 + h_n_2) * (h_n_1 + h_n_2 + h_n_3))
935
+ coeff1 = (2 + (h_n / h_n_1)) / 2 + temp1 + temp2
936
+ coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp1 - (1 + (h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3)))) * temp2
937
+ coeff3 = temp1 * h_n_1 / h_n_2 + ((h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * (1 + h_n_2 / h_n_3)) * temp2
938
+ coeff4 = -temp2 * (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * h_n_1 / h_n_2
939
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2] + coeff4 * buffer_model[-3])
940
+
941
+ if len(buffer_model) == max_order - 1:
942
+ for k in range(max_order - 2):
943
+ buffer_model[k] = buffer_model[k+1]
944
+ buffer_model[-1] = d_cur.detach()
945
+ else:
946
+ buffer_model.append(d_cur.detach())
947
+
948
+ return x_next
949
+
950
+ #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
951
+ #under Apache 2 license
952
+ @torch.no_grad()
953
+ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=3, deis_mode='tab'):
954
+ extra_args = {} if extra_args is None else extra_args
955
+ s_in = x.new_ones([x.shape[0]])
956
+
957
+ x_next = x
958
+ t_steps = sigmas
959
+
960
+ coeff_list = deis.get_deis_coeff_list(t_steps, max_order, deis_mode=deis_mode)
961
+
962
+ buffer_model = []
963
+ for i in trange(len(sigmas) - 1, disable=disable):
964
+ t_cur = sigmas[i]
965
+ t_next = sigmas[i + 1]
966
+
967
+ x_cur = x_next
968
+
969
+ denoised = model(x_cur, t_cur * s_in, **extra_args)
970
+ if callback is not None:
971
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
972
+
973
+ d_cur = (x_cur - denoised) / t_cur
974
+
975
+ order = min(max_order, i+1)
976
+ if t_next <= 0:
977
+ order = 1
978
+
979
+ if order == 1: # First Euler step.
980
+ x_next = x_cur + (t_next - t_cur) * d_cur
981
+ elif order == 2: # Use one history point.
982
+ coeff_cur, coeff_prev1 = coeff_list[i]
983
+ x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1]
984
+ elif order == 3: # Use two history points.
985
+ coeff_cur, coeff_prev1, coeff_prev2 = coeff_list[i]
986
+ x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1] + coeff_prev2 * buffer_model[-2]
987
+ elif order == 4: # Use three history points.
988
+ coeff_cur, coeff_prev1, coeff_prev2, coeff_prev3 = coeff_list[i]
989
+ x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1] + coeff_prev2 * buffer_model[-2] + coeff_prev3 * buffer_model[-3]
990
+
991
+ if len(buffer_model) == max_order - 1:
992
+ for k in range(max_order - 2):
993
+ buffer_model[k] = buffer_model[k+1]
994
+ buffer_model[-1] = d_cur.detach()
995
+ else:
996
+ buffer_model.append(d_cur.detach())
997
+
998
+ return x_next
999
+
1000
+ @torch.no_grad()
1001
+ def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
1002
+ extra_args = {} if extra_args is None else extra_args
1003
+
1004
+ temp = [0]
1005
+ def post_cfg_function(args):
1006
+ temp[0] = args["uncond_denoised"]
1007
+ return args["denoised"]
1008
+
1009
+ model_options = extra_args.get("model_options", {}).copy()
1010
+ extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
1011
+
1012
+ s_in = x.new_ones([x.shape[0]])
1013
+ for i in trange(len(sigmas) - 1, disable=disable):
1014
+ sigma_hat = sigmas[i]
1015
+ denoised = model(x, sigma_hat * s_in, **extra_args)
1016
+ d = to_d(x, sigma_hat, temp[0])
1017
+ if callback is not None:
1018
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
1019
+ dt = sigmas[i + 1] - sigma_hat
1020
+ # Euler method
1021
+ x = denoised + d * sigmas[i + 1]
1022
+ return x
1023
+
1024
+ @torch.no_grad()
1025
+ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
1026
+ """Ancestral sampling with Euler method steps."""
1027
+ extra_args = {} if extra_args is None else extra_args
1028
+ noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
1029
+
1030
+ temp = [0]
1031
+ def post_cfg_function(args):
1032
+ temp[0] = args["uncond_denoised"]
1033
+ return args["denoised"]
1034
+
1035
+ model_options = extra_args.get("model_options", {}).copy()
1036
+ extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
1037
+
1038
+ s_in = x.new_ones([x.shape[0]])
1039
+ for i in trange(len(sigmas) - 1, disable=disable):
1040
+ denoised = model(x, sigmas[i] * s_in, **extra_args)
1041
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
1042
+ if callback is not None:
1043
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
1044
+ d = to_d(x, sigmas[i], temp[0])
1045
+ # Euler method
1046
+ dt = sigma_down - sigmas[i]
1047
+ x = denoised + d * sigma_down
1048
+ if sigmas[i + 1] > 0:
1049
+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
1050
+ return x
ComfyUI/comfy/k_diffusion/utils.py ADDED
@@ -0,0 +1,313 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import contextmanager
2
+ import hashlib
3
+ import math
4
+ from pathlib import Path
5
+ import shutil
6
+ import urllib
7
+ import warnings
8
+
9
+ from PIL import Image
10
+ import torch
11
+ from torch import nn, optim
12
+ from torch.utils import data
13
+
14
+
15
+ def hf_datasets_augs_helper(examples, transform, image_key, mode='RGB'):
16
+ """Apply passed in transforms for HuggingFace Datasets."""
17
+ images = [transform(image.convert(mode)) for image in examples[image_key]]
18
+ return {image_key: images}
19
+
20
+
21
+ def append_dims(x, target_dims):
22
+ """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
23
+ dims_to_append = target_dims - x.ndim
24
+ if dims_to_append < 0:
25
+ raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
26
+ expanded = x[(...,) + (None,) * dims_to_append]
27
+ # MPS will get inf values if it tries to index into the new axes, but detaching fixes this.
28
+ # https://github.com/pytorch/pytorch/issues/84364
29
+ return expanded.detach().clone() if expanded.device.type == 'mps' else expanded
30
+
31
+
32
+ def n_params(module):
33
+ """Returns the number of trainable parameters in a module."""
34
+ return sum(p.numel() for p in module.parameters())
35
+
36
+
37
+ def download_file(path, url, digest=None):
38
+ """Downloads a file if it does not exist, optionally checking its SHA-256 hash."""
39
+ path = Path(path)
40
+ path.parent.mkdir(parents=True, exist_ok=True)
41
+ if not path.exists():
42
+ with urllib.request.urlopen(url) as response, open(path, 'wb') as f:
43
+ shutil.copyfileobj(response, f)
44
+ if digest is not None:
45
+ file_digest = hashlib.sha256(open(path, 'rb').read()).hexdigest()
46
+ if digest != file_digest:
47
+ raise OSError(f'hash of {path} (url: {url}) failed to validate')
48
+ return path
49
+
50
+
51
+ @contextmanager
52
+ def train_mode(model, mode=True):
53
+ """A context manager that places a model into training mode and restores
54
+ the previous mode on exit."""
55
+ modes = [module.training for module in model.modules()]
56
+ try:
57
+ yield model.train(mode)
58
+ finally:
59
+ for i, module in enumerate(model.modules()):
60
+ module.training = modes[i]
61
+
62
+
63
+ def eval_mode(model):
64
+ """A context manager that places a model into evaluation mode and restores
65
+ the previous mode on exit."""
66
+ return train_mode(model, False)
67
+
68
+
69
+ @torch.no_grad()
70
+ def ema_update(model, averaged_model, decay):
71
+ """Incorporates updated model parameters into an exponential moving averaged
72
+ version of a model. It should be called after each optimizer step."""
73
+ model_params = dict(model.named_parameters())
74
+ averaged_params = dict(averaged_model.named_parameters())
75
+ assert model_params.keys() == averaged_params.keys()
76
+
77
+ for name, param in model_params.items():
78
+ averaged_params[name].mul_(decay).add_(param, alpha=1 - decay)
79
+
80
+ model_buffers = dict(model.named_buffers())
81
+ averaged_buffers = dict(averaged_model.named_buffers())
82
+ assert model_buffers.keys() == averaged_buffers.keys()
83
+
84
+ for name, buf in model_buffers.items():
85
+ averaged_buffers[name].copy_(buf)
86
+
87
+
88
+ class EMAWarmup:
89
+ """Implements an EMA warmup using an inverse decay schedule.
90
+ If inv_gamma=1 and power=1, implements a simple average. inv_gamma=1, power=2/3 are
91
+ good values for models you plan to train for a million or more steps (reaches decay
92
+ factor 0.999 at 31.6K steps, 0.9999 at 1M steps), inv_gamma=1, power=3/4 for models
93
+ you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 at
94
+ 215.4k steps).
95
+ Args:
96
+ inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
97
+ power (float): Exponential factor of EMA warmup. Default: 1.
98
+ min_value (float): The minimum EMA decay rate. Default: 0.
99
+ max_value (float): The maximum EMA decay rate. Default: 1.
100
+ start_at (int): The epoch to start averaging at. Default: 0.
101
+ last_epoch (int): The index of last epoch. Default: 0.
102
+ """
103
+
104
+ def __init__(self, inv_gamma=1., power=1., min_value=0., max_value=1., start_at=0,
105
+ last_epoch=0):
106
+ self.inv_gamma = inv_gamma
107
+ self.power = power
108
+ self.min_value = min_value
109
+ self.max_value = max_value
110
+ self.start_at = start_at
111
+ self.last_epoch = last_epoch
112
+
113
+ def state_dict(self):
114
+ """Returns the state of the class as a :class:`dict`."""
115
+ return dict(self.__dict__.items())
116
+
117
+ def load_state_dict(self, state_dict):
118
+ """Loads the class's state.
119
+ Args:
120
+ state_dict (dict): scaler state. Should be an object returned
121
+ from a call to :meth:`state_dict`.
122
+ """
123
+ self.__dict__.update(state_dict)
124
+
125
+ def get_value(self):
126
+ """Gets the current EMA decay rate."""
127
+ epoch = max(0, self.last_epoch - self.start_at)
128
+ value = 1 - (1 + epoch / self.inv_gamma) ** -self.power
129
+ return 0. if epoch < 0 else min(self.max_value, max(self.min_value, value))
130
+
131
+ def step(self):
132
+ """Updates the step count."""
133
+ self.last_epoch += 1
134
+
135
+
136
+ class InverseLR(optim.lr_scheduler._LRScheduler):
137
+ """Implements an inverse decay learning rate schedule with an optional exponential
138
+ warmup. When last_epoch=-1, sets initial lr as lr.
139
+ inv_gamma is the number of steps/epochs required for the learning rate to decay to
140
+ (1 / 2)**power of its original value.
141
+ Args:
142
+ optimizer (Optimizer): Wrapped optimizer.
143
+ inv_gamma (float): Inverse multiplicative factor of learning rate decay. Default: 1.
144
+ power (float): Exponential factor of learning rate decay. Default: 1.
145
+ warmup (float): Exponential warmup factor (0 <= warmup < 1, 0 to disable)
146
+ Default: 0.
147
+ min_lr (float): The minimum learning rate. Default: 0.
148
+ last_epoch (int): The index of last epoch. Default: -1.
149
+ verbose (bool): If ``True``, prints a message to stdout for
150
+ each update. Default: ``False``.
151
+ """
152
+
153
+ def __init__(self, optimizer, inv_gamma=1., power=1., warmup=0., min_lr=0.,
154
+ last_epoch=-1, verbose=False):
155
+ self.inv_gamma = inv_gamma
156
+ self.power = power
157
+ if not 0. <= warmup < 1:
158
+ raise ValueError('Invalid value for warmup')
159
+ self.warmup = warmup
160
+ self.min_lr = min_lr
161
+ super().__init__(optimizer, last_epoch, verbose)
162
+
163
+ def get_lr(self):
164
+ if not self._get_lr_called_within_step:
165
+ warnings.warn("To get the last learning rate computed by the scheduler, "
166
+ "please use `get_last_lr()`.")
167
+
168
+ return self._get_closed_form_lr()
169
+
170
+ def _get_closed_form_lr(self):
171
+ warmup = 1 - self.warmup ** (self.last_epoch + 1)
172
+ lr_mult = (1 + self.last_epoch / self.inv_gamma) ** -self.power
173
+ return [warmup * max(self.min_lr, base_lr * lr_mult)
174
+ for base_lr in self.base_lrs]
175
+
176
+
177
+ class ExponentialLR(optim.lr_scheduler._LRScheduler):
178
+ """Implements an exponential learning rate schedule with an optional exponential
179
+ warmup. When last_epoch=-1, sets initial lr as lr. Decays the learning rate
180
+ continuously by decay (default 0.5) every num_steps steps.
181
+ Args:
182
+ optimizer (Optimizer): Wrapped optimizer.
183
+ num_steps (float): The number of steps to decay the learning rate by decay in.
184
+ decay (float): The factor by which to decay the learning rate every num_steps
185
+ steps. Default: 0.5.
186
+ warmup (float): Exponential warmup factor (0 <= warmup < 1, 0 to disable)
187
+ Default: 0.
188
+ min_lr (float): The minimum learning rate. Default: 0.
189
+ last_epoch (int): The index of last epoch. Default: -1.
190
+ verbose (bool): If ``True``, prints a message to stdout for
191
+ each update. Default: ``False``.
192
+ """
193
+
194
+ def __init__(self, optimizer, num_steps, decay=0.5, warmup=0., min_lr=0.,
195
+ last_epoch=-1, verbose=False):
196
+ self.num_steps = num_steps
197
+ self.decay = decay
198
+ if not 0. <= warmup < 1:
199
+ raise ValueError('Invalid value for warmup')
200
+ self.warmup = warmup
201
+ self.min_lr = min_lr
202
+ super().__init__(optimizer, last_epoch, verbose)
203
+
204
+ def get_lr(self):
205
+ if not self._get_lr_called_within_step:
206
+ warnings.warn("To get the last learning rate computed by the scheduler, "
207
+ "please use `get_last_lr()`.")
208
+
209
+ return self._get_closed_form_lr()
210
+
211
+ def _get_closed_form_lr(self):
212
+ warmup = 1 - self.warmup ** (self.last_epoch + 1)
213
+ lr_mult = (self.decay ** (1 / self.num_steps)) ** self.last_epoch
214
+ return [warmup * max(self.min_lr, base_lr * lr_mult)
215
+ for base_lr in self.base_lrs]
216
+
217
+
218
+ def rand_log_normal(shape, loc=0., scale=1., device='cpu', dtype=torch.float32):
219
+ """Draws samples from an lognormal distribution."""
220
+ return (torch.randn(shape, device=device, dtype=dtype) * scale + loc).exp()
221
+
222
+
223
+ def rand_log_logistic(shape, loc=0., scale=1., min_value=0., max_value=float('inf'), device='cpu', dtype=torch.float32):
224
+ """Draws samples from an optionally truncated log-logistic distribution."""
225
+ min_value = torch.as_tensor(min_value, device=device, dtype=torch.float64)
226
+ max_value = torch.as_tensor(max_value, device=device, dtype=torch.float64)
227
+ min_cdf = min_value.log().sub(loc).div(scale).sigmoid()
228
+ max_cdf = max_value.log().sub(loc).div(scale).sigmoid()
229
+ u = torch.rand(shape, device=device, dtype=torch.float64) * (max_cdf - min_cdf) + min_cdf
230
+ return u.logit().mul(scale).add(loc).exp().to(dtype)
231
+
232
+
233
+ def rand_log_uniform(shape, min_value, max_value, device='cpu', dtype=torch.float32):
234
+ """Draws samples from an log-uniform distribution."""
235
+ min_value = math.log(min_value)
236
+ max_value = math.log(max_value)
237
+ return (torch.rand(shape, device=device, dtype=dtype) * (max_value - min_value) + min_value).exp()
238
+
239
+
240
+ def rand_v_diffusion(shape, sigma_data=1., min_value=0., max_value=float('inf'), device='cpu', dtype=torch.float32):
241
+ """Draws samples from a truncated v-diffusion training timestep distribution."""
242
+ min_cdf = math.atan(min_value / sigma_data) * 2 / math.pi
243
+ max_cdf = math.atan(max_value / sigma_data) * 2 / math.pi
244
+ u = torch.rand(shape, device=device, dtype=dtype) * (max_cdf - min_cdf) + min_cdf
245
+ return torch.tan(u * math.pi / 2) * sigma_data
246
+
247
+
248
+ def rand_split_log_normal(shape, loc, scale_1, scale_2, device='cpu', dtype=torch.float32):
249
+ """Draws samples from a split lognormal distribution."""
250
+ n = torch.randn(shape, device=device, dtype=dtype).abs()
251
+ u = torch.rand(shape, device=device, dtype=dtype)
252
+ n_left = n * -scale_1 + loc
253
+ n_right = n * scale_2 + loc
254
+ ratio = scale_1 / (scale_1 + scale_2)
255
+ return torch.where(u < ratio, n_left, n_right).exp()
256
+
257
+
258
+ class FolderOfImages(data.Dataset):
259
+ """Recursively finds all images in a directory. It does not support
260
+ classes/targets."""
261
+
262
+ IMG_EXTENSIONS = {'.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', '.webp'}
263
+
264
+ def __init__(self, root, transform=None):
265
+ super().__init__()
266
+ self.root = Path(root)
267
+ self.transform = nn.Identity() if transform is None else transform
268
+ self.paths = sorted(path for path in self.root.rglob('*') if path.suffix.lower() in self.IMG_EXTENSIONS)
269
+
270
+ def __repr__(self):
271
+ return f'FolderOfImages(root="{self.root}", len: {len(self)})'
272
+
273
+ def __len__(self):
274
+ return len(self.paths)
275
+
276
+ def __getitem__(self, key):
277
+ path = self.paths[key]
278
+ with open(path, 'rb') as f:
279
+ image = Image.open(f).convert('RGB')
280
+ image = self.transform(image)
281
+ return image,
282
+
283
+
284
+ class CSVLogger:
285
+ def __init__(self, filename, columns):
286
+ self.filename = Path(filename)
287
+ self.columns = columns
288
+ if self.filename.exists():
289
+ self.file = open(self.filename, 'a')
290
+ else:
291
+ self.file = open(self.filename, 'w')
292
+ self.write(*self.columns)
293
+
294
+ def write(self, *args):
295
+ print(*args, sep=',', file=self.file, flush=True)
296
+
297
+
298
+ @contextmanager
299
+ def tf32_mode(cudnn=None, matmul=None):
300
+ """A context manager that sets whether TF32 is allowed on cuDNN or matmul."""
301
+ cudnn_old = torch.backends.cudnn.allow_tf32
302
+ matmul_old = torch.backends.cuda.matmul.allow_tf32
303
+ try:
304
+ if cudnn is not None:
305
+ torch.backends.cudnn.allow_tf32 = cudnn
306
+ if matmul is not None:
307
+ torch.backends.cuda.matmul.allow_tf32 = matmul
308
+ yield
309
+ finally:
310
+ if cudnn is not None:
311
+ torch.backends.cudnn.allow_tf32 = cudnn_old
312
+ if matmul is not None:
313
+ torch.backends.cuda.matmul.allow_tf32 = matmul_old
ComfyUI/comfy/latent_formats.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ class LatentFormat:
4
+ scale_factor = 1.0
5
+ latent_channels = 4
6
+ latent_rgb_factors = None
7
+ taesd_decoder_name = None
8
+
9
+ def process_in(self, latent):
10
+ return latent * self.scale_factor
11
+
12
+ def process_out(self, latent):
13
+ return latent / self.scale_factor
14
+
15
+ class SD15(LatentFormat):
16
+ def __init__(self, scale_factor=0.18215):
17
+ self.scale_factor = scale_factor
18
+ self.latent_rgb_factors = [
19
+ # R G B
20
+ [ 0.3512, 0.2297, 0.3227],
21
+ [ 0.3250, 0.4974, 0.2350],
22
+ [-0.2829, 0.1762, 0.2721],
23
+ [-0.2120, -0.2616, -0.7177]
24
+ ]
25
+ self.taesd_decoder_name = "taesd_decoder"
26
+
27
+ class SDXL(LatentFormat):
28
+ scale_factor = 0.13025
29
+
30
+ def __init__(self):
31
+ self.latent_rgb_factors = [
32
+ # R G B
33
+ [ 0.3920, 0.4054, 0.4549],
34
+ [-0.2634, -0.0196, 0.0653],
35
+ [ 0.0568, 0.1687, -0.0755],
36
+ [-0.3112, -0.2359, -0.2076]
37
+ ]
38
+ self.taesd_decoder_name = "taesdxl_decoder"
39
+
40
+ class SDXL_Playground_2_5(LatentFormat):
41
+ def __init__(self):
42
+ self.scale_factor = 0.5
43
+ self.latents_mean = torch.tensor([-1.6574, 1.886, -1.383, 2.5155]).view(1, 4, 1, 1)
44
+ self.latents_std = torch.tensor([8.4927, 5.9022, 6.5498, 5.2299]).view(1, 4, 1, 1)
45
+
46
+ self.latent_rgb_factors = [
47
+ # R G B
48
+ [ 0.3920, 0.4054, 0.4549],
49
+ [-0.2634, -0.0196, 0.0653],
50
+ [ 0.0568, 0.1687, -0.0755],
51
+ [-0.3112, -0.2359, -0.2076]
52
+ ]
53
+ self.taesd_decoder_name = "taesdxl_decoder"
54
+
55
+ def process_in(self, latent):
56
+ latents_mean = self.latents_mean.to(latent.device, latent.dtype)
57
+ latents_std = self.latents_std.to(latent.device, latent.dtype)
58
+ return (latent - latents_mean) * self.scale_factor / latents_std
59
+
60
+ def process_out(self, latent):
61
+ latents_mean = self.latents_mean.to(latent.device, latent.dtype)
62
+ latents_std = self.latents_std.to(latent.device, latent.dtype)
63
+ return latent * latents_std / self.scale_factor + latents_mean
64
+
65
+
66
+ class SD_X4(LatentFormat):
67
+ def __init__(self):
68
+ self.scale_factor = 0.08333
69
+ self.latent_rgb_factors = [
70
+ [-0.2340, -0.3863, -0.3257],
71
+ [ 0.0994, 0.0885, -0.0908],
72
+ [-0.2833, -0.2349, -0.3741],
73
+ [ 0.2523, -0.0055, -0.1651]
74
+ ]
75
+
76
+ class SC_Prior(LatentFormat):
77
+ latent_channels = 16
78
+ def __init__(self):
79
+ self.scale_factor = 1.0
80
+ self.latent_rgb_factors = [
81
+ [-0.0326, -0.0204, -0.0127],
82
+ [-0.1592, -0.0427, 0.0216],
83
+ [ 0.0873, 0.0638, -0.0020],
84
+ [-0.0602, 0.0442, 0.1304],
85
+ [ 0.0800, -0.0313, -0.1796],
86
+ [-0.0810, -0.0638, -0.1581],
87
+ [ 0.1791, 0.1180, 0.0967],
88
+ [ 0.0740, 0.1416, 0.0432],
89
+ [-0.1745, -0.1888, -0.1373],
90
+ [ 0.2412, 0.1577, 0.0928],
91
+ [ 0.1908, 0.0998, 0.0682],
92
+ [ 0.0209, 0.0365, -0.0092],
93
+ [ 0.0448, -0.0650, -0.1728],
94
+ [-0.1658, -0.1045, -0.1308],
95
+ [ 0.0542, 0.1545, 0.1325],
96
+ [-0.0352, -0.1672, -0.2541]
97
+ ]
98
+
99
+ class SC_B(LatentFormat):
100
+ def __init__(self):
101
+ self.scale_factor = 1.0 / 0.43
102
+ self.latent_rgb_factors = [
103
+ [ 0.1121, 0.2006, 0.1023],
104
+ [-0.2093, -0.0222, -0.0195],
105
+ [-0.3087, -0.1535, 0.0366],
106
+ [ 0.0290, -0.1574, -0.4078]
107
+ ]
108
+
109
+ class SD3(LatentFormat):
110
+ latent_channels = 16
111
+ def __init__(self):
112
+ self.scale_factor = 1.5305
113
+ self.shift_factor = 0.0609
114
+ self.latent_rgb_factors = [
115
+ [-0.0645, 0.0177, 0.1052],
116
+ [ 0.0028, 0.0312, 0.0650],
117
+ [ 0.1848, 0.0762, 0.0360],
118
+ [ 0.0944, 0.0360, 0.0889],
119
+ [ 0.0897, 0.0506, -0.0364],
120
+ [-0.0020, 0.1203, 0.0284],
121
+ [ 0.0855, 0.0118, 0.0283],
122
+ [-0.0539, 0.0658, 0.1047],
123
+ [-0.0057, 0.0116, 0.0700],
124
+ [-0.0412, 0.0281, -0.0039],
125
+ [ 0.1106, 0.1171, 0.1220],
126
+ [-0.0248, 0.0682, -0.0481],
127
+ [ 0.0815, 0.0846, 0.1207],
128
+ [-0.0120, -0.0055, -0.0867],
129
+ [-0.0749, -0.0634, -0.0456],
130
+ [-0.1418, -0.1457, -0.1259]
131
+ ]
132
+ self.taesd_decoder_name = "taesd3_decoder"
133
+
134
+ def process_in(self, latent):
135
+ return (latent - self.shift_factor) * self.scale_factor
136
+
137
+ def process_out(self, latent):
138
+ return (latent / self.scale_factor) + self.shift_factor
139
+
140
+ class StableAudio1(LatentFormat):
141
+ latent_channels = 64
142
+
143
+ class Flux(SD3):
144
+ def __init__(self):
145
+ self.scale_factor = 0.3611
146
+ self.shift_factor = 0.1159
147
+ self.latent_rgb_factors =[
148
+ [-0.0404, 0.0159, 0.0609],
149
+ [ 0.0043, 0.0298, 0.0850],
150
+ [ 0.0328, -0.0749, -0.0503],
151
+ [-0.0245, 0.0085, 0.0549],
152
+ [ 0.0966, 0.0894, 0.0530],
153
+ [ 0.0035, 0.0399, 0.0123],
154
