Files
aigc-home/docs/hel-892/t8-two-reference/t8two__sampling.py
T
2026-11-01 08:02:36 +08:00

317 lines
12 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
DEFAULT_SAMPLER_NAME = "dual_clock_euler"
DEFAULT_SCHEDULER_NAME = "native_flow"
SAMPLER_OPTIONS = [DEFAULT_SAMPLER_NAME]
if hasattr(comfy.model_sampling, "ModelSamplingAV"):
SAMPLER_OPTIONS.extend(
name for name in comfy.samplers.SAMPLER_NAMES if name != DEFAULT_SAMPLER_NAME
)
SCHEDULER_OPTIONS = [
DEFAULT_SCHEDULER_NAME,
*(name for name in comfy.samplers.SCHEDULER_NAMES if name != DEFAULT_SCHEDULER_NAME),
]
class MiniMaxH3FlowSampling(comfy.model_sampling.ModelSamplingDiscreteFlow, comfy.model_sampling.CONST):
@property
def audio_scale(self):
# The custom sampler advances the unpacked audio stream on its own clock.
# New ComfyUI H3 builds require this property, but must not additionally
# carry/scale audio onto the video clock (that would apply the transform twice).
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 native_flow_sigmas(steps: int, shift_video: float) -> torch.Tensor:
base_sigmas = torch.linspace(1.0, 0.0, steps + 1, dtype=torch.float32)
return shift_sigma(base_sigmas, shift_video)
def _scheduler_sigmas(model_sampling, scheduler: str, steps: int, shift_video: float) -> torch.Tensor:
if scheduler == DEFAULT_SCHEDULER_NAME:
return native_flow_sigmas(steps, shift_video)
if scheduler not in comfy.samplers.SCHEDULER_NAMES:
raise ValueError(
f"Unknown scheduler {scheduler!r}; available schedulers: "
f"{', '.join(SCHEDULER_OPTIONS)}"
)
return comfy.samplers.calculate_sigmas(model_sampling, scheduler, steps).cpu()
def _make_sampling(
model,
original_sampling,
shift_video: float,
shift_audio: float,
use_native_av: bool,
):
if use_native_av:
native_av_cls = getattr(comfy.model_sampling, "ModelSamplingAV", None)
if native_av_cls is None or not model_uses_raw_audio_velocity(model):
raise RuntimeError(
"The selected ComfyUI sampler requires native MiniMax H3 FLOW_AV support. "
"Update ComfyUI, or select dual_clock_euler for cross-version compatibility."
)
class MiniMaxH3NativeAVSampling(native_av_cls, comfy.model_sampling.CONST):
pass
model_sampling = MiniMaxH3NativeAVSampling(model.model.model_config)
model_sampling.set_parameters(shift=shift_video, audio_shift=shift_audio)
else:
model_sampling = MiniMaxH3FlowSampling(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)
return model_sampling
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 _rebase_partial_audio_start(
model,
x: torch.Tensor,
sigma_video,
*,
video_values: int,
shift_video: float,
shift_audio: float,
) -> torch.Tensor:
"""Move a partial custom-sampler restart onto the H3 audio clock.
Comfy's generic KSAMPLER initializes every packed stream with ``sigma_video``.
That is exact at a full ``sigma=1`` start, where H3's video and audio clocks
coincide. A learned-upscale refinement starts below one, however, and the
direct/unscaled audio stream must instead start at ``sigma_audio``. Rebuild
only that slice from KSAMPLER's actual initialization terms so noise_scale and
dtype/device details remain authoritative.
Plain callable models used by unit tests or third-party direct invocation do
not expose KSAMPLER's ``noise``/``latent_image`` state; those calls retain the
historical behavior rather than guessing initialization terms.
