From 965d0ed509ce46a3328c342aee23a234ba6e4f88 Mon Sep 17 00:00:00 2001 From: Ivan Zorin Date: Wed, 21 Jan 2026 01:44:28 +0200 Subject: [PATCH] fix: remove normalization of audio in LTX Mel spectrogram creation (#11990) For LTX Audio VAE, remove normalization of audio during MEL spectrogram creation. This aligs inference with training and prevents loud audio from being attenuated. --- comfy/ldm/lightricks/vae/audio_vae.py | 10 ---------- 1 file changed, 10 deletions(-) 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,