| 1 | import logging |
| 2 | from dataclasses import replace |
| 3 | from functools import partial |
| 4 | from typing import Callable |
| 5 | |
| 6 | import torch |
| 7 | from tqdm import tqdm |
| 8 | |
| 9 | from ltx_core.components.diffusion_steps import Res2sDiffusionStep |
| 10 | from ltx_core.components.protocols import DiffusionStepProtocol |
| 11 | from ltx_core.utils import to_denoised, to_velocity |
| 12 | from ltx_pipelines.utils.helpers import post_process_latent, timesteps_from_mask |
| 13 | from ltx_pipelines.utils.res2s import get_res2s_coefficients |
| 14 | from ltx_pipelines.utils.types import DenoisingFunc, LatentState |
| 15 | |
| 16 | logger = logging.getLogger(__name__) |
| 17 | |
| 18 | |
| 19 | def euler_denoising_loop( |
| 20 | sigmas: torch.Tensor, |
| 21 | video_state: LatentState, |
| 22 | audio_state: LatentState, |
| 23 | stepper: DiffusionStepProtocol, |
| 24 | denoise_fn: DenoisingFunc, |
| 25 | ) -> tuple[LatentState, LatentState]: |
| 26 | """ |
| 27 | Perform the joint audio-video denoising loop over a diffusion schedule. |
| 28 | This function iterates over all but the final value in ``sigmas`` and, at |
| 29 | each diffusion step, calls ``denoise_fn`` to obtain denoised video and |
| 30 | audio latents. The denoised latents are post-processed with their |
| 31 | respective denoise masks and clean latents, then passed to ``stepper`` to |
| 32 | advance the noisy latents one step along the diffusion schedule. |
| 33 | ### Parameters |
| 34 | sigmas: |
| 35 | A 1D tensor of noise levels (diffusion sigmas) defining the sampling |
| 36 | schedule. All steps except the last element are iterated over. |
| 37 | video_state: |
| 38 | The current video :class:`LatentState`, containing the noisy latent, |
| 39 | its clean reference latent, and the denoising mask. |
| 40 | audio_state: |
| 41 | The current audio :class:`LatentState`, analogous to ``video_state`` |
| 42 | but for the audio modality. |
| 43 | stepper: |
| 44 | An implementation of :class:`DiffusionStepProtocol` that updates a |
| 45 | latent given the current latent, its denoised estimate, the full |
| 46 | ``sigmas`` schedule, and the current step index. |
| 47 | denoise_fn: |
| 48 | A callable implementing :class:`DenoisingFunc`. It is invoked as |
| 49 | ``denoise_fn(video_state, audio_state, sigmas, step_index)`` and must |
| 50 | return a tuple ``(denoised_video, denoised_audio)``, where each element |
| 51 | is a tensor with the same shape as the corresponding latent. |
| 52 | ### Returns |
| 53 | tuple[LatentState, LatentState] |
| 54 | A pair ``(video_state, audio_state)`` containing the final video and |
| 55 | audio latent states after completing the denoising loop. |
| 56 | """ |
| 57 | for step_idx, _ in enumerate(tqdm(sigmas[:-1])): |
| 58 | denoised_video, denoised_audio = denoise_fn(video_state, audio_state, sigmas, step_idx) |
| 59 | |
| 60 | denoised_video = post_process_latent(denoised_video, video_state.denoise_mask, video_state.clean_latent) |
| 61 | denoised_audio = post_process_latent(denoised_audio, audio_state.denoise_mask, audio_state.clean_latent) |
| 62 | |
| 63 | video_state = replace(video_state, latent=stepper.step(video_state.latent, denoised_video, sigmas, step_idx)) |
| 64 | audio_state = replace(audio_state, latent=stepper.step(audio_state.latent, denoised_audio, sigmas, step_idx)) |
| 65 | |
| 66 | return (video_state, audio_state) |
| 67 | |
| 68 | |
| 69 | def gradient_estimating_euler_denoising_loop( |
| 70 | sigmas: torch.Tensor, |
| 71 | video_state: LatentState, |
| 72 | audio_state: LatentState, |
| 73 | stepper: DiffusionStepProtocol, |
| 74 | denoise_fn: DenoisingFunc, |
| 75 | ge_gamma: float = 2.0, |