"""
noise = getattr(model, "noise", None)
latent_image = getattr(model, "latent_image", None)
if not isinstance(noise, torch.Tensor) or not isinstance(latent_image, torch.Tensor):
return x
if noise.shape != x.shape or latent_image.shape != x.shape:
return x
sigma_video = torch.as_tensor(sigma_video, device=x.device, dtype=x.dtype)
sigma_audio = torch.as_tensor(
time_shift_sigma(sigma_video, shift_video, shift_audio),
device=x.device,
dtype=x.dtype,
)
if bool(torch.isclose(sigma_video, sigma_audio, rtol=0.0, atol=1.0e-7).all()):
return x
if bool((sigma_video <= 0.0).any()):
raise ValueError("MiniMax H3 partial audio restart requires a positive initial sigma")
audio_x = x[..., video_values:]
audio_latent = latent_image[..., video_values:].to(device=x.device, dtype=x.dtype)
sigma_shape = sigma_video.shape[:1] + (1,) * (audio_x.ndim - 1)
sigma_video = sigma_video.reshape(sigma_shape)
sigma_audio = sigma_audio.reshape(sigma_shape)
# Recover the exact scaled-noise term used by CONST without assuming
# model_sampling.noise_scale, then rebuild x_audio at its own clock.
scaled_noise = (audio_x - (1.0 - sigma_video) * audio_latent) / sigma_video
rebased_audio = sigma_audio * scaled_noise + (1.0 - sigma_audio) * audio_latent
return torch.cat((x[..., :video_values], rebased_audio), dim=-1)
def sample_minimax_h3_dual_clock_euler(
model,
x,
sigmas,
extra_args=None,
callback=None,
disable=None,
*,
video_values: int,
packed_values: int,
shift_video: float,
shift_audio: float,
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 sampler setup: "
f"expected {packed_values} values, got {x.shape[-1]}"
)
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:]
x = _rebase_partial_audio_start(
model,
x,
sigmas[0],
video_values=video_values,
shift_video=shift_video,
shift_audio=shift_audio,
)
s_in = x.new_ones([x.shape[0]])
for step in comfy.utils.model_trange(len(sigmas) - 1, disable=disable):
sigma_video = sigmas[step]
sigma_video_next = sigmas[step + 1]
denoised = model(x, sigma_video * s_in, **extra_args)
derivative = to_d(x, sigma_video, denoised)
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:
# Older ComfyUI H3 builds returned an audio velocity multiplied by
# d(sigma_audio)/d(sigma_video), so their update needs the inverse.
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:
endpoint_scale = _audio_step_scale(
sigma_video,
sigma_audio,
slope_audio,
audio_mask,
audio_velocity_is_raw,
)
denoised[..., video_values:] = x[..., video_values:] + derivative[..., video_values:] * endpoint_scale
callback({
"x": x,
"i": step,
"sigma": sigma_video,
"sigma_hat": sigma_video,
"denoised": denoised,
})
x = torch.cat((
x[..., :video_values] + derivative[..., :video_values] * video_delta,
x[..., video_values:] + derivative[..., video_values:] * audio_delta,
), dim=-1)
return x
def setup_dual_clock_sampling(
model,
av_latent: dict,
steps: int,
shift_video: float,
shift_audio: float,
sampler_name: str = DEFAULT_SAMPLER_NAME,
scheduler: str = DEFAULT_SCHEDULER_NAME,
):
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)}"
)
if sampler_name not in SAMPLER_OPTIONS:
raise ValueError(
f"Unknown sampler {sampler_name!r}; available samplers: "
f"{', '.join(SAMPLER_OPTIONS)}"
)
use_native_av = sampler_name != DEFAULT_SAMPLER_NAME
patched_model = model.clone()
original_sampling = model.get_model_object("model_sampling")
model_sampling = _make_sampling(
model,
original_sampling,
shift_video,
shift_audio,
use_native_av,
)
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
if use_native_av:
sampler = comfy.samplers.sampler_object(sampler_name)
else:
audio_velocity_is_raw = model_uses_raw_audio_velocity(model)
video_values = math.prod(video.shape[1:])
packed_values = video_values + math.prod(audio.shape[1:])
def sampler_function(model_wrap, x, sigmas, extra_args=None, callback=None, disable=None):
return sample_minimax_h3_dual_clock_euler(
model_wrap,
x,
sigmas,
extra_args=extra_args,
callback=callback,
disable=disable,
video_values=video_values,
packed_values=packed_values,
shift_video=shift_video,
shift_audio=shift_audio,
audio_velocity_is_raw=audio_velocity_is_raw,
)
sampler_function.__name__ = "sample_minimax_h3_dual_clock_euler"
sampler = comfy.samplers.KSAMPLER(sampler_function)
sigmas = _scheduler_sigmas(model_sampling, scheduler, steps, shift_video)
return patched_model, sampler, sigmas