+ [ 0.0583, 0.1184, 0.1262],
155
+ [-0.0191, -0.0206, -0.0306],
156
+ [-0.0324, 0.0055, 0.1001],
157
+ [ 0.0955, 0.0659, -0.0545],
158
+ [-0.0504, 0.0231, -0.0013],
159
+ [ 0.0500, -0.0008, -0.0088],
160
+ [ 0.0982, 0.0941, 0.0976],
161
+ [-0.1233, -0.0280, -0.0897],
162
+ [-0.0005, -0.0530, -0.0020],
163
+ [-0.1273, -0.0932, -0.0680]
164
+ ]
165
+
166
+ def process_in(self, latent):
167
+ return (latent - self.shift_factor) * self.scale_factor
168
+
169
+ def process_out(self, latent):
170
+ return (latent / self.scale_factor) + self.shift_factor
ComfyUI/comfy/ldm/audio/autoencoder.py ADDED
@@ -0,0 +1,282 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # code adapted from: https://github.com/Stability-AI/stable-audio-tools
2
+
3
+ import torch
4
+ from torch import nn
5
+ from typing import Literal, Dict, Any
6
+ import math
7
+ import comfy.ops
8
+ ops = comfy.ops.disable_weight_init
9
+
10
+ def vae_sample(mean, scale):
11
+ stdev = nn.functional.softplus(scale) + 1e-4
12
+ var = stdev * stdev
13
+ logvar = torch.log(var)
14
+ latents = torch.randn_like(mean) * stdev + mean
15
+
16
+ kl = (mean * mean + var - logvar - 1).sum(1).mean()
17
+
18
+ return latents, kl
19
+
20
+ class VAEBottleneck(nn.Module):
21
+ def __init__(self):
22
+ super().__init__()
23
+ self.is_discrete = False
24
+
25
+ def encode(self, x, return_info=False, **kwargs):
26
+ info = {}
27
+
28
+ mean, scale = x.chunk(2, dim=1)
29
+
30
+ x, kl = vae_sample(mean, scale)
31
+
32
+ info["kl"] = kl
33
+
34
+ if return_info:
35
+ return x, info
36
+ else:
37
+ return x
38
+
39
+ def decode(self, x):
40
+ return x
41
+
42
+
43
+ def snake_beta(x, alpha, beta):
44
+ return x + (1.0 / (beta + 0.000000001)) * pow(torch.sin(x * alpha), 2)
45
+
46
+ # Adapted from https://github.com/NVIDIA/BigVGAN/blob/main/activations.py under MIT license
47
+ class SnakeBeta(nn.Module):
48
+
49
+ def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=True):
50
+ super(SnakeBeta, self).__init__()
51
+ self.in_features = in_features
52
+
53
+ # initialize alpha
54
+ self.alpha_logscale = alpha_logscale
55
+ if self.alpha_logscale: # log scale alphas initialized to zeros
56
+ self.alpha = nn.Parameter(torch.zeros(in_features) * alpha)
57
+ self.beta = nn.Parameter(torch.zeros(in_features) * alpha)
58
+ else: # linear scale alphas initialized to ones
59
+ self.alpha = nn.Parameter(torch.ones(in_features) * alpha)
60
+ self.beta = nn.Parameter(torch.ones(in_features) * alpha)
61
+
62
+ # self.alpha.requires_grad = alpha_trainable
63
+ # self.beta.requires_grad = alpha_trainable
64
+
65
+ self.no_div_by_zero = 0.000000001
66
+
67
+ def forward(self, x):
68
+ alpha = self.alpha.unsqueeze(0).unsqueeze(-1).to(x.device) # line up with x to [B, C, T]
69
+ beta = self.beta.unsqueeze(0).unsqueeze(-1).to(x.device)
70
+ if self.alpha_logscale:
71
+ alpha = torch.exp(alpha)
72
+ beta = torch.exp(beta)
73
+ x = snake_beta(x, alpha, beta)
74
+
75
+ return x
76
+
77
+ def WNConv1d(*args, **kwargs):
78
+ try:
79
+ return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs))
80
+ except:
81
+ return torch.nn.utils.weight_norm(ops.Conv1d(*args, **kwargs)) #support pytorch 2.1 and older
82
+
83
+ def WNConvTranspose1d(*args, **kwargs):
84
+ try:
85
+ return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs))
86
+ except:
87
+ return torch.nn.utils.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) #support pytorch 2.1 and older
88
+
89
+ def get_activation(activation: Literal["elu", "snake", "none"], antialias=False, channels=None) -> nn.Module:
90
+ if activation == "elu":
91
+ act = torch.nn.ELU()
92
+ elif activation == "snake":
93
+ act = SnakeBeta(channels)
94
+ elif activation == "none":
95
+ act = torch.nn.Identity()
96
+ else:
97
+ raise ValueError(f"Unknown activation {activation}")
98
+
99
+ if antialias:
100
+ act = Activation1d(act)
101
+
102
+ return act
103
+
104
+
105
+ class ResidualUnit(nn.Module):
106
+ def __init__(self, in_channels, out_channels, dilation, use_snake=False, antialias_activation=False):
107
+ super().__init__()
108
+
109
+ self.dilation = dilation
110
+
111
+ padding = (dilation * (7-1)) // 2
112
+
113
+ self.layers = nn.Sequential(
114
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=out_channels),
115
+ WNConv1d(in_channels=in_channels, out_channels=out_channels,
116
+ kernel_size=7, dilation=dilation, padding=padding),
117
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=out_channels),
118
+ WNConv1d(in_channels=out_channels, out_channels=out_channels,
119
+ kernel_size=1)
120
+ )
121
+
122
+ def forward(self, x):
123
+ res = x
124
+
125
+ #x = checkpoint(self.layers, x)
126
+ x = self.layers(x)
127
+
128
+ return x + res
129
+
130
+ class EncoderBlock(nn.Module):
131
+ def __init__(self, in_channels, out_channels, stride, use_snake=False, antialias_activation=False):
132
+ super().__init__()
133
+
134
+ self.layers = nn.Sequential(
135
+ ResidualUnit(in_channels=in_channels,
136
+ out_channels=in_channels, dilation=1, use_snake=use_snake),
137
+ ResidualUnit(in_channels=in_channels,
138
+ out_channels=in_channels, dilation=3, use_snake=use_snake),
139
+ ResidualUnit(in_channels=in_channels,
140
+ out_channels=in_channels, dilation=9, use_snake=use_snake),
141
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=in_channels),
142
+ WNConv1d(in_channels=in_channels, out_channels=out_channels,
143
+ kernel_size=2*stride, stride=stride, padding=math.ceil(stride/2)),
144
+ )
145
+
146
+ def forward(self, x):
147
+ return self.layers(x)
148
+
149
+ class DecoderBlock(nn.Module):
150
+ def __init__(self, in_channels, out_channels, stride, use_snake=False, antialias_activation=False, use_nearest_upsample=False):
151
+ super().__init__()
152
+
153
+ if use_nearest_upsample:
154
+ upsample_layer = nn.Sequential(
155
+ nn.Upsample(scale_factor=stride, mode="nearest"),
156
+ WNConv1d(in_channels=in_channels,
157
+ out_channels=out_channels,
158
+ kernel_size=2*stride,
159
+ stride=1,
160
+ bias=False,
161
+ padding='same')
162
+ )
163
+ else:
164
+ upsample_layer = WNConvTranspose1d(in_channels=in_channels,
165
+ out_channels=out_channels,
166
+ kernel_size=2*stride, stride=stride, padding=math.ceil(stride/2))
167
+
168
+ self.layers = nn.Sequential(
169
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=in_channels),
170
+ upsample_layer,
171
+ ResidualUnit(in_channels=out_channels, out_channels=out_channels,
172
+ dilation=1, use_snake=use_snake),
173
+ ResidualUnit(in_channels=out_channels, out_channels=out_channels,
174
+ dilation=3, use_snake=use_snake),
175
+ ResidualUnit(in_channels=out_channels, out_channels=out_channels,
176
+ dilation=9, use_snake=use_snake),
177
+ )
178
+
179
+ def forward(self, x):
180
+ return self.layers(x)
181
+
182
+ class OobleckEncoder(nn.Module):
183
+ def __init__(self,
184
+ in_channels=2,
185
+ channels=128,
186
+ latent_dim=32,
187
+ c_mults = [1, 2, 4, 8],
188
+ strides = [2, 4, 8, 8],
189
+ use_snake=False,
190
+ antialias_activation=False
191
+ ):
192
+ super().__init__()
193
+
194
+ c_mults = [1] + c_mults
195
+
196
+ self.depth = len(c_mults)
197
+
198
+ layers = [
199
+ WNConv1d(in_channels=in_channels, out_channels=c_mults[0] * channels, kernel_size=7, padding=3)
200
+ ]
201
+
202
+ for i in range(self.depth-1):
203
+ layers += [EncoderBlock(in_channels=c_mults[i]*channels, out_channels=c_mults[i+1]*channels, stride=strides[i], use_snake=use_snake)]
204
+
205
+ layers += [
206
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=c_mults[-1] * channels),
207
+ WNConv1d(in_channels=c_mults[-1]*channels, out_channels=latent_dim, kernel_size=3, padding=1)
208
+ ]
209
+
210
+ self.layers = nn.Sequential(*layers)
211
+
212
+ def forward(self, x):
213
+ return self.layers(x)
214
+
215
+
216
+ class OobleckDecoder(nn.Module):
217
+ def __init__(self,
218
+ out_channels=2,
219
+ channels=128,
220
+ latent_dim=32,
221
+ c_mults = [1, 2, 4, 8],
222
+ strides = [2, 4, 8, 8],
223
+ use_snake=False,
224
+ antialias_activation=False,
225
+ use_nearest_upsample=False,
226
+ final_tanh=True):
227
+ super().__init__()
228
+
229
+ c_mults = [1] + c_mults
230
+
231
+ self.depth = len(c_mults)
232
+
233
+ layers = [
234
+ WNConv1d(in_channels=latent_dim, out_channels=c_mults[-1]*channels, kernel_size=7, padding=3),
235
+ ]
236
+
237
+ for i in range(self.depth-1, 0, -1):
238
+ layers += [DecoderBlock(
239
+ in_channels=c_mults[i]*channels,
240
+ out_channels=c_mults[i-1]*channels,
241
+ stride=strides[i-1],
242
+ use_snake=use_snake,
243
+ antialias_activation=antialias_activation,
244
+ use_nearest_upsample=use_nearest_upsample
245
+ )
246
+ ]
247
+
248
+ layers += [
249
+ get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=c_mults[0] * channels),
250
+ WNConv1d(in_channels=c_mults[0] * channels, out_channels=out_channels, kernel_size=7, padding=3, bias=False),
251
+ nn.Tanh() if final_tanh else nn.Identity()
252
+ ]
253
+
254
+ self.layers = nn.Sequential(*layers)
255
+
256
+ def forward(self, x):
257
+ return self.layers(x)
258
+
259
+
260
+ class AudioOobleckVAE(nn.Module):
261
+ def __init__(self,
262
+ in_channels=2,
263
+ channels=128,
264
+ latent_dim=64,
265
+ c_mults = [1, 2, 4, 8, 16],
266
+ strides = [2, 4, 4, 8, 8],
267
+ use_snake=True,
268
+ antialias_activation=False,
269
+ use_nearest_upsample=False,
270
+ final_tanh=False):
271
+ super().__init__()
272
+ self.encoder = OobleckEncoder(in_channels, channels, latent_dim * 2, c_mults, strides, use_snake, antialias_activation)
273
+ self.decoder = OobleckDecoder(in_channels, channels, latent_dim, c_mults, strides, use_snake, antialias_activation,
274
+ use_nearest_upsample=use_nearest_upsample, final_tanh=final_tanh)
275
+ self.bottleneck = VAEBottleneck()
276
+
277
+ def encode(self, x):
278
+ return self.bottleneck.encode(self.encoder(x))
279
+
280
+ def decode(self, x):
281
+ return self.decoder(self.bottleneck.decode(x))
282
+
ComfyUI/comfy/ldm/audio/dit.py ADDED
@@ -0,0 +1,891 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # code adapted from: https://github.com/Stability-AI/stable-audio-tools
2
+
3
+ from comfy.ldm.modules.attention import optimized_attention
4
+ import typing as tp
5
+
6
+ import torch
7
+
8
+ from einops import rearrange
9
+ from torch import nn
10
+ from torch.nn import functional as F
11
+ import math
12
+ import comfy.ops
13
+
14
+ class FourierFeatures(nn.Module):
15
+ def __init__(self, in_features, out_features, std=1., dtype=None, device=None):
16
+ super().__init__()
17
+ assert out_features % 2 == 0
18
+ self.weight = nn.Parameter(torch.empty(
19
+ [out_features // 2, in_features], dtype=dtype, device=device))
20
+
21
+ def forward(self, input):
22
+ f = 2 * math.pi * input @ comfy.ops.cast_to_input(self.weight.T, input)
23
+ return torch.cat([f.cos(), f.sin()], dim=-1)
24
+
25
+ # norms
26
+ class LayerNorm(nn.Module):
27
+ def __init__(self, dim, bias=False, fix_scale=False, dtype=None, device=None):
28
+ """
29
+ bias-less layernorm has been shown to be more stable. most newer models have moved towards rmsnorm, also bias-less
30
+ """
31
+ super().__init__()
32
+
33
+ self.gamma = nn.Parameter(torch.empty(dim, dtype=dtype, device=device))
34
+
35
+ if bias:
36
+ self.beta = nn.Parameter(torch.empty(dim, dtype=dtype, device=device))
37
+ else:
38
+ self.beta = None
39
+
40
+ def forward(self, x):
41
+ beta = self.beta
42
+ if beta is not None:
43
+ beta = comfy.ops.cast_to_input(beta, x)
44
+ return F.layer_norm(x, x.shape[-1:], weight=comfy.ops.cast_to_input(self.gamma, x), bias=beta)
45
+
46
+ class GLU(nn.Module):
47
+ def __init__(
48
+ self,
49
+ dim_in,
50
+ dim_out,
51
+ activation,
52
+ use_conv = False,
53
+ conv_kernel_size = 3,
54
+ dtype=None,
55
+ device=None,
56
+ operations=None,
57
+ ):
58
+ super().__init__()
59
+ self.act = activation
60
+ self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device) if not use_conv else operations.Conv1d(dim_in, dim_out * 2, conv_kernel_size, padding = (conv_kernel_size // 2), dtype=dtype, device=device)
61
+ self.use_conv = use_conv
62
+
63
+ def forward(self, x):
64
+ if self.use_conv:
65
+ x = rearrange(x, 'b n d -> b d n')
66
+ x = self.proj(x)
67
+ x = rearrange(x, 'b d n -> b n d')
68
+ else:
69
+ x = self.proj(x)
70
+
71
+ x, gate = x.chunk(2, dim = -1)
72
+ return x * self.act(gate)
73
+
74
+ class AbsolutePositionalEmbedding(nn.Module):
75
+ def __init__(self, dim, max_seq_len):
76
+ super().__init__()
77
+ self.scale = dim ** -0.5
78
+ self.max_seq_len = max_seq_len
79
+ self.emb = nn.Embedding(max_seq_len, dim)
80
+
81
+ def forward(self, x, pos = None, seq_start_pos = None):
82
+ seq_len, device = x.shape[1], x.device
83
+ assert seq_len <= self.max_seq_len, f'you are passing in a sequence length of {seq_len} but your absolute positional embedding has a max sequence length of {self.max_seq_len}'
84
+
85
+ if pos is None:
86
+ pos = torch.arange(seq_len, device = device)
87
+
88
+ if seq_start_pos is not None:
89
+ pos = (pos - seq_start_pos[..., None]).clamp(min = 0)
90
+
91
+ pos_emb = self.emb(pos)
92
+ pos_emb = pos_emb * self.scale
93
+ return pos_emb
94
+
95
+ class ScaledSinusoidalEmbedding(nn.Module):
96
+ def __init__(self, dim, theta = 10000):
97
+ super().__init__()
98
+ assert (dim % 2) == 0, 'dimension must be divisible by 2'
99
+ self.scale = nn.Parameter(torch.ones(1) * dim ** -0.5)
100
+
101
+ half_dim = dim // 2
102
+ freq_seq = torch.arange(half_dim).float() / half_dim
103
+ inv_freq = theta ** -freq_seq
104
+ self.register_buffer('inv_freq', inv_freq, persistent = False)
105
+
106
+ def forward(self, x, pos = None, seq_start_pos = None):
107
+ seq_len, device = x.shape[1], x.device
108
+
109
+ if pos is None:
110
+ pos = torch.arange(seq_len, device = device)
111
+
112
+ if seq_start_pos is not None:
113
+ pos = pos - seq_start_pos[..., None]
114
+
115
+ emb = torch.einsum('i, j -> i j', pos, self.inv_freq)
116
+ emb = torch.cat((emb.sin(), emb.cos()), dim = -1)
117
+ return emb * self.scale
118
+
119
+ class RotaryEmbedding(nn.Module):
120
+ def __init__(
121
+ self,
122
+ dim,
123
+ use_xpos = False,
124
+ scale_base = 512,
125
+ interpolation_factor = 1.,
126
+ base = 10000,
127
+ base_rescale_factor = 1.,
128
+ dtype=None,
129
+ device=None,
130
+ ):
131
+ super().__init__()
132
+ # proposed by reddit user bloc97, to rescale rotary embeddings to longer sequence length without fine-tuning
133
+ # has some connection to NTK literature
134
+ # https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/
135
+ base *= base_rescale_factor ** (dim / (dim - 2))
136
+
137
+ # inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
138
+ self.register_buffer('inv_freq', torch.empty((dim // 2,), device=device, dtype=dtype))
139
+
140
+ assert interpolation_factor >= 1.
141
+ self.interpolation_factor = interpolation_factor
142
+
143
+ if not use_xpos:
144
+ self.register_buffer('scale', None)
145
+ return
146
+
147
+ scale = (torch.arange(0, dim, 2) + 0.4 * dim) / (1.4 * dim)
148
+
149
+ self.scale_base = scale_base
150
+ self.register_buffer('scale', scale)
151
+
152
+ def forward_from_seq_len(self, seq_len, device, dtype):
153
+ # device = self.inv_freq.device
154
+
155
+ t = torch.arange(seq_len, device=device, dtype=dtype)
156
+ return self.forward(t)
157
+
158
+ def forward(self, t):
159
+ # device = self.inv_freq.device
160
+ device = t.device
161
+ dtype = t.dtype
162
+
163
+ # t = t.to(torch.float32)
164
+
165
+ t = t / self.interpolation_factor
166
+
167
+ freqs = torch.einsum('i , j -> i j', t, comfy.ops.cast_to_input(self.inv_freq, t))
168
+ freqs = torch.cat((freqs, freqs), dim = -1)
169
+
170
+ if self.scale is None:
171
+ return freqs, 1.
172
+
173
+ power = (torch.arange(seq_len, device = device) - (seq_len // 2)) / self.scale_base
174
+ scale = comfy.ops.cast_to_input(self.scale, t) ** rearrange(power, 'n -> n 1')
175
+ scale = torch.cat((scale, scale), dim = -1)
176
+
177
+ return freqs, scale
178
+
179
+ def rotate_half(x):
180
+ x = rearrange(x, '... (j d) -> ... j d', j = 2)
181
+ x1, x2 = x.unbind(dim = -2)
182
+ return torch.cat((-x2, x1), dim = -1)
183
+
184
+ def apply_rotary_pos_emb(t, freqs, scale = 1):
185
+ out_dtype = t.dtype
186
+
187
+ # cast to float32 if necessary for numerical stability
188
+ dtype = t.dtype #reduce(torch.promote_types, (t.dtype, freqs.dtype, torch.float32))
189
+ rot_dim, seq_len = freqs.shape[-1], t.shape[-2]
190
+ freqs, t = freqs.to(dtype), t.to(dtype)
191
+ freqs = freqs[-seq_len:, :]
192
+
193
+ if t.ndim == 4 and freqs.ndim == 3:
194
+ freqs = rearrange(freqs, 'b n d -> b 1 n d')
195
+
196
+ # partial rotary embeddings, Wang et al. GPT-J
197
+ t, t_unrotated = t[..., :rot_dim], t[..., rot_dim:]
198
+ t = (t * freqs.cos() * scale) + (rotate_half(t) * freqs.sin() * scale)
199
+
200
+ t, t_unrotated = t.to(out_dtype), t_unrotated.to(out_dtype)
201
+
202
+ return torch.cat((t, t_unrotated), dim = -1)
203
+
204
+ class FeedForward(nn.Module):
205
+ def __init__(
206
+ self,
207
+ dim,
208
+ dim_out = None,
209
+ mult = 4,
210
+ no_bias = False,
211
+ glu = True,
212
+ use_conv = False,
213
+ conv_kernel_size = 3,
214
+ zero_init_output = True,
215
+ dtype=None,
216
+ device=None,
217
+ operations=None,
218
+ ):
219
+ super().__init__()
220
+ inner_dim = int(dim * mult)
221
+
222
+ # Default to SwiGLU
223
+
224
+ activation = nn.SiLU()
225
+
226
+ dim_out = dim if dim_out is None else dim_out
227
+
228
+ if glu:
229
+ linear_in = GLU(dim, inner_dim, activation, dtype=dtype, device=device, operations=operations)
230
+ else:
231
+ linear_in = nn.Sequential(
232
+ Rearrange('b n d -> b d n') if use_conv else nn.Identity(),
233
+ operations.Linear(dim, inner_dim, bias = not no_bias, dtype=dtype, device=device) if not use_conv else operations.Conv1d(dim, inner_dim, conv_kernel_size, padding = (conv_kernel_size // 2), bias = not no_bias, dtype=dtype, device=device),
234
+ Rearrange('b n d -> b d n') if use_conv else nn.Identity(),
235
+ activation
236
+ )
237
+
238
+ linear_out = operations.Linear(inner_dim, dim_out, bias = not no_bias, dtype=dtype, device=device) if not use_conv else operations.Conv1d(inner_dim, dim_out, conv_kernel_size, padding = (conv_kernel_size // 2), bias = not no_bias, dtype=dtype, device=device)
239
+
240
+ # # init last linear layer to 0
241
+ # if zero_init_output:
242
+ # nn.init.zeros_(linear_out.weight)
243
+ # if not no_bias:
244
+ # nn.init.zeros_(linear_out.bias)
245
+
246
+
247
+ self.ff = nn.Sequential(
248
+ linear_in,
249
+ Rearrange('b d n -> b n d') if use_conv else nn.Identity(),
250
+ linear_out,
251
+ Rearrange('b n d -> b d n') if use_conv else nn.Identity(),
252
+ )
253
+
254
+ def forward(self, x):
255
+ return self.ff(x)
256
+
257
+ class Attention(nn.Module):
258
+ def __init__(
259
+ self,
260
+ dim,
261
+ dim_heads = 64,
262
+ dim_context = None,
263
+ causal = False,
264
+ zero_init_output=True,
265
+ qk_norm = False,
266
+ natten_kernel_size = None,
267
+ dtype=None,
268
+ device=None,
269
+ operations=None,
270
+ ):
271
+ super().__init__()
272
+ self.dim = dim
273
+ self.dim_heads = dim_heads
274
+ self.causal = causal
275
+
276
+ dim_kv = dim_context if dim_context is not None else dim
277
+
278
+ self.num_heads = dim // dim_heads
279
+ self.kv_heads = dim_kv // dim_heads
280
+
281
+ if dim_context is not None:
282
+ self.to_q = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
283
+ self.to_kv = operations.Linear(dim_kv, dim_kv * 2, bias=False, dtype=dtype, device=device)
284
+ else:
285
+ self.to_qkv = operations.Linear(dim, dim * 3, bias=False, dtype=dtype, device=device)
286
+
287
+ self.to_out = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
288
+
289
+ # if zero_init_output:
290
+ # nn.init.zeros_(self.to_out.weight)
291
+
292
+ self.qk_norm = qk_norm
293
+
294
+
295
+ def forward(
296
+ self,
297
+ x,
298
+ context = None,
299
+ mask = None,
300
+ context_mask = None,
301
+ rotary_pos_emb = None,
302
+ causal = None
303
+ ):
304
+ h, kv_h, has_context = self.num_heads, self.kv_heads, context is not None
305
+
306
+ kv_input = context if has_context else x
307
+
308
+ if hasattr(self, 'to_q'):
309
+ # Use separate linear projections for q and k/v
310
+ q = self.to_q(x)
311
+ q = rearrange(q, 'b n (h d) -> b h n d', h = h)
312
+
313
+ k, v = self.to_kv(kv_input).chunk(2, dim=-1)
314
+
315
+ k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = kv_h), (k, v))
316
+ else:
317
+ # Use fused linear projection
318
+ q, k, v = self.to_qkv(x).chunk(3, dim=-1)
319
+ q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q, k, v))
320
+
321
+ # Normalize q and k for cosine sim attention
322
+ if self.qk_norm:
323
+ q = F.normalize(q, dim=-1)
324
+ k = F.normalize(k, dim=-1)
325
+
326
+ if rotary_pos_emb is not None and not has_context:
327
+ freqs, _ = rotary_pos_emb
328
+
329
+ q_dtype = q.dtype
330
+ k_dtype = k.dtype
331
+
332
+ q = q.to(torch.float32)
333
+ k = k.to(torch.float32)
334
+ freqs = freqs.to(torch.float32)
335
+
336
+ q = apply_rotary_pos_emb(q, freqs)
337
+ k = apply_rotary_pos_emb(k, freqs)
338
+
339
+ q = q.to(q_dtype)
340
+ k = k.to(k_dtype)
341
+
342
+ input_mask = context_mask
343
+
344
+ if input_mask is None and not has_context:
345
+ input_mask = mask
346
+
347
+ # determine masking
348
+ masks = []
349
+ final_attn_mask = None # The mask that will be applied to the attention matrix, taking all masks into account
350
+
351
+ if input_mask is not None:
352
+ input_mask = rearrange(input_mask, 'b j -> b 1 1 j')
353
+ masks.append(~input_mask)
354
+
355
+ # Other masks will be added here later
356
+
357
+ if len(masks) > 0:
358
+ final_attn_mask = ~or_reduce(masks)
359
+
360
+ n, device = q.shape[-2], q.device
361
+
362
+ causal = self.causal if causal is None else causal
363
+
364
+ if n == 1 and causal:
365
+ causal = False
366
+
367
+ if h != kv_h:
368
+ # Repeat interleave kv_heads to match q_heads
369
+ heads_per_kv_head = h // kv_h
370
+ k, v = map(lambda t: t.repeat_interleave(heads_per_kv_head, dim = 1), (k, v))
371
+
372
+ out = optimized_attention(q, k, v, h, skip_reshape=True)
373
+ out = self.to_out(out)
374
+
375
+ if mask is not None:
376
+ mask = rearrange(mask, 'b n -> b n 1')
377
+ out = out.masked_fill(~mask, 0.)
