File size: 18,846 Bytes
c785bc6 | 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 | """Chunk-based diffusion model (no history re-noising).
Config: history_len=m, chunk_size=n, steps=T
- Global time t ∈ [0, num_chunks), where num_chunks = 1 + ceil((N - (m+n)) / n)
- Schedule: before window → 1.0, history → 1.0 (clean), target → frac(t), after → 0.0
- Inference: history stays clean, only target frames are denoised
- First chunk uses GT history frames as conditioning
"""
import math
import numpy as np
import torch
from .diffusion_forcing_wan import DiffForcingWanModel
EPSILON = 0.05
class ChunkDiffusionScheduler:
def __init__(self, config):
self.steps = config["steps"]
self.chunk_size = config["chunk_size"] # n
self.history_len = config.get("history_len", 0) # m
self.window_size = self.history_len + self.chunk_size # m+n
self.noise_type = config.get("noise_type", "linear")
self.sigma_type = config.get("sigma_type", "zero")
self.random_epsilon = config.get("random_epsilon", 0.0)
self.content_len = config.get("content_len", None)
if self.noise_type in ("exponential", "exponential_rev"):
self.exp_max = config.get("exp_max", 5.0)
elif self.noise_type == "diffusion":
self.T = config.get("T", 1000)
self.beta_start = config.get("beta_start", 0.0001)
self.beta_end = config.get("beta_end", 0.02)
if self.sigma_type == "memoryless":
self.sigma_scale = config.get("sigma_scale", 1.0)
# ----------------------------------------------------------------
# Chunks
# ----------------------------------------------------------------
def _num_chunks(self, seq_len):
if seq_len <= self.window_size:
return 1
return 1 + math.ceil((seq_len - self.window_size) / self.chunk_size)
def _window_range(self, seq_len, chunk_idx, training=False):
"""Return (input_start, input_end, output_start, output_end) for a chunk."""
if chunk_idx == 0:
os_ = self.history_len # First m frames are always GT history
oe_ = min(self.window_size, seq_len)
is_ = 0
else:
os_ = self.window_size + (chunk_idx - 1) * self.chunk_size
oe_ = min(os_ + self.chunk_size, seq_len)
is_ = os_ - self.history_len
if self.content_len is not None:
is_ = max(is_, oe_ - self.content_len)
# output always covers target only (excludes history)
return is_, oe_, os_, oe_
# ----------------------------------------------------------------
# Scheduler interface
# ----------------------------------------------------------------
def get_total_steps(self, seq_len):
return self._num_chunks(seq_len) * self.steps
def get_time_steps(self, device, valid_len, current_step=None):
time_steps = []
if current_step is None:
for i in range(len(valid_len)):
max_time = self._num_chunks(valid_len[i])
time_steps.append(
torch.tensor(np.random.uniform(0, max_time), device=device)
)
elif isinstance(current_step, int):
for i in range(len(valid_len)):
t = current_step * (1.0 / self.steps)
time_steps.append(torch.tensor(t, device=device))
elif isinstance(current_step, list):
for i in range(len(valid_len)):
t = current_step[i] * (1.0 / self.steps)
time_steps.append(torch.tensor(t, device=device))
return time_steps
def get_time_schedules(self, device, valid_len, time_steps, training=False):
time_schedules = []
time_schedules_derivative = []
for i in range(len(valid_len)):
t = time_steps[i].item()
chunk_idx = min(int(t), self._num_chunks(valid_len[i]) - 1)
t_frac = t - chunk_idx
is_, ie_, os_, oe_ = self._window_range(valid_len[i], chunk_idx)
ts = torch.zeros(valid_len[i], device=device)
# Before window → 1.0 (clean)
ts[:is_] = 1.0
if training:
# Training: entire window uses t_frac
ts[is_:ie_] = t_frac
else:
# Inference: history → 1.0 (clean, no renoise), target → t_frac
ts[is_:os_] = 1.0
