| 1 | """ |
| 2 | Bidirectional pipelines for memory-conditioned DMD. |
| 3 | """ |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from typing import Any, Callable, Dict, Optional, Tuple |
| 8 | |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | |
| 12 | |
| 13 | class BidirectionalMemoryVideoTrajectoryPipeline: |
| 14 | """ |
| 15 | Few-step backward simulation for video-only memory-conditioned DMD. |
| 16 | """ |
| 17 | |
| 18 | def __init__( |
| 19 | self, |
| 20 | generator: nn.Module, |
| 21 | add_noise_fn, |
| 22 | denoising_sigmas: torch.Tensor, |
| 23 | memory_downscale_factor: int = 1, |
| 24 | audio_latent_clamp: float = 0.0, |
| 25 | ) -> None: |
| 26 | self.generator = generator |
| 27 | self.add_noise_fn = add_noise_fn |
| 28 | self.denoising_sigmas = denoising_sigmas |
| 29 | self.memory_downscale_factor = int(memory_downscale_factor) |
| 30 | self.audio_latent_clamp = float(audio_latent_clamp) |
| 31 | |
| 32 | @torch.no_grad() |
| 33 | def inference_with_trajectory( |
| 34 | self, |
| 35 | video_noise: torch.Tensor, |
| 36 | conditional_dict: Dict[str, Any], |
| 37 | memory_video: torch.Tensor, |
| 38 | ) -> torch.Tensor: |
| 39 | batch_size = video_noise.shape[0] |
| 40 | num_frames = video_noise.shape[1] |
| 41 | device = video_noise.device |
| 42 | dtype = video_noise.dtype |
| 43 | memory_video = memory_video.to(device=device, dtype=dtype) |
| 44 | |
| 45 | trajectory = [video_noise] |
| 46 | noisy_video = video_noise |
| 47 | |
| 48 | for idx, sigma in enumerate(self.denoising_sigmas[:-1]): |
| 49 | video_sigma = sigma * torch.ones([batch_size, num_frames], device=device, dtype=dtype) |
| 50 | pred_video, _ = self.generator( |
| 51 | noisy_image_or_video=noisy_video, |
| 52 | conditional_dict=conditional_dict, |
| 53 | timestep=video_sigma, |
| 54 | noisy_audio=None, |
| 55 | audio_timestep=None, |
| 56 | memory_video=memory_video, |
| 57 | memory_downscale_factor=self.memory_downscale_factor, |
| 58 | ) |
| 59 | pred_video = pred_video.to(dtype=dtype) |
| 60 | |
| 61 | next_sigma = self.denoising_sigmas[idx + 1] |
| 62 | if next_sigma > 0: |
| 63 | fresh_noise = torch.randn_like(video_noise) |
| 64 | next_video_sigma = next_sigma * torch.ones([batch_size, num_frames], device=device, dtype=dtype) |
| 65 | noisy_video = self.add_noise_fn( |
| 66 | pred_video.flatten(0, 1), |
| 67 | fresh_noise.flatten(0, 1), |
| 68 | next_video_sigma.flatten(0, 1), |
| 69 | ).unflatten(0, (batch_size, num_frames)).to(dtype=dtype) |
| 70 | else: |
| 71 | noisy_video = pred_video |
| 72 | |
| 73 | trajectory.append(noisy_video) |
| 74 | |
| 75 | return torch.stack(trajectory, dim=1) |
| 76 | |
| 77 | |
| 78 | class BidirectionalMemoryVideoInferencePipeline: |
| 79 | """ |
| 80 | Few-step inference pipeline for video-only memory-conditioned DMD. |
| 81 | """ |
| 82 | |
| 83 | def __init__( |
| 84 | self, |
| 85 | generator: nn.Module, |
| 86 | add_noise_fn, |
| 87 | denoising_sigmas: torch.Tensor, |
| 88 | memory_downscale_factor: int = 1, |
| 89 | trace_fn: Optional[Callable[[Dict[str, Any]], None]] = None, |
| 90 | ) -> None: |
| 91 | self.generator = generator |
| 92 | self.add_noise_fn = add_noise_fn |
| 93 | self.denoising_sigmas = denoising_sigmas |
| 94 | self.memory_downscale_factor = int(memory_downscale_factor) |
| 95 | self.trace_fn = trace_fn |
| 96 | |
| 97 | def _emit_trace( |
| 98 | self, |
