File size: 17,865 Bytes
690f890
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# -*- coding: utf-8 -*-
# Copyright (c) Alibaba, Inc. and its affiliates.
import os
import sys
import gc
import math
import random
import types
import logging
from contextlib import contextmanager
from functools import partial

from PIL import Image
import torchvision.transforms.functional as TF
import torch
import torch.nn.functional as F
import torch.cuda.amp as amp
import torch.distributed as dist
from tqdm import tqdm

from wan.text2video import (WanT2V, T5EncoderModel, WanVAE, shard_model, FlowDPMSolverMultistepScheduler,
                               get_sampling_sigmas, retrieve_timesteps, FlowUniPCMultistepScheduler)
from .modules.model import VaceWanModel
from ..utils.preprocessor import VaceVideoProcessor


class WanVace(WanT2V):
    def __init__(
        self,
        config,
        checkpoint_dir,
        device_id=0,
        rank=0,
        t5_fsdp=False,
        dit_fsdp=False,
        use_usp=False,
        t5_cpu=False,
    ):
        r"""
        Initializes the Wan text-to-video generation model components.

        Args:
            config (EasyDict):
                Object containing model parameters initialized from config.py
            checkpoint_dir (`str`):
                Path to directory containing model checkpoints
            device_id (`int`,  *optional*, defaults to 0):
                Id of target GPU device
            rank (`int`,  *optional*, defaults to 0):
                Process rank for distributed training
            t5_fsdp (`bool`, *optional*, defaults to False):
                Enable FSDP sharding for T5 model
            dit_fsdp (`bool`, *optional*, defaults to False):
                Enable FSDP sharding for DiT model
            use_usp (`bool`, *optional*, defaults to False):
                Enable distribution strategy of USP.
            t5_cpu (`bool`, *optional*, defaults to False):
                Whether to place T5 model on CPU. Only works without t5_fsdp.
        """
        self.device = torch.device(f"cuda:{device_id}")
        self.config = config
        self.rank = rank
        self.t5_cpu = t5_cpu

        self.num_train_timesteps = config.num_train_timesteps
        self.param_dtype = config.param_dtype

        shard_fn = partial(shard_model, device_id=device_id)
        self.text_encoder = T5EncoderModel(
            text_len=config.text_len,
            dtype=config.t5_dtype,
            device=torch.device('cpu'),
            checkpoint_path=os.path.join(checkpoint_dir, config.t5_checkpoint),
            tokenizer_path=os.path.join(checkpoint_dir, config.t5_tokenizer),
            shard_fn=shard_fn if t5_fsdp else None)

        self.vae_stride = config.vae_stride
        self.patch_size = config.patch_size
        self.vae = WanVAE(
            vae_pth=os.path.join(checkpoint_dir, config.vae_checkpoint),
            device=self.device)

        logging.info(f"Creating VaceWanModel from {checkpoint_dir}")
        self.model = VaceWanModel.from_pretrained(checkpoint_dir)
        self.model.eval().requires_grad_(False)

        if use_usp:
            from xfuser.core.distributed import \
                get_sequence_parallel_world_size

            from .distributed.xdit_context_parallel import (usp_attn_forward,
                                                            usp_dit_forward,
                                                            usp_dit_forward_vace)
            for block in self.model.blocks:
                block.self_attn.forward = types.MethodType(
                    usp_attn_forward, block.self_attn)
            for block in self.model.vace_blocks:
                block.self_attn.forward = types.MethodType(
                    usp_attn_forward, block.self_attn)
            self.model.forward = types.MethodType(usp_dit_forward, self.model)
            self.model.forward_vace = types.MethodType(usp_dit_forward_vace, self.model)
            self.sp_size = get_sequence_parallel_world_size()
        else:
            self.sp_size = 1

        if dist.is_initialized():
            dist.barrier()
        if dit_fsdp:
            self.model = shard_fn(self.model)
        else:
            self.model.to(self.device)