378
+
379
+ return out
380
+
381
+ class ConformerModule(nn.Module):
382
+ def __init__(
383
+ self,
384
+ dim,
385
+ norm_kwargs = {},
386
+ ):
387
+
388
+ super().__init__()
389
+
390
+ self.dim = dim
391
+
392
+ self.in_norm = LayerNorm(dim, **norm_kwargs)
393
+ self.pointwise_conv = nn.Conv1d(dim, dim, kernel_size=1, bias=False)
394
+ self.glu = GLU(dim, dim, nn.SiLU())
395
+ self.depthwise_conv = nn.Conv1d(dim, dim, kernel_size=17, groups=dim, padding=8, bias=False)
396
+ self.mid_norm = LayerNorm(dim, **norm_kwargs) # This is a batch norm in the original but I don't like batch norm
397
+ self.swish = nn.SiLU()
398
+ self.pointwise_conv_2 = nn.Conv1d(dim, dim, kernel_size=1, bias=False)
399
+
400
+ def forward(self, x):
401
+ x = self.in_norm(x)
402
+ x = rearrange(x, 'b n d -> b d n')
403
+ x = self.pointwise_conv(x)
404
+ x = rearrange(x, 'b d n -> b n d')
405
+ x = self.glu(x)
406
+ x = rearrange(x, 'b n d -> b d n')
407
+ x = self.depthwise_conv(x)
408
+ x = rearrange(x, 'b d n -> b n d')
409
+ x = self.mid_norm(x)
410
+ x = self.swish(x)
411
+ x = rearrange(x, 'b n d -> b d n')
412
+ x = self.pointwise_conv_2(x)
413
+ x = rearrange(x, 'b d n -> b n d')
414
+
415
+ return x
416
+
417
+ class TransformerBlock(nn.Module):
418
+ def __init__(
419
+ self,
420
+ dim,
421
+ dim_heads = 64,
422
+ cross_attend = False,
423
+ dim_context = None,
424
+ global_cond_dim = None,
425
+ causal = False,
426
+ zero_init_branch_outputs = True,
427
+ conformer = False,
428
+ layer_ix = -1,
429
+ remove_norms = False,
430
+ attn_kwargs = {},
431
+ ff_kwargs = {},
432
+ norm_kwargs = {},
433
+ dtype=None,
434
+ device=None,
435
+ operations=None,
436
+ ):
437
+
438
+ super().__init__()
439
+ self.dim = dim
440
+ self.dim_heads = dim_heads
441
+ self.cross_attend = cross_attend
442
+ self.dim_context = dim_context
443
+ self.causal = causal
444
+
445
+ self.pre_norm = LayerNorm(dim, dtype=dtype, device=device, **norm_kwargs) if not remove_norms else nn.Identity()
446
+
447
+ self.self_attn = Attention(
448
+ dim,
449
+ dim_heads = dim_heads,
450
+ causal = causal,
451
+ zero_init_output=zero_init_branch_outputs,
452
+ dtype=dtype,
453
+ device=device,
454
+ operations=operations,
455
+ **attn_kwargs
456
+ )
457
+
458
+ if cross_attend:
459
+ self.cross_attend_norm = LayerNorm(dim, dtype=dtype, device=device, **norm_kwargs) if not remove_norms else nn.Identity()
460
+ self.cross_attn = Attention(
461
+ dim,
462
+ dim_heads = dim_heads,
463
+ dim_context=dim_context,
464
+ causal = causal,
465
+ zero_init_output=zero_init_branch_outputs,
466
+ dtype=dtype,
467
+ device=device,
468
+ operations=operations,
469
+ **attn_kwargs
470
+ )
471
+
472
+ self.ff_norm = LayerNorm(dim, dtype=dtype, device=device, **norm_kwargs) if not remove_norms else nn.Identity()
473
+ self.ff = FeedForward(dim, zero_init_output=zero_init_branch_outputs, dtype=dtype, device=device, operations=operations,**ff_kwargs)
474
+
475
+ self.layer_ix = layer_ix
476
+
477
+ self.conformer = ConformerModule(dim, norm_kwargs=norm_kwargs) if conformer else None
478
+
479
+ self.global_cond_dim = global_cond_dim
480
+
481
+ if global_cond_dim is not None:
482
+ self.to_scale_shift_gate = nn.Sequential(
483
+ nn.SiLU(),
484
+ nn.Linear(global_cond_dim, dim * 6, bias=False)
485
+ )
486
+
487
+ nn.init.zeros_(self.to_scale_shift_gate[1].weight)
488
+ #nn.init.zeros_(self.to_scale_shift_gate_self[1].bias)
489
+
490
+ def forward(
491
+ self,
492
+ x,
493
+ context = None,
494
+ global_cond=None,
495
+ mask = None,
496
+ context_mask = None,
497
+ rotary_pos_emb = None
498
+ ):
499
+ if self.global_cond_dim is not None and self.global_cond_dim > 0 and global_cond is not None:
500
+
501
+ scale_self, shift_self, gate_self, scale_ff, shift_ff, gate_ff = self.to_scale_shift_gate(global_cond).unsqueeze(1).chunk(6, dim = -1)
502
+
503
+ # self-attention with adaLN
504
+ residual = x
505
+ x = self.pre_norm(x)
506
+ x = x * (1 + scale_self) + shift_self
507
+ x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb)
508
+ x = x * torch.sigmoid(1 - gate_self)
509
+ x = x + residual
510
+
511
+ if context is not None:
512
+ x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
513
+
514
+ if self.conformer is not None:
515
+ x = x + self.conformer(x)
516
+
517
+ # feedforward with adaLN
518
+ residual = x
519
+ x = self.ff_norm(x)
520
+ x = x * (1 + scale_ff) + shift_ff
521
+ x = self.ff(x)
522
+ x = x * torch.sigmoid(1 - gate_ff)
523
+ x = x + residual
524
+
525
+ else:
526
+ x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb)
527
+
528
+ if context is not None:
529
+ x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
530
+
531
+ if self.conformer is not None:
532
+ x = x + self.conformer(x)
533
+
534
+ x = x + self.ff(self.ff_norm(x))
535
+
536
+ return x
537
+
538
+ class ContinuousTransformer(nn.Module):
539
+ def __init__(
540
+ self,
541
+ dim,
542
+ depth,
543
+ *,
544
+ dim_in = None,
545
+ dim_out = None,
546
+ dim_heads = 64,
547
+ cross_attend=False,
548
+ cond_token_dim=None,
549
+ global_cond_dim=None,
550
+ causal=False,
551
+ rotary_pos_emb=True,
552
+ zero_init_branch_outputs=True,
553
+ conformer=False,
554
+ use_sinusoidal_emb=False,
555
+ use_abs_pos_emb=False,
556
+ abs_pos_emb_max_length=10000,
557
+ dtype=None,
558
+ device=None,
559
+ operations=None,
560
+ **kwargs
561
+ ):
562
+
563
+ super().__init__()
564
+
565
+ self.dim = dim
566
+ self.depth = depth
567
+ self.causal = causal
568
+ self.layers = nn.ModuleList([])
569
+
570
+ self.project_in = operations.Linear(dim_in, dim, bias=False, dtype=dtype, device=device) if dim_in is not None else nn.Identity()
571
+ self.project_out = operations.Linear(dim, dim_out, bias=False, dtype=dtype, device=device) if dim_out is not None else nn.Identity()
572
+
573
+ if rotary_pos_emb:
574
+ self.rotary_pos_emb = RotaryEmbedding(max(dim_heads // 2, 32), device=device, dtype=dtype)
575
+ else:
576
+ self.rotary_pos_emb = None
577
+
578
+ self.use_sinusoidal_emb = use_sinusoidal_emb
579
+ if use_sinusoidal_emb:
580
+ self.pos_emb = ScaledSinusoidalEmbedding(dim)
581
+
582
+ self.use_abs_pos_emb = use_abs_pos_emb
583
+ if use_abs_pos_emb:
584
+ self.pos_emb = AbsolutePositionalEmbedding(dim, abs_pos_emb_max_length)
585
+
586
+ for i in range(depth):
587
+ self.layers.append(
588
+ TransformerBlock(
589
+ dim,
590
+ dim_heads = dim_heads,
591
+ cross_attend = cross_attend,
592
+ dim_context = cond_token_dim,
593
+ global_cond_dim = global_cond_dim,
594
+ causal = causal,
595
+ zero_init_branch_outputs = zero_init_branch_outputs,
596
+ conformer=conformer,
597
+ layer_ix=i,
598
+ dtype=dtype,
599
+ device=device,
600
+ operations=operations,
601
+ **kwargs
602
+ )
603
+ )
604
+
605
+ def forward(
606
+ self,
607
+ x,
608
+ mask = None,
609
+ prepend_embeds = None,
610
+ prepend_mask = None,
611
+ global_cond = None,
612
+ return_info = False,
613
+ **kwargs
614
+ ):
615
+ batch, seq, device = *x.shape[:2], x.device
616
+
617
+ info = {
618
+ "hidden_states": [],
619
+ }
620
+
621
+ x = self.project_in(x)
622
+
623
+ if prepend_embeds is not None:
624
+ prepend_length, prepend_dim = prepend_embeds.shape[1:]
625
+
626
+ assert prepend_dim == x.shape[-1], 'prepend dimension must match sequence dimension'
627
+
628
+ x = torch.cat((prepend_embeds, x), dim = -2)
629
+
630
+ if prepend_mask is not None or mask is not None:
631
+ mask = mask if mask is not None else torch.ones((batch, seq), device = device, dtype = torch.bool)
632
+ prepend_mask = prepend_mask if prepend_mask is not None else torch.ones((batch, prepend_length), device = device, dtype = torch.bool)
633
+
634
+ mask = torch.cat((prepend_mask, mask), dim = -1)
635
+
636
+ # Attention layers
637
+
638
+ if self.rotary_pos_emb is not None:
639
+ rotary_pos_emb = self.rotary_pos_emb.forward_from_seq_len(x.shape[1], dtype=x.dtype, device=x.device)
640
+ else:
641
+ rotary_pos_emb = None
642
+
643
+ if self.use_sinusoidal_emb or self.use_abs_pos_emb:
644
+ x = x + self.pos_emb(x)
645
+
646
+ # Iterate over the transformer layers
647
+ for layer in self.layers:
648
+ x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, **kwargs)
649
+ # x = checkpoint(layer, x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, **kwargs)
650
+
651
+ if return_info:
652
+ info["hidden_states"].append(x)
653
+
654
+ x = self.project_out(x)
655
+
656
+ if return_info:
657
+ return x, info
658
+
659
+ return x
660
+
661
+ class AudioDiffusionTransformer(nn.Module):
662
+ def __init__(self,
663
+ io_channels=64,
664
+ patch_size=1,
665
+ embed_dim=1536,
666
+ cond_token_dim=768,
667
+ project_cond_tokens=False,
668
+ global_cond_dim=1536,
669
+ project_global_cond=True,
670
+ input_concat_dim=0,
671
+ prepend_cond_dim=0,
672
+ depth=24,
673
+ num_heads=24,
674
+ transformer_type: tp.Literal["continuous_transformer"] = "continuous_transformer",
675
+ global_cond_type: tp.Literal["prepend", "adaLN"] = "prepend",
676
+ audio_model="",
677
+ dtype=None,
678
+ device=None,
679
+ operations=None,
680
+ **kwargs):
681
+
682
+ super().__init__()
683
+
684
+ self.dtype = dtype
685
+ self.cond_token_dim = cond_token_dim
686
+
687
+ # Timestep embeddings
688
+ timestep_features_dim = 256
689
+
690
+ self.timestep_features = FourierFeatures(1, timestep_features_dim, dtype=dtype, device=device)
691
+
692
+ self.to_timestep_embed = nn.Sequential(
693
+ operations.Linear(timestep_features_dim, embed_dim, bias=True, dtype=dtype, device=device),
694
+ nn.SiLU(),
695
+ operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device),
696
+ )
697
+
698
+ if cond_token_dim > 0:
699
+ # Conditioning tokens
700
+
701
+ cond_embed_dim = cond_token_dim if not project_cond_tokens else embed_dim
702
+ self.to_cond_embed = nn.Sequential(
703
+ operations.Linear(cond_token_dim, cond_embed_dim, bias=False, dtype=dtype, device=device),
704
+ nn.SiLU(),
705
+ operations.Linear(cond_embed_dim, cond_embed_dim, bias=False, dtype=dtype, device=device)
706
+ )
707
+ else:
708
+ cond_embed_dim = 0
709
+
710
+ if global_cond_dim > 0:
711
+ # Global conditioning
712
+ global_embed_dim = global_cond_dim if not project_global_cond else embed_dim
713
+ self.to_global_embed = nn.Sequential(
714
+ operations.Linear(global_cond_dim, global_embed_dim, bias=False, dtype=dtype, device=device),
715
+ nn.SiLU(),
716
+ operations.Linear(global_embed_dim, global_embed_dim, bias=False, dtype=dtype, device=device)
717
+ )
718
+
719
+ if prepend_cond_dim > 0:
720
+ # Prepend conditioning
721
+ self.to_prepend_embed = nn.Sequential(
722
+ operations.Linear(prepend_cond_dim, embed_dim, bias=False, dtype=dtype, device=device),
723
+ nn.SiLU(),
724
+ operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
725
+ )
726
+
727
+ self.input_concat_dim = input_concat_dim
728
+
729
+ dim_in = io_channels + self.input_concat_dim
730
+
731
+ self.patch_size = patch_size
732
+
733
+ # Transformer
734
+
735
+ self.transformer_type = transformer_type
736
+
737
+ self.global_cond_type = global_cond_type
738
+
739
+ if self.transformer_type == "continuous_transformer":
740
+
741
+ global_dim = None
742
+
743
+ if self.global_cond_type == "adaLN":
744
+ # The global conditioning is projected to the embed_dim already at this point
745
+ global_dim = embed_dim
746
+
747
+ self.transformer = ContinuousTransformer(
748
+ dim=embed_dim,
749
+ depth=depth,
750
+ dim_heads=embed_dim // num_heads,
751
+ dim_in=dim_in * patch_size,
752
+ dim_out=io_channels * patch_size,
753
+ cross_attend = cond_token_dim > 0,
754
+ cond_token_dim = cond_embed_dim,
755
+ global_cond_dim=global_dim,
756
+ dtype=dtype,
757
+ device=device,
758
+ operations=operations,
759
+ **kwargs
760
+ )
761
+ else:
762
+ raise ValueError(f"Unknown transformer type: {self.transformer_type}")
763
+
764
+ self.preprocess_conv = operations.Conv1d(dim_in, dim_in, 1, bias=False, dtype=dtype, device=device)
765
+ self.postprocess_conv = operations.Conv1d(io_channels, io_channels, 1, bias=False, dtype=dtype, device=device)
766
+
767
+ def _forward(
768
+ self,
769
+ x,
770
+ t,
771
+ mask=None,
772
+ cross_attn_cond=None,
773
+ cross_attn_cond_mask=None,
774
+ input_concat_cond=None,
775
+ global_embed=None,
776
+ prepend_cond=None,
777
+ prepend_cond_mask=None,
778
+ return_info=False,
779
+ **kwargs):
780
+
781
+ if cross_attn_cond is not None:
782
+ cross_attn_cond = self.to_cond_embed(cross_attn_cond)
783
+
784
+ if global_embed is not None:
785
+ # Project the global conditioning to the embedding dimension
786
+ global_embed = self.to_global_embed(global_embed)
787
+
788
+ prepend_inputs = None
789
+ prepend_mask = None
790
+ prepend_length = 0
791
+ if prepend_cond is not None:
792
+ # Project the prepend conditioning to the embedding dimension
793
+ prepend_cond = self.to_prepend_embed(prepend_cond)
794
+
795
+ prepend_inputs = prepend_cond
796
+ if prepend_cond_mask is not None:
797
+ prepend_mask = prepend_cond_mask
798
+
799
+ if input_concat_cond is not None:
800
+
801
+ # Interpolate input_concat_cond to the same length as x
802
+ if input_concat_cond.shape[2] != x.shape[2]:
803
+ input_concat_cond = F.interpolate(input_concat_cond, (x.shape[2], ), mode='nearest')
804
+
805
+ x = torch.cat([x, input_concat_cond], dim=1)
806
+
807
+ # Get the batch of timestep embeddings
808
+ timestep_embed = self.to_timestep_embed(self.timestep_features(t[:, None]).to(x.dtype)) # (b, embed_dim)
809
+
810
+ # Timestep embedding is considered a global embedding. Add to the global conditioning if it exists
811
+ if global_embed is not None:
812
+ global_embed = global_embed + timestep_embed
813
+ else:
814
+ global_embed = timestep_embed
815
+
816
+ # Add the global_embed to the prepend inputs if there is no global conditioning support in the transformer
817
+ if self.global_cond_type == "prepend":
818
+ if prepend_inputs is None:
819
+ # Prepend inputs are just the global embed, and the mask is all ones
820
+ prepend_inputs = global_embed.unsqueeze(1)
821
+ prepend_mask = torch.ones((x.shape[0], 1), device=x.device, dtype=torch.bool)
822
+ else:
823
+ # Prepend inputs are the prepend conditioning + the global embed
824
+ prepend_inputs = torch.cat([prepend_inputs, global_embed.unsqueeze(1)], dim=1)
825
+ prepend_mask = torch.cat([prepend_mask, torch.ones((x.shape[0], 1), device=x.device, dtype=torch.bool)], dim=1)
826
+
827
+ prepend_length = prepend_inputs.shape[1]
828
+
829
+ x = self.preprocess_conv(x) + x
830
+
831
+ x = rearrange(x, "b c t -> b t c")
832
+
833
+ extra_args = {}
834
+
835
+ if self.global_cond_type == "adaLN":
836
+ extra_args["global_cond"] = global_embed
837
+
838
+ if self.patch_size > 1:
839
+ x = rearrange(x, "b (t p) c -> b t (c p)", p=self.patch_size)
840
+
841
+ if self.transformer_type == "x-transformers":
842
+ output = self.transformer(x, prepend_embeds=prepend_inputs, context=cross_attn_cond, context_mask=cross_attn_cond_mask, mask=mask, prepend_mask=prepend_mask, **extra_args, **kwargs)
843
+ elif self.transformer_type == "continuous_transformer":
844
+ output = self.transformer(x, prepend_embeds=prepend_inputs, context=cross_attn_cond, context_mask=cross_attn_cond_mask, mask=mask, prepend_mask=prepend_mask, return_info=return_info, **extra_args, **kwargs)
845
+
846
+ if return_info:
847
+ output, info = output
848
+ elif self.transformer_type == "mm_transformer":
849
+ output = self.transformer(x, context=cross_attn_cond, mask=mask, context_mask=cross_attn_cond_mask, **extra_args, **kwargs)
850
+
851
+ output = rearrange(output, "b t c -> b c t")[:,:,prepend_length:]
852
+
853
+ if self.patch_size > 1:
854
+ output = rearrange(output, "b (c p) t -> b c (t p)", p=self.patch_size)
855
+
856
+ output = self.postprocess_conv(output) + output
857
+
858
+ if return_info:
859
+ return output, info
860
+
861
+ return output
862
+
863
+ def forward(
864
+ self,
865
+ x,
866
+ timestep,
867
+ context=None,
868
+ context_mask=None,
869
+ input_concat_cond=None,
870
+ global_embed=None,
871
+ negative_global_embed=None,
872
+ prepend_cond=None,
873
+ prepend_cond_mask=None,
874
+ mask=None,
875
+ return_info=False,
876
+ control=None,
877
+ transformer_options={},
878
+ **kwargs):
879
+ return self._forward(
880
+ x,
881
+ timestep,
882
+ cross_attn_cond=context,
883
+ cross_attn_cond_mask=context_mask,
884
+ input_concat_cond=input_concat_cond,
885
+ global_embed=global_embed,
886
+ prepend_cond=prepend_cond,
887
+ prepend_cond_mask=prepend_cond_mask,
888
+ mask=mask,
889
+ return_info=return_info,
890
+ **kwargs
891
+ )
ComfyUI/comfy/ldm/audio/embedders.py ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # code adapted from: https://github.com/Stability-AI/stable-audio-tools
2
+
3
+ import torch
4
+ import torch.nn as nn
5
+ from torch import Tensor, einsum
6
+ from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, TypeVar, Union
7
+ from einops import rearrange
8
+ import math
9
+ import comfy.ops
10
+
11
+ class LearnedPositionalEmbedding(nn.Module):
12
+ """Used for continuous time"""
13
+
14
+ def __init__(self, dim: int):
15
+ super().__init__()
16
+ assert (dim % 2) == 0
17
+ half_dim = dim // 2
18
+ self.weights = nn.Parameter(torch.empty(half_dim))
19
+
20
+ def forward(self, x: Tensor) -> Tensor:
21
+ x = rearrange(x, "b -> b 1")
22
+ freqs = x * rearrange(self.weights, "d -> 1 d") * 2 * math.pi
23
+ fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
24
+ fouriered = torch.cat((x, fouriered), dim=-1)
25
+ return fouriered
26
+
27
+ def TimePositionalEmbedding(dim: int, out_features: int) -> nn.Module:
28
+ return nn.Sequential(
29
+ LearnedPositionalEmbedding(dim),
30
+ comfy.ops.manual_cast.Linear(in_features=dim + 1, out_features=out_features),
31
+ )
32
+
33
+
34
+ class NumberEmbedder(nn.Module):
35
+ def __init__(
36
+ self,
37
+ features: int,
38
+ dim: int = 256,
39
+ ):
40
+ super().__init__()
41
+ self.features = features
42
+ self.embedding = TimePositionalEmbedding(dim=dim, out_features=features)
43
+
44
+ def forward(self, x: Union[List[float], Tensor]) -> Tensor:
45
+ if not torch.is_tensor(x):
46
+ device = next(self.embedding.parameters()).device
47
+ x = torch.tensor(x, device=device)
48
+ assert isinstance(x, Tensor)
49
+ shape = x.shape
50
+ x = rearrange(x, "... -> (...)")
51
+ embedding = self.embedding(x)
52
+ x = embedding.view(*shape, self.features)
53
+ return x # type: ignore
54
+
55
+
56
+ class Conditioner(nn.Module):
57
+ def __init__(
58
+ self,
59
+ dim: int,
60
+ output_dim: int,
61
+ project_out: bool = False
62
+ ):
63
+
64
+ super().__init__()
65
+
66
+ self.dim = dim
67
+ self.output_dim = output_dim
68
+ self.proj_out = nn.Linear(dim, output_dim) if (dim != output_dim or project_out) else nn.Identity()
69
+
70
+ def forward(self, x):
71
+ raise NotImplementedError()
72
+
73
+ class NumberConditioner(Conditioner):
74
+ '''
75
+ Conditioner that takes a list of floats, normalizes them for a given range, and returns a list of embeddings
76
+ '''
77
+ def __init__(self,
78
+ output_dim: int,
79
+ min_val: float=0,
80
+ max_val: float=1
81
+ ):
82
+ super().__init__(output_dim, output_dim)
83
+
84
+ self.min_val = min_val
85
+ self.max_val = max_val
86
+
87
+ self.embedder = NumberEmbedder(features=output_dim)
88
+
89
+ def forward(self, floats, device=None):
90
+ # Cast the inputs to floats
91
+ floats = [float(x) for x in floats]
92
+
93
+ if device is None:
94
+ device = next(self.embedder.parameters()).device
95
+
96
+ floats = torch.tensor(floats).to(device)
97
+
98
+ floats = floats.clamp(self.min_val, self.max_val)
99
+
100
+ normalized_floats = (floats - self.min_val) / (self.max_val - self.min_val)
101
+
102
+ # Cast floats to same type as embedder
103
+ embedder_dtype = next(self.embedder.parameters()).dtype
104
+ normalized_floats = normalized_floats.to(embedder_dtype)
105
+
106
+ float_embeds = self.embedder(normalized_floats).unsqueeze(1)
107
+
108
+ return [float_embeds, torch.ones(float_embeds.shape[0], 1).to(device)]
ComfyUI/comfy/ldm/aura/mmdit.py ADDED
@@ -0,0 +1,478 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #AuraFlow MMDiT
2
+ #Originally written by the AuraFlow Authors
3
+
4
+ import math
5
+
6
+ import torch
7
+ import torch.nn as nn
8
+ import torch.nn.functional as F
9
+
10
+ from comfy.ldm.modules.attention import optimized_attention
11
+ import comfy.ops
12
+ import comfy.ldm.common_dit
13
+
14
+ def modulate(x, shift, scale):
15
+ return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
16
+
17
+
18
+ def find_multiple(n: int, k: int) -> int:
19
+ if n % k == 0:
20
+ return n
21
+ return n + k - (n % k)
22
+
23
+
24
+ class MLP(nn.Module):
25
+ def __init__(self, dim, hidden_dim=None, dtype=None, device=None, operations=None) -> None:
26
+ super().__init__()
27
+ if hidden_dim is None:
28
+ hidden_dim = 4 * dim
29
+
30
+ n_hidden = int(2 * hidden_dim / 3)
31
+ n_hidden = find_multiple(n_hidden, 256)
32
+
33
+ self.c_fc1 = operations.Linear(dim, n_hidden, bias=False, dtype=dtype, device=device)
34
+ self.c_fc2 = operations.Linear(dim, n_hidden, bias=False, dtype=dtype, device=device)
35
+ self.c_proj = operations.Linear(n_hidden, dim, bias=False, dtype=dtype, device=device)
36
+
37
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
38
+ x = F.silu(self.c_fc1(x)) * self.c_fc2(x)
39
+ x = self.c_proj(x)
40
+ return x
41
+
42
+
43
+ class MultiHeadLayerNorm(nn.Module):
44
+ def __init__(self, hidden_size=None, eps=1e-5, dtype=None, device=None):
45
+ # Copy pasta from https://github.com/huggingface/transformers/blob/e5f71ecaae50ea476d1e12351003790273c4b2ed/src/transformers/models/cohere/modeling_cohere.py#L78
46
+
47
+ super().__init__()
48
+ self.weight = nn.Parameter(torch.empty(hidden_size, dtype=dtype, device=device))
49
+ self.variance_epsilon = eps
50
+
51
+ def forward(self, hidden_states):
52
+ input_dtype = hidden_states.dtype
53
+ hidden_states = hidden_states.to(torch.float32)
54
+ mean = hidden_states.mean(-1, keepdim=True)
55
+ variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
56
+ hidden_states = (hidden_states - mean) * torch.rsqrt(
57
+ variance + self.variance_epsilon
58
+ )
59
+ hidden_states = self.weight.to(torch.float32) * hidden_states
60
+ return hidden_states.to(input_dtype)
61
+
62
+ class SingleAttention(nn.Module):
63
+ def __init__(self, dim, n_heads, mh_qknorm=False, dtype=None, device=None, operations=None):
64
+ super().__init__()
65
+
66
+ self.n_heads = n_heads
67
+ self.head_dim = dim // n_heads
68
+
69
+ # this is for cond
70
+ self.w1q = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
71
+ self.w1k = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
72
+ self.w1v = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
73
+ self.w1o = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
74
+
75
+ self.q_norm1 = (
76
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
77
+ if mh_qknorm
78
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
79
+ )
80
+ self.k_norm1 = (
81
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
82
+ if mh_qknorm
83
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
84
+ )
85
+
86
+ #@torch.compile()
87
+ def forward(self, c):
88
+
89
+ bsz, seqlen1, _ = c.shape
90
+
91
+ q, k, v = self.w1q(c), self.w1k(c), self.w1v(c)
92
+ q = q.view(bsz, seqlen1, self.n_heads, self.head_dim)
93
+ k = k.view(bsz, seqlen1, self.n_heads, self.head_dim)
94
+ v = v.view(bsz, seqlen1, self.n_heads, self.head_dim)
95
+ q, k = self.q_norm1(q), self.k_norm1(k)
96
+
97
+ output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
98
+ c = self.w1o(output)
99
+ return c
100
+
101
+
102
+
103
+ class DoubleAttention(nn.Module):
104
+ def __init__(self, dim, n_heads, mh_qknorm=False, dtype=None, device=None, operations=None):
105
+ super().__init__()
106
+
107
+ self.n_heads = n_heads
108
+ self.head_dim = dim // n_heads
109
+
110
+ # this is for cond
111
+ self.w1q = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
112
+ self.w1k = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
113
+ self.w1v = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
114
+ self.w1o = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
115
+
116
+ # this is for x
117
+ self.w2q = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
118
+ self.w2k = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
119
+ self.w2v = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
120
+ self.w2o = operations.Linear(dim, dim, bias=False, dtype=dtype, device=device)
121
+
122
+ self.q_norm1 = (
123
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
124
+ if mh_qknorm
125
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
126
+ )
127
+ self.k_norm1 = (
128
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
129
+ if mh_qknorm
130
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
131
+ )
132
+
133
+ self.q_norm2 = (
134
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
135
+ if mh_qknorm
136
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
137
+ )
138
+ self.k_norm2 = (
139
+ MultiHeadLayerNorm((self.n_heads, self.head_dim), dtype=dtype, device=device)
140
+ if mh_qknorm
141
+ else operations.LayerNorm(self.head_dim, elementwise_affine=False, dtype=dtype, device=device)
142
+ )
143
+
144
+
145
+ #@torch.compile()
146
+ def forward(self, c, x):
147
+
148
+ bsz, seqlen1, _ = c.shape
149
+ bsz, seqlen2, _ = x.shape
150
+ seqlen = seqlen1 + seqlen2
151
+
152
+ cq, ck, cv = self.w1q(c), self.w1k(c), self.w1v(c)
153
+ cq = cq.view(bsz, seqlen1, self.n_heads, self.head_dim)
154
+ ck = ck.view(bsz, seqlen1, self.n_heads, self.head_dim)
155
+ cv = cv.view(bsz, seqlen1, self.n_heads, self.head_dim)
156
+ cq, ck = self.q_norm1(cq), self.k_norm1(ck)
157
+
158
+ xq, xk, xv = self.w2q(x), self.w2k(x), self.w2v(x)
159
+ xq = xq.view(bsz, seqlen2, self.n_heads, self.head_dim)
160
+ xk = xk.view(bsz, seqlen2, self.n_heads, self.head_dim)
161
+ xv = xv.view(bsz, seqlen2, self.n_heads, self.head_dim)
162
+ xq, xk = self.q_norm2(xq), self.k_norm2(xk)
163
+
164
+ # concat all
165
+ q, k, v = (
166
+ torch.cat([cq, xq], dim=1),
167
+ torch.cat([ck, xk], dim=1),
168
+ torch.cat([cv, xv], dim=1),
169
+ )
170
+
171
+ output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
172
+
173
+ c, x = output.split([seqlen1, seqlen2], dim=1)
174
+ c = self.w1o(c)
175
+ x = self.w2o(x)
176
+
177
+ return c, x
178
+
179
+
180
+ class MMDiTBlock(nn.Module):
181
+ def __init__(self, dim, heads=8, global_conddim=1024, is_last=False, dtype=None, device=None, operations=None):
182
+ super().__init__()
183
+
184
+ self.normC1 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
185
+ self.normC2 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
186
+ if not is_last:
187
+ self.mlpC = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
188
+ self.modC = nn.Sequential(
189
+ nn.SiLU(),
190
+ operations.Linear(global_conddim, 6 * dim, bias=False, dtype=dtype, device=device),
191
+ )
192
+ else:
193
+ self.modC = nn.Sequential(
194
+ nn.SiLU(),
195
+ operations.Linear(global_conddim, 2 * dim, bias=False, dtype=dtype, device=device),
196
+ )
197
+
198
+ self.normX1 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
199
+ self.normX2 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
200
+ self.mlpX = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
201
+ self.modX = nn.Sequential(
202
+ nn.SiLU(),
203
+ operations.Linear(global_conddim, 6 * dim, bias=False, dtype=dtype, device=device),
204
+ )
205
+
206
+ self.attn = DoubleAttention(dim, heads, dtype=dtype, device=device, operations=operations)
207
+ self.is_last = is_last
208
+
209
+ #@torch.compile()
210
+ def forward(self, c, x, global_cond, **kwargs):
211
+
212
+ cres, xres = c, x
213
+
214
+ cshift_msa, cscale_msa, cgate_msa, cshift_mlp, cscale_mlp, cgate_mlp = (
215
+ self.modC(global_cond).chunk(6, dim=1)
216
+ )
217
+
218
+ c = modulate(self.normC1(c), cshift_msa, cscale_msa)
219
+
220
+ # xpath
221
+ xshift_msa, xscale_msa, xgate_msa, xshift_mlp, xscale_mlp, xgate_mlp = (
222
+ self.modX(global_cond).chunk(6, dim=1)
223
+ )
224
+
225
+ x = modulate(self.normX1(x), xshift_msa, xscale_msa)
226
+
227
+ # attention
228
+ c, x = self.attn(c, x)
229
+
230
+
231
+ c = self.normC2(cres + cgate_msa.unsqueeze(1) * c)
232
+ c = cgate_mlp.unsqueeze(1) * self.mlpC(modulate(c, cshift_mlp, cscale_mlp))
233
+ c = cres + c
234
+
235
+ x = self.normX2(xres + xgate_msa.unsqueeze(1) * x)
236
+ x = xgate_mlp.unsqueeze(1) * self.mlpX(modulate(x, xshift_mlp, xscale_mlp))
237
+ x = xres + x
238
+
239
+ return c, x
240
+
241
+ class DiTBlock(nn.Module):
242
+ # like MMDiTBlock, but it only has X
243
+ def __init__(self, dim, heads=8, global_conddim=1024, dtype=None, device=None, operations=None):
244
+ super().__init__()
245
+
246
+ self.norm1 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
247
+ self.norm2 = operations.LayerNorm(dim, elementwise_affine=False, dtype=dtype, device=device)
248
+
249
+ self.modCX = nn.Sequential(
250
+ nn.SiLU(),
251
+ operations.Linear(global_conddim, 6 * dim, bias=False, dtype=dtype, device=device),
252
+ )
253
+
254
+ self.attn = SingleAttention(dim, heads, dtype=dtype, device=device, operations=operations)
255
+ self.mlp = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
256
+
257
+ #@torch.compile()
258
+ def forward(self, cx, global_cond, **kwargs):
259
+ cxres = cx
260
+ shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.modCX(
261
+ global_cond
262
+ ).chunk(6, dim=1)
263
+ cx = modulate(self.norm1(cx), shift_msa, scale_msa)
264
+ cx = self.attn(cx)
265
+ cx = self.norm2(cxres + gate_msa.unsqueeze(1) * cx)
266
+ mlpout = self.mlp(modulate(cx, shift_mlp, scale_mlp))
267
+ cx = gate_mlp.unsqueeze(1) * mlpout
268
+
269
+ cx = cxres + cx
270
+
271
+ return cx
272
+
273
+
274
+
275
+ class TimestepEmbedder(nn.Module):
276
+ def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
277
+ super().__init__()
278
+ self.mlp = nn.Sequential(
279
+ operations.Linear(frequency_embedding_size, hidden_size, dtype=dtype, device=device),
280
+ nn.SiLU(),
281
+ operations.Linear(hidden_size, hidden_size, dtype=dtype, device=device),
282
+ )
283
+ self.frequency_embedding_size = frequency_embedding_size
284
+
285
+ @staticmethod
286
+ def timestep_embedding(t, dim, max_period=10000):
287
+ half = dim // 2
288
+ freqs = 1000 * torch.exp(
289
+ -math.log(max_period) * torch.arange(start=0, end=half) / half
290
+ ).to(t.device)
291
+ args = t[:, None] * freqs[None]
292
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
293
+ if dim % 2:
294
+ embedding = torch.cat(
295
+ [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
296
+ )
297
+ return embedding
298
+
299
+ #@torch.compile()
300
+ def forward(self, t, dtype):
301
+ t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(dtype)
302
+ t_emb = self.mlp(t_freq)
303
+ return t_emb
304
+
305
+
306
+ class MMDiT(nn.Module):
307
+ def __init__(
308
+ self,
309
+ in_channels=4,
310
+ out_channels=4,
311
+ patch_size=2,
312
+ dim=3072,
313
+ n_layers=36,
314
+ n_double_layers=4,
315
+ n_heads=12,
316
+ global_conddim=3072,
317
+ cond_seq_dim=2048,
318
+ max_seq=32 * 32,
319
+ device=None,
320
+ dtype=None,
321
+ operations=None,
322
+ ):
323
+ super().__init__()
324
+ self.dtype = dtype
325
+
326
+ self.t_embedder = TimestepEmbedder(global_conddim, dtype=dtype, device=device, operations=operations)
327
+
328
+ self.cond_seq_linear = operations.Linear(
329
+ cond_seq_dim, dim, bias=False, dtype=dtype, device=device
330
+ ) # linear for something like text sequence.