| 99 | event: str, |
| 100 | tensor: torch.Tensor, |
| 101 | *, |
| 102 | sigma_idx: Optional[int] = None, |
| 103 | sigma: Optional[torch.Tensor] = None, |
| 104 | ) -> None: |
| 105 | if self.trace_fn is None: |
| 106 | return |
| 107 | values = tensor.detach().float() |
| 108 | if values.numel() == 0: |
| 109 | stats = { |
| 110 | "mean": 0.0, |
| 111 | "std": 0.0, |
| 112 | "min": 0.0, |
| 113 | "max": 0.0, |
| 114 | "absmax": 0.0, |
| 115 | "nonzero_frac": 0.0, |
| 116 | } |
| 117 | else: |
| 118 | stats = { |
| 119 | "mean": values.mean().item(), |
| 120 | "std": values.std(unbiased=False).item() if values.numel() > 1 else 0.0, |
| 121 | "min": values.min().item(), |
| 122 | "max": values.max().item(), |
| 123 | "absmax": values.abs().max().item(), |
| 124 | "nonzero_frac": values.ne(0).float().mean().item(), |
| 125 | } |
| 126 | payload: Dict[str, Any] = { |
| 127 | "phase": "bootstrap", |
| 128 | "event": event, |
| 129 | "shape": list(tensor.shape), |
| 130 | **stats, |
| 131 | } |
| 132 | if sigma_idx is not None: |
| 133 | payload["sigma_idx"] = int(sigma_idx) |
| 134 | if sigma is not None: |
| 135 | payload["sigma"] = float(sigma.detach().float().item()) |
| 136 | self.trace_fn(payload) |
| 137 | |
| 138 | @torch.no_grad() |
| 139 | def generate( |
| 140 | self, |
| 141 | video_shape: Tuple[int, ...], |
| 142 | conditional_dict: Dict[str, Any], |
| 143 | memory_video: torch.Tensor, |
| 144 | seed: Optional[int] = None, |
| 145 | ) -> torch.Tensor: |
| 146 | batch_size = video_shape[0] |
| 147 | num_frames = video_shape[1] |
| 148 | |
| 149 | if seed is not None: |
| 150 | torch.manual_seed(seed) |
| 151 | |
| 152 | device = next(self.generator.parameters()).device |
| 153 | dtype = next(self.generator.parameters()).dtype |
| 154 | |
| 155 | video = torch.randn(video_shape, device=device, dtype=dtype) |
| 156 | memory_video = memory_video.to(device=device, dtype=dtype) |
| 157 | self._emit_trace("initial_noise", video) |
| 158 | |
| 159 | for idx, sigma in enumerate(self.denoising_sigmas[:-1]): |
| 160 | video_sigma = sigma * torch.ones([batch_size, num_frames], device=device, dtype=dtype) |
| 161 | |
| 162 | pred_video, _ = self.generator( |
| 163 | noisy_image_or_video=video, |
| 164 | conditional_dict=conditional_dict, |
| 165 | timestep=video_sigma, |
| 166 | noisy_audio=None, |
| 167 | audio_timestep=None, |
| 168 | memory_video=memory_video, |
| 169 | memory_downscale_factor=self.memory_downscale_factor, |
| 170 | ) |
| 171 | pred_video = pred_video.to(dtype=dtype) |
| 172 | self._emit_trace("pred_x0", pred_video, sigma_idx=idx, sigma=sigma) |
| 173 | |
| 174 | next_sigma = self.denoising_sigmas[idx + 1] |
| 175 | if next_sigma > 0: |
| 176 | fresh_noise = torch.randn_like(video) |
| 177 | next_video_sigma = next_sigma * torch.ones([batch_size, num_frames], device=device, dtype=dtype) |
| 178 | video = self.add_noise_fn( |
| 179 | pred_video.flatten(0, 1), |
| 180 | fresh_noise.flatten(0, 1), |
| 181 | next_video_sigma.flatten(0, 1), |
| 182 | ).unflatten(0, (batch_size, num_frames)).to(dtype=dtype) |
| 183 | else: |
| 184 | video = pred_video |
| 185 | self._emit_trace("updated_video", video, sigma_idx=idx + 1, sigma=next_sigma) |
| 186 | |
| 187 | return video |
| 188 | |
| 189 | |
| 190 | class BidirectionalMemoryAVTrajectoryPipeline: |
| 191 | """ |
| 192 | Few-step backward simulation for video-memory-conditioned AV DMD. |
