Files
aigc-home/docs/hel-892/t8-two-reference/t8base__sampling_multirate_exp.py
T
2026-11-01 08:02:36 +08:00

258 lines
9.3 KiB
Python

from __future__ import annotations
import math
import torch
import comfy.model_sampling
import comfy.samplers
import comfy.utils
from comfy.k_diffusion.sampling import to_d
from .core import nested_av_parts
class MiniMaxH3MultiRateFlowSamplingEXP(
comfy.model_sampling.ModelSamplingDiscreteFlow,
comfy.model_sampling.CONST,
):
@property
def audio_scale(self):
# This sampler owns the separate audio clock, so ComfyUI must not also
# carry/scale the audio stream onto the video schedule.
return 1.0
def model_uses_raw_audio_velocity(model) -> bool:
"""Detect the post-2026-08-06 ComfyUI MiniMax H3 sampling protocol."""
base_model = getattr(model, "model", None)
return callable(getattr(base_model, "audio_scale", None))
def shift_sigma(base_sigma, shift: float):
return shift * base_sigma / (1.0 + (shift - 1.0) * base_sigma)
def time_shift_sigma(sigma, from_shift: float, to_shift: float):
base_sigma = sigma / (from_shift + sigma * (1.0 - from_shift))
return shift_sigma(base_sigma, to_shift)
def time_shift_slope(sigma, from_shift: float, to_shift: float):
base_sigma = sigma / (from_shift + sigma * (1.0 - from_shift))
numerator = to_shift * (1.0 + (from_shift - 1.0) * base_sigma) ** 2
denominator = from_shift * (1.0 + (to_shift - 1.0) * base_sigma) ** 2
return numerator / denominator
def balanced_microstep_counts(video_steps: int, audio_steps: int) -> tuple[int, ...]:
if video_steps < 1:
raise ValueError("video_steps must be at least 1")
if audio_steps < video_steps:
raise ValueError(
"audio_steps must be greater than or equal to video_steps; "
"the EXP sampler needs at least one joint H3 call per video macro step"
)
return tuple(
((macro + 1) * audio_steps) // video_steps
- (macro * audio_steps) // video_steps
for macro in range(video_steps)
)
def multirate_flow_schedule(
video_steps: int,
audio_steps: int,
shift_video: float,
) -> tuple[torch.Tensor, tuple[int, ...], tuple[int, ...]]:
counts = balanced_microstep_counts(video_steps, audio_steps)
base_sigmas = [1.0]
macro_indices = [0]
for macro, microsteps in enumerate(counts):
macro_start = 1.0 - macro / video_steps
macro_end = 1.0 - (macro + 1) / video_steps
for micro in range(1, microsteps + 1):
fraction = micro / microsteps
base_sigmas.append(macro_start + (macro_end - macro_start) * fraction)
macro_indices.append(len(base_sigmas) - 1)
sigmas = shift_sigma(torch.tensor(base_sigmas, dtype=torch.float32), shift_video)
return sigmas, tuple(macro_indices), counts
def _audio_step_scale(
sigma_video,
sigma_audio,
slope_audio,
denoise_mask,
audio_velocity_is_raw: bool,
):
flat_scale = -sigma_video
dual_scale = -sigma_audio if audio_velocity_is_raw else -sigma_audio / slope_audio
if denoise_mask is None:
return dual_scale
return flat_scale + denoise_mask * (dual_scale - flat_scale)
def sample_minimax_h3_multirate_exp_euler(
model,
x,
sigmas,
extra_args=None,
callback=None,
disable=None,
*,
video_values: int,
packed_values: int,
shift_video: float,
shift_audio: float,
macro_indices: tuple[int, ...],
audio_velocity_is_raw: bool = False,
):
extra_args = {} if extra_args is None else extra_args
if x.shape[-1] != packed_values:
raise ValueError(
"MiniMax H3 packed latent changed after EXP sampler setup: "
f"expected {packed_values} values, got {x.shape[-1]}"
)
if len(macro_indices) < 2 or macro_indices[0] != 0:
raise ValueError("Invalid MiniMax H3 EXP macro-step boundaries")
if macro_indices[-1] != len(sigmas) - 1:
raise ValueError(
"MiniMax H3 EXP sigma schedule changed after sampler setup: "
f"expected {macro_indices[-1] + 1} values, got {len(sigmas)}"
)
denoise_mask = extra_args.get("denoise_mask")
audio_mask = None
if denoise_mask is not None:
if denoise_mask.shape[-1] != packed_values:
raise ValueError("MiniMax H3 denoise mask does not match the packed AV latent")
audio_mask = denoise_mask[..., video_values:]
video = x[..., :video_values]
audio = x[..., video_values:]
s_in = x.new_ones([x.shape[0]])
