wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-04-15 13:02:35 +08:00
Code Issues Actions 17 Packages Projects Releases Wiki Activity

Fix freenoise application for LTXAV context windows, fix audio mapping to context windows

Browse Source
This commit is contained in:
ozbayb 2026-03-27 11:24:40 -06:00
parent 115dbb69d1
commit ef61ddfaed
2 changed files with 32 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
comfy/context_windows.py
View File

@ -619,7 +619,28 @@ def _sampler_sample_wrapper(executor, guider, sigmas, extra_args, callback, nois
raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.") raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.")
if not handler.freenoise: if not handler.freenoise:
return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)
noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"])
# For packed multimodal tensors (e.g. LTXAV), noise is [B, 1, flat] and FreeNoise
# must only shuffle the video portion. Unpack, apply to video, repack.
latent_shapes = None
try:
latent_shapes = guider.conds['positive'][0]['model_conds']['latent_shapes'].cond
except (KeyError, IndexError, AttributeError):
pass
if latent_shapes is not None and len(latent_shapes) > 1:
modalities = comfy.utils.unpack_latents(noise, latent_shapes)
video_total = latent_shapes[0][handler.dim]
modalities[0] = apply_freenoise(modalities[0], handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"])
for i in range(1, len(modalities)):
mod_total = latent_shapes[i][handler.dim]
ratio = mod_total / video_total if video_total > 0 else 1
mod_ctx_len = max(round(handler.context_length * ratio), 1)
mod_ctx_overlap = max(round(handler.context_overlap * ratio), 0)
modalities[i] = apply_freenoise(modalities[i], handler.dim, mod_ctx_len, mod_ctx_overlap, extra_args["seed"])
noise, _ = comfy.utils.pack_latents(modalities)
else:
noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"])
return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)

19
comfy/model_base.py
View File

@ -1197,17 +1197,18 @@ class LTXAV(BaseModel):
return result return result
video_total = latent_shapes[0][dim] video_total = latent_shapes[0][dim]
audio_total = latent_shapes[1][dim] video_window_len = len(primary_indices)
# Proportional mapping — video and audio cover same real-time duration for i in range(1, len(latent_shapes)):
v_start, v_end = min(primary_indices), max(primary_indices) + 1 mod_total = latent_shapes[i][dim]
a_start = round(v_start * audio_total / video_total) # Length proportional to video window frame count (not index span)
a_end = round(v_end * audio_total / video_total) mod_window_len = max(round(video_window_len * mod_total / video_total), 1)
audio_indices = list(range(a_start, min(a_end, audio_total))) # Anchor to end of video range
if not audio_indices: v_end = max(primary_indices) + 1
audio_indices = [min(a_start, audio_total - 1)] mod_end = min(round(v_end * mod_total / video_total), mod_total)
mod_start = max(mod_end - mod_window_len, 0)
result.append(list(range(mod_start, min(mod_start + mod_window_len, mod_total))))
result.append(audio_indices)
return result return result
def get_guide_frame_count(self, x, conds): def get_guide_frame_count(self, x, conds):