        self.sample_neg_prompt = config.sample_neg_prompt

        self.vid_proc = VaceVideoProcessor(downsample=tuple([x * y for x, y in zip(config.vae_stride, self.patch_size)]),
                                           min_area=480*832,
                                           max_area=480*832,
                                           min_fps=config.sample_fps,
                                           max_fps=config.sample_fps,
                                           zero_start=True,
                                           seq_len=32760,
                                           keep_last=True)

    def vace_encode_frames(self, frames, ref_images, masks=None):
        if ref_images is None:
            ref_images = [None] * len(frames)
        else:
            assert len(frames) == len(ref_images)

        if masks is None:
            latents = self.vae.encode(frames)
        else:
            inactive = [i * (1 - m) + 0 * m for i, m in zip(frames, masks)]
            reactive = [i * m + 0 * (1 - m) for i, m in zip(frames, masks)]
            inactive = self.vae.encode(inactive)
            reactive = self.vae.encode(reactive)
            latents = [torch.cat((u, c), dim=0) for u, c in zip(inactive, reactive)]

        cat_latents = []
        for latent, refs in zip(latents, ref_images):
            if refs is not None:
                if masks is None:
                    ref_latent = self.vae.encode(refs)
                else:
                    ref_latent = self.vae.encode(refs)
                    ref_latent = [torch.cat((u, torch.zeros_like(u)), dim=0) for u in ref_latent]
                assert all([x.shape[1] == 1 for x in ref_latent])
                latent = torch.cat([*ref_latent, latent], dim=1)
            cat_latents.append(latent)
        return cat_latents

    def vace_encode_masks(self, masks, ref_images=None):
        if ref_images is None:
            ref_images = [None] * len(masks)
        else:
            assert len(masks) == len(ref_images)

        result_masks = []
        for mask, refs in zip(masks, ref_images):
            c, depth, height, width = mask.shape
            new_depth = int((depth + 3) // self.vae_stride[0])
            height = 2 * (int(height) // (self.vae_stride[1] * 2))
            width = 2 * (int(width) // (self.vae_stride[2] * 2))

            # reshape
            mask = mask[0, :, :, :]
            mask = mask.view(
                depth, height, self.vae_stride[1], width, self.vae_stride[1]
            )  # depth, height, 8, width, 8
            mask = mask.permute(2, 4, 0, 1, 3)  # 8, 8, depth, height, width
            mask = mask.reshape(
                self.vae_stride[1] * self.vae_stride[2], depth, height, width
            )  # 8*8, depth, height, width

            # interpolation
            mask = F.interpolate(mask.unsqueeze(0), size=(new_depth, height, width), mode='nearest-exact').squeeze(0)

            if refs is not None:
                length = len(refs)
                mask_pad = torch.zeros_like(mask[:, :length, :, :])
                mask = torch.cat((mask_pad, mask), dim=1)
            result_masks.append(mask)
        return result_masks

    def vace_latent(self, z, m):
        return [torch.cat([zz, mm], dim=0) for zz, mm in zip(z, m)]

    def prepare_source(self, src_video, src_mask, src_ref_images, num_frames, image_size, device):
        image_sizes = []
        for i, (sub_src_video, sub_src_mask) in enumerate(zip(src_video, src_mask)):
            if sub_src_mask is not None and sub_src_video is not None:
                src_video[i], src_mask[i], _, _, _ = self.vid_proc.load_video_pair(sub_src_video, sub_src_mask)
                src_video[i] = src_video[i].to(device)
                src_mask[i] = src_mask[i].to(device)
                src_mask[i] = torch.clamp((src_mask[i][:1, :, :, :] + 1) / 2, min=0, max=1)
                image_sizes.append(src_video[i].shape[2:])
            elif sub_src_video is None:
                src_video[i] = torch.zeros((3, num_frames, image_size[0], image_size[1]), device=device)
                src_mask[i] = torch.ones_like(src_video[i], device=device)
                image_sizes.append(image_size)
            else:
                src_video[i], _, _, _ = self.vid_proc.load_video(sub_src_video)
                src_video[i] = src_video[i].to(device)
                src_mask[i] = torch.ones_like(src_video[i], device=device)
                image_sizes.append(src_video[i].shape[2:])