ts[os_:oe_] = t_frac
tsd = torch.full((valid_len[i],), 1.0 / self.steps, device=device)
if training:
ts = torch.clamp(
ts + torch.randn_like(ts) * self.random_epsilon,
min=0.0, max=1.0,
)
time_schedules.append(ts)
time_schedules_derivative.append(tsd)
return time_schedules, time_schedules_derivative
def get_windows(self, valid_len, time_steps, training=False):
input_start, input_end, output_start, output_end = [], [], [], []
for i in range(len(time_steps)):
t = time_steps[i].item()
chunk_idx = min(int(t), self._num_chunks(valid_len[i]) - 1)
is_, ie_, os_, oe_ = self._window_range(valid_len[i], chunk_idx, training=training)
input_start.append(is_)
input_end.append(ie_)
output_start.append(os_)
output_end.append(oe_)
return input_start, input_end, output_start, output_end
def get_noise_levels(self, device, valid_len, time_schedules, training=False):
alpha, dalpha, dlog_alpha = [], [], []
beta, dbeta, dlog_beta = [], [], []
sigma = []
for i in range(len(valid_len)):
t = time_schedules[i]
if self.noise_type == "linear":
alpha_i = t
dalpha_i = torch.ones_like(alpha_i)
dlog_alpha_i = dalpha_i / torch.clamp(alpha_i, min=EPSILON)
beta_i = 1 - t
dbeta_i = -torch.ones_like(beta_i)
dlog_beta_i = dbeta_i / torch.clamp(beta_i, min=EPSILON)
elif self.noise_type == "exponential":
k = self.exp_max
alpha_i = torch.exp(-k * (1 - t))
dalpha_i = k * alpha_i
dlog_alpha_i = k * torch.ones_like(alpha_i)
beta_i = 1 - alpha_i
dbeta_i = -dalpha_i
dlog_beta_i = dbeta_i / torch.clamp(beta_i, min=EPSILON)
elif self.noise_type == "exponential_rev":
k = self.exp_max
beta_i = torch.exp(-k * t)
dbeta_i = -k * beta_i
dlog_beta_i = -k * torch.ones_like(beta_i)
alpha_i = 1 - beta_i
dalpha_i = -dbeta_i
dlog_alpha_i = dalpha_i / torch.clamp(alpha_i, min=EPSILON)
elif self.noise_type == "diffusion":
t_rev = 1.0 - t
beta_rate = (self.beta_start + t_rev * (self.beta_end - self.beta_start)) * self.T
Gamma = (self.beta_start * t_rev + 0.5 * (self.beta_end - self.beta_start) * t_rev * t_rev) * self.T
alpha_i = torch.exp(-0.5 * Gamma)
dalpha_i = 0.5 * beta_rate * alpha_i
dlog_alpha_i = 0.5 * beta_rate
beta_i = torch.sqrt(torch.clamp(1 - torch.exp(-Gamma), min=0.0))
dbeta_i = -0.5 * torch.exp(-Gamma) * beta_rate / torch.clamp(beta_i, min=EPSILON)
dlog_beta_i = dbeta_i / torch.clamp(beta_i, min=EPSILON)
else:
raise ValueError(f"Unknown noise type: {self.noise_type}")
alpha.append(torch.clamp(alpha_i, min=0.0, max=1.0))
dalpha.append(dalpha_i)
dlog_alpha.append(dlog_alpha_i)
beta.append(torch.clamp(beta_i, min=0.0, max=1.0))
dbeta.append(dbeta_i)
dlog_beta.append(dlog_beta_i)
if self.sigma_type == "zero":
sigma_i = torch.zeros_like(t)
elif self.sigma_type == "memoryless":
if self.noise_type in ("linear", "exponential", "exponential_rev"):
sigma_i = self.sigma_scale * torch.sqrt(torch.clamp(2 * dlog_alpha_i * beta_i, min=0.0))
elif self.noise_type == "diffusion":
sigma_i = self.sigma_scale * torch.sqrt(torch.clamp(2 * dlog_alpha_i, min=0.0))
else:
sigma_i = self.sigma_scale * torch.sqrt(torch.clamp(2 * beta_i * (dlog_alpha_i * beta_i - dbeta_i), min=0.0))
sigma.append(sigma_i)
return alpha, dalpha, beta, dbeta, sigma, dlog_alpha, dlog_beta
def add_noise(self, x, alpha, beta, input_start, input_end, output_start, output_end, training=False, noise=None):
x0, eps, xt = [], [], []
if training:
for i in range(len(x)):
noise_i = noise[i] if noise is not None else torch.randn_like(x[i])
alpha_i = alpha[i][None, :, None, None]
beta_i = beta[i][None, :, None, None]
noisy_x_i = x[i] * alpha_i + noise_i * beta_i
x0.append(x[i][:, output_start[i]:output_end[i], ...])
