File size: 9,401 Bytes
7da7768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
 
 
 
 
7da7768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
 
 
 
28425ed
47163f5
 
 
 
 
 
 
 
928a499
758cfc5
47163f5
 
7e98f35
7093637
7e98f35
c6e2b5c
5914cbe
7beb833
5914cbe
 
7093637
5914cbe
c6e2b5c
5914cbe
7beb833
47163f5
1eefa67
7093637
47163f5
c6e2b5c
 
47163f5
 
 
 
c6e2b5c
7093637
47163f5
 
03bc7e7
c6e2b5c
 
 
47163f5
 
 
acce61f
a976ca4
 
 
 
 
 
 
 
 
47163f5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1eefa67
 
47163f5
 
 
aff3d3b
47163f5
 
 
 
1eefa67
47163f5
 
 
 
 
 
 
 
 
 
 
 
1eefa67
47163f5
5f8442e
47163f5
 
 
 
c6e2b5c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c242bd5
c6e2b5c
 
 
 
 
 
 
 
 
 
 
 
 
 
2b80ac9
c6e2b5c
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import os
import os.path as osp

import cv2
import numpy as np
import torch
from basicsr.utils import img2tensor, tensor2img
from pytorch_lightning import seed_everything
from ldm.models.diffusion.plms import PLMSSampler
from ldm.modules.encoders.adapter import Adapter
from ldm.util import instantiate_from_config
from model_edge import pidinet
import gradio as gr
from omegaconf import OmegaConf

import pathlib
import random
import shlex
import subprocess
import sys

sys.path.append('T2I-Adapter')

config_path =  'https://github.com/TencentARC/T2I-Adapter/raw/main/configs/stable-diffusion/'
model_path = 'https://github.com/TencentARC/T2I-Adapter/raw/main/models/'

def load_model_from_config(config, ckpt, verbose=False):
    print(f"Loading model from {ckpt}")
    pl_sd = torch.load(ckpt, map_location="cpu")
    if "global_step" in pl_sd:
        print(f"Global Step: {pl_sd['global_step']}")
    if "state_dict" in pl_sd:
        sd = pl_sd["state_dict"]
    else:
        sd = pl_sd
    model = instantiate_from_config(config.model)
    m, u = model.load_state_dict(sd, strict=False)
    # if len(m) > 0 and verbose:
    #     print("missing keys:")
    #     print(m)
    # if len(u) > 0 and verbose:
    #     print("unexpected keys:")
    #     print(u)

    model.cuda()
    model.eval()
    return model

class Model:
    def __init__(self,
                 model_config_path: str = 'ControlNet/models/cldm_v15.yaml',
                 model_dir: str = 'models',
                 use_lightweight: bool = True):
        self.device = torch.device(
            'cuda:0' if torch.cuda.is_available() else 'cpu')    
        self.model_dir = pathlib.Path(model_dir)

        self.download_models()



    def download_models(self) -> None:
        self.model_dir.mkdir(exist_ok=True, parents=True)
        device = 'cuda'
    
        config = OmegaConf.load("configs/stable-diffusion/test_sketch.yaml")
        config.model.params.cond_stage_config.params.device = device

        base_model_file = "https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt"
        base_model_file_anything = "https://huggingface.co/andite/anything-v4.0/resolve/main/anything-v4.0-pruned.ckpt"
        sketch_adapter_file = "https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_sketch_sd14v1.pth"
        pose_adapter_file = "https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_keypose_sd14v1.pth" 
        pidinet_file = model_path+"table5_pidinet.pth"
        clip_file = "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/*"
        
        subprocess.run(shlex.split(f'wget {base_model_file} -O models/sd-v1-4.ckpt'))
        subprocess.run(shlex.split(f'wget {base_model_file_anything} -O models/anything-v4.0-pruned.ckpt'))
        subprocess.run(shlex.split(f'wget {sketch_adapter_file} -O models/t2iadapter_sketch_sd14v1.pth'))
        subprocess.run(shlex.split(f'wget {pose_adapter_file} -O models/t2iadapter_keypose_sd14v1.pth'))
        subprocess.run(shlex.split(f'wget {pidinet_file} -O models/table5_pidinet.pth'))

        
        self.model = load_model_from_config(config, "models/sd-v1-4.ckpt").to(device)
        self.model_anything = load_model_from_config(config, "models/anything-v4.0-pruned.ckpt").to(device)
        current_base = 'sd-v1-4.ckpt'
        self.model_ad_sketch = Adapter(channels=[320, 640, 1280, 1280][:4], nums_rb=2, ksize=1, sk=True, use_conv=False).to(device)
        self.model_ad_sketch.load_state_dict(torch.load("models/t2iadapter_sketch_sd14v1.pth"))
        net_G = pidinet()
        ckp = torch.load('models/table5_pidinet.pth', map_location='cpu')['state_dict']
        net_G.load_state_dict({k.replace('module.',''):v for k, v in ckp.items()})
        net_G.to(device)
        self.sampler= PLMSSampler(self.model)
        self.sampler_anything= PLMSSampler(self.model_anything)
        save_memory=True

        self.model_ad_pose = Adapter(cin=int(3*64),channels=[320, 640, 1280, 1280][:4], nums_rb=2, ksize=1, sk=True, use_conv=False).to(device)
        self.model_ad_pose.load_state_dict(torch.load("models/t2iadapter_keypose_sd14v1.pth"))