| 193 | """ |
| 194 | |
| 195 | def __init__( |
| 196 | self, |
| 197 | generator: nn.Module, |
| 198 | add_noise_fn, |
| 199 | denoising_sigmas: torch.Tensor, |
| 200 | memory_downscale_factor: int = 1, |
| 201 | audio_latent_clamp: float = 0.0, |
| 202 | ) -> None: |
| 203 | self.generator = generator |
| 204 | self.add_noise_fn = add_noise_fn |
| 205 | self.denoising_sigmas = denoising_sigmas |
| 206 | self.memory_downscale_factor = int(memory_downscale_factor) |
| 207 | self.audio_latent_clamp = float(audio_latent_clamp) |
| 208 | |
| 209 | @torch.no_grad() |
| 210 | def inference_with_trajectory( |
| 211 | self, |
| 212 | video_noise: torch.Tensor, |
| 213 | audio_noise: torch.Tensor, |
| 214 | conditional_dict: Dict[str, Any], |
| 215 | memory_video: torch.Tensor, |
| 216 | memory_audio: Optional[torch.Tensor] = None, |
| 217 | memory_audio_timestep: Optional[torch.Tensor] = None, |
| 218 | memory_audio_segment_lengths: tuple[tuple[int, ...], ...] | None = None, |
| 219 | memory_position_mode: str = "reference", |
| 220 | memory_position_offset: float = 0.0, |
| 221 | memory_position_slot_stride: float = 50.0, |
| 222 | ) -> Tuple[torch.Tensor, torch.Tensor]: |
| 223 | batch_size = video_noise.shape[0] |
| 224 | num_video_frames = video_noise.shape[1] |
| 225 | num_audio_frames = audio_noise.shape[1] |
| 226 | device = video_noise.device |
| 227 | dtype = video_noise.dtype |
| 228 | memory_video = memory_video.to(device=device, dtype=dtype) |
| 229 | if memory_audio is not None: |
| 230 | memory_audio = memory_audio.to(device=device, dtype=dtype) |
| 231 | if memory_audio_timestep is not None: |
| 232 | memory_audio_timestep = memory_audio_timestep.to(device=device, dtype=dtype) |
| 233 | |
| 234 | video_trajectory = [video_noise] |
| 235 | audio_trajectory = [audio_noise] |
| 236 | noisy_video = video_noise |
| 237 | noisy_audio = audio_noise |
| 238 | memory_audio_kwargs = ( |
| 239 | { |
| 240 | "memory_audio": memory_audio, |
| 241 | "memory_audio_timestep": memory_audio_timestep, |
| 242 | "memory_audio_segment_lengths": memory_audio_segment_lengths, |
| 243 | } |
| 244 | if memory_audio is not None or memory_audio_timestep is not None |
| 245 | else {} |
| 246 | ) |
| 247 | |
| 248 | for idx, sigma in enumerate(self.denoising_sigmas[:-1]): |
| 249 | video_sigma = sigma * torch.ones([batch_size, num_video_frames], device=device, dtype=dtype) |
| 250 | audio_sigma = sigma * torch.ones([batch_size, num_audio_frames], device=device, dtype=dtype) |
| 251 | |
| 252 | pred_video, pred_audio = self.generator( |
| 253 | noisy_image_or_video=noisy_video, |
| 254 | conditional_dict=conditional_dict, |
| 255 | timestep=video_sigma, |
| 256 | noisy_audio=noisy_audio, |
| 257 | audio_timestep=audio_sigma, |
| 258 | memory_video=memory_video, |
| 259 | memory_downscale_factor=self.memory_downscale_factor, |
| 260 | memory_position_mode=memory_position_mode, |
| 261 | memory_position_offset=memory_position_offset, |
| 262 | memory_position_slot_stride=memory_position_slot_stride, |
| 263 | **memory_audio_kwargs, |
| 264 | ) |
| 265 | pred_video = pred_video.to(dtype=dtype) |
| 266 | pred_audio = pred_audio.to(dtype=dtype) |
| 267 | if self.audio_latent_clamp > 0: |
| 268 | pred_audio = pred_audio.clamp(-self.audio_latent_clamp, self.audio_latent_clamp) |
| 269 | |
| 270 | next_sigma = self.denoising_sigmas[idx + 1] |