eps.append(noise_i[:, output_start[i]:output_end[i], ...])
xt.append(noisy_x_i[:, input_start[i]:input_end[i], ...])
else:
# No re-noising: history frames stay as-is, target frames stay as-is
for i in range(len(x)):
xt.append(x[i][:, input_start[i]:input_end[i], ...])
return x0, eps, xt
def prepare(self, x, device, valid_len, training=True, current_step=None):
"""Single call replacing 5 separate scheduler calls."""
time_steps = self.get_time_steps(device, valid_len, current_step)
time_schedules, time_schedules_derivative = self.get_time_schedules(
device, valid_len, time_steps, training=training
)
alpha, dalpha, beta, dbeta, sigma, dlog_alpha, dlog_beta = \
self.get_noise_levels(device, valid_len, time_schedules, training=training)
input_start, input_end, output_start, output_end = \
self.get_windows(valid_len, time_steps, training=training)
x0, eps, xt = self.add_noise(
x, alpha, beta, input_start, input_end,
output_start, output_end, training=training
)
# Slice all coefficients to their respective windows
batch_size = len(valid_len)
time_schedules = [time_schedules[i][input_start[i]:input_end[i]] for i in range(batch_size)]
time_schedules_derivative = [time_schedules_derivative[i][output_start[i]:output_end[i]] for i in range(batch_size)]
alpha = [alpha[i][output_start[i]:output_end[i]] for i in range(batch_size)]
dalpha = [dalpha[i][output_start[i]:output_end[i]] for i in range(batch_size)]
beta = [beta[i][output_start[i]:output_end[i]] for i in range(batch_size)]
dbeta = [dbeta[i][output_start[i]:output_end[i]] for i in range(batch_size)]
sigma = [sigma[i][output_start[i]:output_end[i]] for i in range(batch_size)]
dlog_alpha = [dlog_alpha[i][output_start[i]:output_end[i]] for i in range(batch_size)]
dlog_beta = [dlog_beta[i][output_start[i]:output_end[i]] for i in range(batch_size)]
return {
"time_schedules": time_schedules,
"time_schedules_derivative": time_schedules_derivative,
"input_start": input_start,
"input_end": input_end,
"output_start": output_start,
"output_end": output_end,
"alpha": alpha,
"dalpha": dalpha,
"beta": beta,
"dbeta": dbeta,
"sigma": sigma,
"dlog_alpha": dlog_alpha,
"dlog_beta": dlog_beta,
"xt": xt,
"x0": x0,
"eps": eps,
}
# ----------------------------------------------------------------
# Streaming support
# ----------------------------------------------------------------
def get_committable(self, total_frames):
if total_frames < self.window_size:
return 0, 0
committed = self.window_size
committable_steps = self.steps
remaining = total_frames - self.window_size
extra_chunks = remaining // self.chunk_size
committed += extra_chunks * self.chunk_size
committable_steps += extra_chunks * self.steps
return committed, committable_steps
def get_step_rollback(self, seq_len):
if seq_len < self.window_size:
return 0
completed = 1
remaining = seq_len - self.window_size
completed += remaining // self.chunk_size
return completed * self.steps
class ChunkDiffWanModel(DiffForcingWanModel):
"""Chunk-based diffusion model with clean history conditioning.
First chunk: GT history (history_len frames) + noisy target.
Subsequent chunks: previously generated frames as history + noisy target.
History is never re-noised.
"""
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.time_scheduler = ChunkDiffusionScheduler(self.schedule_config)
def generate(self, x):
x = self._extract_inputs(x)
extra_len = self.schedule_config.get("extra_len", 0)
feature_length = x["feature_length"]
batch_size = len(feature_length)
seq_len = max(feature_length).item() + extra_len
device = next(self.parameters()).device
valid_len = [min(fl.item(), seq_len) for fl in feature_length]
generated_len = [seq_len] * batch_size
history_len = self.time_scheduler.history_len
# Initialize entire sequence as pure noise
generated = torch.randn(
batch_size, seq_len, *self.spatial_shape, self.input_dim, device=device
)
generated = [generated[i] for i in range(batch_size)]
generated = self.preprocess(generated)
# Inject GT history into the first history_len frames
if "feature" in x:
gt_feature = x["feature"]
gt_feature = (gt_feature - self.mean) / self.std
gt_list = []
for i in range(batch_size):
gt_list.append(gt_feature[i, :valid_len[i], ...])
gt_list = self.preprocess(gt_list)
for i in range(batch_size):
h = min(history_len, gt_list[i].shape[1])
generated[i][:, :h, ...] = gt_list[i][:, :h, ...]