| 76 | ) -> tuple[LatentState, LatentState]: |
| 77 | """ |
| 78 | Perform the joint audio-video denoising loop using gradient-estimation sampling. |
| 79 | This function is similar to :func:`euler_denoising_loop`, but applies |
| 80 | gradient estimation to improve the denoised estimates by tracking velocity |
| 81 | changes across steps. See the referenced function for detailed parameter |
| 82 | documentation. |
| 83 | ### Parameters |
| 84 | ge_gamma: |
| 85 | Gradient estimation coefficient controlling the velocity correction term. |
| 86 | Default is 2.0. Paper: https://openreview.net/pdf?id=o2ND9v0CeK |
| 87 | sigmas, video_state, audio_state, stepper, denoise_fn: |
| 88 | See :func:`euler_denoising_loop` for parameter descriptions. |
| 89 | ### Returns |
| 90 | tuple[LatentState, LatentState] |
| 91 | See :func:`euler_denoising_loop` for return value description. |
| 92 | """ |
| 93 | |
| 94 | previous_audio_velocity = None |
| 95 | previous_video_velocity = None |
| 96 | |
| 97 | def update_velocity_and_sample( |
| 98 | noisy_sample: torch.Tensor, denoised_sample: torch.Tensor, sigma: float, previous_velocity: torch.Tensor | None |
| 99 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 100 | current_velocity = to_velocity(noisy_sample, sigma, denoised_sample) |
| 101 | if previous_velocity is not None: |
| 102 | delta_v = current_velocity - previous_velocity |
| 103 | total_velocity = ge_gamma * delta_v + previous_velocity |
| 104 | denoised_sample = to_denoised(noisy_sample, total_velocity, sigma) |
| 105 | return current_velocity, denoised_sample |
| 106 | |
| 107 | for step_idx, _ in enumerate(tqdm(sigmas[:-1])): |
| 108 | denoised_video, denoised_audio = denoise_fn(video_state, audio_state, sigmas, step_idx) |
| 109 | |
| 110 | denoised_video = post_process_latent(denoised_video, video_state.denoise_mask, video_state.clean_latent) |
| 111 | denoised_audio = post_process_latent(denoised_audio, audio_state.denoise_mask, audio_state.clean_latent) |
| 112 | |
| 113 | if sigmas[step_idx + 1] == 0: |
| 114 | return replace(video_state, latent=denoised_video), replace(audio_state, latent=denoised_audio) |
| 115 | |
| 116 | previous_video_velocity, denoised_video = update_velocity_and_sample( |
| 117 | video_state.latent, denoised_video, sigmas[step_idx], previous_video_velocity |
| 118 | ) |
| 119 | previous_audio_velocity, denoised_audio = update_velocity_and_sample( |
| 120 | audio_state.latent, denoised_audio, sigmas[step_idx], previous_audio_velocity |
| 121 | ) |
| 122 | |
| 123 | video_state = replace(video_state, latent=stepper.step(video_state.latent, denoised_video, sigmas, step_idx)) |
| 124 | audio_state = replace(audio_state, latent=stepper.step(audio_state.latent, denoised_audio, sigmas, step_idx)) |
| 125 | |
| 126 | return (video_state, audio_state) |
| 127 | |
| 128 | |
| 129 | def _channelwise_normalize(x: torch.Tensor) -> torch.Tensor: |
| 130 | return x.sub_(x.mean(dim=(-2, -1), keepdim=True)).div_(x.std(dim=(-2, -1), keepdim=True)) |
| 131 | |
| 132 | |
| 133 | def _get_new_noise(x: torch.Tensor, generator: torch.Generator) -> torch.Tensor: |
| 134 | noise = torch.randn(x.shape, generator=generator, dtype=torch.float64, device=generator.device) |
| 135 | noise = (noise - noise.mean()) / noise.std() |
| 136 | return _channelwise_normalize(noise) |
| 137 | |
| 138 | |
| 139 | def _inject_sde_noise( |
| 140 | state: LatentState, |
| 141 | sample: torch.Tensor, |
| 142 | denoised_sample: torch.Tensor, |
| 143 | step_noise_generator: torch.Generator, |
| 144 | new_noise_fn: Callable[[torch.Tensor, torch.Generator], torch.Tensor], |
| 145 | stepper: DiffusionStepProtocol, |
| 146 | sigmas: torch.Tensor, |
