wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-07-18 20:38:15 +08:00
Code Issues Actions 27 Packages Projects Releases Wiki Activity

Fix StableAudio extra_conds using wrong temporal downscale ratio

Browse Source 
Replace hardcoded magic numbers (21.53 for SA1, 21.5332 for SA3) with
self.latent_format.temporal_downscale_ratio (2048 for SA1, 4096 for SA3).
Also switch from int() truncation to round() to avoid off-by-one errors.

This fixes SA3 generating truncated audio because extra_conds divided by
SA1's rate after fix_empty_latent_channels had already resized the latent
to SA3's native rate.

Refs: #14825
Signed-off-by: Yuchen Fan <functionhx@gmail.com>
This commit is contained in:
Functionhx 2026-07-10 02:21:29 +08:00 committed by Yuchen Fan
parent 04a30fb375
commit 0975da3d75
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/model_base.py
View File

@ -799,7 +799,8 @@ class StableAudio1(BaseModel):
device = kwargs["device"]
seconds_start = kwargs.get("seconds_start", 0)
seconds_total = kwargs.get("seconds_total", int(noise.shape[-1] / 21.53))
ratio = self.latent_format.temporal_downscale_ratio
seconds_total = kwargs.get("seconds_total", round(noise.shape[-1] * ratio / 44100))
seconds_start_embed = self.seconds_start_embedder([seconds_start])[0].to(device)
seconds_total_embed = self.seconds_total_embedder([seconds_total])[0].to(device)
@ -866,7 +867,8 @@ class StableAudio3(BaseModel):
noise = kwargs.get("noise", None)
device = kwargs["device"]
seconds_total = kwargs.get("seconds_total", int(noise.shape[-1] / 10.7666))
ratio = self.latent_format.temporal_downscale_ratio
seconds_total = kwargs.get("seconds_total", round(noise.shape[-1] * ratio / 44100))
seconds_total_embed = self.seconds_total_embedder([seconds_total])[0].to(device)
global_embed = seconds_total_embed.reshape((1, -1))