331
+ self.init_x_linear = operations.Linear(
332
+ patch_size * patch_size * in_channels, dim, dtype=dtype, device=device
333
+ ) # init linear for patchified image.
334
+
335
+ self.positional_encoding = nn.Parameter(torch.empty(1, max_seq, dim, dtype=dtype, device=device))
336
+ self.register_tokens = nn.Parameter(torch.empty(1, 8, dim, dtype=dtype, device=device))
337
+
338
+ self.double_layers = nn.ModuleList([])
339
+ self.single_layers = nn.ModuleList([])
340
+
341
+
342
+ for idx in range(n_double_layers):
343
+ self.double_layers.append(
344
+ MMDiTBlock(dim, n_heads, global_conddim, is_last=(idx == n_layers - 1), dtype=dtype, device=device, operations=operations)
345
+ )
346
+
347
+ for idx in range(n_double_layers, n_layers):
348
+ self.single_layers.append(
349
+ DiTBlock(dim, n_heads, global_conddim, dtype=dtype, device=device, operations=operations)
350
+ )
351
+
352
+
353
+ self.final_linear = operations.Linear(
354
+ dim, patch_size * patch_size * out_channels, bias=False, dtype=dtype, device=device
355
+ )
356
+
357
+ self.modF = nn.Sequential(
358
+ nn.SiLU(),
359
+ operations.Linear(global_conddim, 2 * dim, bias=False, dtype=dtype, device=device),
360
+ )
361
+
362
+ self.out_channels = out_channels
363
+ self.patch_size = patch_size
364
+ self.n_double_layers = n_double_layers
365
+ self.n_layers = n_layers
366
+
367
+ self.h_max = round(max_seq**0.5)
368
+ self.w_max = round(max_seq**0.5)
369
+
370
+ @torch.no_grad()
371
+ def extend_pe(self, init_dim=(16, 16), target_dim=(64, 64)):
372
+ # extend pe
373
+ pe_data = self.positional_encoding.data.squeeze(0)[: init_dim[0] * init_dim[1]]
374
+
375
+ pe_as_2d = pe_data.view(init_dim[0], init_dim[1], -1).permute(2, 0, 1)
376
+
377
+ # now we need to extend this to target_dim. for this we will use interpolation.
378
+ # we will use torch.nn.functional.interpolate
379
+ pe_as_2d = F.interpolate(
380
+ pe_as_2d.unsqueeze(0), size=target_dim, mode="bilinear"
381
+ )
382
+ pe_new = pe_as_2d.squeeze(0).permute(1, 2, 0).flatten(0, 1)
383
+ self.positional_encoding.data = pe_new.unsqueeze(0).contiguous()
384
+ self.h_max, self.w_max = target_dim
385
+ print("PE extended to", target_dim)
386
+
387
+ def pe_selection_index_based_on_dim(self, h, w):
388
+ h_p, w_p = h // self.patch_size, w // self.patch_size
389
+ original_pe_indexes = torch.arange(self.positional_encoding.shape[1])
390
+ original_pe_indexes = original_pe_indexes.view(self.h_max, self.w_max)
391
+ starth = self.h_max // 2 - h_p // 2
392
+ endh =starth + h_p
393
+ startw = self.w_max // 2 - w_p // 2
394
+ endw = startw + w_p
395
+ original_pe_indexes = original_pe_indexes[
396
+ starth:endh, startw:endw
397
+ ]
398
+ return original_pe_indexes.flatten()
399
+
400
+ def unpatchify(self, x, h, w):
401
+ c = self.out_channels
402
+ p = self.patch_size
403
+
404
+ x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
405
+ x = torch.einsum("nhwpqc->nchpwq", x)
406
+ imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
407
+ return imgs
408
+
409
+ def patchify(self, x):
410
+ B, C, H, W = x.size()
411
+ x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
412
+ x = x.view(
413
+ B,
414
+ C,
415
+ (H + 1) // self.patch_size,
416
+ self.patch_size,
417
+ (W + 1) // self.patch_size,
418
+ self.patch_size,
419
+ )
420
+ x = x.permute(0, 2, 4, 1, 3, 5).flatten(-3).flatten(1, 2)
421
+ return x
422
+
423
+ def apply_pos_embeds(self, x, h, w):
424
+ h = (h + 1) // self.patch_size
425
+ w = (w + 1) // self.patch_size
426
+ max_dim = max(h, w)
427
+
428
+ cur_dim = self.h_max
429
+ pos_encoding = comfy.ops.cast_to_input(self.positional_encoding.reshape(1, cur_dim, cur_dim, -1), x)
430
+
431
+ if max_dim > cur_dim:
432
+ pos_encoding = F.interpolate(pos_encoding.movedim(-1, 1), (max_dim, max_dim), mode="bilinear").movedim(1, -1)
433
+ cur_dim = max_dim
434
+
435
+ from_h = (cur_dim - h) // 2
436
+ from_w = (cur_dim - w) // 2
437
+ pos_encoding = pos_encoding[:,from_h:from_h+h,from_w:from_w+w]
438
+ return x + pos_encoding.reshape(1, -1, self.positional_encoding.shape[-1])
439
+
440
+ def forward(self, x, timestep, context, **kwargs):
441
+ # patchify x, add PE
442
+ b, c, h, w = x.shape
443
+
444
+ # pe_indexes = self.pe_selection_index_based_on_dim(h, w)
445
+ # print(pe_indexes, pe_indexes.shape)
446
+
447
+ x = self.init_x_linear(self.patchify(x)) # B, T_x, D
448
+ x = self.apply_pos_embeds(x, h, w)
449
+ # x = x + self.positional_encoding[:, : x.size(1)].to(device=x.device, dtype=x.dtype)
450
+ # x = x + self.positional_encoding[:, pe_indexes].to(device=x.device, dtype=x.dtype)
451
+
452
+ # process conditions for MMDiT Blocks
453
+ c_seq = context # B, T_c, D_c
454
+ t = timestep
455
+
456
+ c = self.cond_seq_linear(c_seq) # B, T_c, D
457
+ c = torch.cat([comfy.ops.cast_to_input(self.register_tokens, c).repeat(c.size(0), 1, 1), c], dim=1)
458
+
459
+ global_cond = self.t_embedder(t, x.dtype) # B, D
460
+
461
+ if len(self.double_layers) > 0:
462
+ for layer in self.double_layers:
463
+ c, x = layer(c, x, global_cond, **kwargs)
464
+
465
+ if len(self.single_layers) > 0:
466
+ c_len = c.size(1)
467
+ cx = torch.cat([c, x], dim=1)
468
+ for layer in self.single_layers:
469
+ cx = layer(cx, global_cond, **kwargs)
470
+
471
+ x = cx[:, c_len:]
472
+
473
+ fshift, fscale = self.modF(global_cond).chunk(2, dim=1)
474
+
475
+ x = modulate(x, fshift, fscale)
476
+ x = self.final_linear(x)
477
+ x = self.unpatchify(x, (h + 1) // self.patch_size, (w + 1) // self.patch_size)[:,:,:h,:w]
478
+ return x
ComfyUI/comfy/ldm/cascade/common.py ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+
19
+ import torch
20
+ import torch.nn as nn
21
+ from comfy.ldm.modules.attention import optimized_attention
22
+ import comfy.ops
23
+
24
+ class OptimizedAttention(nn.Module):
25
+ def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None):
26
+ super().__init__()
27
+ self.heads = nhead
28
+
29
+ self.to_q = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
30
+ self.to_k = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
31
+ self.to_v = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
32
+
33
+ self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
34
+
35
+ def forward(self, q, k, v):
36
+ q = self.to_q(q)
37
+ k = self.to_k(k)
38
+ v = self.to_v(v)
39
+
40
+ out = optimized_attention(q, k, v, self.heads)
41
+
42
+ return self.out_proj(out)
43
+
44
+ class Attention2D(nn.Module):
45
+ def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None):
46
+ super().__init__()
47
+ self.attn = OptimizedAttention(c, nhead, dtype=dtype, device=device, operations=operations)
48
+ # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True, dtype=dtype, device=device)
49
+
50
+ def forward(self, x, kv, self_attn=False):
51
+ orig_shape = x.shape
52
+ x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1) # Bx4xHxW -> Bx(HxW)x4
53
+ if self_attn:
54
+ kv = torch.cat([x, kv], dim=1)
55
+ # x = self.attn(x, kv, kv, need_weights=False)[0]
56
+ x = self.attn(x, kv, kv)
57
+ x = x.permute(0, 2, 1).view(*orig_shape)
58
+ return x
59
+
60
+
61
+ def LayerNorm2d_op(operations):
62
+ class LayerNorm2d(operations.LayerNorm):
63
+ def __init__(self, *args, **kwargs):
64
+ super().__init__(*args, **kwargs)
65
+
66
+ def forward(self, x):
67
+ return super().forward(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
68
+ return LayerNorm2d
69
+
70
+ class GlobalResponseNorm(nn.Module):
71
+ "from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105"
72
+ def __init__(self, dim, dtype=None, device=None):
73
+ super().__init__()
74
+ self.gamma = nn.Parameter(torch.empty(1, 1, 1, dim, dtype=dtype, device=device))
75
+ self.beta = nn.Parameter(torch.empty(1, 1, 1, dim, dtype=dtype, device=device))
76
+
77
+ def forward(self, x):
78
+ Gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
79
+ Nx = Gx / (Gx.mean(dim=-1, keepdim=True) + 1e-6)
80
+ return comfy.ops.cast_to_input(self.gamma, x) * (x * Nx) + comfy.ops.cast_to_input(self.beta, x) + x
81
+
82
+
83
+ class ResBlock(nn.Module):
84
+ def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0, dtype=None, device=None, operations=None): # , num_heads=4, expansion=2):
85
+ super().__init__()
86
+ self.depthwise = operations.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c, dtype=dtype, device=device)
87
+ # self.depthwise = SAMBlock(c, num_heads, expansion)
88
+ self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
89
+ self.channelwise = nn.Sequential(
90
+ operations.Linear(c + c_skip, c * 4, dtype=dtype, device=device),
91
+ nn.GELU(),
92
+ GlobalResponseNorm(c * 4, dtype=dtype, device=device),
93
+ nn.Dropout(dropout),
94
+ operations.Linear(c * 4, c, dtype=dtype, device=device)
95
+ )
96
+
97
+ def forward(self, x, x_skip=None):
98
+ x_res = x
99
+ x = self.norm(self.depthwise(x))
100
+ if x_skip is not None:
101
+ x = torch.cat([x, x_skip], dim=1)
102
+ x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
103
+ return x + x_res
104
+
105
+
106
+ class AttnBlock(nn.Module):
107
+ def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0, dtype=None, device=None, operations=None):
108
+ super().__init__()
109
+ self.self_attn = self_attn
110
+ self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
111
+ self.attention = Attention2D(c, nhead, dropout, dtype=dtype, device=device, operations=operations)
112
+ self.kv_mapper = nn.Sequential(
113
+ nn.SiLU(),
114
+ operations.Linear(c_cond, c, dtype=dtype, device=device)
115
+ )
116
+
117
+ def forward(self, x, kv):
118
+ kv = self.kv_mapper(kv)
119
+ x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn)
120
+ return x
121
+
122
+
123
+ class FeedForwardBlock(nn.Module):
124
+ def __init__(self, c, dropout=0.0, dtype=None, device=None, operations=None):
125
+ super().__init__()
126
+ self.norm = LayerNorm2d_op(operations)(c, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
127
+ self.channelwise = nn.Sequential(
128
+ operations.Linear(c, c * 4, dtype=dtype, device=device),
129
+ nn.GELU(),
130
+ GlobalResponseNorm(c * 4, dtype=dtype, device=device),
131
+ nn.Dropout(dropout),
132
+ operations.Linear(c * 4, c, dtype=dtype, device=device)
133
+ )
134
+
135
+ def forward(self, x):
136
+ x = x + self.channelwise(self.norm(x).permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
137
+ return x
138
+
139
+
140
+ class TimestepBlock(nn.Module):
141
+ def __init__(self, c, c_timestep, conds=['sca'], dtype=None, device=None, operations=None):
142
+ super().__init__()
143
+ self.mapper = operations.Linear(c_timestep, c * 2, dtype=dtype, device=device)
144
+ self.conds = conds
145
+ for cname in conds:
146
+ setattr(self, f"mapper_{cname}", operations.Linear(c_timestep, c * 2, dtype=dtype, device=device))
147
+
148
+ def forward(self, x, t):
149
+ t = t.chunk(len(self.conds) + 1, dim=1)
150
+ a, b = self.mapper(t[0])[:, :, None, None].chunk(2, dim=1)
151
+ for i, c in enumerate(self.conds):
152
+ ac, bc = getattr(self, f"mapper_{c}")(t[i + 1])[:, :, None, None].chunk(2, dim=1)
153
+ a, b = a + ac, b + bc
154
+ return x * (1 + a) + b
ComfyUI/comfy/ldm/cascade/controlnet.py ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+
19
+ import torch
20
+ import torchvision
21
+ from torch import nn
22
+ from .common import LayerNorm2d_op
23
+
24
+
25
+ class CNetResBlock(nn.Module):
26
+ def __init__(self, c, dtype=None, device=None, operations=None):
27
+ super().__init__()
28
+ self.blocks = nn.Sequential(
29
+ LayerNorm2d_op(operations)(c, dtype=dtype, device=device),
30
+ nn.GELU(),
31
+ operations.Conv2d(c, c, kernel_size=3, padding=1),
32
+ LayerNorm2d_op(operations)(c, dtype=dtype, device=device),
33
+ nn.GELU(),
34
+ operations.Conv2d(c, c, kernel_size=3, padding=1),
35
+ )
36
+
37
+ def forward(self, x):
38
+ return x + self.blocks(x)
39
+
40
+
41
+ class ControlNet(nn.Module):
42
+ def __init__(self, c_in=3, c_proj=2048, proj_blocks=None, bottleneck_mode=None, dtype=None, device=None, operations=nn):
43
+ super().__init__()
44
+ if bottleneck_mode is None:
45
+ bottleneck_mode = 'effnet'
46
+ self.proj_blocks = proj_blocks
47
+ if bottleneck_mode == 'effnet':
48
+ embd_channels = 1280
49
+ self.backbone = torchvision.models.efficientnet_v2_s().features.eval()
50
+ if c_in != 3:
51
+ in_weights = self.backbone[0][0].weight.data
52
+ self.backbone[0][0] = operations.Conv2d(c_in, 24, kernel_size=3, stride=2, bias=False, dtype=dtype, device=device)
53
+ if c_in > 3:
54
+ # nn.init.constant_(self.backbone[0][0].weight, 0)
55
+ self.backbone[0][0].weight.data[:, :3] = in_weights[:, :3].clone()
56
+ else:
57
+ self.backbone[0][0].weight.data = in_weights[:, :c_in].clone()
58
+ elif bottleneck_mode == 'simple':
59
+ embd_channels = c_in
60
+ self.backbone = nn.Sequential(
61
+ operations.Conv2d(embd_channels, embd_channels * 4, kernel_size=3, padding=1, dtype=dtype, device=device),
62
+ nn.LeakyReLU(0.2, inplace=True),
63
+ operations.Conv2d(embd_channels * 4, embd_channels, kernel_size=3, padding=1, dtype=dtype, device=device),
64
+ )
65
+ elif bottleneck_mode == 'large':
66
+ self.backbone = nn.Sequential(
67
+ operations.Conv2d(c_in, 4096 * 4, kernel_size=1, dtype=dtype, device=device),
68
+ nn.LeakyReLU(0.2, inplace=True),
69
+ operations.Conv2d(4096 * 4, 1024, kernel_size=1, dtype=dtype, device=device),
70
+ *[CNetResBlock(1024, dtype=dtype, device=device, operations=operations) for _ in range(8)],
71
+ operations.Conv2d(1024, 1280, kernel_size=1, dtype=dtype, device=device),
72
+ )
73
+ embd_channels = 1280
74
+ else:
75
+ raise ValueError(f'Unknown bottleneck mode: {bottleneck_mode}')
76
+ self.projections = nn.ModuleList()
77
+ for _ in range(len(proj_blocks)):
78
+ self.projections.append(nn.Sequential(
79
+ operations.Conv2d(embd_channels, embd_channels, kernel_size=1, bias=False, dtype=dtype, device=device),
80
+ nn.LeakyReLU(0.2, inplace=True),
81
+ operations.Conv2d(embd_channels, c_proj, kernel_size=1, bias=False, dtype=dtype, device=device),
82
+ ))
83
+ # nn.init.constant_(self.projections[-1][-1].weight, 0) # zero output projection
84
+ self.xl = False
85
+ self.input_channels = c_in
86
+ self.unshuffle_amount = 8
87
+
88
+ def forward(self, x):
89
+ x = self.backbone(x)
90
+ proj_outputs = [None for _ in range(max(self.proj_blocks) + 1)]
91
+ for i, idx in enumerate(self.proj_blocks):
92
+ proj_outputs[idx] = self.projections[i](x)
93
+ return {"input": proj_outputs[::-1]}
ComfyUI/comfy/ldm/cascade/stage_a.py ADDED
@@ -0,0 +1,255 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+
19
+ import torch
20
+ from torch import nn
21
+ from torch.autograd import Function
22
+
23
+ class vector_quantize(Function):
24
+ @staticmethod
25
+ def forward(ctx, x, codebook):
26
+ with torch.no_grad():
27
+ codebook_sqr = torch.sum(codebook ** 2, dim=1)
28
+ x_sqr = torch.sum(x ** 2, dim=1, keepdim=True)
29
+
30
+ dist = torch.addmm(codebook_sqr + x_sqr, x, codebook.t(), alpha=-2.0, beta=1.0)
31
+ _, indices = dist.min(dim=1)
32
+
33
+ ctx.save_for_backward(indices, codebook)
34
+ ctx.mark_non_differentiable(indices)
35
+
36
+ nn = torch.index_select(codebook, 0, indices)
37
+ return nn, indices
38
+
39
+ @staticmethod
40
+ def backward(ctx, grad_output, grad_indices):
41
+ grad_inputs, grad_codebook = None, None
42
+
43
+ if ctx.needs_input_grad[0]:
44
+ grad_inputs = grad_output.clone()
45
+ if ctx.needs_input_grad[1]:
46
+ # Gradient wrt. the codebook
47
+ indices, codebook = ctx.saved_tensors
48
+
49
+ grad_codebook = torch.zeros_like(codebook)
50
+ grad_codebook.index_add_(0, indices, grad_output)
51
+
52
+ return (grad_inputs, grad_codebook)
53
+
54
+
55
+ class VectorQuantize(nn.Module):
56
+ def __init__(self, embedding_size, k, ema_decay=0.99, ema_loss=False):
57
+ """
58
+ Takes an input of variable size (as long as the last dimension matches the embedding size).