| 147 | step_idx: int, |
| 148 | legacy_mode: bool = False, |
| 149 | ) -> torch.Tensor: |
| 150 | sigmas_copy = sigmas.clone() |
| 151 | new_noise = new_noise_fn(state.latent, step_noise_generator) |
| 152 | if not legacy_mode: |
| 153 | timesteps = timesteps_from_mask(state.denoise_mask.double(), sigmas_copy[step_idx].double()) |
| 154 | next_timesteps = timesteps_from_mask(state.denoise_mask.double(), sigmas_copy[step_idx + 1].double()) |
| 155 | sigmas = torch.stack([timesteps, next_timesteps]) |
| 156 | step_idx = 0 |
| 157 | x_next = stepper.step( |
| 158 | sample=sample, |
| 159 | denoised_sample=denoised_sample, |
| 160 | sigmas=sigmas, |
| 161 | step_index=step_idx, |
| 162 | noise=new_noise, |
| 163 | ) |
| 164 | |
| 165 | if legacy_mode: |
| 166 | x_next = post_process_latent(x_next, state.denoise_mask, state.clean_latent) |
| 167 | |
| 168 | return x_next |
| 169 | |
| 170 | |
| 171 | def res2s_audio_video_denoising_loop( # noqa: PLR0913,PLR0915 |
| 172 | sigmas: torch.Tensor, |
| 173 | video_state: LatentState, |
| 174 | audio_state: LatentState, |
| 175 | stepper: DiffusionStepProtocol, |
| 176 | denoise_fn: DenoisingFunc, |
| 177 | noise_seed: int = -1, |
| 178 | noise_seed_substep: int | None = None, |
| 179 | bongmath: bool = True, |
| 180 | bongmath_max_iter: int = 100, |
| 181 | new_noise_fn: Callable[[torch.Tensor, torch.Generator], torch.Tensor] = _get_new_noise, |
| 182 | model_dtype: torch.dtype = torch.bfloat16, |
| 183 | legacy_mode: bool = True, |
| 184 | ) -> tuple[LatentState, LatentState]: |
| 185 | """ |
| 186 | Joint audio-video denoising loop using the res_2s second-order sampler. |
| 187 | Iterates over the diffusion schedule with a two-stage Runge-Kutta step: |
| 188 | evaluates the denoiser at the current point and at a midpoint (with SDE |
| 189 | noise), then combines both with RK coefficients. Supports anchor-point |
| 190 | refinement (bong iteration) and optional SDE noise injection. Requires |
| 191 | :class:`Res2sDiffusionStep` as ``stepper``. |
| 192 | ### Parameters |
| 193 | sigmas: |
| 194 | A 1D tensor of noise levels defining the sampling schedule. |
| 195 | video_state: |
| 196 | Current video :class:`LatentState` (noisy latent, clean reference, mask). |
| 197 | audio_state: |
| 198 | Current audio :class:`LatentState`, same structure as ``video_state``. |
| 199 | stepper: |
| 200 | Must be an instance of :class:`Res2sDiffusionStep`; performs SDE step |
| 201 | with noise injection. |
| 202 | denoise_fn: |
| 203 | Callable ``(video_state, audio_state, sigmas, step_index)`` returning |
| 204 | ``(denoised_video, denoised_audio)``. |
| 205 | noise_seed: |
| 206 | Seed for step-level SDE noise; substep seed defaults to ``noise_seed + 10000``. |
| 207 | noise_seed_substep: |
| 208 | Optional seed for substep SDE noise; if None, derived from ``noise_seed``. |
| 209 | bongmath: |
| 210 | Whether to run iterative anchor refinement (bong iteration) when step size is small. |
| 211 | bongmath_max_iter: |
| 212 | Max iterations for bong refinement when enabled. |
| 213 | new_noise_fn: |
| 214 | Callable ``(latent, generator) -> noise`` for SDE injection; default |
| 215 | uses normalized channel-wise Gaussian noise. |
| 216 | model_dtype: |
| 217 | Dtype for latent state updates (e.g. bfloat16). |
| 218 | ### Returns |
| 219 | tuple[LatentState, LatentState] |
| 220 | Final ``(video_state, audio_state)`` after the denoising loop. |
| 221 | """ |
| 222 | # Initialize noise generators with different seeds |
| 223 | if noise_seed_substep is None: |
| 224 | noise_seed_substep = noise_seed + 10000 # Offset to ensure different seeds |
| 225 | step_noise_generator = torch.Generator(device=video_state.latent.device).manual_seed(noise_seed) |