| 271 | if next_sigma > 0: |
| 272 | fresh_noise_video = torch.randn_like(video_noise) |
| 273 | fresh_noise_audio = torch.randn_like(audio_noise) |
| 274 | next_video_sigma = next_sigma * torch.ones([batch_size, num_video_frames], device=device, dtype=dtype) |
| 275 | next_audio_sigma = next_sigma * torch.ones([batch_size, num_audio_frames], device=device, dtype=dtype) |
| 276 | noisy_video = self.add_noise_fn( |
| 277 | pred_video.flatten(0, 1), |
| 278 | fresh_noise_video.flatten(0, 1), |
| 279 | next_video_sigma.flatten(0, 1), |
| 280 | ).unflatten(0, (batch_size, num_video_frames)).to(dtype=dtype) |
| 281 | noisy_audio = self.add_noise_fn( |
| 282 | pred_audio, |
| 283 | fresh_noise_audio, |
| 284 | next_audio_sigma, |
| 285 | ).to(dtype=dtype) |
| 286 | else: |
| 287 | noisy_video = pred_video |
| 288 | noisy_audio = pred_audio |
| 289 | |
| 290 | video_trajectory.append(noisy_video) |
| 291 | audio_trajectory.append(noisy_audio) |
| 292 | |
| 293 | return torch.stack(video_trajectory, dim=1), torch.stack(audio_trajectory, dim=1) |
| 294 | |
| 295 | |
| 296 | class BidirectionalMemoryAVInferencePipeline: |
| 297 | """ |
| 298 | Few-step inference pipeline for video-memory-conditioned AV generation. |
| 299 | """ |
| 300 | |
| 301 | def __init__( |
| 302 | self, |
| 303 | generator: nn.Module, |
| 304 | add_noise_fn, |
| 305 | denoising_sigmas: torch.Tensor, |
| 306 | memory_downscale_factor: int = 1, |
| 307 | ) -> None: |
| 308 | self.generator = generator |
| 309 | self.add_noise_fn = add_noise_fn |
| 310 | self.denoising_sigmas = denoising_sigmas |
| 311 | self.memory_downscale_factor = int(memory_downscale_factor) |
| 312 | |
| 313 | @torch.no_grad() |
| 314 | def generate( |
| 315 | self, |
| 316 | video_shape: Tuple[int, ...], |
| 317 | audio_shape: Tuple[int, ...], |
| 318 | conditional_dict: Dict[str, Any], |
| 319 | memory_video: Optional[torch.Tensor] = None, |
| 320 | memory_audio: Optional[torch.Tensor] = None, |
| 321 | memory_audio_timestep: Optional[torch.Tensor] = None, |
| 322 | memory_audio_segment_lengths: tuple[tuple[int, ...], ...] | None = None, |
| 323 | memory_position_mode: str = "reference", |
| 324 | memory_position_offset: float = 0.0, |
| 325 | memory_position_slot_stride: float = 50.0, |
| 326 | seed: Optional[int] = None, |
| 327 | first_frame_latent: Optional[torch.Tensor] = None, |
| 328 | ) -> Tuple[torch.Tensor, torch.Tensor]: |
| 329 | batch_size = video_shape[0] |
| 330 | num_video_frames = video_shape[1] |
| 331 | num_audio_frames = audio_shape[1] |
| 332 | |
| 333 | if seed is not None: |
| 334 | torch.manual_seed(seed) |
| 335 | |
| 336 | device = next(self.generator.parameters()).device |
| 337 | dtype = next(self.generator.parameters()).dtype |
| 338 | |
| 339 | video = torch.randn(video_shape, device=device, dtype=dtype) |
| 340 | audio = torch.randn(audio_shape, device=device, dtype=dtype) |
| 341 | if first_frame_latent is not None: |
| 342 | if first_frame_latent.shape[1] != 1: |
| 343 | raise ValueError( |
| 344 | "first_frame_latent must contain one frame, got " |
| 345 | f"{tuple(first_frame_latent.shape)}" |
| 346 | ) |
| 347 | if tuple(first_frame_latent.shape[2:]) != tuple(video_shape[2:]): |
| 348 | raise ValueError( |
| 349 | "first-frame latent shape does not match requested output: " |
| 350 | f"{tuple(first_frame_latent.shape)} vs {tuple(video_shape)}" |
