Spaces:
Running
on
Zero
Running
on
Zero
Use diffusers implementation
Browse files- .gitmodules +0 -3
- .vscode/settings.json +18 -0
- Attend-and-Excite +0 -1
- app.py +6 -11
- model.py +45 -55
- requirements.txt +5 -8
.gitmodules
CHANGED
@@ -1,3 +0,0 @@
|
|
1 |
-
[submodule "Attend-and-Excite"]
|
2 |
-
path = Attend-and-Excite
|
3 |
-
url = https://github.com/AttendAndExcite/Attend-and-Excite
|
|
|
|
|
|
|
|
.vscode/settings.json
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"python.linting.enabled": true,
|
3 |
+
"python.linting.flake8Enabled": true,
|
4 |
+
"python.linting.pylintEnabled": false,
|
5 |
+
"python.linting.lintOnSave": true,
|
6 |
+
"python.formatting.provider": "yapf",
|
7 |
+
"python.formatting.yapfArgs": [
|
8 |
+
"--style={based_on_style: pep8, indent_width: 4, blank_line_before_nested_class_or_def: false, spaces_before_comment: 2, split_before_logical_operator: true}"
|
9 |
+
],
|
10 |
+
"[python]": {
|
11 |
+
"editor.formatOnType": true,
|
12 |
+
"editor.codeActionsOnSave": {
|
13 |
+
"source.organizeImports": true
|
14 |
+
}
|
15 |
+
},
|
16 |
+
"editor.formatOnSave": true,
|
17 |
+
"files.insertFinalNewline": true
|
18 |
+
}
|
Attend-and-Excite
DELETED
@@ -1 +0,0 @@
|
|
1 |
-
Subproject commit 41620338367f980b9d73752360ffd2557d8ddf5d
|
|
|
|
app.py
CHANGED
@@ -24,12 +24,11 @@ def process_example(
|
|
24 |
seed: int,
|
25 |
apply_attend_and_excite: bool,
|
26 |
) -> tuple[list[tuple[int, str]], PIL.Image.Image]:
|
27 |
-
model_id = 'CompVis/stable-diffusion-v1-4'
|
28 |
num_steps = 50
|
29 |
guidance_scale = 7.5
|
30 |
|
31 |
-
token_table = model.get_token_table(
|
32 |
-
result = model.run(
|
33 |
apply_attend_and_excite, num_steps, guidance_scale)
|
34 |
return token_table, result
|
35 |
|
@@ -39,9 +38,6 @@ with gr.Blocks(css='style.css') as demo:
|
|
39 |
|
40 |
with gr.Row():
|
41 |
with gr.Column():
|
42 |
-
model_id = gr.Text(label='Model ID',
|
43 |
-
value='CompVis/stable-diffusion-v1-4',
|
44 |
-
visible=False)
|
45 |
prompt = gr.Text(
|
46 |
label='Prompt',
|
47 |
max_lines=1,
|
@@ -171,13 +167,12 @@ with gr.Blocks(css='style.css') as demo:
|
|
171 |
|
172 |
show_token_indices_button.click(
|
173 |
fn=model.get_token_table,
|
174 |
-
inputs=
|
175 |
outputs=token_indices_table,
|
176 |
queue=False,
|
177 |
)
|
178 |
|
179 |
inputs = [
|
180 |
-
model_id,
|
181 |
prompt,
|
182 |
token_indices_str,
|
183 |
seed,
|
@@ -187,7 +182,7 @@ with gr.Blocks(css='style.css') as demo:
|
|
187 |
]
|
188 |
prompt.submit(
|
189 |
fn=model.get_token_table,
|
190 |
-
inputs=
|
191 |
outputs=token_indices_table,
|
192 |
queue=False,
|
193 |
).then(
|
@@ -197,7 +192,7 @@ with gr.Blocks(css='style.css') as demo:
|
|
197 |
)
|
198 |
token_indices_str.submit(
|
199 |
fn=model.get_token_table,
|
200 |
-
inputs=
|
201 |
outputs=token_indices_table,
|
202 |
queue=False,
|
203 |
).then(
|
@@ -207,7 +202,7 @@ with gr.Blocks(css='style.css') as demo:
|
|
207 |
)
|
208 |
run_button.click(
|
209 |
fn=model.get_token_table,
|
210 |
-
inputs=
|
211 |
outputs=token_indices_table,
|
212 |
queue=False,
|
213 |
).then(
|
|
|
24 |
seed: int,
|
25 |
apply_attend_and_excite: bool,
