Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

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

Ivan Zorin965d0ed509ce
diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py
index 29d9e6c29..55a074661 100644
--- a/comfy/ldm/lightricks/vae/audio_vae.py
+++ b/comfy/ldm/lightricks/vae/audio_vae.py
@@ -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,