| 226 | substep_noise_generator = torch.Generator(device=video_state.latent.device).manual_seed(noise_seed_substep) |
| 227 | sde_noise_injecting_fn = partial( |
| 228 | _inject_sde_noise, stepper=stepper, new_noise_fn=new_noise_fn, legacy_mode=legacy_mode |
| 229 | ) |
| 230 | step_noise_injecting_fn = partial(sde_noise_injecting_fn, step_noise_generator=step_noise_generator) |
| 231 | substep_noise_injecting_fn = partial(sde_noise_injecting_fn, step_noise_generator=substep_noise_generator) |
| 232 | |
| 233 | if not isinstance(stepper, Res2sDiffusionStep): |
| 234 | raise ValueError("stepper must be an instance of Res2sDiffusionStep") |
| 235 | |
| 236 | n_full_steps = len(sigmas) - 1 |
| 237 | # inject minimal sigma value to avoid division by zero |
| 238 | if sigmas[-1] == 0: |
| 239 | sigmas = torch.cat([sigmas[:-1], torch.tensor([0.0011, 0.0], device=sigmas.device)], dim=0) |
| 240 | # Compute step sizes in hyperbolic space |
| 241 | hs = -torch.log(sigmas[1:].double().cpu() / (sigmas[:-1].double().cpu())) |
| 242 | |
| 243 | # Initialize phi cache for reuse across loop iterations |
| 244 | # Cache key: (j, neg_h) where j is phi order and neg_h is negative step value |
| 245 | phi_cache = {} |
| 246 | c2 = 0.5 # Midpoint for res_2s |
| 247 | |
| 248 | # Progress bar shows only full two-stage steps; final (sigma_next==0) step is done silently |
| 249 | |
| 250 | for step_idx in tqdm(range(n_full_steps)): |
| 251 | sigma = sigmas[step_idx].double() |
| 252 | sigma_next = sigmas[step_idx + 1].double() |
| 253 | |
| 254 | # Initialize anchor point |
| 255 | x_anchor_video = video_state.latent.clone().double() |
| 256 | x_anchor_audio = audio_state.latent.clone().double() |
| 257 | |
| 258 | # ==================================================================== |
| 259 | # STAGE 1: Evaluate at current point |
| 260 | # ==================================================================== |
| 261 | denoised_video_1, denoised_audio_1 = denoise_fn(video_state, audio_state, sigmas, step_idx) |
| 262 | denoised_video_1 = post_process_latent(denoised_video_1, video_state.denoise_mask, video_state.clean_latent) |
| 263 | denoised_audio_1 = post_process_latent(denoised_audio_1, audio_state.denoise_mask, audio_state.clean_latent) |
| 264 | |
| 265 | h = hs[step_idx].item() |
| 266 | |
| 267 | # Compute RK coefficients (pass phi_cache for caching) |
| 268 | a21, b1, b2 = get_res2s_coefficients(h, phi_cache, c2) |
| 269 | |
| 270 | # Compute substep sigma, sqrt is a hardcode for c2 = 0.5 |
| 271 | sub_sigma = torch.sqrt(sigma * sigma_next) |
| 272 | |
| 273 | # ==================================================================== |
| 274 | # Compute substep x using RK coefficient a21 |
| 275 | # ==================================================================== |
| 276 | eps_1_video = denoised_video_1.double() - x_anchor_video |
| 277 | eps_1_audio = denoised_audio_1.double() - x_anchor_audio |
| 278 | |
| 279 | x_mid_video = x_anchor_video.double() + h * a21 * eps_1_video |
| 280 | x_mid_audio = x_anchor_audio.double() + h * a21 * eps_1_audio |
| 281 | |
| 282 | # ==================================================================== |
| 283 | # SDE noise injection at substep |
| 284 | # ==================================================================== |
| 285 | x_mid_video = substep_noise_injecting_fn( |
| 286 | state=video_state, |
| 287 | sample=x_anchor_video, |
| 288 | denoised_sample=x_mid_video, |
| 289 | sigmas=torch.stack([sigma, sub_sigma]), |
| 290 | step_idx=0, |
| 291 | ) |
| 292 | x_mid_audio = substep_noise_injecting_fn( |
| 293 | state=audio_state, |
| 294 | sample=x_anchor_audio, |