|
26 |
) -> tuple[list[tuple[int, str]], PIL.Image.Image]:
|
|
|
27 |
num_steps = 50
|
28 |
guidance_scale = 7.5
|
29 |
|
30 |
+
token_table = model.get_token_table(prompt)
|
31 |
+
result = model.run(prompt, indices_to_alter_str, seed,
|
32 |
apply_attend_and_excite, num_steps, guidance_scale)
|
33 |
return token_table, result
|
34 |
|
|
|
38 |
|
39 |
with gr.Row():
|
40 |
with gr.Column():
|
|
|
|
|
|
|
41 |
prompt = gr.Text(
|
42 |
label='Prompt',
|
43 |
max_lines=1,
|
|
|
167 |
|
168 |
show_token_indices_button.click(
|
169 |
fn=model.get_token_table,
|
170 |
+
inputs=prompt,
|
171 |
outputs=token_indices_table,
|
172 |
queue=False,
|
173 |
)
|
174 |
|
175 |
inputs = [
|
|
|
176 |
prompt,
|
177 |
token_indices_str,
|
178 |
seed,
|
|
|
182 |
]
|
183 |
prompt.submit(
|
184 |
fn=model.get_token_table,
|
185 |
+
inputs=prompt,
|
186 |
outputs=token_indices_table,
|
187 |
queue=False,
|
188 |
).then(
|
|
|
192 |
)
|
193 |
token_indices_str.submit(
|
194 |
fn=model.get_token_table,
|
195 |
+
inputs=prompt,
|
196 |
outputs=token_indices_table,
|
197 |
queue=False,
|
198 |
).then(
|
|
|
202 |
)
|
203 |
run_button.click(
|
204 |
fn=model.get_token_table,
|
205 |
+
inputs=prompt,
|
206 |
outputs=token_indices_table,
|
207 |
queue=False,
|
208 |
).then(
|
model.py
CHANGED
@@ -1,83 +1,73 @@
|
|
1 |
from __future__ import annotations
|
2 |
|
3 |
-
import sys
|
4 |
-
|
5 |
-
import gradio as gr
|
6 |
import PIL.Image
|
7 |
import torch
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
from config import RunConfig
|
12 |
-
from pipeline_attend_and_excite import AttendAndExcitePipeline
|
13 |
-
from run import run_on_prompt
|
14 |
-
from utils.ptp_utils import AttentionStore
|
15 |
|
16 |
|
17 |
class Model:
|
18 |
def __init__(self):
|
19 |
self.device = torch.device(
|
20 |
'cuda:0' if torch.cuda.is_available() else 'cpu')
|
21 |
-
|
22 |
-
self.
|
23 |
-
|
24 |
-
|
25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
-
def
|
28 |
-
if model_id == self.model_id:
|
29 |
-
return
|
30 |
-
self.model = AttendAndExcitePipeline.from_pretrained(model_id).to(
|
31 |
-
self.device)
|
32 |
-
self.tokenizer = self.model.tokenizer
|
33 |
-
self.model_id = model_id
|
34 |
-
|
35 |
-
def get_token_table(self, model_id: str, prompt: str):
|
36 |
-
self.load_model(model_id)
|
37 |
tokens = [
|
38 |
-
self.tokenizer.decode(t)
|
39 |
-
for t in self.tokenizer(prompt)['input_ids']
|
40 |
]
|
41 |
tokens = tokens[1:-1]
|
42 |
return list(enumerate(tokens, start=1))
|
43 |
|
44 |
def run(
|
45 |
self,
|
46 |
-
model_id: str,
|
47 |
prompt: str,
|
48 |
indices_to_alter_str: str,
|
49 |
-
seed: int,
|
50 |
-
apply_attend_and_excite: bool,
|
51 |
-
num_steps: int,
|
52 |
-
guidance_scale: float,
|
53 |
scale_factor: int = 20,
|
54 |
thresholds: dict[int, float] = {
|
55 |
10: 0.5,
|
56 |
-
20: 0.8
|
57 |
},
|
58 |
max_iter_to_alter: int = 25,
|
59 |
) -> PIL.Image.Image:
|
60 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
61 |
-
try:
|
62 |
-
indices_to_alter = list(map(int, indices_to_alter_str.split(',')))
|
63 |
-
except:
|
64 |
-
raise gr.Error('Invalid token indices.')
