File size: 18,757 Bytes
5eb8981
 
 
 
 
 
 
 
 
 
 
b34b4e8
5eb8981
 
 
 
 
648acf7
5eb8981
 
 
 
 
 
 
 
 
4809200
b34b4e8
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4809200
 
5eb8981
 
 
 
 
 
 
 
 
 
 
b34b4e8
 
 
 
 
 
 
 
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b34b4e8
5eb8981
 
 
 
 
 
 
 
 
b34b4e8
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b34b4e8
 
 
 
 
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b34b4e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5eb8981
 
b34b4e8
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b34b4e8
 
5eb8981
 
 
b34b4e8
5eb8981
 
b34b4e8
 
5eb8981
b34b4e8
 
 
 
 
5eb8981
 
 
b34b4e8
 
 
 
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b34b4e8
 
 
 
 
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
b34b4e8
5eb8981
 
b34b4e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5eb8981
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
import glob
import os
import numpy as np
import cv2
from pathlib import Path
import torch
import torch.nn as nn
import torchvision.transforms as T
import argparse
from PIL import Image
import yaml
import inspect
from tqdm import tqdm
from transformers import logging
from diffusers import DDIMScheduler, StableDiffusionPipeline

from tokenflow_utils import *
from utils import save_video, seed_everything

# suppress partial model loading warning
logging.set_verbosity_error()

VAE_BATCH_SIZE = 10


class TokenFlow(nn.Module):
    def __init__(self, config, 
                 pipe,
                 frames = None,
                 inverted_latents = None, #X0,...,XT,
                 zs = None):
        super().__init__()
        self.config = config
        self.device = config["device"]
        
        sd_version = config["sd_version"]
        self.sd_version = sd_version
        if sd_version == '2.1':
            model_key = "stabilityai/stable-diffusion-2-1-base"
        elif sd_version == '2.0':
            model_key = "stabilityai/stable-diffusion-2-base"
        elif sd_version == '1.5':
            model_key = "runwayml/stable-diffusion-v1-5"
        elif sd_version == 'depth':
            model_key = "stabilityai/stable-diffusion-2-depth"
        else:
            raise ValueError(f'Stable-diffusion version {sd_version} not supported.')

        # Create SD models
        print('Loading SD model')

        # pipe = StableDiffusionPipeline.from_pretrained(model_key, torch_dtype=torch.float16).to("cuda")
        # pipe.enable_xformers_memory_efficient_attention()

        self.vae = pipe.vae
        self.tokenizer = pipe.tokenizer
        self.text_encoder = pipe.text_encoder
        self.unet = pipe.unet

        self.scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
        self.scheduler.set_timesteps(config["n_timesteps"], device=self.device)
        print('SD model loaded')
        
        # data
        self.inversion = config['inversion']
        if self.inversion == 'ddpm':
            self.skip_steps = config['skip_steps']
            self.eta = 1.0
        else:
            self.eta = 0.0
        self.extra_step_kwargs = self.prepare_extra_step_kwargs(self.eta)
        
        # data
        self.frames, self.inverted_latents, self.zs  = frames, inverted_latents, zs
        self.latents_path = self.get_latents_path()
        
        # load frames
        self.paths, self.frames, self.latents, self.eps = self.get_data()
        
        if self.sd_version == 'depth':
            self.depth_maps = self.prepare_depth_maps()

        self.text_embeds = self.get_text_embeds(config["prompt"], config["negative_prompt"])
        # pnp_inversion_prompt = self.get_pnp_inversion_prompt()
        self.pnp_guidance_embeds = self.get_text_embeds(config["pnp_inversion_prompt"], config["pnp_inversion_prompt"]).chunk(2)[0]
    
    @torch.no_grad()   
    def prepare_depth_maps(self, model_type='DPT_Large', device='cuda'):
        depth_maps = []
        midas = torch.hub.load("intel-isl/MiDaS", model_type)
        midas.to(device)
        midas.eval()

        midas_transforms = torch.hub.load("intel-isl/MiDaS", "transforms")

        if model_type == "DPT_Large" or model_type == "DPT_Hybrid":
            transform = midas_transforms.dpt_transform
        else:
            transform = midas_transforms.small_transform

        for i in range(len(self.paths)):
            img = cv2.imread(self.paths[i])
            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
            
            latent_h = img.shape[0] // 8
            latent_w = img.shape[1] // 8
            
            input_batch = transform(img).to(device)
            prediction = midas(input_batch)