59
+ Returns one tensor containing the nearest neigbour embeddings to each of the inputs,
60
+ with the same size as the input, vq and commitment components for the loss as a touple
61
+ in the second output and the indices of the quantized vectors in the third:
62
+ quantized, (vq_loss, commit_loss), indices
63
+ """
64
+ super(VectorQuantize, self).__init__()
65
+
66
+ self.codebook = nn.Embedding(k, embedding_size)
67
+ self.codebook.weight.data.uniform_(-1./k, 1./k)
68
+ self.vq = vector_quantize.apply
69
+
70
+ self.ema_decay = ema_decay
71
+ self.ema_loss = ema_loss
72
+ if ema_loss:
73
+ self.register_buffer('ema_element_count', torch.ones(k))
74
+ self.register_buffer('ema_weight_sum', torch.zeros_like(self.codebook.weight))
75
+
76
+ def _laplace_smoothing(self, x, epsilon):
77
+ n = torch.sum(x)
78
+ return ((x + epsilon) / (n + x.size(0) * epsilon) * n)
79
+
80
+ def _updateEMA(self, z_e_x, indices):
81
+ mask = nn.functional.one_hot(indices, self.ema_element_count.size(0)).float()
82
+ elem_count = mask.sum(dim=0)
83
+ weight_sum = torch.mm(mask.t(), z_e_x)
84
+
85
+ self.ema_element_count = (self.ema_decay * self.ema_element_count) + ((1-self.ema_decay) * elem_count)
86
+ self.ema_element_count = self._laplace_smoothing(self.ema_element_count, 1e-5)
87
+ self.ema_weight_sum = (self.ema_decay * self.ema_weight_sum) + ((1-self.ema_decay) * weight_sum)
88
+
89
+ self.codebook.weight.data = self.ema_weight_sum / self.ema_element_count.unsqueeze(-1)
90
+
91
+ def idx2vq(self, idx, dim=-1):
92
+ q_idx = self.codebook(idx)
93
+ if dim != -1:
94
+ q_idx = q_idx.movedim(-1, dim)
95
+ return q_idx
96
+
97
+ def forward(self, x, get_losses=True, dim=-1):
98
+ if dim != -1:
99
+ x = x.movedim(dim, -1)
100
+ z_e_x = x.contiguous().view(-1, x.size(-1)) if len(x.shape) > 2 else x
101
+ z_q_x, indices = self.vq(z_e_x, self.codebook.weight.detach())
102
+ vq_loss, commit_loss = None, None
103
+ if self.ema_loss and self.training:
104
+ self._updateEMA(z_e_x.detach(), indices.detach())
105
+ # pick the graded embeddings after updating the codebook in order to have a more accurate commitment loss
106
+ z_q_x_grd = torch.index_select(self.codebook.weight, dim=0, index=indices)
107
+ if get_losses:
108
+ vq_loss = (z_q_x_grd - z_e_x.detach()).pow(2).mean()
109
+ commit_loss = (z_e_x - z_q_x_grd.detach()).pow(2).mean()
110
+
111
+ z_q_x = z_q_x.view(x.shape)
112
+ if dim != -1:
113
+ z_q_x = z_q_x.movedim(-1, dim)
114
+ return z_q_x, (vq_loss, commit_loss), indices.view(x.shape[:-1])
115
+
116
+
117
+ class ResBlock(nn.Module):
118
+ def __init__(self, c, c_hidden):
119
+ super().__init__()
120
+ # depthwise/attention
121
+ self.norm1 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
122
+ self.depthwise = nn.Sequential(
123
+ nn.ReplicationPad2d(1),
124
+ nn.Conv2d(c, c, kernel_size=3, groups=c)
125
+ )
126
+
127
+ # channelwise
128
+ self.norm2 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
129
+ self.channelwise = nn.Sequential(
130
+ nn.Linear(c, c_hidden),
131
+ nn.GELU(),
132
+ nn.Linear(c_hidden, c),
133
+ )
134
+
135
+ self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True)
136
+
137
+ # Init weights
138
+ def _basic_init(module):
139
+ if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
140
+ torch.nn.init.xavier_uniform_(module.weight)
141
+ if module.bias is not None:
142
+ nn.init.constant_(module.bias, 0)
143
+
144
+ self.apply(_basic_init)
145
+
146
+ def _norm(self, x, norm):
147
+ return norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
148
+
149
+ def forward(self, x):
150
+ mods = self.gammas
151
+
152
+ x_temp = self._norm(x, self.norm1) * (1 + mods[0]) + mods[1]
153
+ try:
154
+ x = x + self.depthwise(x_temp) * mods[2]
155
+ except: #operation not implemented for bf16
156
+ x_temp = self.depthwise[0](x_temp.float()).to(x.dtype)
157
+ x = x + self.depthwise[1](x_temp) * mods[2]
158
+
159
+ x_temp = self._norm(x, self.norm2) * (1 + mods[3]) + mods[4]
160
+ x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5]
161
+
162
+ return x
163
+
164
+
165
+ class StageA(nn.Module):
166
+ def __init__(self, levels=2, bottleneck_blocks=12, c_hidden=384, c_latent=4, codebook_size=8192):
167
+ super().__init__()
168
+ self.c_latent = c_latent
169
+ c_levels = [c_hidden // (2 ** i) for i in reversed(range(levels))]
170
+
171
+ # Encoder blocks
172
+ self.in_block = nn.Sequential(
173
+ nn.PixelUnshuffle(2),
174
+ nn.Conv2d(3 * 4, c_levels[0], kernel_size=1)
175
+ )
176
+ down_blocks = []
177
+ for i in range(levels):
178
+ if i > 0:
179
+ down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1))
180
+ block = ResBlock(c_levels[i], c_levels[i] * 4)
181
+ down_blocks.append(block)
182
+ down_blocks.append(nn.Sequential(
183
+ nn.Conv2d(c_levels[-1], c_latent, kernel_size=1, bias=False),
184
+ nn.BatchNorm2d(c_latent), # then normalize them to have mean 0 and std 1
185
+ ))
186
+ self.down_blocks = nn.Sequential(*down_blocks)
187
+ self.down_blocks[0]
188
+
189
+ self.codebook_size = codebook_size
190
+ self.vquantizer = VectorQuantize(c_latent, k=codebook_size)
191
+
192
+ # Decoder blocks
193
+ up_blocks = [nn.Sequential(
194
+ nn.Conv2d(c_latent, c_levels[-1], kernel_size=1)
195
+ )]
196
+ for i in range(levels):
197
+ for j in range(bottleneck_blocks if i == 0 else 1):
198
+ block = ResBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4)
199
+ up_blocks.append(block)
200
+ if i < levels - 1:
201
+ up_blocks.append(
202
+ nn.ConvTranspose2d(c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2,
203
+ padding=1))
204
+ self.up_blocks = nn.Sequential(*up_blocks)
205
+ self.out_block = nn.Sequential(
206
+ nn.Conv2d(c_levels[0], 3 * 4, kernel_size=1),
207
+ nn.PixelShuffle(2),
208
+ )
209
+
210
+ def encode(self, x, quantize=False):
211
+ x = self.in_block(x)
212
+ x = self.down_blocks(x)
213
+ if quantize:
214
+ qe, (vq_loss, commit_loss), indices = self.vquantizer.forward(x, dim=1)
215
+ return qe, x, indices, vq_loss + commit_loss * 0.25
216
+ else:
217
+ return x
218
+
219
+ def decode(self, x):
220
+ x = self.up_blocks(x)
221
+ x = self.out_block(x)
222
+ return x
223
+
224
+ def forward(self, x, quantize=False):
225
+ qe, x, _, vq_loss = self.encode(x, quantize)
226
+ x = self.decode(qe)
227
+ return x, vq_loss
228
+
229
+
230
+ class Discriminator(nn.Module):
231
+ def __init__(self, c_in=3, c_cond=0, c_hidden=512, depth=6):
232
+ super().__init__()
233
+ d = max(depth - 3, 3)
234
+ layers = [
235
+ nn.utils.spectral_norm(nn.Conv2d(c_in, c_hidden // (2 ** d), kernel_size=3, stride=2, padding=1)),
236
+ nn.LeakyReLU(0.2),
237
+ ]
238
+ for i in range(depth - 1):
239
+ c_in = c_hidden // (2 ** max((d - i), 0))
240
+ c_out = c_hidden // (2 ** max((d - 1 - i), 0))
241
+ layers.append(nn.utils.spectral_norm(nn.Conv2d(c_in, c_out, kernel_size=3, stride=2, padding=1)))
242
+ layers.append(nn.InstanceNorm2d(c_out))
243
+ layers.append(nn.LeakyReLU(0.2))
244
+ self.encoder = nn.Sequential(*layers)
245
+ self.shuffle = nn.Conv2d((c_hidden + c_cond) if c_cond > 0 else c_hidden, 1, kernel_size=1)
246
+ self.logits = nn.Sigmoid()
247
+
248
+ def forward(self, x, cond=None):
249
+ x = self.encoder(x)
250
+ if cond is not None:
251
+ cond = cond.view(cond.size(0), cond.size(1), 1, 1, ).expand(-1, -1, x.size(-2), x.size(-1))
252
+ x = torch.cat([x, cond], dim=1)
253
+ x = self.shuffle(x)
254
+ x = self.logits(x)
255
+ return x
ComfyUI/comfy/ldm/cascade/stage_b.py ADDED
@@ -0,0 +1,256 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+
19
+ import math
20
+ import torch
21
+ from torch import nn
22
+ from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock
23
+
24
+ class StageB(nn.Module):
25
+ def __init__(self, c_in=4, c_out=4, c_r=64, patch_size=2, c_cond=1280, c_hidden=[320, 640, 1280, 1280],
26
+ nhead=[-1, -1, 20, 20], blocks=[[2, 6, 28, 6], [6, 28, 6, 2]],
27
+ block_repeat=[[1, 1, 1, 1], [3, 3, 2, 2]], level_config=['CT', 'CT', 'CTA', 'CTA'], c_clip=1280,
28
+ c_clip_seq=4, c_effnet=16, c_pixels=3, kernel_size=3, dropout=[0, 0, 0.0, 0.0], self_attn=True,
29
+ t_conds=['sca'], stable_cascade_stage=None, dtype=None, device=None, operations=None):
30
+ super().__init__()
31
+ self.dtype = dtype
32
+ self.c_r = c_r
33
+ self.t_conds = t_conds
34
+ self.c_clip_seq = c_clip_seq
35
+ if not isinstance(dropout, list):
36
+ dropout = [dropout] * len(c_hidden)
37
+ if not isinstance(self_attn, list):
38
+ self_attn = [self_attn] * len(c_hidden)
39
+
40
+ # CONDITIONING
41
+ self.effnet_mapper = nn.Sequential(
42
+ operations.Conv2d(c_effnet, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device),
43
+ nn.GELU(),
44
+ operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device),
45
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
46
+ )
47
+ self.pixels_mapper = nn.Sequential(
48
+ operations.Conv2d(c_pixels, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device),
49
+ nn.GELU(),
50
+ operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device),
51
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
52
+ )
53
+ self.clip_mapper = operations.Linear(c_clip, c_cond * c_clip_seq, dtype=dtype, device=device)
54
+ self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
55
+
56
+ self.embedding = nn.Sequential(
57
+ nn.PixelUnshuffle(patch_size),
58
+ operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device),
59
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
60
+ )
61
+
62
+ def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True):
63
+ if block_type == 'C':
64
+ return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations)
65
+ elif block_type == 'A':
66
+ return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations)
67
+ elif block_type == 'F':
68
+ return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations)
69
+ elif block_type == 'T':
70
+ return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations)
71
+ else:
72
+ raise Exception(f'Block type {block_type} not supported')
73
+
74
+ # BLOCKS
75
+ # -- down blocks
76
+ self.down_blocks = nn.ModuleList()
77
+ self.down_downscalers = nn.ModuleList()
78
+ self.down_repeat_mappers = nn.ModuleList()
79
+ for i in range(len(c_hidden)):
80
+ if i > 0:
81
+ self.down_downscalers.append(nn.Sequential(
82
+ LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
83
+ operations.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2, dtype=dtype, device=device),
84
+ ))
85
+ else:
86
+ self.down_downscalers.append(nn.Identity())
87
+ down_block = nn.ModuleList()
88
+ for _ in range(blocks[0][i]):
89
+ for block_type in level_config[i]:
90
+ block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i])
91
+ down_block.append(block)
92
+ self.down_blocks.append(down_block)
93
+ if block_repeat is not None:
94
+ block_repeat_mappers = nn.ModuleList()
95
+ for _ in range(block_repeat[0][i] - 1):
96
+ block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
97
+ self.down_repeat_mappers.append(block_repeat_mappers)
98
+
99
+ # -- up blocks
100
+ self.up_blocks = nn.ModuleList()
101
+ self.up_upscalers = nn.ModuleList()
102
+ self.up_repeat_mappers = nn.ModuleList()
103
+ for i in reversed(range(len(c_hidden))):
104
+ if i > 0:
105
+ self.up_upscalers.append(nn.Sequential(
106
+ LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
107
+ operations.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2, dtype=dtype, device=device),
108
+ ))
109
+ else:
110
+ self.up_upscalers.append(nn.Identity())
111
+ up_block = nn.ModuleList()
112
+ for j in range(blocks[1][::-1][i]):
113
+ for k, block_type in enumerate(level_config[i]):
114
+ c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0
115
+ block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i],
116
+ self_attn=self_attn[i])
117
+ up_block.append(block)
118
+ self.up_blocks.append(up_block)
119
+ if block_repeat is not None:
120
+ block_repeat_mappers = nn.ModuleList()
121
+ for _ in range(block_repeat[1][::-1][i] - 1):
122
+ block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
123
+ self.up_repeat_mappers.append(block_repeat_mappers)
124
+
125
+ # OUTPUT
126
+ self.clf = nn.Sequential(
127
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
128
+ operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device),
129
+ nn.PixelShuffle(patch_size),
130
+ )
131
+
132
+ # --- WEIGHT INIT ---
133
+ # self.apply(self._init_weights) # General init
134
+ # nn.init.normal_(self.clip_mapper.weight, std=0.02) # conditionings
135
+ # nn.init.normal_(self.effnet_mapper[0].weight, std=0.02) # conditionings
136
+ # nn.init.normal_(self.effnet_mapper[2].weight, std=0.02) # conditionings
137
+ # nn.init.normal_(self.pixels_mapper[0].weight, std=0.02) # conditionings
138
+ # nn.init.normal_(self.pixels_mapper[2].weight, std=0.02) # conditionings
139
+ # torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs
140
+ # nn.init.constant_(self.clf[1].weight, 0) # outputs
141
+ #
142
+ # # blocks
143
+ # for level_block in self.down_blocks + self.up_blocks:
144
+ # for block in level_block:
145
+ # if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock):
146
+ # block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0]))
147
+ # elif isinstance(block, TimestepBlock):
148
+ # for layer in block.modules():
149
+ # if isinstance(layer, nn.Linear):
150
+ # nn.init.constant_(layer.weight, 0)
151
+ #
152
+ # def _init_weights(self, m):
153
+ # if isinstance(m, (nn.Conv2d, nn.Linear)):
154
+ # torch.nn.init.xavier_uniform_(m.weight)
155
+ # if m.bias is not None:
156
+ # nn.init.constant_(m.bias, 0)
157
+
158
+ def gen_r_embedding(self, r, max_positions=10000):
159
+ r = r * max_positions
160
+ half_dim = self.c_r // 2
161
+ emb = math.log(max_positions) / (half_dim - 1)
162
+ emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
163
+ emb = r[:, None] * emb[None, :]
164
+ emb = torch.cat([emb.sin(), emb.cos()], dim=1)
165
+ if self.c_r % 2 == 1: # zero pad
166
+ emb = nn.functional.pad(emb, (0, 1), mode='constant')
167
+ return emb
168
+
169
+ def gen_c_embeddings(self, clip):
170
+ if len(clip.shape) == 2:
171
+ clip = clip.unsqueeze(1)
172
+ clip = self.clip_mapper(clip).view(clip.size(0), clip.size(1) * self.c_clip_seq, -1)
173
+ clip = self.clip_norm(clip)
174
+ return clip
175
+
176
+ def _down_encode(self, x, r_embed, clip):
177
+ level_outputs = []
178
+ block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
179
+ for down_block, downscaler, repmap in block_group:
180
+ x = downscaler(x)
181
+ for i in range(len(repmap) + 1):
182
+ for block in down_block:
183
+ if isinstance(block, ResBlock) or (
184
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
185
+ ResBlock)):
186
+ x = block(x)
187
+ elif isinstance(block, AttnBlock) or (
188
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
189
+ AttnBlock)):
190
+ x = block(x, clip)
191
+ elif isinstance(block, TimestepBlock) or (
192
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
193
+ TimestepBlock)):
194
+ x = block(x, r_embed)
195
+ else:
196
+ x = block(x)
197
+ if i < len(repmap):
198
+ x = repmap[i](x)
199
+ level_outputs.insert(0, x)
200
+ return level_outputs
201
+
202
+ def _up_decode(self, level_outputs, r_embed, clip):
203
+ x = level_outputs[0]
204
+ block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
205
+ for i, (up_block, upscaler, repmap) in enumerate(block_group):
206
+ for j in range(len(repmap) + 1):
207
+ for k, block in enumerate(up_block):
208
+ if isinstance(block, ResBlock) or (
209
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
210
+ ResBlock)):
211
+ skip = level_outputs[i] if k == 0 and i > 0 else None
212
+ if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
213
+ x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear',
214
+ align_corners=True)
215
+ x = block(x, skip)
216
+ elif isinstance(block, AttnBlock) or (
217
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
218
+ AttnBlock)):
219
+ x = block(x, clip)
220
+ elif isinstance(block, TimestepBlock) or (
221
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
222
+ TimestepBlock)):
223
+ x = block(x, r_embed)
224
+ else:
225
+ x = block(x)
226
+ if j < len(repmap):
227
+ x = repmap[j](x)
228
+ x = upscaler(x)
229
+ return x
230
+
231
+ def forward(self, x, r, effnet, clip, pixels=None, **kwargs):
232
+ if pixels is None:
233
+ pixels = x.new_zeros(x.size(0), 3, 8, 8)
234
+
235
+ # Process the conditioning embeddings
236
+ r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
237
+ for c in self.t_conds:
238
+ t_cond = kwargs.get(c, torch.zeros_like(r))
239
+ r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1)
240
+ clip = self.gen_c_embeddings(clip)
241
+
242
+ # Model Blocks
243
+ x = self.embedding(x)
244
+ x = x + self.effnet_mapper(
245
+ nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True))
246
+ x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear',
247
+ align_corners=True)
248
+ level_outputs = self._down_encode(x, r_embed, clip)
249
+ x = self._up_decode(level_outputs, r_embed, clip)
250
+ return self.clf(x)
251
+
252
+ def update_weights_ema(self, src_model, beta=0.999):
253
+ for self_params, src_params in zip(self.parameters(), src_model.parameters()):
254
+ self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta)
255
+ for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()):
256
+ self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta)
ComfyUI/comfy/ldm/cascade/stage_c.py ADDED
@@ -0,0 +1,273 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+
19
+ import torch
20
+ from torch import nn
21
+ import math
22
+ from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock
23
+ # from .controlnet import ControlNetDeliverer
24
+
25
+ class UpDownBlock2d(nn.Module):
26
+ def __init__(self, c_in, c_out, mode, enabled=True, dtype=None, device=None, operations=None):
27
+ super().__init__()
28
+ assert mode in ['up', 'down']
29
+ interpolation = nn.Upsample(scale_factor=2 if mode == 'up' else 0.5, mode='bilinear',
30
+ align_corners=True) if enabled else nn.Identity()
31
+ mapping = operations.Conv2d(c_in, c_out, kernel_size=1, dtype=dtype, device=device)
32
+ self.blocks = nn.ModuleList([interpolation, mapping] if mode == 'up' else [mapping, interpolation])
33
+
34
+ def forward(self, x):
35
+ for block in self.blocks:
36
+ x = block(x)
37
+ return x
38
+
39
+
40
+ class StageC(nn.Module):
41
+ def __init__(self, c_in=16, c_out=16, c_r=64, patch_size=1, c_cond=2048, c_hidden=[2048, 2048], nhead=[32, 32],
42
+ blocks=[[8, 24], [24, 8]], block_repeat=[[1, 1], [1, 1]], level_config=['CTA', 'CTA'],
43
+ c_clip_text=1280, c_clip_text_pooled=1280, c_clip_img=768, c_clip_seq=4, kernel_size=3,
44
+ dropout=[0.0, 0.0], self_attn=True, t_conds=['sca', 'crp'], switch_level=[False], stable_cascade_stage=None,
45
+ dtype=None, device=None, operations=None):
46
+ super().__init__()
47
+ self.dtype = dtype
48
+ self.c_r = c_r
49
+ self.t_conds = t_conds
50
+ self.c_clip_seq = c_clip_seq
51
+ if not isinstance(dropout, list):
52
+ dropout = [dropout] * len(c_hidden)
53
+ if not isinstance(self_attn, list):
54
+ self_attn = [self_attn] * len(c_hidden)
55
+
56
+ # CONDITIONING
57
+ self.clip_txt_mapper = operations.Linear(c_clip_text, c_cond, dtype=dtype, device=device)
58
+ self.clip_txt_pooled_mapper = operations.Linear(c_clip_text_pooled, c_cond * c_clip_seq, dtype=dtype, device=device)
59
+ self.clip_img_mapper = operations.Linear(c_clip_img, c_cond * c_clip_seq, dtype=dtype, device=device)
60
+ self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
61
+
62
+ self.embedding = nn.Sequential(
63
+ nn.PixelUnshuffle(patch_size),
64
+ operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device),
65
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6)
66
+ )
67
+
68
+ def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True):
69
+ if block_type == 'C':
70
+ return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations)
71
+ elif block_type == 'A':
72
+ return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations)
73
+ elif block_type == 'F':
74
+ return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations)
75
+ elif block_type == 'T':
76
+ return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations)
77
+ else:
78
+ raise Exception(f'Block type {block_type} not supported')
79
+
80
+ # BLOCKS
81
+ # -- down blocks
82
+ self.down_blocks = nn.ModuleList()
83
+ self.down_downscalers = nn.ModuleList()
84
+ self.down_repeat_mappers = nn.ModuleList()
85
+ for i in range(len(c_hidden)):
86
+ if i > 0:
87
+ self.down_downscalers.append(nn.Sequential(
88
+ LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6),
89
+ UpDownBlock2d(c_hidden[i - 1], c_hidden[i], mode='down', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations)
90
+ ))
91
+ else:
92
+ self.down_downscalers.append(nn.Identity())
93
+ down_block = nn.ModuleList()
94
+ for _ in range(blocks[0][i]):
95
+ for block_type in level_config[i]:
96
+ block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i])
97
+ down_block.append(block)
98
+ self.down_blocks.append(down_block)
99
+ if block_repeat is not None:
100
+ block_repeat_mappers = nn.ModuleList()
101
+ for _ in range(block_repeat[0][i] - 1):
102
+ block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
103
+ self.down_repeat_mappers.append(block_repeat_mappers)
104
+
105
+ # -- up blocks
106
+ self.up_blocks = nn.ModuleList()
107
+ self.up_upscalers = nn.ModuleList()
108
+ self.up_repeat_mappers = nn.ModuleList()
109
+ for i in reversed(range(len(c_hidden))):
110
+ if i > 0:
111
+ self.up_upscalers.append(nn.Sequential(
112
+ LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6),
113
+ UpDownBlock2d(c_hidden[i], c_hidden[i - 1], mode='up', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations)
114
+ ))
115
+ else:
116
+ self.up_upscalers.append(nn.Identity())
117
+ up_block = nn.ModuleList()
118
+ for j in range(blocks[1][::-1][i]):
119
+ for k, block_type in enumerate(level_config[i]):
120
+ c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0
121
+ block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i],
122
+ self_attn=self_attn[i])
123
+ up_block.append(block)
124
+ self.up_blocks.append(up_block)
125
+ if block_repeat is not None:
126
+ block_repeat_mappers = nn.ModuleList()
127
+ for _ in range(block_repeat[1][::-1][i] - 1):
128
+ block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
129
+ self.up_repeat_mappers.append(block_repeat_mappers)
130
+
131
+ # OUTPUT
132
+ self.clf = nn.Sequential(
133
+ LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
134
+ operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device),
135
+ nn.PixelShuffle(patch_size),
136
+ )
137
+
138
+ # --- WEIGHT INIT ---
139
+ # self.apply(self._init_weights) # General init
140
+ # nn.init.normal_(self.clip_txt_mapper.weight, std=0.02) # conditionings
141
+ # nn.init.normal_(self.clip_txt_pooled_mapper.weight, std=0.02) # conditionings
142
+ # nn.init.normal_(self.clip_img_mapper.weight, std=0.02) # conditionings
143
+ # torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs
144
+ # nn.init.constant_(self.clf[1].weight, 0) # outputs
145
+ #
146
+ # # blocks
147
+ # for level_block in self.down_blocks + self.up_blocks:
148
+ # for block in level_block:
149
+ # if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock):
150
+ # block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0]))
151
+ # elif isinstance(block, TimestepBlock):
152
+ # for layer in block.modules():
153
+ # if isinstance(layer, nn.Linear):
154
+ # nn.init.constant_(layer.weight, 0)
155
+ #
156
+ # def _init_weights(self, m):
157
+ # if isinstance(m, (nn.Conv2d, nn.Linear)):
158
+ # torch.nn.init.xavier_uniform_(m.weight)
159
+ # if m.bias is not None:
160
+ # nn.init.constant_(m.bias, 0)
161
+
162
+ def gen_r_embedding(self, r, max_positions=10000):
163
+ r = r * max_positions
164
+ half_dim = self.c_r // 2
165
+ emb = math.log(max_positions) / (half_dim - 1)
166
+ emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
167
+ emb = r[:, None] * emb[None, :]
168
+ emb = torch.cat([emb.sin(), emb.cos()], dim=1)
169
+ if self.c_r % 2 == 1: # zero pad
170
+ emb = nn.functional.pad(emb, (0, 1), mode='constant')
171
+ return emb
172
+
173
+ def gen_c_embeddings(self, clip_txt, clip_txt_pooled, clip_img):
174
+ clip_txt = self.clip_txt_mapper(clip_txt)
175
+ if len(clip_txt_pooled.shape) == 2:
176
+ clip_txt_pooled = clip_txt_pooled.unsqueeze(1)
177
+ if len(clip_img.shape) == 2:
178
+ clip_img = clip_img.unsqueeze(1)
179
+ clip_txt_pool = self.clip_txt_pooled_mapper(clip_txt_pooled).view(clip_txt_pooled.size(0), clip_txt_pooled.size(1) * self.c_clip_seq, -1)
180
+ clip_img = self.clip_img_mapper(clip_img).view(clip_img.size(0), clip_img.size(1) * self.c_clip_seq, -1)
181
+ clip = torch.cat([clip_txt, clip_txt_pool, clip_img], dim=1)
182
+ clip = self.clip_norm(clip)
183
+ return clip
184
+
185
+ def _down_encode(self, x, r_embed, clip, cnet=None):
186
+ level_outputs = []
187
+ block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
188
+ for down_block, downscaler, repmap in block_group:
189
+ x = downscaler(x)
190
+ for i in range(len(repmap) + 1):
191
+ for block in down_block:
192
+ if isinstance(block, ResBlock) or (
193
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
194
+ ResBlock)):
195
+ if cnet is not None:
196
+ next_cnet = cnet.pop()
197
+ if next_cnet is not None:
198
+ x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear',
199
+ align_corners=True).to(x.dtype)
200
+ x = block(x)
201
+ elif isinstance(block, AttnBlock) or (
202
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
203
+ AttnBlock)):
204
+ x = block(x, clip)
205
+ elif isinstance(block, TimestepBlock) or (
206
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
207
+ TimestepBlock)):
208
+ x = block(x, r_embed)
209
+ else:
210
+ x = block(x)
211
+ if i < len(repmap):
212
+ x = repmap[i](x)
213
+ level_outputs.insert(0, x)
214
+ return level_outputs
215
+
216
+ def _up_decode(self, level_outputs, r_embed, clip, cnet=None):
217
+ x = level_outputs[0]
218
+ block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
219
+ for i, (up_block, upscaler, repmap) in enumerate(block_group):
220
+ for j in range(len(repmap) + 1):
221
+ for k, block in enumerate(up_block):
222
+ if isinstance(block, ResBlock) or (
223
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
224
+ ResBlock)):
225
+ skip = level_outputs[i] if k == 0 and i > 0 else None
226
+ if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
227
+ x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear',
228
+ align_corners=True)
229
+ if cnet is not None:
230
+ next_cnet = cnet.pop()
231
+ if next_cnet is not None:
232
+ x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear',
233
+ align_corners=True).to(x.dtype)
234
+ x = block(x, skip)
235
+ elif isinstance(block, AttnBlock) or (
236
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
237
+ AttnBlock)):
238
+ x = block(x, clip)
239
+ elif isinstance(block, TimestepBlock) or (
240
+ hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
241
+ TimestepBlock)):
242
+ x = block(x, r_embed)
243
+ else:
244
+ x = block(x)
245
+ if j < len(repmap):
246
+ x = repmap[j](x)
247
+ x = upscaler(x)
248
+ return x
249
+
250
+ def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, **kwargs):
251
+ # Process the conditioning embeddings
252
+ r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
253
+ for c in self.t_conds:
254
+ t_cond = kwargs.get(c, torch.zeros_like(r))
255
+ r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1)
256
+ clip = self.gen_c_embeddings(clip_text, clip_text_pooled, clip_img)
257
+
258
+ if control is not None:
259
+ cnet = control.get("input")
260
+ else:
261
+ cnet = None
262
+
263
+ # Model Blocks
264
+ x = self.embedding(x)
265
+ level_outputs = self._down_encode(x, r_embed, clip, cnet)
266
+ x = self._up_decode(level_outputs, r_embed, clip, cnet)
267
+ return self.clf(x)
268
+
269
+ def update_weights_ema(self, src_model, beta=0.999):
270
+ for self_params, src_params in zip(self.parameters(), src_model.parameters()):
271
+ self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta)
272
+ for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()):
273
+ self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta)
ComfyUI/comfy/ldm/cascade/stage_c_coder.py ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This file is part of ComfyUI.
3
+ Copyright (C) 2024 Stability AI
4
+
5
+ This program is free software: you can redistribute it and/or modify
6
+ it under the terms of the GNU General Public License as published by
7
+ the Free Software Foundation, either version 3 of the License, or
8
+ (at your option) any later version.
9
+
10
+ This program is distributed in the hope that it will be useful,
11
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
12
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
+ GNU General Public License for more details.
14
+
15
+ You should have received a copy of the GNU General Public License
16
+ along with this program. If not, see <https://www.gnu.org/licenses/>.
17
+ """
18
+ import torch
19
+ import torchvision
20
+ from torch import nn
21
+
22
+
23
+ # EfficientNet
24
+ class EfficientNetEncoder(nn.Module):
25
+ def __init__(self, c_latent=16):
26
+ super().__init__()
27
+ self.backbone = torchvision.models.efficientnet_v2_s().features.eval()
28
+ self.mapper = nn.Sequential(
29
+ nn.Conv2d(1280, c_latent, kernel_size=1, bias=False),
30
+ nn.BatchNorm2d(c_latent, affine=False), # then normalize them to have mean 0 and std 1
31
+ )
32
+ self.mean = nn.Parameter(torch.tensor([0.485, 0.456, 0.406]))
33
+ self.std = nn.Parameter(torch.tensor([0.229, 0.224, 0.225]))
34
+
35
+ def forward(self, x):
36
+ x = x * 0.5 + 0.5
37
+ x = (x - self.mean.view([3,1,1])) / self.std.view([3,1,1])
38
+ o = self.mapper(self.backbone(x))
39
+ return o
40
+
41
+
42
+ # Fast Decoder for Stage C latents. E.g. 16 x 24 x 24 -> 3 x 192 x 192
43
+ class Previewer(nn.Module):
44
+ def __init__(self, c_in=16, c_hidden=512, c_out=3):
45
+ super().__init__()
46
+ self.blocks = nn.Sequential(
47
+ nn.Conv2d(c_in, c_hidden, kernel_size=1), # 16 channels to 512 channels
48
+ nn.GELU(),
49
+ nn.BatchNorm2d(c_hidden),
50
+
51
+ nn.Conv2d(c_hidden, c_hidden, kernel_size=3, padding=1),
52
+ nn.GELU(),
53
+ nn.BatchNorm2d(c_hidden),
54
+
55
+ nn.ConvTranspose2d(c_hidden, c_hidden // 2, kernel_size=2, stride=2), # 16 -> 32
56
+ nn.GELU(),
57
+ nn.BatchNorm2d(c_hidden // 2),
58
+
59
+ nn.Conv2d(c_hidden // 2, c_hidden // 2, kernel_size=3, padding=1),
60
+ nn.GELU(),
61
+ nn.BatchNorm2d(c_hidden // 2),
62
+
63
+ nn.ConvTranspose2d(c_hidden // 2, c_hidden // 4, kernel_size=2, stride=2), # 32 -> 64
64
+ nn.GELU(),
65
+ nn.BatchNorm2d(c_hidden // 4),
66
+
67
+ nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
68
+ nn.GELU(),
69
+ nn.BatchNorm2d(c_hidden // 4),
70
+
71
+ nn.ConvTranspose2d(c_hidden // 4, c_hidden // 4, kernel_size=2, stride=2), # 64 -> 128
72
+ nn.GELU(),
73
+ nn.BatchNorm2d(c_hidden // 4),
74
+
75
+ nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
76
+ nn.GELU(),
77
+ nn.BatchNorm2d(c_hidden // 4),
78
+
79
+ nn.Conv2d(c_hidden // 4, c_out, kernel_size=1),
80
+ )
81
+
82
+ def forward(self, x):
83
+ return (self.blocks(x) - 0.5) * 2.0
84
+
85
+ class StageC_coder(nn.Module):
86
+ def __init__(self):
87
+ super().__init__()
88
+ self.previewer = Previewer()
89
+ self.encoder = EfficientNetEncoder()
90
+
91
+ def encode(self, x):
92
+ return self.encoder(x)
93
+
94
+ def decode(self, x):
95
+ return self.previewer(x)
ComfyUI/comfy/ldm/common_dit.py ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
4
+ if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
5
+ padding_mode = "reflect"
6
+ pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
7
+ pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
8
+ return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)
ComfyUI/comfy/ldm/flux/layers.py ADDED
@@ -0,0 +1,263 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from dataclasses import dataclass
3
+
4
+ import torch
5
+ from einops import rearrange
6
+ from torch import Tensor, nn
7
+
8
+ from .math import attention, rope
9
+ import comfy.ops
10
+
11
+ class EmbedND(nn.Module):
12
+ def __init__(self, dim: int, theta: int, axes_dim: list):
13
+ super().__init__()
14
+ self.dim = dim
15
+ self.theta = theta
16
+ self.axes_dim = axes_dim
17
+
18
+ def forward(self, ids: Tensor) -> Tensor:
19
+ n_axes = ids.shape[-1]
20
+ emb = torch.cat(
21
+ [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
22
+ dim=-3,
23
+ )
24
+
25
+ return emb.unsqueeze(1)
26
+
27
+
28
+ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
29
+ """
30
+ Create sinusoidal timestep embeddings.
31
+ :param t: a 1-D Tensor of N indices, one per batch element.
32
+ These may be fractional.
33
+ :param dim: the dimension of the output.
34
+ :param max_period: controls the minimum frequency of the embeddings.
35
+ :return: an (N, D) Tensor of positional embeddings.