|
65 |
-
|
66 |
-
self.load_model(model_id)
|
67 |
-
|
68 |
-
controller = AttentionStore()
|
69 |
-
config = RunConfig(prompt=prompt,
|
70 |
-
n_inference_steps=num_steps,
|
71 |
-
guidance_scale=guidance_scale,
|
72 |
-
run_standard_sd=not apply_attend_and_excite,
|
73 |
-
scale_factor=scale_factor,
|
74 |
-
thresholds=thresholds,
|
75 |
-
max_iter_to_alter=max_iter_to_alter)
|
76 |
-
image = run_on_prompt(model=self.model,
|
77 |
-
prompt=[prompt],
|
78 |
-
controller=controller,
|
79 |
-
token_indices=indices_to_alter,
|
80 |
-
seed=generator,
|
81 |
-
config=config)
|
82 |
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from __future__ import annotations
|
2 |
|
|
|
|
|
|
|
3 |
import PIL.Image
|
4 |
import torch
|
5 |
+
from diffusers import (StableDiffusionAttendAndExcitePipeline,
|
6 |
+
StableDiffusionPipeline)
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
|
9 |
class Model:
|
10 |
def __init__(self):
|
11 |
self.device = torch.device(
|
12 |
'cuda:0' if torch.cuda.is_available() else 'cpu')
|
13 |
+
model_id = 'CompVis/stable-diffusion-v1-4'
|
14 |
+
if self.device.type == 'cuda':
|
15 |
+
self.ax_pipe = StableDiffusionAttendAndExcitePipeline.from_pretrained(
|
16 |
+
model_id, torch_dtype=torch.float16)
|
17 |
+
self.ax_pipe.to(self.device)
|
18 |
+
self.sd_pipe = StableDiffusionPipeline.from_pretrained(
|
19 |
+
model_id, torch_dtype=torch.float16)
|
20 |
+
self.sd_pipe.to(self.device)
|
21 |
+
else:
|
22 |
+
self.ax_pipe = StableDiffusionAttendAndExcitePipeline.from_pretrained(
|
23 |
+
model_id)
|
24 |
+
self.sd_pipe = StableDiffusionPipeline.from_pretrained(model_id)
|
25 |
|
26 |
+
def get_token_table(self, prompt: str):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
tokens = [
|
28 |
+
self.ax_pipe.tokenizer.decode(t)
|
29 |
+
for t in self.ax_pipe.tokenizer(prompt)['input_ids']
|
30 |
]
|
31 |
tokens = tokens[1:-1]
|
32 |
return list(enumerate(tokens, start=1))
|
33 |
|
34 |
def run(
|
35 |
self,
|
|
|
36 |
prompt: str,
|
37 |
indices_to_alter_str: str,
|
38 |
+
seed: int = 0,
|
39 |
+
apply_attend_and_excite: bool = True,
|
40 |
+
num_steps: int = 50,
|
41 |
+
guidance_scale: float = 7.5,
|
42 |
scale_factor: int = 20,
|
43 |
thresholds: dict[int, float] = {
|
44 |
10: 0.5,
|
45 |
+
20: 0.8,
|
46 |
},
|
47 |
max_iter_to_alter: int = 25,
|
48 |
) -> PIL.Image.Image:
|
49 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
|
51 |
+
if apply_attend_and_excite:
|
52 |
+
try:
|
53 |
+
token_indices = list(map(int, indices_to_alter_str.split(',')))
|
54 |
+
except Exception:
|
55 |
+
raise ValueError('Invalid token indices.')
|
56 |
+
out = self.ax_pipe(
|
57 |
+
prompt=prompt,
|
58 |
+
token_indices=token_indices,
|
59 |
+
guidance_scale=guidance_scale,
|
60 |
+
generator=generator,
|
61 |
+
num_inference_steps=num_steps,
|
62 |
+
max_iter_to_alter=max_iter_to_alter,
|
63 |
+
thresholds=thresholds,
|
64 |
+
scale_factor=scale_factor,
|
65 |
+
)
|
66 |
+
else:
|
67 |
+
out = self.sd_pipe(
|
68 |
+
prompt=prompt,
|
69 |
+
guidance_scale=guidance_scale,
|
70 |
+
generator=generator,
|
71 |
+
num_inference_steps=num_steps,
|
72 |
+
)
|
73 |
+
return out.images[0]
|
requirements.txt
CHANGED
@@ -1,8 +1,5 @@
|
|
1 |
-
accelerate==0.
|
2 |
-
diffusers==0.
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
pyrallis==0.3.1
|
7 |
-
torch==1.13.1
|
8 |
-
transformers==4.29.2
|
|
|
1 |
+
accelerate==0.20.3
|
2 |
+
diffusers==0.17.0
|
3 |
+
Pillow==9.5.0
|
4 |
+
torch==2.0.1
|
5 |
+
transformers==4.30.1
|
|
|
|
|
|