progress = iter(comfy.utils.model_trange(len(sigmas) - 1, disable=disable))
for macro in range(len(macro_indices) - 1):
first_micro = macro_indices[macro]
after_last_micro = macro_indices[macro + 1]
sigma_macro_start = sigmas[first_micro]
sigma_macro_end = sigmas[after_last_micro]
video_start = video
video_derivative = None
for expected_step in range(first_micro, after_last_micro):
step = next(progress)
if step != expected_step:
raise RuntimeError("MiniMax H3 EXP progress iterator returned an unexpected step")
sigma_video = sigmas[step]
sigma_video_next = sigmas[step + 1]
if video_derivative is None:
video_eval = video_start
else:
video_eval = video_start + video_derivative * (sigma_video - sigma_macro_start)
x_eval = torch.cat((video_eval, audio), dim=-1)
denoised = model(x_eval, sigma_video * s_in, **extra_args)
derivative = to_d(x_eval, sigma_video, denoised)
if video_derivative is None:
video_derivative = derivative[..., :video_values]
audio_derivative = derivative[..., video_values:]
sigma_audio = time_shift_sigma(sigma_video, shift_video, shift_audio)
sigma_audio_next = time_shift_sigma(sigma_video_next, shift_video, shift_audio)
slope_audio = time_shift_slope(sigma_video, shift_video, shift_audio)
video_delta = sigma_video_next - sigma_video
audio_delta = sigma_audio_next - sigma_audio
if not audio_velocity_is_raw:
# Legacy H3 returned slope-scaled audio velocity; current H3
# returns the raw audio-clock velocity.
audio_delta = audio_delta / slope_audio
if audio_mask is not None:
audio_delta = video_delta + audio_mask * (audio_delta - video_delta)
if callback is not None:
video_endpoint = video_start - video_derivative * sigma_macro_start
audio_endpoint_scale = _audio_step_scale(
sigma_video,
sigma_audio,
slope_audio,
audio_mask,
audio_velocity_is_raw,
)
audio_endpoint = audio + audio_derivative * audio_endpoint_scale
callback({
"x": x_eval,
"i": step,
"sigma": sigma_video,
"sigma_hat": sigma_video,
"denoised": torch.cat((video_endpoint, audio_endpoint), dim=-1),
})
audio = audio + audio_derivative * audio_delta
video = video_start + video_derivative * (sigma_macro_end - sigma_macro_start)
return torch.cat((video, audio), dim=-1)
def setup_multirate_exp_sampling(
model,
av_latent: dict,
video_steps: int,
audio_steps: int,
shift_video: float,
shift_audio: float,
):
video, audio = nested_av_parts(av_latent)
if video.shape[1] != 24 or audio.shape[1] != 32 or audio.shape[2] != 2:
raise ValueError(
"Unexpected MiniMax H3 AV channels: "
f"video={tuple(video.shape)}, audio={tuple(audio.shape)}"
)
sigmas, macro_indices, _counts = multirate_flow_schedule(
video_steps, audio_steps, shift_video
)
patched_model = model.clone()
audio_velocity_is_raw = model_uses_raw_audio_velocity(model)
original_sampling = model.get_model_object("model_sampling")
model_sampling = MiniMaxH3MultiRateFlowSamplingEXP(model.model.model_config)
model_sampling.set_parameters(shift=shift_video)
if hasattr(original_sampling, "noise_scale"):
model_sampling.set_noise_scale(original_sampling.noise_scale)
patched_model.add_object_patch("model_sampling", model_sampling)
transformer_options = patched_model.model_options.get("transformer_options", {}).copy()
transformer_options["minimax_h3_sigma_shift_video"] = shift_video
transformer_options["minimax_h3_sigma_shift_audio"] = shift_audio
patched_model.model_options["transformer_options"] = transformer_options
video_values = math.prod(video.shape[1:])
packed_values = video_values + math.prod(audio.shape[1:])
def sampler_function(model_wrap, x, schedule, extra_args=None, callback=None, disable=None):
return sample_minimax_h3_multirate_exp_euler(
model_wrap,
x,
schedule,
extra_args=extra_args,
callback=callback,
disable=disable,
video_values=video_values,
packed_values=packed_values,
shift_video=shift_video,
shift_audio=shift_audio,
macro_indices=macro_indices,
audio_velocity_is_raw=audio_velocity_is_raw,
)
sampler_function.__name__ = "sample_minimax_h3_multirate_exp_euler"
sampler = comfy.samplers.KSAMPLER(sampler_function)
return patched_model, sampler, sigmas