36
+ """
37
+ t = time_factor * t
38
+ half = dim // 2
39
+ freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
40
+ t.device
41
+ )
42
+
43
+ args = t[:, None].float() * freqs[None]
44
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
45
+ if dim % 2:
46
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
47
+ if torch.is_floating_point(t):
48
+ embedding = embedding.to(t)
49
+ return embedding
50
+
51
+
52
+ class MLPEmbedder(nn.Module):
53
+ def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
54
+ super().__init__()
55
+ self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
56
+ self.silu = nn.SiLU()
57
+ self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
58
+
59
+ def forward(self, x: Tensor) -> Tensor:
60
+ return self.out_layer(self.silu(self.in_layer(x)))
61
+
62
+
63
+ class RMSNorm(torch.nn.Module):
64
+ def __init__(self, dim: int, dtype=None, device=None, operations=None):
65
+ super().__init__()
66
+ self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
67
+
68
+ def forward(self, x: Tensor):
69
+ x_dtype = x.dtype
70
+ x = x.float()
71
+ rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
72
+ return (x * rrms).to(dtype=x_dtype) * comfy.ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
73
+
74
+
75
+ class QKNorm(torch.nn.Module):
76
+ def __init__(self, dim: int, dtype=None, device=None, operations=None):
77
+ super().__init__()
78
+ self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
79
+ self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
80
+
81
+ def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
82
+ q = self.query_norm(q)
83
+ k = self.key_norm(k)
84
+ return q.to(v), k.to(v)
85
+
86
+
87
+ class SelfAttention(nn.Module):
88
+ def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
89
+ super().__init__()
90
+ self.num_heads = num_heads
91
+ head_dim = dim // num_heads
92
+
93
+ self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
94
+ self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
95
+ self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
96
+
97
+ def forward(self, x: Tensor, pe: Tensor) -> Tensor:
98
+ qkv = self.qkv(x)
99
+ q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
100
+ q, k = self.norm(q, k, v)
101
+ x = attention(q, k, v, pe=pe)
102
+ x = self.proj(x)
103
+ return x
104
+
105
+
106
+ @dataclass
107
+ class ModulationOut:
108
+ shift: Tensor
109
+ scale: Tensor
110
+ gate: Tensor
111
+
112
+
113
+ class Modulation(nn.Module):
114
+ def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
115
+ super().__init__()
116
+ self.is_double = double
117
+ self.multiplier = 6 if double else 3
118
+ self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
119
+
120
+ def forward(self, vec: Tensor) -> tuple:
121
+ out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
122
+
123
+ return (
124
+ ModulationOut(*out[:3]),
125
+ ModulationOut(*out[3:]) if self.is_double else None,
126
+ )
127
+
128
+
129
+ class DoubleStreamBlock(nn.Module):
130
+ def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, dtype=None, device=None, operations=None):
131
+ super().__init__()
132
+
133
+ mlp_hidden_dim = int(hidden_size * mlp_ratio)
134
+ self.num_heads = num_heads
135
+ self.hidden_size = hidden_size
136
+ self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
137
+ self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
138
+ self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
139
+
140
+ self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
141
+ self.img_mlp = nn.Sequential(
142
+ operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
143
+ nn.GELU(approximate="tanh"),
144
+ operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
145
+ )
146
+
147
+ self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
148
+ self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
149
+ self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
150
+
151
+ self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
152
+ self.txt_mlp = nn.Sequential(
153
+ operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
154
+ nn.GELU(approximate="tanh"),
155
+ operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
156
+ )
157
+
158
+ def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor):
159
+ img_mod1, img_mod2 = self.img_mod(vec)
160
+ txt_mod1, txt_mod2 = self.txt_mod(vec)
161
+
162
+ # prepare image for attention
163
+ img_modulated = self.img_norm1(img)
164
+ img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
165
+ img_qkv = self.img_attn.qkv(img_modulated)
166
+ img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
167
+ img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
168
+
169
+ # prepare txt for attention
170
+ txt_modulated = self.txt_norm1(txt)
171
+ txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
172
+ txt_qkv = self.txt_attn.qkv(txt_modulated)
173
+ txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
174
+ txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
175
+
176
+ # run actual attention
177
+ q = torch.cat((txt_q, img_q), dim=2)
178
+ k = torch.cat((txt_k, img_k), dim=2)
179
+ v = torch.cat((txt_v, img_v), dim=2)
180
+
181
+ attn = attention(q, k, v, pe=pe)
182
+ txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
183
+
184
+ # calculate the img bloks
185
+ img = img + img_mod1.gate * self.img_attn.proj(img_attn)
186
+ img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
187
+
188
+ # calculate the txt bloks
189
+ txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
190
+ txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
191
+
192
+ if txt.dtype == torch.float16:
193
+ txt = txt.clip(-65504, 65504)
194
+
195
+ return img, txt
196
+
197
+
198
+ class SingleStreamBlock(nn.Module):
199
+ """
200
+ A DiT block with parallel linear layers as described in
201
+ https://arxiv.org/abs/2302.05442 and adapted modulation interface.
202
+ """
203
+
204
+ def __init__(
205
+ self,
206
+ hidden_size: int,
207
+ num_heads: int,
208
+ mlp_ratio: float = 4.0,
209
+ qk_scale: float = None,
210
+ dtype=None,
211
+ device=None,
212
+ operations=None
213
+ ):
214
+ super().__init__()
215
+ self.hidden_dim = hidden_size
216
+ self.num_heads = num_heads
217
+ head_dim = hidden_size // num_heads
218
+ self.scale = qk_scale or head_dim**-0.5
219
+
220
+ self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
221
+ # qkv and mlp_in
222
+ self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
223
+ # proj and mlp_out
224
+ self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
225
+
226
+ self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
227
+
228
+ self.hidden_size = hidden_size
229
+ self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
230
+
231
+ self.mlp_act = nn.GELU(approximate="tanh")
232
+ self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
233
+
234
+ def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
235
+ mod, _ = self.modulation(vec)
236
+ x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
237
+ qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
238
+
239
+ q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
240
+ q, k = self.norm(q, k, v)
241
+
242
+ # compute attention
243
+ attn = attention(q, k, v, pe=pe)
244
+ # compute activation in mlp stream, cat again and run second linear layer
245
+ output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
246
+ x = x + mod.gate * output
247
+ if x.dtype == torch.float16:
248
+ x = x.clip(-65504, 65504)
249
+ return x
250
+
251
+
252
+ class LastLayer(nn.Module):
253
+ def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
254
+ super().__init__()
255
+ self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
256
+ self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
257
+ self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
258
+
259
+ def forward(self, x: Tensor, vec: Tensor) -> Tensor:
260
+ shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
261
+ x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
262
+ x = self.linear(x)
263
+ return x
ComfyUI/comfy/ldm/flux/math.py ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from einops import rearrange
3
+ from torch import Tensor
4
+ from comfy.ldm.modules.attention import optimized_attention
5
+ import comfy.model_management
6
+
7
+ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
8
+ q, k = apply_rope(q, k, pe)
9
+
10
+ heads = q.shape[1]
11
+ x = optimized_attention(q, k, v, heads, skip_reshape=True)
12
+ return x
13
+
14
+
15
+ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
16
+ assert dim % 2 == 0
17
+ if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
18
+ device = torch.device("cpu")
19
+ else:
20
+ device = pos.device
21
+
22
+ scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device)
23
+ omega = 1.0 / (theta**scale)
24
+ out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega)
25
+ out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
26
+ out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
27
+ return out.to(dtype=torch.float32, device=pos.device)
28
+
29
+
30
+ def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
31
+ xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
32
+ xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
33
+ xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
34
+ xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
35
+ return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
ComfyUI/comfy/ldm/flux/model.py ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #Original code can be found on: https://github.com/black-forest-labs/flux
2
+
3
+ from dataclasses import dataclass
4
+
5
+ import torch
6
+ from torch import Tensor, nn
7
+
8
+ from .layers import (
9
+ DoubleStreamBlock,
10
+ EmbedND,
11
+ LastLayer,
12
+ MLPEmbedder,
13
+ SingleStreamBlock,
14
+ timestep_embedding,
15
+ )
16
+
17
+ from einops import rearrange, repeat
18
+ import comfy.ldm.common_dit
19
+
20
+ @dataclass
21
+ class FluxParams:
22
+ in_channels: int
23
+ vec_in_dim: int
24
+ context_in_dim: int
25
+ hidden_size: int
26
+ mlp_ratio: float
27
+ num_heads: int
28
+ depth: int
29
+ depth_single_blocks: int
30
+ axes_dim: list
31
+ theta: int
32
+ qkv_bias: bool
33
+ guidance_embed: bool
34
+
35
+
36
+ class Flux(nn.Module):
37
+ """
38
+ Transformer model for flow matching on sequences.
39
+ """
40
+
41
+ def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
42
+ super().__init__()
43
+ self.dtype = dtype
44
+ params = FluxParams(**kwargs)
45
+ self.params = params
46
+ self.in_channels = params.in_channels * 2 * 2
47
+ self.out_channels = self.in_channels
48
+ if params.hidden_size % params.num_heads != 0:
49
+ raise ValueError(
50
+ f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
51
+ )
52
+ pe_dim = params.hidden_size // params.num_heads
53
+ if sum(params.axes_dim) != pe_dim:
54
+ raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
55
+ self.hidden_size = params.hidden_size
56
+ self.num_heads = params.num_heads
57
+ self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
58
+ self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
59
+ self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
60
+ self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
61
+ self.guidance_in = (
62
+ MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
63
+ )
64
+ self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
65
+
66
+ self.double_blocks = nn.ModuleList(
67
+ [
68
+ DoubleStreamBlock(
69
+ self.hidden_size,
70
+ self.num_heads,
71
+ mlp_ratio=params.mlp_ratio,
72
+ qkv_bias=params.qkv_bias,
73
+ dtype=dtype, device=device, operations=operations
74
+ )
75
+ for _ in range(params.depth)
76
+ ]
77
+ )
78
+
79
+ self.single_blocks = nn.ModuleList(
80
+ [
81
+ SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
82
+ for _ in range(params.depth_single_blocks)
83
+ ]
84
+ )
85
+
86
+ self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
87
+
88
+ def forward_orig(
89
+ self,
90
+ img: Tensor,
91
+ img_ids: Tensor,
92
+ txt: Tensor,
93
+ txt_ids: Tensor,
94
+ timesteps: Tensor,
95
+ y: Tensor,
96
+ guidance: Tensor = None,
97
+ ) -> Tensor:
98
+ if img.ndim != 3 or txt.ndim != 3:
99
+ raise ValueError("Input img and txt tensors must have 3 dimensions.")
100
+
101
+ # running on sequences img
102
+ img = self.img_in(img)
103
+ vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
104
+ if self.params.guidance_embed:
105
+ if guidance is None:
106
+ raise ValueError("Didn't get guidance strength for guidance distilled model.")
107
+ vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
108
+
109
+ vec = vec + self.vector_in(y)
110
+ txt = self.txt_in(txt)
111
+
112
+ ids = torch.cat((txt_ids, img_ids), dim=1)
113
+ pe = self.pe_embedder(ids)
114
+
115
+ for block in self.double_blocks:
116
+ img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
117
+
118
+ img = torch.cat((txt, img), 1)
119
+ for block in self.single_blocks:
120
+ img = block(img, vec=vec, pe=pe)
121
+ img = img[:, txt.shape[1] :, ...]
122
+
123
+ img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
124
+ return img
125
+
126
+ def forward(self, x, timestep, context, y, guidance, **kwargs):
127
+ bs, c, h, w = x.shape
128
+ patch_size = 2
129
+ x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
130
+
131
+ img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
132
+
133
+ h_len = ((h + (patch_size // 2)) // patch_size)
134
+ w_len = ((w + (patch_size // 2)) // patch_size)
135
+ img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
136
+ img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
137
+ img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
138
+ img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
139
+
140
+ txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
141
+ out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance)
142
+ return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]
ComfyUI/comfy/ldm/hydit/attn_layers.py ADDED
@@ -0,0 +1,219 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from typing import Tuple, Union, Optional
4
+ from comfy.ldm.modules.attention import optimized_attention
5
+
6
+
7
+ def reshape_for_broadcast(freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], x: torch.Tensor, head_first=False):
8
+ """
9
+ Reshape frequency tensor for broadcasting it with another tensor.
10
+
11
+ This function reshapes the frequency tensor to have the same shape as the target tensor 'x'
12
+ for the purpose of broadcasting the frequency tensor during element-wise operations.
13
+
14
+ Args:
15
+ freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Frequency tensor to be reshaped.
16
+ x (torch.Tensor): Target tensor for broadcasting compatibility.
17
+ head_first (bool): head dimension first (except batch dim) or not.
18
+
19
+ Returns:
20
+ torch.Tensor: Reshaped frequency tensor.
21
+
22
+ Raises:
23
+ AssertionError: If the frequency tensor doesn't match the expected shape.
24
+ AssertionError: If the target tensor 'x' doesn't have the expected number of dimensions.
25
+ """
26
+ ndim = x.ndim
27
+ assert 0 <= 1 < ndim
28
+
29
+ if isinstance(freqs_cis, tuple):
30
+ # freqs_cis: (cos, sin) in real space
31
+ if head_first:
32
+ assert freqs_cis[0].shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}'
33
+ shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
34
+ else:
35
+ assert freqs_cis[0].shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}'
36
+ shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
37
+ return freqs_cis[0].view(*shape), freqs_cis[1].view(*shape)
38
+ else:
39
+ # freqs_cis: values in complex space
40
+ if head_first:
41
+ assert freqs_cis.shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}'
42
+ shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
43
+ else:
44
+ assert freqs_cis.shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}'
45
+ shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
46
+ return freqs_cis.view(*shape)
47
+
48
+
49
+ def rotate_half(x):
50
+ x_real, x_imag = x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2]
51
+ return torch.stack([-x_imag, x_real], dim=-1).flatten(3)
52
+
53
+
54
+ def apply_rotary_emb(
55
+ xq: torch.Tensor,
56
+ xk: Optional[torch.Tensor],
57
+ freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
58
+ head_first: bool = False,
59
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
60
+ """
61
+ Apply rotary embeddings to input tensors using the given frequency tensor.
62
+
63
+ This function applies rotary embeddings to the given query 'xq' and key 'xk' tensors using the provided
64
+ frequency tensor 'freqs_cis'. The input tensors are reshaped as complex numbers, and the frequency tensor
65
+ is reshaped for broadcasting compatibility. The resulting tensors contain rotary embeddings and are
66
+ returned as real tensors.
67
+
68
+ Args:
69
+ xq (torch.Tensor): Query tensor to apply rotary embeddings. [B, S, H, D]
70
+ xk (torch.Tensor): Key tensor to apply rotary embeddings. [B, S, H, D]
71
+ freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Precomputed frequency tensor for complex exponentials.
72
+ head_first (bool): head dimension first (except batch dim) or not.
73
+
74
+ Returns:
75
+ Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
76
+
77
+ """
78
+ xk_out = None
79
+ if isinstance(freqs_cis, tuple):
80
+ cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first) # [S, D]
81
+ cos, sin = cos.to(xq.device), sin.to(xq.device)
82
+ xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).type_as(xq)
83
+ if xk is not None:
84
+ xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).type_as(xk)
85
+ else:
86
+ xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) # [B, S, H, D//2]
87
+ freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to(xq.device) # [S, D//2] --> [1, S, 1, D//2]
88
+ xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3).type_as(xq)
89
+ if xk is not None:
90
+ xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) # [B, S, H, D//2]
91
+ xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3).type_as(xk)
92
+
93
+ return xq_out, xk_out
94
+
95
+
96
+
97
+ class CrossAttention(nn.Module):
98
+ """
99
+ Use QK Normalization.
100
+ """
101
+ def __init__(self,
102
+ qdim,
103
+ kdim,
104
+ num_heads,
105
+ qkv_bias=True,
106
+ qk_norm=False,
107
+ attn_drop=0.0,
108
+ proj_drop=0.0,
109
+ attn_precision=None,
110
+ device=None,
111
+ dtype=None,
112
+ operations=None,
113
+ ):
114
+ factory_kwargs = {'device': device, 'dtype': dtype}
115
+ super().__init__()
116
+ self.attn_precision = attn_precision
117
+ self.qdim = qdim
118
+ self.kdim = kdim
119
+ self.num_heads = num_heads
120
+ assert self.qdim % num_heads == 0, "self.qdim must be divisible by num_heads"
121
+ self.head_dim = self.qdim // num_heads
122
+ assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8"
123
+ self.scale = self.head_dim ** -0.5
124
+
125
+ self.q_proj = operations.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs)
126
+ self.kv_proj = operations.Linear(kdim, 2 * qdim, bias=qkv_bias, **factory_kwargs)
127
+
128
+ # TODO: eps should be 1 / 65530 if using fp16
129
+ self.q_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
130
+ self.k_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
131
+ self.attn_drop = nn.Dropout(attn_drop)
132
+ self.out_proj = operations.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs)
133
+ self.proj_drop = nn.Dropout(proj_drop)
134
+
135
+ def forward(self, x, y, freqs_cis_img=None):
136
+ """
137
+ Parameters
138
+ ----------
139
+ x: torch.Tensor
140
+ (batch, seqlen1, hidden_dim) (where hidden_dim = num heads * head dim)
141
+ y: torch.Tensor
142
+ (batch, seqlen2, hidden_dim2)
143
+ freqs_cis_img: torch.Tensor
144
+ (batch, hidden_dim // 2), RoPE for image
145
+ """
146
+ b, s1, c = x.shape # [b, s1, D]
147
+ _, s2, c = y.shape # [b, s2, 1024]
148
+
149
+ q = self.q_proj(x).view(b, s1, self.num_heads, self.head_dim) # [b, s1, h, d]
150
+ kv = self.kv_proj(y).view(b, s2, 2, self.num_heads, self.head_dim) # [b, s2, 2, h, d]
151
+ k, v = kv.unbind(dim=2) # [b, s, h, d]
152
+ q = self.q_norm(q)
153
+ k = self.k_norm(k)
154
+
155
+ # Apply RoPE if needed
156
+ if freqs_cis_img is not None:
157
+ qq, _ = apply_rotary_emb(q, None, freqs_cis_img)
158
+ assert qq.shape == q.shape, f'qq: {qq.shape}, q: {q.shape}'
159
+ q = qq
160
+
161
+ q = q.transpose(-2, -3).contiguous() # q -> B, L1, H, C - B, H, L1, C
162
+ k = k.transpose(-2, -3).contiguous() # k -> B, L2, H, C - B, H, C, L2
163
+ v = v.transpose(-2, -3).contiguous()
164
+
165
+ context = optimized_attention(q, k, v, self.num_heads, skip_reshape=True, attn_precision=self.attn_precision)
166
+
167
+ out = self.out_proj(context) # context.reshape - B, L1, -1
168
+ out = self.proj_drop(out)
169
+
170
+ out_tuple = (out,)
171
+
172
+ return out_tuple
173
+
174
+
175
+ class Attention(nn.Module):
176
+ """
177
+ We rename some layer names to align with flash attention
178
+ """
179
+ def __init__(self, dim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0., proj_drop=0., attn_precision=None, dtype=None, device=None, operations=None):
180
+ super().__init__()
181
+ self.attn_precision = attn_precision
182
+ self.dim = dim
183
+ self.num_heads = num_heads
184
+ assert self.dim % num_heads == 0, 'dim should be divisible by num_heads'
185
+ self.head_dim = self.dim // num_heads
186
+ # This assertion is aligned with flash attention
187
+ assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8"
188
+ self.scale = self.head_dim ** -0.5
189
+
190
+ # qkv --> Wqkv
191
+ self.Wqkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
192
+ # TODO: eps should be 1 / 65530 if using fp16
193
+ self.q_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
194
+ self.k_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
195
+ self.attn_drop = nn.Dropout(attn_drop)
196
+ self.out_proj = operations.Linear(dim, dim, dtype=dtype, device=device)
197
+ self.proj_drop = nn.Dropout(proj_drop)
198
+
199
+ def forward(self, x, freqs_cis_img=None):
200
+ B, N, C = x.shape
201
+ qkv = self.Wqkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) # [3, b, h, s, d]
202
+ q, k, v = qkv.unbind(0) # [b, h, s, d]
203
+ q = self.q_norm(q) # [b, h, s, d]
204
+ k = self.k_norm(k) # [b, h, s, d]
205
+
206
+ # Apply RoPE if needed
207
+ if freqs_cis_img is not None:
208
+ qq, kk = apply_rotary_emb(q, k, freqs_cis_img, head_first=True)
209
+ assert qq.shape == q.shape and kk.shape == k.shape, \
210
+ f'qq: {qq.shape}, q: {q.shape}, kk: {kk.shape}, k: {k.shape}'
211
+ q, k = qq, kk
212
+
213
+ x = optimized_attention(q, k, v, self.num_heads, skip_reshape=True, attn_precision=self.attn_precision)
214
+ x = self.out_proj(x)
215
+ x = self.proj_drop(x)
216
+
217
+ out_tuple = (x,)
218
+
219
+ return out_tuple
ComfyUI/comfy/ldm/hydit/models.py ADDED
@@ -0,0 +1,405 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Any
2
+
3
+ import torch
4
+ import torch.nn as nn
5
+ import torch.nn.functional as F
6
+
7
+ import comfy.ops
8
+ from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed, RMSNorm
9
+ from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
10
+ from torch.utils import checkpoint
11
+
12
+ from .attn_layers import Attention, CrossAttention
13
+ from .poolers import AttentionPool
14
+ from .posemb_layers import get_2d_rotary_pos_embed, get_fill_resize_and_crop
15
+
16
+ def calc_rope(x, patch_size, head_size):
17
+ th = (x.shape[2] + (patch_size // 2)) // patch_size
18
+ tw = (x.shape[3] + (patch_size // 2)) // patch_size
19
+ base_size = 512 // 8 // patch_size
20
+ start, stop = get_fill_resize_and_crop((th, tw), base_size)
21
+ sub_args = [start, stop, (th, tw)]
22
+ # head_size = HUNYUAN_DIT_CONFIG['DiT-g/2']['hidden_size'] // HUNYUAN_DIT_CONFIG['DiT-g/2']['num_heads']
23
+ rope = get_2d_rotary_pos_embed(head_size, *sub_args)
24
+ return rope
25
+
26
+
27
+ def modulate(x, shift, scale):
28
+ return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
29
+
30
+
31
+ class HunYuanDiTBlock(nn.Module):
32
+ """
33
+ A HunYuanDiT block with `add` conditioning.
34
+ """
35
+ def __init__(self,
36
+ hidden_size,
37
+ c_emb_size,
38
+ num_heads,
39
+ mlp_ratio=4.0,
40
+ text_states_dim=1024,
41
+ qk_norm=False,
42
+ norm_type="layer",
43
+ skip=False,
44
+ attn_precision=None,
45
+ dtype=None,
46
+ device=None,
47
+ operations=None,
48
+ ):
49
+ super().__init__()
50
+ use_ele_affine = True
51
+
52
+ if norm_type == "layer":
53
+ norm_layer = operations.LayerNorm
54
+ elif norm_type == "rms":
55
+ norm_layer = RMSNorm
56
+ else:
57
+ raise ValueError(f"Unknown norm_type: {norm_type}")
58
+
59
+ # ========================= Self-Attention =========================
60
+ self.norm1 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6, dtype=dtype, device=device)
61
+ self.attn1 = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=qk_norm, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
62
+
63
+ # ========================= FFN =========================
64
+ self.norm2 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6, dtype=dtype, device=device)
65
+ mlp_hidden_dim = int(hidden_size * mlp_ratio)
66
+ approx_gelu = lambda: nn.GELU(approximate="tanh")
67
+ self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0, dtype=dtype, device=device, operations=operations)
68
+
69
+ # ========================= Add =========================
70
+ # Simply use add like SDXL.
71
+ self.default_modulation = nn.Sequential(
72
+ nn.SiLU(),
73
+ operations.Linear(c_emb_size, hidden_size, bias=True, dtype=dtype, device=device)
74
+ )
75
+
76
+ # ========================= Cross-Attention =========================
77
+ self.attn2 = CrossAttention(hidden_size, text_states_dim, num_heads=num_heads, qkv_bias=True,
78
+ qk_norm=qk_norm, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
79
+ self.norm3 = norm_layer(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
80
+
81
+ # ========================= Skip Connection =========================
82
+ if skip:
83
+ self.skip_norm = norm_layer(2 * hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
84
+ self.skip_linear = operations.Linear(2 * hidden_size, hidden_size, dtype=dtype, device=device)
85
+ else:
86
+ self.skip_linear = None
87
+
88
+ self.gradient_checkpointing = False
89
+
90
+ def _forward(self, x, c=None, text_states=None, freq_cis_img=None, skip=None):
91
+ # Long Skip Connection
92
+ if self.skip_linear is not None:
93
+ cat = torch.cat([x, skip], dim=-1)
94
+ cat = self.skip_norm(cat)
95
+ x = self.skip_linear(cat)
96
+
97
+ # Self-Attention
98
+ shift_msa = self.default_modulation(c).unsqueeze(dim=1)
99
+ attn_inputs = (
100
+ self.norm1(x) + shift_msa, freq_cis_img,
101
+ )
102
+ x = x + self.attn1(*attn_inputs)[0]
103
+
104
+ # Cross-Attention
105
+ cross_inputs = (
106
+ self.norm3(x), text_states, freq_cis_img
107
+ )
108
+ x = x + self.attn2(*cross_inputs)[0]
109
+
110
+ # FFN Layer
111
+ mlp_inputs = self.norm2(x)
112
+ x = x + self.mlp(mlp_inputs)
113
+
114
+ return x
115
+
116
+ def forward(self, x, c=None, text_states=None, freq_cis_img=None, skip=None):
117
+ if self.gradient_checkpointing and self.training:
118
+ return checkpoint.checkpoint(self._forward, x, c, text_states, freq_cis_img, skip)
119
+ return self._forward(x, c, text_states, freq_cis_img, skip)
120
+
121
+
122
+ class FinalLayer(nn.Module):
123
+ """
124
+ The final layer of HunYuanDiT.
125
+ """
126
+ def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels, dtype=None, device=None, operations=None):
127
+ super().__init__()
128
+ self.norm_final = operations.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
129
+ self.linear = operations.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
130
+ self.adaLN_modulation = nn.Sequential(
131
+ nn.SiLU(),
132
+ operations.Linear(c_emb_size, 2 * final_hidden_size, bias=True, dtype=dtype, device=device)
133
+ )
134
+
135
+ def forward(self, x, c):
136
+ shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
137
+ x = modulate(self.norm_final(x), shift, scale)
138
+ x = self.linear(x)
139
+ return x
140
+
141
+
142
+ class HunYuanDiT(nn.Module):
143
+ """
144
+ HunYuanDiT: Diffusion model with a Transformer backbone.
145
+
146
+ Inherit ModelMixin and ConfigMixin to be compatible with the sampler StableDiffusionPipeline of diffusers.
147
+
148
+ Inherit PeftAdapterMixin to be compatible with the PEFT training pipeline.
149
+
150
+ Parameters
151
+ ----------
152
+ args: argparse.Namespace
153
+ The arguments parsed by argparse.
154
+ input_size: tuple
155
+ The size of the input image.
156
+ patch_size: int
157
+ The size of the patch.
158
+ in_channels: int
159
+ The number of input channels.
160
+ hidden_size: int
161
+ The hidden size of the transformer backbone.
162
+ depth: int
163
+ The number of transformer blocks.
164
+ num_heads: int
165
+ The number of attention heads.
166
+ mlp_ratio: float
167
+ The ratio of the hidden size of the MLP in the transformer block.
168
+ log_fn: callable
169
+ The logging function.
170
+ """
171
+ #@register_to_config
172
+ def __init__(self,
173
+ input_size: tuple = 32,
174
+ patch_size: int = 2,
175
+ in_channels: int = 4,
176
+ hidden_size: int = 1152,
177
+ depth: int = 28,
178
+ num_heads: int = 16,
179
+ mlp_ratio: float = 4.0,
180
+ text_states_dim = 1024,
181
+ text_states_dim_t5 = 2048,
182
+ text_len = 77,
183
+ text_len_t5 = 256,
184
+ qk_norm = True,# See http://arxiv.org/abs/2302.05442 for details.