        for i, ref_images in enumerate(src_ref_images):
            if ref_images is not None:
                image_size = image_sizes[i]
                for j, ref_img in enumerate(ref_images):
                    if ref_img is not None:
                        ref_img = Image.open(ref_img).convert("RGB")
                        ref_img = TF.to_tensor(ref_img).sub_(0.5).div_(0.5).unsqueeze(1)
                        if ref_img.shape[-2:] != image_size:
                            canvas_height, canvas_width = image_size
                            ref_height, ref_width = ref_img.shape[-2:]
                            white_canvas = torch.ones((3, 1, canvas_height, canvas_width), device=device) # [-1, 1]
                            scale = min(canvas_height / ref_height, canvas_width / ref_width)
                            new_height = int(ref_height * scale)
                            new_width = int(ref_width * scale)
                            resized_image = F.interpolate(ref_img.squeeze(1).unsqueeze(0), size=(new_height, new_width), mode='bilinear', align_corners=False).squeeze(0).unsqueeze(1)
                            top = (canvas_height - new_height) // 2
                            left = (canvas_width - new_width) // 2
                            white_canvas[:, :, top:top + new_height, left:left + new_width] = resized_image
                            ref_img = white_canvas
                        src_ref_images[i][j] = ref_img.to(device)
        return src_video, src_mask, src_ref_images

    def decode_latent(self, zs, ref_images=None):
        if ref_images is None:
            ref_images = [None] * len(zs)
        else:
            assert len(zs) == len(ref_images)

        trimed_zs = []
        for z, refs in zip(zs, ref_images):
            if refs is not None:
                z = z[:, len(refs):, :, :]
            trimed_zs.append(z)

        return self.vae.decode(trimed_zs)



    def generate(self,
                 input_prompt,
                 input_frames,
                 input_masks,
                 input_ref_images,
                 size=(1280, 720),
                 frame_num=81,
                 context_scale=1.0,
                 shift=5.0,
                 sample_solver='unipc',
                 sampling_steps=50,
                 guide_scale=5.0,
                 n_prompt="",
                 seed=-1,
                 offload_model=True):
        r"""
        Generates video frames from text prompt using diffusion process.

        Args:
            input_prompt (`str`):
                Text prompt for content generation
            size (tupele[`int`], *optional*, defaults to (1280,720)):
                Controls video resolution, (width,height).
            frame_num (`int`, *optional*, defaults to 81):
                How many frames to sample from a video. The number should be 4n+1
            shift (`float`, *optional*, defaults to 5.0):
                Noise schedule shift parameter. Affects temporal dynamics
            sample_solver (`str`, *optional*, defaults to 'unipc'):
                Solver used to sample the video.
            sampling_steps (`int`, *optional*, defaults to 40):
                Number of diffusion sampling steps. Higher values improve quality but slow generation
            guide_scale (`float`, *optional*, defaults 5.0):
                Classifier-free guidance scale. Controls prompt adherence vs. creativity
            n_prompt (`str`, *optional*, defaults to ""):
                Negative prompt for content exclusion. If not given, use `config.sample_neg_prompt`
            seed (`int`, *optional*, defaults to -1):
                Random seed for noise generation. If -1, use random seed.
            offload_model (`bool`, *optional*, defaults to True):
                If True, offloads models to CPU during generation to save VRAM

        Returns:
            torch.Tensor:
                Generated video frames tensor. Dimensions: (C, N H, W) where:
                - C: Color channels (3 for RGB)
                - N: Number of frames (81)
                - H: Frame height (from size)
                - W: Frame width from size)
        """
        # preprocess
        # F = frame_num
        # target_shape = (self.vae.model.z_dim, (F - 1) // self.vae_stride[0] + 1,
        #                 size[1] // self.vae_stride[1],
        #                 size[0] // self.vae_stride[2])
        #
        # seq_len = math.ceil((target_shape[2] * target_shape[3]) /
        #                     (self.patch_size[1] * self.patch_size[2]) *
        #                     target_shape[1] / self.sp_size) * self.sp_size