            depth_map = torch.nn.functional.interpolate(
                prediction.unsqueeze(1),
                size=(latent_h, latent_w),
                mode="bicubic",
                align_corners=False,
            )
            depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
            depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
            depth_map = 2.0 * (depth_map - depth_min) / (depth_max - depth_min) - 1.0
            depth_maps.append(depth_map)

        return torch.cat(depth_maps).to(torch.float16).to(self.device)
    
    def get_pnp_inversion_prompt(self):
        inv_prompts_path = os.path.join(str(Path(self.latents_path).parent), 'inversion_prompt.txt')
        # read inversion prompt
        with open(inv_prompts_path, 'r') as f:
            inv_prompt = f.read()
        return inv_prompt

    def get_latents_path(self):
        read_from_files = self.frames is None
        if read_from_files: 
            latents_path = os.path.join(self.config["latents_path"], f'sd_{self.config["sd_version"]}',
                                 Path(self.config["data_path"]).stem, f'steps_{self.config["n_inversion_steps"]}')
            latents_path = [x for x in glob.glob(f'{latents_path}/*') if '.' not in Path(x).name]
            n_frames = [int([x for x in latents_path[i].split('/') if 'nframes' in x][0].split('_')[1]) for i in range(len(latents_path))]
            latents_path = latents_path[np.argmax(n_frames)]

            self.config["n_frames"] = min(max(n_frames), self.config["n_frames"])
            
        else:
            n_frames = self.frames.shape[0]
            self.config["n_frames"] = min(n_frames, self.config["n_frames"])
        
        if self.config["n_frames"] % self.config["batch_size"] != 0:
            # make n_frames divisible by batch_size
            self.config["n_frames"] = self.config["n_frames"] - (self.config["n_frames"] % self.config["batch_size"])

        if read_from_files:
            return os.path.join(latents_path, 'latents')
        else:
            return None

    @torch.no_grad()
    def get_text_embeds(self, prompt, negative_prompt, batch_size=1):
        # Tokenize text and get embeddings
        text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
                                    truncation=True, return_tensors='pt')
        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]

        # Do the same for unconditional embeddings
        uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
                                      return_tensors='pt')

        uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]

        # Cat for final embeddings
        text_embeddings = torch.cat([uncond_embeddings] * batch_size + [text_embeddings] * batch_size)
        return text_embeddings

    @torch.no_grad()
    def encode_imgs(self, imgs, batch_size=VAE_BATCH_SIZE, deterministic=False):
        imgs = 2 * imgs - 1
        latents = []
        for i in range(0, len(imgs), batch_size):
            posterior = self.vae.encode(imgs[i:i + batch_size]).latent_dist
            latent = posterior.mean if deterministic else posterior.sample()
            latents.append(latent * 0.18215)
        latents = torch.cat(latents)
        return latents

    @torch.no_grad()
    def decode_latents(self, latents, batch_size=VAE_BATCH_SIZE):
        latents = 1 / 0.18215 * latents
        imgs = []
        for i in range(0, len(latents), batch_size):
            imgs.append(self.vae.decode(latents[i:i + batch_size]).sample)
        imgs = torch.cat(imgs)
        imgs = (imgs / 2 + 0.5).clamp(0, 1)
        return imgs

    
    def get_data(self):
        read_from_files = self.frames is None
        # read_from_files = True
        if read_from_files:
            # load frames
            paths = [os.path.join(self.config["data_path"], "%05d.jpg" % idx) for idx in
                                   range(self.config["n_frames"])]
            if not os.path.exists(paths[0]):
                paths = [os.path.join(self.config["data_path"], "%05d.png" % idx) for idx in
                                       range(self.config["n_frames"])]
            frames = [Image.open(paths[idx]).convert('RGB') for idx in range(self.config["n_frames"])]
            if frames[0].size[0] == frames[0].size[1]:
                frames = [frame.resize((512, 512), resample=Image.Resampling.LANCZOS) for frame in frames]
            frames = torch.stack([T.ToTensor()(frame) for frame in frames]).to(torch.float16).to(self.device)
            save_video(frames, f'{self.config["output_path"]}/input_fps10.mp4', fps=10)
            save_video(frames, f'{self.config["output_path"]}/input_fps20.mp4', fps=20)
            save_video(frames, f'{self.config["output_path"]}/input_fps30.mp4', fps=30)
        else:
            frames = self.frames
        # encode to latents
        latents = self.encode_imgs(frames, deterministic=True).to(torch.float16).to(self.device)
        # get noise
        if self.inversion == 'ddim':
            eps = self.get_ddim_eps(latents, range(self.config["n_frames"])).to(torch.float16).to(self.device)
        elif self.inversion == 'ddpm':
            eps = self.get_ddpm_noise()
        else:
            raise NotImplementedError()
        
