Comfy-Org/ComfyUI · error · ValueError

Input audio must have {expected_channels} channels, got {wav

Error message

Input audio must have {expected_channels} channels, got {waveform.shape[1]}

What it means

The audio VAE encoder has a fixed channel count (self.autoencoder.encoder.in_channels, typically 2 for stereo). A mono waveform is auto-expanded, but anything with a different channel count (3+, or 0) is rejected because the mel preprocessor would produce a malformed spectrogram.

Source

Thrown at comfy/ldm/lightricks/vae/audio_vae.py:148

        self.preprocessor = AudioPreprocessor(
            target_sample_rate=autoencoder_config["sampling_rate"],
            mel_bins=autoencoder_config["mel_bins"],
            mel_hop_length=autoencoder_config["mel_hop_length"],
            n_fft=autoencoder_config["n_fft"],
        )

    def encode(self, audio, sample_rate=44100) -> torch.Tensor:
        """Encode a waveform dictionary into normalized latent tensors."""

        waveform = audio
        waveform_sample_rate = sample_rate
        input_device = waveform.device
        expected_channels = self.autoencoder.encoder.in_channels
        if waveform.shape[1] != expected_channels:
            if waveform.shape[1] == 1:
                waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:])
            else:
                raise ValueError(
                    f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
                )

        mel_spec = self.preprocessor.waveform_to_mel(
            waveform, waveform_sample_rate, device=waveform.device
        )

        latents = self.autoencoder.encode(mel_spec)
        posterior = DiagonalGaussianDistribution(latents)
        latent_mode = posterior.mode()

        normalized = self.normalizer.normalize(latent_mode)
        return normalized.to(input_device)

    def decode(self, latents: torch.Tensor) -> torch.Tensor:
        """Decode normalized latent tensors into an audio waveform."""
        original_shape = latents.shape

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Convert to stereo first: waveform.convert(2) via torchaudio or torch.mean/tile to 2 channels
2. Mono is fine as-is (auto-expanded); only other counts fail
3. Check the tensor is (B, C, T) — transpose if you built (B, T, C)

Example fix

# before
latents = audio_vae.encode(waveform_6ch)  # 6 channels -> ValueError
# after
waveform_2ch = waveform_6ch[:, :2]  # or downmix
latents = audio_vae.encode(waveform_2ch)

C = waveform.shape[1]
expected = audio_vae.autoencoder.encoder.in_channels
assert C == 1 or C == expected, (C, expected)

def audio_channels_ok(waveform: torch.Tensor, expected: int) -> bool:
    return waveform.dim() >= 2 and waveform.shape[1] in (1, expected)

When it happens

Trigger: vae.encode(waveform) where waveform is (B, C, T) with C not in {1, expected_channels}; e.g. passing 5.1 audio (6ch), or a tensor laid out (B, T, C) so the channel axis is misread.

Common situations: Loading multichannel files, wrong tensor layout from a custom loader, or batching audio with a channels-first vs channels-last mismatch.

Related errors

• Unsupported spatial_scale {scale}. Choose from {list(mapping
• Unsupported audio_channels: {audio_channels}
• Vocoder is missing upsample_factor; cannot infer output samp
• {cls.__name__} does not have a NONE member to map None to
• Invalid {cls.__name__} string: '{value}'. Valid values are:

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/6a38194981bc018c. Report an issue: GitHub.