185
+ size_cond = False,
186
+ use_style_cond = False,
187
+ learn_sigma = True,
188
+ norm = "layer",
189
+ log_fn: callable = print,
190
+ attn_precision=None,
191
+ dtype=None,
192
+ device=None,
193
+ operations=None,
194
+ **kwargs,
195
+ ):
196
+ super().__init__()
197
+ self.log_fn = log_fn
198
+ self.depth = depth
199
+ self.learn_sigma = learn_sigma
200
+ self.in_channels = in_channels
201
+ self.out_channels = in_channels * 2 if learn_sigma else in_channels
202
+ self.patch_size = patch_size
203
+ self.num_heads = num_heads
204
+ self.hidden_size = hidden_size
205
+ self.text_states_dim = text_states_dim
206
+ self.text_states_dim_t5 = text_states_dim_t5
207
+ self.text_len = text_len
208
+ self.text_len_t5 = text_len_t5
209
+ self.size_cond = size_cond
210
+ self.use_style_cond = use_style_cond
211
+ self.norm = norm
212
+ self.dtype = dtype
213
+ #import pdb
214
+ #pdb.set_trace()
215
+
216
+ self.mlp_t5 = nn.Sequential(
217
+ operations.Linear(self.text_states_dim_t5, self.text_states_dim_t5 * 4, bias=True, dtype=dtype, device=device),
218
+ nn.SiLU(),
219
+ operations.Linear(self.text_states_dim_t5 * 4, self.text_states_dim, bias=True, dtype=dtype, device=device),
220
+ )
221
+ # learnable replace
222
+ self.text_embedding_padding = nn.Parameter(
223
+ torch.empty(self.text_len + self.text_len_t5, self.text_states_dim, dtype=dtype, device=device))
224
+
225
+ # Attention pooling
226
+ pooler_out_dim = 1024
227
+ self.pooler = AttentionPool(self.text_len_t5, self.text_states_dim_t5, num_heads=8, output_dim=pooler_out_dim, dtype=dtype, device=device, operations=operations)
228
+
229
+ # Dimension of the extra input vectors
230
+ self.extra_in_dim = pooler_out_dim
231
+
232
+ if self.size_cond:
233
+ # Image size and crop size conditions
234
+ self.extra_in_dim += 6 * 256
235
+
236
+ if self.use_style_cond:
237
+ # Here we use a default learned embedder layer for future extension.
238
+ self.style_embedder = operations.Embedding(1, hidden_size, dtype=dtype, device=device)
239
+ self.extra_in_dim += hidden_size
240
+
241
+ # Text embedding for `add`
242
+ self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, dtype=dtype, device=device, operations=operations)
243
+ self.t_embedder = TimestepEmbedder(hidden_size, dtype=dtype, device=device, operations=operations)
244
+ self.extra_embedder = nn.Sequential(
245
+ operations.Linear(self.extra_in_dim, hidden_size * 4, dtype=dtype, device=device),
246
+ nn.SiLU(),
247
+ operations.Linear(hidden_size * 4, hidden_size, bias=True, dtype=dtype, device=device),
248
+ )
249
+
250
+ # Image embedding
251
+ num_patches = self.x_embedder.num_patches
252
+
253
+ # HUnYuanDiT Blocks
254
+ self.blocks = nn.ModuleList([
255
+ HunYuanDiTBlock(hidden_size=hidden_size,
256
+ c_emb_size=hidden_size,
257
+ num_heads=num_heads,
258
+ mlp_ratio=mlp_ratio,
259
+ text_states_dim=self.text_states_dim,
260
+ qk_norm=qk_norm,
261
+ norm_type=self.norm,
262
+ skip=layer > depth // 2,
263
+ attn_precision=attn_precision,
264
+ dtype=dtype,
265
+ device=device,
266
+ operations=operations,
267
+ )
268
+ for layer in range(depth)
269
+ ])
270
+
271
+ self.final_layer = FinalLayer(hidden_size, hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
272
+ self.unpatchify_channels = self.out_channels
273
+
274
+
275
+
276
+ def forward(self,
277
+ x,
278
+ t,
279
+ context,#encoder_hidden_states=None,
280
+ text_embedding_mask=None,
281
+ encoder_hidden_states_t5=None,
282
+ text_embedding_mask_t5=None,
283
+ image_meta_size=None,
284
+ style=None,
285
+ return_dict=False,
286
+ control=None,
287
+ transformer_options=None,
288
+ ):
289
+ """
290
+ Forward pass of the encoder.
291
+
292
+ Parameters
293
+ ----------
294
+ x: torch.Tensor
295
+ (B, D, H, W)
296
+ t: torch.Tensor
297
+ (B)
298
+ encoder_hidden_states: torch.Tensor
299
+ CLIP text embedding, (B, L_clip, D)
300
+ text_embedding_mask: torch.Tensor
301
+ CLIP text embedding mask, (B, L_clip)
302
+ encoder_hidden_states_t5: torch.Tensor
303
+ T5 text embedding, (B, L_t5, D)
304
+ text_embedding_mask_t5: torch.Tensor
305
+ T5 text embedding mask, (B, L_t5)
306
+ image_meta_size: torch.Tensor
307
+ (B, 6)
308
+ style: torch.Tensor
309
+ (B)
310
+ cos_cis_img: torch.Tensor
311
+ sin_cis_img: torch.Tensor
312
+ return_dict: bool
313
+ Whether to return a dictionary.
314
+ """
315
+ #import pdb
316
+ #pdb.set_trace()
317
+ encoder_hidden_states = context
318
+ text_states = encoder_hidden_states # 2,77,1024
319
+ text_states_t5 = encoder_hidden_states_t5 # 2,256,2048
320
+ text_states_mask = text_embedding_mask.bool() # 2,77
321
+ text_states_t5_mask = text_embedding_mask_t5.bool() # 2,256
322
+ b_t5, l_t5, c_t5 = text_states_t5.shape
323
+ text_states_t5 = self.mlp_t5(text_states_t5.view(-1, c_t5)).view(b_t5, l_t5, -1)
324
+
325
+ padding = comfy.ops.cast_to_input(self.text_embedding_padding, text_states)
326
+
327
+ text_states[:,-self.text_len:] = torch.where(text_states_mask[:,-self.text_len:].unsqueeze(2), text_states[:,-self.text_len:], padding[:self.text_len])
328
+ text_states_t5[:,-self.text_len_t5:] = torch.where(text_states_t5_mask[:,-self.text_len_t5:].unsqueeze(2), text_states_t5[:,-self.text_len_t5:], padding[self.text_len:])
329
+
330
+ text_states = torch.cat([text_states, text_states_t5], dim=1) # 2,205,1024
331
+ # clip_t5_mask = torch.cat([text_states_mask, text_states_t5_mask], dim=-1)
332
+
333
+ _, _, oh, ow = x.shape
334
+ th, tw = (oh + (self.patch_size // 2)) // self.patch_size, (ow + (self.patch_size // 2)) // self.patch_size
335
+
336
+
337
+ # Get image RoPE embedding according to `reso`lution.
338
+ freqs_cis_img = calc_rope(x, self.patch_size, self.hidden_size // self.num_heads) #(cos_cis_img, sin_cis_img)
339
+
340
+ # ========================= Build time and image embedding =========================
341
+ t = self.t_embedder(t, dtype=x.dtype)
342
+ x = self.x_embedder(x)
343
+
344
+ # ========================= Concatenate all extra vectors =========================
345
+ # Build text tokens with pooling
346
+ extra_vec = self.pooler(encoder_hidden_states_t5)
347
+
348
+ # Build image meta size tokens if applicable
349
+ if self.size_cond:
350
+ image_meta_size = timestep_embedding(image_meta_size.view(-1), 256).to(x.dtype) # [B * 6, 256]
351
+ image_meta_size = image_meta_size.view(-1, 6 * 256)
352
+ extra_vec = torch.cat([extra_vec, image_meta_size], dim=1) # [B, D + 6 * 256]
353
+
354
+ # Build style tokens
355
+ if self.use_style_cond:
356
+ if style is None:
357
+ style = torch.zeros((extra_vec.shape[0],), device=x.device, dtype=torch.int)
358
+ style_embedding = self.style_embedder(style, out_dtype=x.dtype)
359
+ extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
360
+
361
+ # Concatenate all extra vectors
362
+ c = t + self.extra_embedder(extra_vec) # [B, D]
363
+
364
+ controls = None
365
+ # ========================= Forward pass through HunYuanDiT blocks =========================
366
+ skips = []
367
+ for layer, block in enumerate(self.blocks):
368
+ if layer > self.depth // 2:
369
+ if controls is not None:
370
+ skip = skips.pop() + controls.pop()
371
+ else:
372
+ skip = skips.pop()
373
+ x = block(x, c, text_states, freqs_cis_img, skip) # (N, L, D)
374
+ else:
375
+ x = block(x, c, text_states, freqs_cis_img) # (N, L, D)
376
+
377
+ if layer < (self.depth // 2 - 1):
378
+ skips.append(x)
379
+ if controls is not None and len(controls) != 0:
380
+ raise ValueError("The number of controls is not equal to the number of skip connections.")
381
+
382
+ # ========================= Final layer =========================
383
+ x = self.final_layer(x, c) # (N, L, patch_size ** 2 * out_channels)
384
+ x = self.unpatchify(x, th, tw) # (N, out_channels, H, W)
385
+
386
+ if return_dict:
387
+ return {'x': x}
388
+ if self.learn_sigma:
389
+ return x[:,:self.out_channels // 2,:oh,:ow]
390
+ return x[:,:,:oh,:ow]
391
+
392
+ def unpatchify(self, x, h, w):
393
+ """
394
+ x: (N, T, patch_size**2 * C)
395
+ imgs: (N, H, W, C)
396
+ """
397
+ c = self.unpatchify_channels
398
+ p = self.x_embedder.patch_size[0]
399
+ # h = w = int(x.shape[1] ** 0.5)
400
+ assert h * w == x.shape[1]
401
+
402
+ x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
403
+ x = torch.einsum('nhwpqc->nchpwq', x)
404
+ imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
405
+ return imgs
ComfyUI/comfy/ldm/hydit/poolers.py ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ import torch.nn.functional as F
4
+ from comfy.ldm.modules.attention import optimized_attention
5
+ import comfy.ops
6
+
7
+ class AttentionPool(nn.Module):
8
+ def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None, dtype=None, device=None, operations=None):
9
+ super().__init__()
10
+ self.positional_embedding = nn.Parameter(torch.empty(spacial_dim + 1, embed_dim, dtype=dtype, device=device))
11
+ self.k_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
12
+ self.q_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
13
+ self.v_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
14
+ self.c_proj = operations.Linear(embed_dim, output_dim or embed_dim, dtype=dtype, device=device)
15
+ self.num_heads = num_heads
16
+ self.embed_dim = embed_dim
17
+
18
+ def forward(self, x):
19
+ x = x[:,:self.positional_embedding.shape[0] - 1]
20
+ x = x.permute(1, 0, 2) # NLC -> LNC
21
+ x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (L+1)NC
22
+ x = x + comfy.ops.cast_to_input(self.positional_embedding[:, None, :], x) # (L+1)NC
23
+
24
+ q = self.q_proj(x[:1])
25
+ k = self.k_proj(x)
26
+ v = self.v_proj(x)
27
+
28
+ batch_size = q.shape[1]
29
+ head_dim = self.embed_dim // self.num_heads
30
+ q = q.view(1, batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
31
+ k = k.view(k.shape[0], batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
32
+ v = v.view(v.shape[0], batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
33
+
34
+ attn_output = optimized_attention(q, k, v, self.num_heads, skip_reshape=True).transpose(0, 1)
35
+
36
+ attn_output = self.c_proj(attn_output)
37
+ return attn_output.squeeze(0)
ComfyUI/comfy/ldm/hydit/posemb_layers.py ADDED
@@ -0,0 +1,224 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import numpy as np
3
+ from typing import Union
4
+
5
+
6
+ def _to_tuple(x):
7
+ if isinstance(x, int):
8
+ return x, x
9
+ else:
10
+ return x
11
+
12
+
13
+ def get_fill_resize_and_crop(src, tgt):
14
+ th, tw = _to_tuple(tgt)
15
+ h, w = _to_tuple(src)
16
+
17
+ tr = th / tw # base resolution
18
+ r = h / w # target resolution
19
+
20
+ # resize
21
+ if r > tr:
22
+ resize_height = th
23
+ resize_width = int(round(th / h * w))
24
+ else:
25
+ resize_width = tw
26
+ resize_height = int(round(tw / w * h)) # resize the target resolution down based on the base resolution
27
+
28
+ crop_top = int(round((th - resize_height) / 2.0))
29
+ crop_left = int(round((tw - resize_width) / 2.0))
30
+
31
+ return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
32
+
33
+
34
+ def get_meshgrid(start, *args):
35
+ if len(args) == 0:
36
+ # start is grid_size
37
+ num = _to_tuple(start)
38
+ start = (0, 0)
39
+ stop = num
40
+ elif len(args) == 1:
41
+ # start is start, args[0] is stop, step is 1
42
+ start = _to_tuple(start)
43
+ stop = _to_tuple(args[0])
44
+ num = (stop[0] - start[0], stop[1] - start[1])
45
+ elif len(args) == 2:
46
+ # start is start, args[0] is stop, args[1] is num
47
+ start = _to_tuple(start)
48
+ stop = _to_tuple(args[0])
49
+ num = _to_tuple(args[1])
50
+ else:
51
+ raise ValueError(f"len(args) should be 0, 1 or 2, but got {len(args)}")
52
+
53
+ grid_h = np.linspace(start[0], stop[0], num[0], endpoint=False, dtype=np.float32)
54
+ grid_w = np.linspace(start[1], stop[1], num[1], endpoint=False, dtype=np.float32)
55
+ grid = np.meshgrid(grid_w, grid_h) # here w goes first
56
+ grid = np.stack(grid, axis=0) # [2, W, H]
57
+ return grid
58
+
59
+ #################################################################################
60
+ # Sine/Cosine Positional Embedding Functions #
61
+ #################################################################################
62
+ # https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
63
+
64
+ def get_2d_sincos_pos_embed(embed_dim, start, *args, cls_token=False, extra_tokens=0):
65
+ """
66
+ grid_size: int of the grid height and width
67
+ return:
68
+ pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
69
+ """
70
+ grid = get_meshgrid(start, *args) # [2, H, w]
71
+ # grid_h = np.arange(grid_size, dtype=np.float32)
72
+ # grid_w = np.arange(grid_size, dtype=np.float32)
73
+ # grid = np.meshgrid(grid_w, grid_h) # here w goes first
74
+ # grid = np.stack(grid, axis=0) # [2, W, H]
75
+
76
+ grid = grid.reshape([2, 1, *grid.shape[1:]])
77
+ pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
78
+ if cls_token and extra_tokens > 0:
79
+ pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
80
+ return pos_embed
81
+
82
+
83
+ def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
84
+ assert embed_dim % 2 == 0
85
+
86
+ # use half of dimensions to encode grid_h
87
+ emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
88
+ emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
89
+
90
+ emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
91
+ return emb
92
+
93
+
94
+ def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
95
+ """
96
+ embed_dim: output dimension for each position
97
+ pos: a list of positions to be encoded: size (W,H)
98
+ out: (M, D)
99
+ """
100
+ assert embed_dim % 2 == 0
101
+ omega = np.arange(embed_dim // 2, dtype=np.float64)
102
+ omega /= embed_dim / 2.
103
+ omega = 1. / 10000**omega # (D/2,)
104
+
105
+ pos = pos.reshape(-1) # (M,)
106
+ out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
107
+
108
+ emb_sin = np.sin(out) # (M, D/2)
109
+ emb_cos = np.cos(out) # (M, D/2)
110
+
111
+ emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
112
+ return emb
113
+
114
+
115
+ #################################################################################
116
+ # Rotary Positional Embedding Functions #
117
+ #################################################################################
118
+ # https://github.com/facebookresearch/llama/blob/main/llama/model.py#L443
119
+
120
+ def get_2d_rotary_pos_embed(embed_dim, start, *args, use_real=True):
121
+ """
122
+ This is a 2d version of precompute_freqs_cis, which is a RoPE for image tokens with 2d structure.
123
+
124
+ Parameters
125
+ ----------
126
+ embed_dim: int
127
+ embedding dimension size
128
+ start: int or tuple of int
129
+ If len(args) == 0, start is num; If len(args) == 1, start is start, args[0] is stop, step is 1;
130
+ If len(args) == 2, start is start, args[0] is stop, args[1] is num.
131
+ use_real: bool
132
+ If True, return real part and imaginary part separately. Otherwise, return complex numbers.
133
+
134
+ Returns
135
+ -------
136
+ pos_embed: torch.Tensor
137
+ [HW, D/2]
138
+ """
139
+ grid = get_meshgrid(start, *args) # [2, H, w]
140
+ grid = grid.reshape([2, 1, *grid.shape[1:]]) # Returns a sampling matrix with the same resolution as the target resolution
141
+ pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
142
+ return pos_embed
143
+
144
+
145
+ def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
146
+ assert embed_dim % 4 == 0
147
+
148
+ # use half of dimensions to encode grid_h
149
+ emb_h = get_1d_rotary_pos_embed(embed_dim // 2, grid[0].reshape(-1), use_real=use_real) # (H*W, D/4)
150
+ emb_w = get_1d_rotary_pos_embed(embed_dim // 2, grid[1].reshape(-1), use_real=use_real) # (H*W, D/4)
151
+
152
+ if use_real:
153
+ cos = torch.cat([emb_h[0], emb_w[0]], dim=1) # (H*W, D/2)
154
+ sin = torch.cat([emb_h[1], emb_w[1]], dim=1) # (H*W, D/2)
155
+ return cos, sin
156
+ else:
157
+ emb = torch.cat([emb_h, emb_w], dim=1) # (H*W, D/2)
158
+ return emb
159
+
160
+
161
+ def get_1d_rotary_pos_embed(dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False):
162
+ """
163
+ Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
164
+
165
+ This function calculates a frequency tensor with complex exponentials using the given dimension 'dim'
166
+ and the end index 'end'. The 'theta' parameter scales the frequencies.
167
+ The returned tensor contains complex values in complex64 data type.
168
+
169
+ Args:
170
+ dim (int): Dimension of the frequency tensor.
171
+ pos (np.ndarray, int): Position indices for the frequency tensor. [S] or scalar
172
+ theta (float, optional): Scaling factor for frequency computation. Defaults to 10000.0.
173
+ use_real (bool, optional): If True, return real part and imaginary part separately.
174
+ Otherwise, return complex numbers.
175
+
176
+ Returns:
177
+ torch.Tensor: Precomputed frequency tensor with complex exponentials. [S, D/2]
178
+
179
+ """
180
+ if isinstance(pos, int):
181
+ pos = np.arange(pos)
182
+ freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) # [D/2]
183
+ t = torch.from_numpy(pos).to(freqs.device) # type: ignore # [S]
184
+ freqs = torch.outer(t, freqs).float() # type: ignore # [S, D/2]
185
+ if use_real:
186
+ freqs_cos = freqs.cos().repeat_interleave(2, dim=1) # [S, D]
187
+ freqs_sin = freqs.sin().repeat_interleave(2, dim=1) # [S, D]
188
+ return freqs_cos, freqs_sin
189
+ else:
190
+ freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 # [S, D/2]
191
+ return freqs_cis
192
+
193
+
194
+
195
+ def calc_sizes(rope_img, patch_size, th, tw):
196
+ if rope_img == 'extend':
197
+ # Expansion mode
198
+ sub_args = [(th, tw)]
199
+ elif rope_img.startswith('base'):
200
+ # Based on the specified dimensions, other dimensions are obtained through interpolation.
201
+ base_size = int(rope_img[4:]) // 8 // patch_size
202
+ start, stop = get_fill_resize_and_crop((th, tw), base_size)
203
+ sub_args = [start, stop, (th, tw)]
204
+ else:
205
+ raise ValueError(f"Unknown rope_img: {rope_img}")
206
+ return sub_args
207
+
208
+
209
+ def init_image_posemb(rope_img,
210
+ resolutions,
211
+ patch_size,
212
+ hidden_size,
213
+ num_heads,
214
+ log_fn,
215
+ rope_real=True,
216
+ ):
217
+ freqs_cis_img = {}
218
+ for reso in resolutions:
219
+ th, tw = reso.height // 8 // patch_size, reso.width // 8 // patch_size
220
+ sub_args = calc_sizes(rope_img, patch_size, th, tw)
221
+ freqs_cis_img[str(reso)] = get_2d_rotary_pos_embed(hidden_size // num_heads, *sub_args, use_real=rope_real)
222
+ log_fn(f" Using image RoPE ({rope_img}) ({'real' if rope_real else 'complex'}): {sub_args} | ({reso}) "
223
+ f"{freqs_cis_img[str(reso)][0].shape if rope_real else freqs_cis_img[str(reso)].shape}")
224
+ return freqs_cis_img
ComfyUI/comfy/ldm/models/autoencoder.py ADDED
@@ -0,0 +1,226 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from contextlib import contextmanager
3
+ from typing import Any, Dict, List, Optional, Tuple, Union
4
+
5
+ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistribution
6
+
7
+ from comfy.ldm.util import instantiate_from_config
8
+ from comfy.ldm.modules.ema import LitEma
9
+ import comfy.ops
10
+
11
+ class DiagonalGaussianRegularizer(torch.nn.Module):
12
+ def __init__(self, sample: bool = True):
13
+ super().__init__()
14
+ self.sample = sample
15
+
16
+ def get_trainable_parameters(self) -> Any:
17
+ yield from ()
18
+
19
+ def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
20
+ log = dict()
21
+ posterior = DiagonalGaussianDistribution(z)
22
+ if self.sample:
23
+ z = posterior.sample()
24
+ else:
25
+ z = posterior.mode()
26
+ kl_loss = posterior.kl()
27
+ kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
28
+ log["kl_loss"] = kl_loss
29
+ return z, log
30
+
31
+
32
+ class AbstractAutoencoder(torch.nn.Module):
33
+ """
34
+ This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators,
35
+ unCLIP models, etc. Hence, it is fairly general, and specific features
36
+ (e.g. discriminator training, encoding, decoding) must be implemented in subclasses.
37
+ """
38
+
39
+ def __init__(
40
+ self,
41
+ ema_decay: Union[None, float] = None,
42
+ monitor: Union[None, str] = None,
43
+ input_key: str = "jpg",
44
+ **kwargs,
45
+ ):
46
+ super().__init__()
47
+
48
+ self.input_key = input_key
49
+ self.use_ema = ema_decay is not None
50
+ if monitor is not None:
51
+ self.monitor = monitor
52
+
53
+ if self.use_ema:
54
+ self.model_ema = LitEma(self, decay=ema_decay)
55
+ logpy.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
56
+
57
+ def get_input(self, batch) -> Any:
58
+ raise NotImplementedError()
59
+
60
+ def on_train_batch_end(self, *args, **kwargs):
61
+ # for EMA computation
62
+ if self.use_ema:
63
+ self.model_ema(self)
64
+
65
+ @contextmanager
66
+ def ema_scope(self, context=None):
67
+ if self.use_ema:
68
+ self.model_ema.store(self.parameters())
69
+ self.model_ema.copy_to(self)
70
+ if context is not None:
71
+ logpy.info(f"{context}: Switched to EMA weights")
72
+ try:
73
+ yield None
74
+ finally:
75
+ if self.use_ema:
76
+ self.model_ema.restore(self.parameters())
77
+ if context is not None:
78
+ logpy.info(f"{context}: Restored training weights")
79
+
80
+ def encode(self, *args, **kwargs) -> torch.Tensor:
81
+ raise NotImplementedError("encode()-method of abstract base class called")
82
+
83
+ def decode(self, *args, **kwargs) -> torch.Tensor:
84
+ raise NotImplementedError("decode()-method of abstract base class called")
85
+
86
+ def instantiate_optimizer_from_config(self, params, lr, cfg):
87
+ logpy.info(f"loading >>> {cfg['target']} <<< optimizer from config")
88
+ return get_obj_from_str(cfg["target"])(
89
+ params, lr=lr, **cfg.get("params", dict())
90
+ )
91
+
92
+ def configure_optimizers(self) -> Any:
93
+ raise NotImplementedError()
94
+
95
+
96
+ class AutoencodingEngine(AbstractAutoencoder):
97
+ """
98
+ Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL
99
+ (we also restore them explicitly as special cases for legacy reasons).
100
+ Regularizations such as KL or VQ are moved to the regularizer class.