        if not read_from_files:
            return None, frames, latents, eps
        return paths, frames, latents, eps

    def get_ddim_eps(self, latent, indices):
        read_from_files = self.inverted_latents is None
        if read_from_files:
            noisest = max([int(x.split('_')[-1].split('.')[0]) for x in glob.glob(os.path.join(self.latents_path, f'noisy_latents_*.pt'))])
            latents_path = os.path.join(self.latents_path, f'noisy_latents_{noisest}.pt')
            noisy_latent = torch.load(latents_path)[indices].to(self.device)
        else:
            noisest = max([int(key.split("_")[-1]) for key in self.inverted_latents.keys()])
            noisy_latent = self.inverted_latents[f'noisy_latents_{noisest}'][indices]

        alpha_prod_T = self.scheduler.alphas_cumprod[noisest]
        mu_T, sigma_T = alpha_prod_T ** 0.5, (1 - alpha_prod_T) ** 0.5
        eps = (noisy_latent - mu_T * latent) / sigma_T
        return eps
    
    def get_ddpm_noise(self):
        read_from_files = self.inverted_latents is None
        idx_to_t = {int(k): int(v) for k, v in enumerate(self.scheduler.timesteps)}
        t = idx_to_t[self.skip_steps]
        if read_from_files:
            x0_path = os.path.join(self.latents_path, f'noisy_latents_{t}.pt')
            zs_path = os.path.join(self.latents_path, f'noise_total.pt')
            x0 = torch.load(x0_path)[:self.config["n_frames"]].to(self.device)
            zs = torch.load(zs_path)[self.skip_steps:, :self.config["n_frames"]].to(self.device)
        else:
            x0 = self.inverted_latents[f'noisy_latents_{t}'][:self.config["n_frames"]].to(self.device)
            zs = self.zs[self.skip_steps:, :self.config["n_frames"]].to(self.device)
        return x0, zs

    def prepare_extra_step_kwargs(self, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        return extra_step_kwargs

    @torch.no_grad()
    def denoise_step(self, x, t, indices, zs=None):
        # register the time step and features in pnp injection modules
        read_files = self.inverted_latents is None

        if read_files:
            source_latents = load_source_latents_t(t, self.latents_path)[indices]

        else:
            source_latents = self.inverted_latents[f'noisy_latents_{t}'][indices]

        latent_model_input = torch.cat([source_latents] + ([x] * 2))
        if self.sd_version == 'depth':
            latent_model_input = torch.cat([latent_model_input, torch.cat([self.depth_maps[indices]] * 3)], dim=1)

        register_time(self, t.item())

        # compute text embeddings
        text_embed_input = torch.cat([self.pnp_guidance_embeds.repeat(len(indices), 1, 1),
                                      torch.repeat_interleave(self.text_embeds, len(indices), dim=0)])

        # apply the denoising network
        noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input)['sample']

        # perform guidance
        _, noise_pred_uncond, noise_pred_cond = noise_pred.chunk(3)
        noise_pred = noise_pred_uncond + self.config["guidance_scale"] * (noise_pred_cond - noise_pred_uncond)

        # compute the denoising step with the reference model
        denoised_latent = self.scheduler.step(noise_pred, t, x, variance_noise=zs, **self.extra_step_kwargs)[
            'prev_sample']
        return denoised_latent
    
    @torch.autocast(dtype=torch.float16, device_type='cuda')
    def batched_denoise_step(self, x, t, indices, zs=None):
        batch_size = self.config["batch_size"]
        denoised_latents = []
        pivotal_idx = torch.randint(batch_size, (len(x) // batch_size,)) + torch.arange(0, len(x), batch_size)

        register_pivotal(self, True)
        if zs is None:
            zs_input = None
        else:
            zs_input = zs[pivotal_idx]
        self.denoise_step(x[pivotal_idx], t, indices[pivotal_idx], zs_input)
        register_pivotal(self, False)
        for i, b in enumerate(range(0, len(x), batch_size)):
            register_batch_idx(self, i)
            if zs is None:
                zs_input = None
            else:
                zs_input = zs[b:b + batch_size]
            denoised_latents.append(self.denoise_step(x[b:b + batch_size], t, indices[b:b + batch_size]
                                                      , zs_input))
        denoised_latents = torch.cat(denoised_latents)
        return denoised_latents