        if n_prompt == "":
            n_prompt = self.sample_neg_prompt
        seed = seed if seed >= 0 else random.randint(0, sys.maxsize)
        seed_g = torch.Generator(device=self.device)
        seed_g.manual_seed(seed)

        if not self.t5_cpu:
            self.text_encoder.model.to(self.device)
            context = self.text_encoder([input_prompt], self.device)
            context_null = self.text_encoder([n_prompt], self.device)
            if offload_model:
                self.text_encoder.model.cpu()
        else:
            context = self.text_encoder([input_prompt], torch.device('cpu'))
            context_null = self.text_encoder([n_prompt], torch.device('cpu'))
            context = [t.to(self.device) for t in context]
            context_null = [t.to(self.device) for t in context_null]

        # vace context encode
        z0 = self.vace_encode_frames(input_frames, input_ref_images, masks=input_masks)
        m0 = self.vace_encode_masks(input_masks, input_ref_images)
        z = self.vace_latent(z0, m0)

        target_shape = list(z0[0].shape)
        target_shape[0] = int(target_shape[0] / 2)
        noise = [
            torch.randn(
                target_shape[0],
                target_shape[1],
                target_shape[2],
                target_shape[3],
                dtype=torch.float32,
                device=self.device,
                generator=seed_g)
        ]
        seq_len = math.ceil((target_shape[2] * target_shape[3]) /
                            (self.patch_size[1] * self.patch_size[2]) *
                            target_shape[1] / self.sp_size) * self.sp_size

        @contextmanager
        def noop_no_sync():
            yield

        no_sync = getattr(self.model, 'no_sync', noop_no_sync)

        # evaluation mode
        with amp.autocast(dtype=self.param_dtype), torch.no_grad(), no_sync():

            if sample_solver == 'unipc':
                sample_scheduler = FlowUniPCMultistepScheduler(
                    num_train_timesteps=self.num_train_timesteps,
                    shift=1,
                    use_dynamic_shifting=False)
                sample_scheduler.set_timesteps(
                    sampling_steps, device=self.device, shift=shift)
                timesteps = sample_scheduler.timesteps
            elif sample_solver == 'dpm++':
                sample_scheduler = FlowDPMSolverMultistepScheduler(
                    num_train_timesteps=self.num_train_timesteps,
                    shift=1,
                    use_dynamic_shifting=False)
                sampling_sigmas = get_sampling_sigmas(sampling_steps, shift)
                timesteps, _ = retrieve_timesteps(
                    sample_scheduler,
                    device=self.device,
                    sigmas=sampling_sigmas)
            else:
                raise NotImplementedError("Unsupported solver.")

            # sample videos
            latents = noise

            arg_c = {'context': context, 'seq_len': seq_len}
            arg_null = {'context': context_null, 'seq_len': seq_len}

            for _, t in enumerate(tqdm(timesteps)):
                latent_model_input = latents
                timestep = [t]

                timestep = torch.stack(timestep)

                self.model.to(self.device)
                noise_pred_cond = self.model(
                    latent_model_input, t=timestep, vace_context=z, vace_context_scale=context_scale, **arg_c)[0]
                noise_pred_uncond = self.model(
                    latent_model_input, t=timestep, vace_context=z, vace_context_scale=context_scale,**arg_null)[0]

                noise_pred = noise_pred_uncond + guide_scale * (
                    noise_pred_cond - noise_pred_uncond)

                temp_x0 = sample_scheduler.step(
                    noise_pred.unsqueeze(0),
                    t,
                    latents[0].unsqueeze(0),
                    return_dict=False,
                    generator=seed_g)[0]
                latents = [temp_x0.squeeze(0)]

            x0 = latents
            if offload_model:
                self.model.cpu()
                torch.cuda.empty_cache()
            if self.rank == 0:
                videos = self.decode_latent(x0, input_ref_images)

        del noise, latents
        del sample_scheduler
        if offload_model:
            gc.collect()
            torch.cuda.synchronize()
        if dist.is_initialized():
            dist.barrier()

        return videos[0] if self.rank == 0 else None