    @torch.inference_mode()
    def process_sketch(self, input_img, type_in, color_back, prompt, neg_prompt, fix_sample, scale, con_strength, base_model):    
        global current_base
        device = 'cuda' 
        # if current_base != base_model:
        #     ckpt = os.path.join("models", base_model)
        #     pl_sd = torch.load(ckpt, map_location="cpu")
        #     if "state_dict" in pl_sd:
        #         sd = pl_sd["state_dict"]
        #     else:
        #         sd = pl_sd
        #     model.load_state_dict(sd, strict=False) #load_model_from_config(config, os.path.join("models", base_model)).to(device)
        #     current_base = base_model
        con_strength = int((1-con_strength)*50)
        if fix_sample == 'True':
            seed_everything(42)
        
        im = cv2.resize(input_img,(512,512))
    
        if type_in == 'Sketch':
            # net_G = net_G.cpu()
            if color_back == 'White':
                im = 255-im
            im_edge = im.copy()
            im = img2tensor(im)[0].unsqueeze(0).unsqueeze(0)/255.
            # edge = 1-edge # for white background
            im = im>0.5
            im = im.float()
        elif type_in == 'Image':
            im = img2tensor(im).unsqueeze(0)/255.
            im = net_G(im.to(device))[-1]
            im = im>0.5
            im = im.float()
            im_edge = tensor2img(im)
    
        c = self.model.get_learned_conditioning([prompt])
        nc = self.model.get_learned_conditioning([neg_prompt])
        
        with torch.no_grad():
            # extract condition features
            features_adapter = self.model_ad_sketch(im.to(device))
    
        shape = [4, 64, 64]
    
        # sampling
        samples_ddim, _ = self.sampler.sample(S=50,
                                        conditioning=c,
                                        batch_size=1,
                                        shape=shape,
                                        verbose=False,
                                        unconditional_guidance_scale=scale,
                                        unconditional_conditioning=nc,
                                        eta=0.0,
                                        x_T=None,
                                        features_adapter1=features_adapter,
                                        mode = 'sketch',
                                        con_strength = con_strength)
    
        x_samples_ddim = self.model.decode_first_stage(samples_ddim)
        x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
        x_samples_ddim = x_samples_ddim.permute(0, 2, 3, 1).cpu().numpy()[0]
        x_samples_ddim = 255.*x_samples_ddim
        x_samples_ddim = x_samples_ddim.astype(np.uint8)
    
        return [im_edge, x_samples_ddim]

    @torch.inference_mode()
    def process_pose(self, input_img, prompt, neg_prompt, fix_sample, scale, con_strength, base_model):    
        global current_base
        device = 'cuda' 
        # if current_base != base_model:
        #     ckpt = os.path.join("models", base_model)
        #     pl_sd = torch.load(ckpt, map_location="cpu")
        #     if "state_dict" in pl_sd:
        #         sd = pl_sd["state_dict"]
        #     else:
        #         sd = pl_sd
        #     model.load_state_dict(sd, strict=False) #load_model_from_config(config, os.path.join("models", base_model)).to(device)
        #     current_base = base_model
        con_strength = int((1-con_strength)*50)
        if fix_sample == 'True':
            seed_everything(42)
        
        im = cv2.resize(input_img,(512,512))
        pose = img2tensor(im, bgr2rgb=True, float32=True)/255.
        pose = pose.unsqueeze(0)

        im_pose = tensor2img(pose)
    
        c = self.model.get_learned_conditioning([prompt])
        nc = self.model.get_learned_conditioning([neg_prompt])
        
        with torch.no_grad():
            # extract condition features
            features_adapter = self.model_ad_pose(pose.to(device))
    
        shape = [4, 64, 64]
    
        # sampling
        samples_ddim, _ = self.sampler.sample(S=50,
                                        conditioning=c,
                                        batch_size=1,
                                        shape=shape,
                                        verbose=False,
                                        unconditional_guidance_scale=scale,
                                        unconditional_conditioning=nc,
                                        eta=0.0,
                                        x_T=None,
                                        features_adapter1=features_adapter,
                                        mode = 'sketch',
                                        con_strength = con_strength)
    
        x_samples_ddim = self.model.decode_first_stage(samples_ddim)
        x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
        x_samples_ddim = x_samples_ddim.permute(0, 2, 3, 1).cpu().numpy()[0]
        x_samples_ddim = 255.*x_samples_ddim
        x_samples_ddim = x_samples_ddim.astype(np.uint8)
    
        return [im_pose, x_samples_ddim]