101
+ """
102
+
103
+ def __init__(
104
+ self,
105
+ *args,
106
+ encoder_config: Dict,
107
+ decoder_config: Dict,
108
+ regularizer_config: Dict,
109
+ **kwargs,
110
+ ):
111
+ super().__init__(*args, **kwargs)
112
+
113
+ self.encoder: torch.nn.Module = instantiate_from_config(encoder_config)
114
+ self.decoder: torch.nn.Module = instantiate_from_config(decoder_config)
115
+ self.regularization: AbstractRegularizer = instantiate_from_config(
116
+ regularizer_config
117
+ )
118
+
119
+ def get_last_layer(self):
120
+ return self.decoder.get_last_layer()
121
+
122
+ def encode(
123
+ self,
124
+ x: torch.Tensor,
125
+ return_reg_log: bool = False,
126
+ unregularized: bool = False,
127
+ ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
128
+ z = self.encoder(x)
129
+ if unregularized:
130
+ return z, dict()
131
+ z, reg_log = self.regularization(z)
132
+ if return_reg_log:
133
+ return z, reg_log
134
+ return z
135
+
136
+ def decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor:
137
+ x = self.decoder(z, **kwargs)
138
+ return x
139
+
140
+ def forward(
141
+ self, x: torch.Tensor, **additional_decode_kwargs
142
+ ) -> Tuple[torch.Tensor, torch.Tensor, dict]:
143
+ z, reg_log = self.encode(x, return_reg_log=True)
144
+ dec = self.decode(z, **additional_decode_kwargs)
145
+ return z, dec, reg_log
146
+
147
+
148
+ class AutoencodingEngineLegacy(AutoencodingEngine):
149
+ def __init__(self, embed_dim: int, **kwargs):
150
+ self.max_batch_size = kwargs.pop("max_batch_size", None)
151
+ ddconfig = kwargs.pop("ddconfig")
152
+ super().__init__(
153
+ encoder_config={
154
+ "target": "comfy.ldm.modules.diffusionmodules.model.Encoder",
155
+ "params": ddconfig,
156
+ },
157
+ decoder_config={
158
+ "target": "comfy.ldm.modules.diffusionmodules.model.Decoder",
159
+ "params": ddconfig,
160
+ },
161
+ **kwargs,
162
+ )
163
+ self.quant_conv = comfy.ops.disable_weight_init.Conv2d(
164
+ (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
165
+ (1 + ddconfig["double_z"]) * embed_dim,
166
+ 1,
167
+ )
168
+ self.post_quant_conv = comfy.ops.disable_weight_init.Conv2d(embed_dim, ddconfig["z_channels"], 1)
169
+ self.embed_dim = embed_dim
170
+
171
+ def get_autoencoder_params(self) -> list:
172
+ params = super().get_autoencoder_params()
173
+ return params
174
+
175
+ def encode(
176
+ self, x: torch.Tensor, return_reg_log: bool = False
177
+ ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
178
+ if self.max_batch_size is None:
179
+ z = self.encoder(x)
180
+ z = self.quant_conv(z)
181
+ else:
182
+ N = x.shape[0]
183
+ bs = self.max_batch_size
184
+ n_batches = int(math.ceil(N / bs))
185
+ z = list()
186
+ for i_batch in range(n_batches):
187
+ z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs])
188
+ z_batch = self.quant_conv(z_batch)
189
+ z.append(z_batch)
190
+ z = torch.cat(z, 0)
191
+
192
+ z, reg_log = self.regularization(z)
193
+ if return_reg_log:
194
+ return z, reg_log
195
+ return z
196
+
197
+ def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
198
+ if self.max_batch_size is None:
199
+ dec = self.post_quant_conv(z)
200
+ dec = self.decoder(dec, **decoder_kwargs)
201
+ else:
202
+ N = z.shape[0]
203
+ bs = self.max_batch_size
204
+ n_batches = int(math.ceil(N / bs))
205
+ dec = list()
206
+ for i_batch in range(n_batches):
207
+ dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs])
208
+ dec_batch = self.decoder(dec_batch, **decoder_kwargs)
209
+ dec.append(dec_batch)
210
+ dec = torch.cat(dec, 0)
211
+
212
+ return dec
213
+
214
+
215
+ class AutoencoderKL(AutoencodingEngineLegacy):
216
+ def __init__(self, **kwargs):
217
+ if "lossconfig" in kwargs:
218
+ kwargs["loss_config"] = kwargs.pop("lossconfig")
219
+ super().__init__(
220
+ regularizer_config={
221
+ "target": (
222
+ "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"
223
+ )
224
+ },
225
+ **kwargs,
226
+ )
ComfyUI/comfy/ldm/modules/attention.py ADDED
@@ -0,0 +1,865 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ import torch.nn.functional as F
4
+ from torch import nn, einsum
5
+ from einops import rearrange, repeat
6
+ from typing import Optional
7
+ import logging
8
+
9
+ from .diffusionmodules.util import AlphaBlender, timestep_embedding
10
+ from .sub_quadratic_attention import efficient_dot_product_attention
11
+
12
+ from comfy import model_management
13
+
14
+ if model_management.xformers_enabled():
15
+ import xformers
16
+ import xformers.ops
17
+
18
+ from comfy.cli_args import args
19
+ import comfy.ops
20
+ ops = comfy.ops.disable_weight_init
21
+
22
+ FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype()
23
+
24
+ def get_attn_precision(attn_precision):
25
+ if args.dont_upcast_attention:
26
+ return None
27
+ if FORCE_UPCAST_ATTENTION_DTYPE is not None:
28
+ return FORCE_UPCAST_ATTENTION_DTYPE
29
+ return attn_precision
30
+
31
+ def exists(val):
32
+ return val is not None
33
+
34
+
35
+ def uniq(arr):
36
+ return{el: True for el in arr}.keys()
37
+
38
+
39
+ def default(val, d):
40
+ if exists(val):
41
+ return val
42
+ return d
43
+
44
+
45
+ def max_neg_value(t):
46
+ return -torch.finfo(t.dtype).max
47
+
48
+
49
+ def init_(tensor):
50
+ dim = tensor.shape[-1]
51
+ std = 1 / math.sqrt(dim)
52
+ tensor.uniform_(-std, std)
53
+ return tensor
54
+
55
+
56
+ # feedforward
57
+ class GEGLU(nn.Module):
58
+ def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
59
+ super().__init__()
60
+ self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device)
61
+
62
+ def forward(self, x):
63
+ x, gate = self.proj(x).chunk(2, dim=-1)
64
+ return x * F.gelu(gate)
65
+
66
+
67
+ class FeedForward(nn.Module):
68
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=ops):
69
+ super().__init__()
70
+ inner_dim = int(dim * mult)
71
+ dim_out = default(dim_out, dim)
72
+ project_in = nn.Sequential(
73
+ operations.Linear(dim, inner_dim, dtype=dtype, device=device),
74
+ nn.GELU()
75
+ ) if not glu else GEGLU(dim, inner_dim, dtype=dtype, device=device, operations=operations)
76
+
77
+ self.net = nn.Sequential(
78
+ project_in,
79
+ nn.Dropout(dropout),
80
+ operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
81
+ )
82
+
83
+ def forward(self, x):
84
+ return self.net(x)
85
+
86
+ def Normalize(in_channels, dtype=None, device=None):
87
+ return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
88
+
89
+ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
90
+ attn_precision = get_attn_precision(attn_precision)
91
+
92
+ if skip_reshape:
93
+ b, _, _, dim_head = q.shape
94
+ else:
95
+ b, _, dim_head = q.shape
96
+ dim_head //= heads
97
+
98
+ scale = dim_head ** -0.5
99
+
100
+ h = heads
101
+ if skip_reshape:
102
+ q, k, v = map(
103
+ lambda t: t.reshape(b * heads, -1, dim_head),
104
+ (q, k, v),
105
+ )
106
+ else:
107
+ q, k, v = map(
108
+ lambda t: t.unsqueeze(3)
109
+ .reshape(b, -1, heads, dim_head)
110
+ .permute(0, 2, 1, 3)
111
+ .reshape(b * heads, -1, dim_head)
112
+ .contiguous(),
113
+ (q, k, v),
114
+ )
115
+
116
+ # force cast to fp32 to avoid overflowing
117
+ if attn_precision == torch.float32:
118
+ sim = einsum('b i d, b j d -> b i j', q.float(), k.float()) * scale
119
+ else:
120
+ sim = einsum('b i d, b j d -> b i j', q, k) * scale
121
+
122
+ del q, k
123
+
124
+ if exists(mask):
125
+ if mask.dtype == torch.bool:
126
+ mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention
127
+ max_neg_value = -torch.finfo(sim.dtype).max
128
+ mask = repeat(mask, 'b j -> (b h) () j', h=h)
129
+ sim.masked_fill_(~mask, max_neg_value)
130
+ else:
131
+ if len(mask.shape) == 2:
132
+ bs = 1
133
+ else:
134
+ bs = mask.shape[0]
135
+ mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
136
+ sim.add_(mask)
137
+
138
+ # attention, what we cannot get enough of
139
+ sim = sim.softmax(dim=-1)
140
+
141
+ out = einsum('b i j, b j d -> b i d', sim.to(v.dtype), v)
142
+ out = (
143
+ out.unsqueeze(0)
144
+ .reshape(b, heads, -1, dim_head)
145
+ .permute(0, 2, 1, 3)
146
+ .reshape(b, -1, heads * dim_head)
147
+ )
148
+ return out
149
+
150
+
151
+ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False):
152
+ attn_precision = get_attn_precision(attn_precision)
153
+
154
+ if skip_reshape:
155
+ b, _, _, dim_head = query.shape
156
+ else:
157
+ b, _, dim_head = query.shape
158
+ dim_head //= heads
159
+
160
+ scale = dim_head ** -0.5
161
+
162
+ if skip_reshape:
163
+ query = query.reshape(b * heads, -1, dim_head)
164
+ value = value.reshape(b * heads, -1, dim_head)
165
+ key = key.reshape(b * heads, -1, dim_head).movedim(1, 2)
166
+ else:
167
+ query = query.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
168
+ value = value.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
169
+ key = key.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 3, 1).reshape(b * heads, dim_head, -1)
170
+
171
+
172
+ dtype = query.dtype
173
+ upcast_attention = attn_precision == torch.float32 and query.dtype != torch.float32
174
+ if upcast_attention:
175
+ bytes_per_token = torch.finfo(torch.float32).bits//8
176
+ else:
177
+ bytes_per_token = torch.finfo(query.dtype).bits//8
178
+ batch_x_heads, q_tokens, _ = query.shape
179
+ _, _, k_tokens = key.shape
180
+ qk_matmul_size_bytes = batch_x_heads * bytes_per_token * q_tokens * k_tokens
181
+
182
+ mem_free_total, mem_free_torch = model_management.get_free_memory(query.device, True)
183
+
184
+ kv_chunk_size_min = None
185
+ kv_chunk_size = None
186
+ query_chunk_size = None
187
+
188
+ for x in [4096, 2048, 1024, 512, 256]:
189
+ count = mem_free_total / (batch_x_heads * bytes_per_token * x * 4.0)
190
+ if count >= k_tokens:
191
+ kv_chunk_size = k_tokens
192
+ query_chunk_size = x
193
+ break
194
+
195
+ if query_chunk_size is None:
196
+ query_chunk_size = 512
197
+
198
+ if mask is not None:
199
+ if len(mask.shape) == 2:
200
+ bs = 1
201
+ else:
202
+ bs = mask.shape[0]
203
+ mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
204
+
205
+ hidden_states = efficient_dot_product_attention(
206
+ query,
207
+ key,
208
+ value,
209
+ query_chunk_size=query_chunk_size,
210
+ kv_chunk_size=kv_chunk_size,
211
+ kv_chunk_size_min=kv_chunk_size_min,
212
+ use_checkpoint=False,
213
+ upcast_attention=upcast_attention,
214
+ mask=mask,
215
+ )
216
+
217
+ hidden_states = hidden_states.to(dtype)
218
+
219
+ hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
220
+ return hidden_states
221
+
222
+ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
223
+ attn_precision = get_attn_precision(attn_precision)
224
+
225
+ if skip_reshape:
226
+ b, _, _, dim_head = q.shape
227
+ else:
228
+ b, _, dim_head = q.shape
229
+ dim_head //= heads
230
+
231
+ scale = dim_head ** -0.5
232
+
233
+ h = heads
234
+ if skip_reshape:
235
+ q, k, v = map(
236
+ lambda t: t.reshape(b * heads, -1, dim_head),
237
+ (q, k, v),
238
+ )
239
+ else:
240
+ q, k, v = map(
241
+ lambda t: t.unsqueeze(3)
242
+ .reshape(b, -1, heads, dim_head)
243
+ .permute(0, 2, 1, 3)
244
+ .reshape(b * heads, -1, dim_head)
245
+ .contiguous(),
246
+ (q, k, v),
247
+ )
248
+
249
+ r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
250
+
251
+ mem_free_total = model_management.get_free_memory(q.device)
252
+
253
+ if attn_precision == torch.float32:
254
+ element_size = 4
255
+ upcast = True
256
+ else:
257
+ element_size = q.element_size()
258
+ upcast = False
259
+
260
+ gb = 1024 ** 3
261
+ tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size
262
+ modifier = 3
263
+ mem_required = tensor_size * modifier
264
+ steps = 1
265
+
266
+
267
+ if mem_required > mem_free_total:
268
+ steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
269
+ # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
270
+ # f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
271
+
272
+ if steps > 64:
273
+ max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
274
+ raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
275
+ f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free')
276
+
277
+ if mask is not None:
278
+ if len(mask.shape) == 2:
279
+ bs = 1
280
+ else:
281
+ bs = mask.shape[0]
282
+ mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
283
+
284
+ # print("steps", steps, mem_required, mem_free_total, modifier, q.element_size(), tensor_size)
285
+ first_op_done = False
286
+ cleared_cache = False
287
+ while True:
288
+ try:
289
+ slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
290
+ for i in range(0, q.shape[1], slice_size):
291
+ end = i + slice_size
292
+ if upcast:
293
+ with torch.autocast(enabled=False, device_type = 'cuda'):
294
+ s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scale
295
+ else:
296
+ s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * scale
297
+
298
+ if mask is not None:
299
+ if len(mask.shape) == 2:
300
+ s1 += mask[i:end]
301
+ else:
302
+ s1 += mask[:, i:end]
303
+
304
+ s2 = s1.softmax(dim=-1).to(v.dtype)
305
+ del s1
306
+ first_op_done = True
307
+
308
+ r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
309
+ del s2
310
+ break
311
+ except model_management.OOM_EXCEPTION as e:
312
+ if first_op_done == False:
313
+ model_management.soft_empty_cache(True)
314
+ if cleared_cache == False:
315
+ cleared_cache = True
316
+ logging.warning("out of memory error, emptying cache and trying again")
317
+ continue
318
+ steps *= 2
319
+ if steps > 64:
320
+ raise e
321
+ logging.warning("out of memory error, increasing steps and trying again {}".format(steps))
322
+ else:
323
+ raise e
324
+
325
+ del q, k, v
326
+
327
+ r1 = (
328
+ r1.unsqueeze(0)
329
+ .reshape(b, heads, -1, dim_head)
330
+ .permute(0, 2, 1, 3)
331
+ .reshape(b, -1, heads * dim_head)
332
+ )
333
+ return r1
334
+
335
+ BROKEN_XFORMERS = False
336
+ try:
337
+ x_vers = xformers.__version__
338
+ # XFormers bug confirmed on all versions from 0.0.21 to 0.0.26 (q with bs bigger than 65535 gives CUDA error)
339
+ BROKEN_XFORMERS = x_vers.startswith("0.0.2") and not x_vers.startswith("0.0.20")
340
+ except:
341
+ pass
342
+
343
+ def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
344
+ if skip_reshape:
345
+ b, _, _, dim_head = q.shape
346
+ else:
347
+ b, _, dim_head = q.shape
348
+ dim_head //= heads
349
+
350
+ disabled_xformers = False
351
+
352
+ if BROKEN_XFORMERS:
353
+ if b * heads > 65535:
354
+ disabled_xformers = True
355
+
356
+ if not disabled_xformers:
357
+ if torch.jit.is_tracing() or torch.jit.is_scripting():
358
+ disabled_xformers = True
359
+
360
+ if disabled_xformers:
361
+ return attention_pytorch(q, k, v, heads, mask)
362
+
363
+ if skip_reshape:
364
+ q, k, v = map(
365
+ lambda t: t.reshape(b * heads, -1, dim_head),
366
+ (q, k, v),
367
+ )
368
+ else:
369
+ q, k, v = map(
370
+ lambda t: t.reshape(b, -1, heads, dim_head),
371
+ (q, k, v),
372
+ )
373
+
374
+ if mask is not None:
375
+ pad = 8 - q.shape[1] % 8
376
+ mask_out = torch.empty([q.shape[0], q.shape[1], q.shape[1] + pad], dtype=q.dtype, device=q.device)
377
+ mask_out[:, :, :mask.shape[-1]] = mask
378
+ mask = mask_out[:, :, :mask.shape[-1]]
379
+
380
+ out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=mask)
381
+
382
+ if skip_reshape:
383
+ out = (
384
+ out.unsqueeze(0)
385
+ .reshape(b, heads, -1, dim_head)
386
+ .permute(0, 2, 1, 3)
387
+ .reshape(b, -1, heads * dim_head)
388
+ )
389
+ else:
390
+ out = (
391
+ out.reshape(b, -1, heads * dim_head)
392
+ )
393
+
394
+ return out
395
+
396
+ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
397
+ if skip_reshape:
398
+ b, _, _, dim_head = q.shape
399
+ else:
400
+ b, _, dim_head = q.shape
401
+ dim_head //= heads
402
+ q, k, v = map(
403
+ lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
404
+ (q, k, v),
405
+ )
406
+
407
+ out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
408
+ out = (
409
+ out.transpose(1, 2).reshape(b, -1, heads * dim_head)
410
+ )
411
+ return out
412
+
413
+
414
+ optimized_attention = attention_basic
415
+
416
+ if model_management.xformers_enabled():
417
+ logging.info("Using xformers cross attention")
418
+ optimized_attention = attention_xformers
419
+ elif model_management.pytorch_attention_enabled():
420
+ logging.info("Using pytorch cross attention")
421
+ optimized_attention = attention_pytorch
422
+ else:
423
+ if args.use_split_cross_attention:
424
+ logging.info("Using split optimization for cross attention")
425
+ optimized_attention = attention_split
426
+ else:
427
+ logging.info("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention")
428
+ optimized_attention = attention_sub_quad
429
+
430
+ optimized_attention_masked = optimized_attention
431
+
432
+ def optimized_attention_for_device(device, mask=False, small_input=False):
433
+ if small_input:
434
+ if model_management.pytorch_attention_enabled():
435
+ return attention_pytorch #TODO: need to confirm but this is probably slightly faster for small inputs in all cases
436
+ else:
437
+ return attention_basic
438
+
439
+ if device == torch.device("cpu"):
440
+ return attention_sub_quad
441
+
442
+ if mask:
443
+ return optimized_attention_masked
444
+
445
+ return optimized_attention
446
+
447
+
448
+ class CrossAttention(nn.Module):
449
+ def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=ops):
450
+ super().__init__()
451
+ inner_dim = dim_head * heads
452
+ context_dim = default(context_dim, query_dim)
453
+ self.attn_precision = attn_precision
454
+
455
+ self.heads = heads
456
+ self.dim_head = dim_head
457
+
458
+ self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
459
+ self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
460
+ self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
461
+
462
+ self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
463
+
464
+ def forward(self, x, context=None, value=None, mask=None):
465
+ q = self.to_q(x)
466
+ context = default(context, x)
467
+ k = self.to_k(context)
468
+ if value is not None:
469
+ v = self.to_v(value)
470
+ del value
471
+ else:
472
+ v = self.to_v(context)
473
+
474
+ if mask is None:
475
+ out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
476
+ else:
477
+ out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
478
+ return self.to_out(out)
479
+
480
+
481
+ class BasicTransformerBlock(nn.Module):
482
+ def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, ff_in=False, inner_dim=None,
483
+ disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, attn_precision=None, dtype=None, device=None, operations=ops):
484
+ super().__init__()
485
+
486
+ self.ff_in = ff_in or inner_dim is not None
487
+ if inner_dim is None:
488
+ inner_dim = dim
489
+
490
+ self.is_res = inner_dim == dim
491
+ self.attn_precision = attn_precision
492
+
493
+ if self.ff_in:
494
+ self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device)
495
+ self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
496
+
497
+ self.disable_self_attn = disable_self_attn
498
+ self.attn1 = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
499
+ context_dim=context_dim if self.disable_self_attn else None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) # is a self-attention if not self.disable_self_attn
500
+ self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
501
+
502
+ if disable_temporal_crossattention:
503
+ if switch_temporal_ca_to_sa:
504
+ raise ValueError
505
+ else:
506
+ self.attn2 = None
507
+ else:
508
+ context_dim_attn2 = None
509
+ if not switch_temporal_ca_to_sa:
510
+ context_dim_attn2 = context_dim
511
+
512
+ self.attn2 = CrossAttention(query_dim=inner_dim, context_dim=context_dim_attn2,
513
+ heads=n_heads, dim_head=d_head, dropout=dropout, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) # is self-attn if context is none
514
+ self.norm2 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
515
+
516
+ self.norm1 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
517
+ self.norm3 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
518
+ self.n_heads = n_heads
519
+ self.d_head = d_head
520
+ self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
521
+
522
+ def forward(self, x, context=None, transformer_options={}):
523
+ extra_options = {}
524
+ block = transformer_options.get("block", None)
525
+ block_index = transformer_options.get("block_index", 0)
526
+ transformer_patches = {}
527
+ transformer_patches_replace = {}
528
+
529
+ for k in transformer_options:
530
+ if k == "patches":
531
+ transformer_patches = transformer_options[k]
532
+ elif k == "patches_replace":
533
+ transformer_patches_replace = transformer_options[k]
534
+ else:
535
+ extra_options[k] = transformer_options[k]
536
+
537
+ extra_options["n_heads"] = self.n_heads
538
+ extra_options["dim_head"] = self.d_head
539
+ extra_options["attn_precision"] = self.attn_precision
540
+
541
+ if self.ff_in:
542
+ x_skip = x
543
+ x = self.ff_in(self.norm_in(x))
544
+ if self.is_res:
545
+ x += x_skip
546
+
547
+ n = self.norm1(x)
548
+ if self.disable_self_attn:
549
+ context_attn1 = context
550
+ else:
551
+ context_attn1 = None
552
+ value_attn1 = None
553
+
554
+ if "attn1_patch" in transformer_patches:
555
+ patch = transformer_patches["attn1_patch"]
556
+ if context_attn1 is None:
557
+ context_attn1 = n
558
+ value_attn1 = context_attn1
559
+ for p in patch:
560
+ n, context_attn1, value_attn1 = p(n, context_attn1, value_attn1, extra_options)
561
+
562
+ if block is not None:
563
+ transformer_block = (block[0], block[1], block_index)
564
+ else:
565
+ transformer_block = None
566
+ attn1_replace_patch = transformer_patches_replace.get("attn1", {})
567
+ block_attn1 = transformer_block
568
+ if block_attn1 not in attn1_replace_patch:
569
+ block_attn1 = block
570
+
571
+ if block_attn1 in attn1_replace_patch:
572
+ if context_attn1 is None:
573
+ context_attn1 = n
574
+ value_attn1 = n
575
+ n = self.attn1.to_q(n)
576
+ context_attn1 = self.attn1.to_k(context_attn1)
577
+ value_attn1 = self.attn1.to_v(value_attn1)
578
+ n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
579
+ n = self.attn1.to_out(n)
580
+ else:
581
+ n = self.attn1(n, context=context_attn1, value=value_attn1)
582
+
583
+ if "attn1_output_patch" in transformer_patches:
584
+ patch = transformer_patches["attn1_output_patch"]
585
+ for p in patch:
586
+ n = p(n, extra_options)
587
+
588
+ x += n
589
+ if "middle_patch" in transformer_patches:
590
+ patch = transformer_patches["middle_patch"]
591
+ for p in patch:
592
+ x = p(x, extra_options)
593
+
594
+ if self.attn2 is not None:
595
+ n = self.norm2(x)
596
+ if self.switch_temporal_ca_to_sa:
597
+ context_attn2 = n
598
+ else:
599
+ context_attn2 = context
600
+ value_attn2 = None
601
+ if "attn2_patch" in transformer_patches:
602
+ patch = transformer_patches["attn2_patch"]
603
+ value_attn2 = context_attn2
604
+ for p in patch:
605
+ n, context_attn2, value_attn2 = p(n, context_attn2, value_attn2, extra_options)
606
+
607
+ attn2_replace_patch = transformer_patches_replace.get("attn2", {})
608
+ block_attn2 = transformer_block
609
+ if block_attn2 not in attn2_replace_patch:
610
+ block_attn2 = block
611
+
612
+ if block_attn2 in attn2_replace_patch:
613
+ if value_attn2 is None:
614
+ value_attn2 = context_attn2
615
+ n = self.attn2.to_q(n)
616
+ context_attn2 = self.attn2.to_k(context_attn2)
617
+ value_attn2 = self.attn2.to_v(value_attn2)
618
+ n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
619
+ n = self.attn2.to_out(n)
620
+ else:
621
+ n = self.attn2(n, context=context_attn2, value=value_attn2)
622
+
623
+ if "attn2_output_patch" in transformer_patches:
624
+ patch = transformer_patches["attn2_output_patch"]
625
+ for p in patch:
626
+ n = p(n, extra_options)
627
+
628
+ x += n
629
+ if self.is_res:
630
+ x_skip = x
631
+ x = self.ff(self.norm3(x))
632
+ if self.is_res:
633
+ x += x_skip
634
+
635
+ return x
636
+
637
+
638
+ class SpatialTransformer(nn.Module):
639
+ """
640
+ Transformer block for image-like data.
641
+ First, project the input (aka embedding)
642
+ and reshape to b, t, d.
643
+ Then apply standard transformer action.
644
+ Finally, reshape to image
645
+ NEW: use_linear for more efficiency instead of the 1x1 convs
646
+ """
647
+ def __init__(self, in_channels, n_heads, d_head,
648
+ depth=1, dropout=0., context_dim=None,
649
+ disable_self_attn=False, use_linear=False,
650
+ use_checkpoint=True, attn_precision=None, dtype=None, device=None, operations=ops):
651
+ super().__init__()
652
+ if exists(context_dim) and not isinstance(context_dim, list):
653
+ context_dim = [context_dim] * depth
654
+ self.in_channels = in_channels
655
+ inner_dim = n_heads * d_head
656
+ self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
657
+ if not use_linear:
658
+ self.proj_in = operations.Conv2d(in_channels,
659
+ inner_dim,
660
+ kernel_size=1,
661
+ stride=1,
662
+ padding=0, dtype=dtype, device=device)
663
+ else:
664
+ self.proj_in = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
665
+
666
+ self.transformer_blocks = nn.ModuleList(
667
+ [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
668
+ disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
669
+ for d in range(depth)]
670
+ )
671
+ if not use_linear:
672
+ self.proj_out = operations.Conv2d(inner_dim,in_channels,
673
+ kernel_size=1,
674
+ stride=1,
675
+ padding=0, dtype=dtype, device=device)
676
+ else:
677
+ self.proj_out = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
678
+ self.use_linear = use_linear
679
+
680
+ def forward(self, x, context=None, transformer_options={}):
681
+ # note: if no context is given, cross-attention defaults to self-attention
682
+ if not isinstance(context, list):
683
+ context = [context] * len(self.transformer_blocks)
684
+ b, c, h, w = x.shape
685
+ x_in = x
686
+ x = self.norm(x)
687
+ if not self.use_linear:
688
+ x = self.proj_in(x)
689
+ x = x.movedim(1, 3).flatten(1, 2).contiguous()
690
+ if self.use_linear:
691
+ x = self.proj_in(x)
692
+ for i, block in enumerate(self.transformer_blocks):
693
+ transformer_options["block_index"] = i
694
+ x = block(x, context=context[i], transformer_options=transformer_options)
695
+ if self.use_linear:
696
+ x = self.proj_out(x)
697
+ x = x.reshape(x.shape[0], h, w, x.shape[-1]).movedim(3, 1).contiguous()
698
+ if not self.use_linear:
699
+ x = self.proj_out(x)
700
+ return x + x_in
701
+
702
+
703
+ class SpatialVideoTransformer(SpatialTransformer):
704
+ def __init__(
705
+ self,
706
+ in_channels,
707
+ n_heads,
708
+ d_head,
709
+ depth=1,
710
+ dropout=0.0,
711
+ use_linear=False,
712
+ context_dim=None,
713
+ use_spatial_context=False,
714
+ timesteps=None,
715
+ merge_strategy: str = "fixed",
716
+ merge_factor: float = 0.5,
717
+ time_context_dim=None,
718
+ ff_in=False,
719
+ checkpoint=False,
720
+ time_depth=1,
721
+ disable_self_attn=False,
722
+ disable_temporal_crossattention=False,
723
+ max_time_embed_period: int = 10000,
724
+ attn_precision=None,
725
+ dtype=None, device=None, operations=ops
726
+ ):
727
+ super().__init__(
728
+ in_channels,
729
+ n_heads,
730
+ d_head,
731
+ depth=depth,
732
+ dropout=dropout,
733
+ use_checkpoint=checkpoint,
734
+ context_dim=context_dim,
735
+ use_linear=use_linear,
736
+ disable_self_attn=disable_self_attn,
737
+ attn_precision=attn_precision,
738
+ dtype=dtype, device=device, operations=operations
739
+ )
740
+ self.time_depth = time_depth
741
+ self.depth = depth
742
+ self.max_time_embed_period = max_time_embed_period
743
+
744
+ time_mix_d_head = d_head
745
+ n_time_mix_heads = n_heads
746
+
747
+ time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads)
748
+
749
+ inner_dim = n_heads * d_head
750
+ if use_spatial_context:
751
+ time_context_dim = context_dim
752
+
753
+ self.time_stack = nn.ModuleList(
754
+ [
755
+ BasicTransformerBlock(
756
+ inner_dim,
757
+ n_time_mix_heads,
758
+ time_mix_d_head,
759
+ dropout=dropout,
760
+ context_dim=time_context_dim,
761
+ # timesteps=timesteps,
762
+ checkpoint=checkpoint,
763
+ ff_in=ff_in,
764
+ inner_dim=time_mix_inner_dim,
765
+ disable_self_attn=disable_self_attn,
766
+ disable_temporal_crossattention=disable_temporal_crossattention,
767
+ attn_precision=attn_precision,
768
+ dtype=dtype, device=device, operations=operations
769
+ )
770
+ for _ in range(self.depth)
771
+ ]
772
+ )
773
+
774
+ assert len(self.time_stack) == len(self.transformer_blocks)
775
+
776
+ self.use_spatial_context = use_spatial_context
777
+ self.in_channels = in_channels
778
+
779
+ time_embed_dim = self.in_channels * 4
780
+ self.time_pos_embed = nn.Sequential(
781
+ operations.Linear(self.in_channels, time_embed_dim, dtype=dtype, device=device),
782
+ nn.SiLU(),
783
+ operations.Linear(time_embed_dim, self.in_channels, dtype=dtype, device=device),
784
+ )
785
+
786
+ self.time_mixer = AlphaBlender(
787
+ alpha=merge_factor, merge_strategy=merge_strategy
788
+ )
789
+
790
+ def forward(
791
+ self,
792
+ x: torch.Tensor,
793
+ context: Optional[torch.Tensor] = None,
794
+ time_context: Optional[torch.Tensor] = None,
795
+ timesteps: Optional[int] = None,
796
+ image_only_indicator: Optional[torch.Tensor] = None,
797
+ transformer_options={}
798
+ ) -> torch.Tensor:
799
+ _, _, h, w = x.shape
800
+ x_in = x
801
+ spatial_context = None
802
+ if exists(context):
803
+ spatial_context = context
804
+
805
+ if self.use_spatial_context:
806
+ assert (
807
+ context.ndim == 3
808
+ ), f"n dims of spatial context should be 3 but are {context.ndim}"
809
+
810
+ if time_context is None:
811
+ time_context = context
812
+ time_context_first_timestep = time_context[::timesteps]
813
+ time_context = repeat(
814
+ time_context_first_timestep, "b ... -> (b n) ...", n=h * w
815
+ )
816
+ elif time_context is not None and not self.use_spatial_context:
817
+ time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w)
818
+ if time_context.ndim == 2:
819
+ time_context = rearrange(time_context, "b c -> b 1 c")
820
+
821
+ x = self.norm(x)
822
+ if not self.use_linear:
823
+ x = self.proj_in(x)
824
+ x = rearrange(x, "b c h w -> b (h w) c")
825
+ if self.use_linear:
826
+ x = self.proj_in(x)
827
+
828
+ num_frames = torch.arange(timesteps, device=x.device)
829
+ num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
830
+ num_frames = rearrange(num_frames, "b t -> (b t)")
831
+ t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False, max_period=self.max_time_embed_period).to(x.dtype)
832
+ emb = self.time_pos_embed(t_emb)
833
+ emb = emb[:, None, :]
834
+
835
+ for it_, (block, mix_block) in enumerate(
836
+ zip(self.transformer_blocks, self.time_stack)
837
+ ):
838
+ transformer_options["block_index"] = it_
839
+ x = block(
840
+ x,
841
+ context=spatial_context,
842
+ transformer_options=transformer_options,
843
+ )
844
+
845
+ x_mix = x
846
+ x_mix = x_mix + emb
847
+
848
+ B, S, C = x_mix.shape
849
+ x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
850
+ x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
851
+ x_mix = rearrange(
852
+ x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
853
+ )
854
+
855
+ x = self.time_mixer(x_spatial=x, x_temporal=x_mix, image_only_indicator=image_only_indicator)
856
+
857
+ if self.use_linear:
858
+ x = self.proj_out(x)
859
+ x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
860
+ if not self.use_linear:
861
+ x = self.proj_out(x)
862
+ out = x + x_in
863
+ return out
864
+
865
+