| 295 | denoised_sample=x_mid_audio, |
| 296 | sigmas=torch.stack([sigma, sub_sigma]), |
| 297 | step_idx=0, |
| 298 | ) |
| 299 | # ==================================================================== |
| 300 | # ITERATIVE REFINEMENT (Bong Iteration) - Stabilize anchor point |
| 301 | # ==================================================================== |
| 302 | if bongmath and h < 0.5 and sigma > 0.03: |
| 303 | for _ in range(bongmath_max_iter): |
| 304 | x_anchor_video = x_mid_video - h * a21 * eps_1_video |
| 305 | eps_1_video = denoised_video_1.double() - x_anchor_video |
| 306 | x_anchor_audio = x_mid_audio - h * a21 * eps_1_audio |
| 307 | eps_1_audio = denoised_audio_1.double() - x_anchor_audio |
| 308 | |
| 309 | # ==================================================================== |
| 310 | # STAGE 2: Evaluate at substep point (WITH NOISE) |
| 311 | # ==================================================================== |
| 312 | mid_video_state = replace(video_state, latent=x_mid_video.to(model_dtype)) |
| 313 | mid_audio_state = replace(audio_state, latent=x_mid_audio.to(model_dtype)) |
| 314 | |
| 315 | denoised_video_2, denoised_audio_2 = denoise_fn( |
| 316 | video_state=mid_video_state, |
| 317 | audio_state=mid_audio_state, |
| 318 | sigmas=torch.stack([sub_sigma]).to(sigmas.device), |
| 319 | step_index=0, |
| 320 | ) |
| 321 | denoised_video_2 = post_process_latent(denoised_video_2, video_state.denoise_mask, video_state.clean_latent) |
| 322 | denoised_audio_2 = post_process_latent(denoised_audio_2, audio_state.denoise_mask, audio_state.clean_latent) |
| 323 | |
| 324 | # ==================================================================== |
| 325 | # FINAL COMBINATION: Compute x_next using RK coefficients |
| 326 | # ==================================================================== |
| 327 | eps_2_video = denoised_video_2.double() - x_anchor_video |
| 328 | eps_2_audio = denoised_audio_2.double() - x_anchor_audio |
| 329 | |
| 330 | x_next_video = x_anchor_video + h * (b1 * eps_1_video + b2 * eps_2_video) |
| 331 | x_next_audio = x_anchor_audio + h * (b1 * eps_1_audio + b2 * eps_2_audio) |
| 332 | |
| 333 | # ==================================================================== |
| 334 | # SDE NOISE INJECTION AT STEP LEVEL |
| 335 | # ==================================================================== |
| 336 | x_next_video = step_noise_injecting_fn( |
| 337 | state=video_state, |
| 338 | sample=x_anchor_video, |
| 339 | denoised_sample=x_next_video, |
| 340 | sigmas=sigmas, |
| 341 | step_idx=step_idx, |
| 342 | ) |
| 343 | x_next_audio = step_noise_injecting_fn( |
| 344 | state=audio_state, |
| 345 | sample=x_anchor_audio, |
| 346 | denoised_sample=x_next_audio, |
| 347 | sigmas=sigmas, |
| 348 | step_idx=step_idx, |
| 349 | ) |
| 350 | |
| 351 | # Update states |
| 352 | video_state = replace(video_state, latent=x_next_video.to(model_dtype)) |
| 353 | audio_state = replace(audio_state, latent=x_next_audio.to(model_dtype)) |
| 354 | |
| 355 | # Final step if we need to fully remove the noise |
| 356 | if sigmas[-1] == 0: |
| 357 | denoised_video_1, denoised_audio_1 = denoise_fn(video_state, audio_state, sigmas, n_full_steps) |
| 358 | denoised_video_1 = post_process_latent(denoised_video_1, video_state.denoise_mask, video_state.clean_latent) |
| 359 | denoised_audio_1 = post_process_latent(denoised_audio_1, audio_state.denoise_mask, audio_state.clean_latent) |
| 360 | video_state = replace(video_state, latent=denoised_video_1.to(model_dtype)) |
| 361 | audio_state = replace(audio_state, latent=denoised_audio_1.to(model_dtype)) |
| 362 | |
| 363 | return video_state, audio_state |
| 364 |