    def init_method(self, conv_injection_t, qk_injection_t):
        self.qk_injection_timesteps = self.scheduler.timesteps[:qk_injection_t] if qk_injection_t >= 0 else []
        self.conv_injection_timesteps = self.scheduler.timesteps[:conv_injection_t] if conv_injection_t >= 0 else []
        register_extended_attention_pnp(self, self.qk_injection_timesteps)
        register_conv_injection(self, self.conv_injection_timesteps)
        set_tokenflow(self.unet)

    def save_vae_recon(self):
        os.makedirs(f'{self.config["output_path"]}/vae_recon', exist_ok=True)
        decoded = self.decode_latents(self.latents)
        for i in range(len(decoded)):
            T.ToPILImage()(decoded[i]).save(f'{self.config["output_path"]}/vae_recon/%05d.png' % i)
        save_video(decoded, f'{self.config["output_path"]}/vae_recon_10.mp4', fps=10)
        save_video(decoded, f'{self.config["output_path"]}/vae_recon_20.mp4', fps=20)
        save_video(decoded, f'{self.config["output_path"]}/vae_recon_30.mp4', fps=30)

    def edit_video(self):
        save_files = self.inverted_latents is None # if we're in the original non-demo setting
        if save_files: 
            os.makedirs(f'{self.config["output_path"]}/img_ode', exist_ok=True)
            self.save_vae_recon()
        # self.save_vae_recon()
        pnp_f_t = int(self.config["n_timesteps"] * self.config["pnp_f_t"])
        pnp_attn_t = int(self.config["n_timesteps"] * self.config["pnp_attn_t"])
        
        self.init_method(conv_injection_t=pnp_f_t, qk_injection_t=pnp_attn_t)
        
        if self.inversion == 'ddim':
            noisy_latents = self.scheduler.add_noise(self.latents, self.eps, self.scheduler.timesteps[0])
        elif self.inversion == 'ddpm':
            noisy_latents = self.eps[0]
        else:
            raise NotImplementedError()
            
        edited_frames = self.sample_loop(noisy_latents, torch.arange(self.config["n_frames"]))
        
        if save_files:
            save_video(edited_frames, f'{self.config["output_path"]}/tokenflow_PnP_fps_10.mp4')
            save_video(edited_frames, f'{self.config["output_path"]}/tokenflow_PnP_fps_20.mp4', fps=20)
            save_video(edited_frames, f'{self.config["output_path"]}/tokenflow_PnP_fps_30.mp4', fps=30)
            print('Done!')
        else:
            return edited_frames

    def sample_loop(self, x, indices):
        save_files  = self.inverted_latents is None # if we're in the original non-demo settinge
        if save_files:
            os.makedirs(f'{self.config["output_path"]}/img_ode', exist_ok=True)
        
        timesteps = self.scheduler.timesteps
        if self.inversion == 'ddpm':
            zs_total = self.eps[1]

            t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs_total.shape[0]:])}
            timesteps = timesteps[-zs_total.shape[0]:]
            
        for i, t in enumerate(tqdm(timesteps, desc="Sampling")):
            if self.inversion == 'ddpm':
                idx = t_to_idx[int(t)]
                zs = zs_total[idx]
            else:
                zs = None
            x = self.batched_denoise_step(x, t, indices, zs)
        
        decoded_latents = self.decode_latents(x)
        if save_files:
            for i in range(len(decoded_latents)):
                T.ToPILImage()(decoded_latents[i]).save(f'{self.config["output_path"]}/img_ode/%05d.png' % i)

        return decoded_latents


# def run(config):
#     seed_everything(config["seed"])
#     print(config)
#     editor = TokenFlow(config)
#     editor.edit_video()


# if __name__ == '__main__':
#     parser = argparse.ArgumentParser()
#     parser.add_argument('--config_path', type=str, default='configs/config_pnp.yaml')
#     opt = parser.parse_args()
#     with open(opt.config_path, "r") as f:
#         config = yaml.safe_load(f)
#     config["output_path"] = os.path.join(config["output_path"] + f'_pnp_SD_{config["sd_version"]}',
#                                              Path(config["data_path"]).stem,
#                                              config["prompt"][:240],
#                                              f'attn_{config["pnp_attn_t"]}_f_{config["pnp_f_t"]}',
#                                              f'batch_size_{str(config["batch_size"])}',
#                                              str(config["n_timesteps"]),
#     )
#     os.makedirs(config["output_path"], exist_ok=True)
#     print(config["data_path"])
#     assert os.path.exists(config["data_path"]), "Data path does not exist"
#     with open(os.path.join(config["output_path"], "config.yaml"), "w") as f:
#         yaml.dump(config, f)
#     run(config)