wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-24 05:10:18 +08:00
Code Issues Actions 14 Packages Projects Releases Wiki Activity

fix: remove normalization of audio in LTX Mel spectrogram creation (#11990)

Browse Source 
For LTX Audio VAE, remove normalization of audio during MEL spectrogram creation.
This aligs inference with training and prevents loud audio from being attenuated.
This commit is contained in:
Ivan Zorin 2026-01-21 01:44:28 +02:00 committed by GitHub
parent ddc541ffda
commit 965d0ed509
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
comfy/ldm/lightricks/vae/audio_vae.py
View File

@ -103,20 +103,10 @@ class AudioPreprocessor:
return waveform
return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
@staticmethod
def normalize_amplitude(
waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5
) -> torch.Tensor:
waveform = waveform - waveform.mean(dim=2, keepdim=True)
peak = torch.max(torch.abs(waveform)) + eps
scale = peak.clamp(max=max_amplitude) / peak
return waveform * scale
def waveform_to_mel(
self, waveform: torch.Tensor, waveform_sample_rate: int, device
) -> torch.Tensor:
waveform = self.resample(waveform, waveform_sample_rate)
waveform = self.normalize_amplitude(waveform)
mel_transform = torchaudio.transforms.MelSpectrogram(
sample_rate=self.target_sample_rate,