# Precompute text and null contexts
text_context, metadata = self.text_module.get_context(
x, generated_len, device, self.param_dtype, training=False,
)
null_context = self.text_module.get_null_context(batch_size, device, self.param_dtype)
full_text = metadata["full_text"]
total_steps = self.time_scheduler.get_total_steps(seq_len)
for step in range(total_steps):
s = self.time_scheduler.prepare(
generated, device, generated_len, training=False, current_step=step
)
time_schedules = s["time_schedules"]
time_schedules_derivative = s["time_schedules_derivative"]
alpha = s["alpha"]
dalpha = s["dalpha"]
beta = s["beta"]
dbeta = s["dbeta"]
sigma = s["sigma"]
dlog_alpha = s["dlog_alpha"]
dlog_beta = s["dlog_beta"]
input_start_index = s["input_start"]
input_end_index = s["input_end"]
output_start_index = s["output_start"]
output_end_index = s["output_end"]
xt = s["xt"]
time_schedules_input = [
time_schedules[i] * self.time_embedding_scale for i in range(batch_size)
]
if isinstance(text_context[0], (list, tuple)):
window_text_context = [
text_context[i][input_start_index[i]:input_end_index[i]]
for i in range(batch_size)
]
else:
window_text_context = text_context
# CFG
pred_text = self.model(xt, time_schedules_input, window_text_context, seq_len, y=None)
pred_null = self.model(xt, time_schedules_input, null_context, seq_len, y=None)
predicted_result = [
self.cfg_config["text_scale"] * pt + self.cfg_config["null_scale"] * pn
for pt, pn in zip(pred_text, pred_null)
]
# SDE update only on output (target) frames
for i in range(batch_size):
os_idx, oe_idx = output_start_index[i], output_end_index[i]
pred_os = os_idx - input_start_index[i]
pred_oe = oe_idx - input_start_index[i]
predicted_result_i = predicted_result[i][:, pred_os:pred_oe, ...]
generated_i = generated[i][:, os_idx:oe_idx, ...]
dt = time_schedules_derivative[i][None, :, None, None]
alpha_i = alpha[i][None, :, None, None]
dalpha_i = dalpha[i][None, :, None, None]
beta_i = beta[i][None, :, None, None]
dbeta_i = dbeta[i][None, :, None, None]
sigma_i = sigma[i][None, :, None, None]
dlog_alpha_i = dlog_alpha[i][None, :, None, None]
dlog_beta_i = dlog_beta[i][None, :, None, None]
if self.prediction_type == "vel":
vel = predicted_result_i
elif self.prediction_type == "x0":
vel = (
predicted_result_i * (-dlog_beta_i * alpha_i + dalpha_i)
+ generated_i * dlog_beta_i
)
elif self.prediction_type == "eps":
vel = (
predicted_result_i * (-dlog_alpha_i * beta_i + dbeta_i)
+ generated_i * dlog_alpha_i
)
st = (vel - generated_i * dlog_alpha_i) / (
(beta_i * dlog_alpha_i - dbeta_i) * beta_i
)
generated[i][:, os_idx:oe_idx, ...] += (
vel * dt
+ st * 0.5 * sigma_i ** 2 * dt
+ sigma_i * torch.sqrt(dt) * torch.randn_like(generated_i)
)
generated = self.postprocess(generated)
y_hat_out = []
for i in range(batch_size):
single_generated = generated[i][:valid_len[i], :] * self.std + self.mean
y_hat_out.append(single_generated)
return {"generated": y_hat_out, "text": full_text}
def init_generated(self, seq_len, batch_size=1, schedule_config={}):
super().init_generated(seq_len, batch_size, schedule_config)
self.time_scheduler = ChunkDiffusionScheduler(self.schedule_config)
|