| 351 | ) |
| 352 | first_frame_latent = first_frame_latent.to(device=device, dtype=dtype) |
| 353 | video[:, :1] = first_frame_latent |
| 354 | if memory_video is not None: |
| 355 | memory_video = memory_video.to(device=device, dtype=dtype) |
| 356 | if memory_audio is not None: |
| 357 | memory_audio = memory_audio.to(device=device, dtype=dtype) |
| 358 | if memory_audio_timestep is not None: |
| 359 | memory_audio_timestep = memory_audio_timestep.to(device=device, dtype=dtype) |
| 360 | memory_audio_kwargs = ( |
| 361 | { |
| 362 | "memory_audio": memory_audio, |
| 363 | "memory_audio_timestep": memory_audio_timestep, |
| 364 | "memory_audio_segment_lengths": memory_audio_segment_lengths, |
| 365 | } |
| 366 | if memory_audio is not None or memory_audio_timestep is not None |
| 367 | else {} |
| 368 | ) |
| 369 | |
| 370 | for idx, sigma in enumerate(self.denoising_sigmas[:-1]): |
| 371 | video_sigma = sigma * torch.ones([batch_size, num_video_frames], device=device, dtype=dtype) |
| 372 | if first_frame_latent is not None: |
| 373 | video_sigma[:, 0] = 0 |
| 374 | audio_sigma = sigma * torch.ones([batch_size, num_audio_frames], device=device, dtype=dtype) |
| 375 | |
| 376 | generator_kwargs: Dict[str, Any] = {} |
| 377 | if memory_video is not None: |
| 378 | generator_kwargs = { |
| 379 | "memory_video": memory_video, |
| 380 | "memory_downscale_factor": self.memory_downscale_factor, |
| 381 | "memory_position_mode": memory_position_mode, |
| 382 | "memory_position_offset": memory_position_offset, |
| 383 | "memory_position_slot_stride": memory_position_slot_stride, |
| 384 | **memory_audio_kwargs, |
| 385 | } |
| 386 | pred_video, pred_audio = self.generator( |
| 387 | noisy_image_or_video=video, |
| 388 | conditional_dict=conditional_dict, |
| 389 | timestep=video_sigma, |
| 390 | noisy_audio=audio, |
| 391 | audio_timestep=audio_sigma, |
| 392 | **generator_kwargs, |
| 393 | ) |
| 394 | pred_video = pred_video.to(dtype=dtype) |
| 395 | pred_audio = pred_audio.to(dtype=dtype) |
| 396 | if first_frame_latent is not None: |
| 397 | pred_video[:, :1] = first_frame_latent |
| 398 | |
| 399 | next_sigma = self.denoising_sigmas[idx + 1] |
| 400 | if next_sigma > 0: |
| 401 | fresh_noise_video = torch.randn_like(video) |
| 402 | fresh_noise_audio = torch.randn_like(audio) |
| 403 | next_video_sigma = next_sigma * torch.ones([batch_size, num_video_frames], device=device, dtype=dtype) |
| 404 | next_audio_sigma = next_sigma * torch.ones([batch_size, num_audio_frames], device=device, dtype=dtype) |
| 405 | video = self.add_noise_fn( |
| 406 | pred_video.flatten(0, 1), |
| 407 | fresh_noise_video.flatten(0, 1), |
| 408 | next_video_sigma.flatten(0, 1), |
| 409 | ).unflatten(0, (batch_size, num_video_frames)).to(dtype=dtype) |
| 410 | audio = self.add_noise_fn(pred_audio, fresh_noise_audio, next_audio_sigma).to(dtype=dtype) |
| 411 | if first_frame_latent is not None: |
| 412 | video[:, :1] = first_frame_latent |
| 413 | else: |
| 414 | video = pred_video |
| 415 | audio = pred_audio |
| 416 | |
| 417 | if first_frame_latent is not None: |
| 418 | video[:, :1] = first_frame_latent |
| 419 | return video, audio |
| 420 | |
| 421 | |
| 422 | class BidirectionalR2VInferencePipeline(BidirectionalMemoryAVInferencePipeline): |
| 423 | """Unified T2V, memory-R2V, first-frame I2V, and combined R2V pipeline.""" |
| 424 |