IAMCCS Audio Normalize For LTX

The one-trick node that prevents a whole class of LTX audio errors

LTX-2 is an audio-synced video model, which means at some point your AUDIO tensor has to survive the trip into LTXVAudioVAEEncode. The problem: audio loaders in the wild return waveforms in several different shapes - [C, S], [S, C], [B, S, C], sometimes plain 1D - and LTX's audio VAE expects a specific layout ([B, C, S], with channels before samples). Feed it the wrong arrangement and you get transposed garbage or an outright error, and the failure usually isn't obvious because it's about ordering, not values.

This node is the normalization shim. It takes any AUDIO and rewrites the waveform into the layout LTX expects, then hands it back. That's the whole job, and it's a genuinely useful one.

What it does, precisely

One required input: audio. One optional: debug_label (default ltx_audio), which names the node's report. The normalization logic handles the real-world mess:

• 1D waveform → padded to [1, 1, S].
• [C, S] or [S, C] → transposed into [B, C, S] (it detects which orientation by channel count - anything with more than 8 leading dimensions is treated as channels).
• [B, S, C] → transposed to [B, C, S], because the LTX encode path calls movedim(1, -1) and expects channels second.

Outputs: audio (the fixed tensor) and report, which confirms the final shape and sample rate - e.g. ltx_audio: shape=(1, 2, 44100) sample_rate=44100. That report is the diagnostic you want when audio VAE errors strike: if it shows a wrong shape, the problem is upstream of this node.

Using it

Drop it right before the LTX audio conditioning/encode stage in a V2V or audio-driven workflow, and feed the output to whatever consumes AUDIO. It's listed in the pack under IAMCCS/V2V and appears in the Cine V2V pipeline, but you can use it anywhere LTX audio encoding happens. If you're debugging "LTXVAudioVAEEncode failed" and you don't have this node in the path, add it - a large fraction of those errors are exactly the layout problem it fixes.

Install is the pack install: ComfyUI Manager (search "IAMCCS") or git clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes, restart. No dependencies beyond torch, which you obviously have. The trap is the opposite of most nodes: it's so simple people skip it and then blame the model. It won't fix clipping, loudness, or a bad mix - that's what the AudioBoard tools are for - but it will make the audio reach the VAE in the shape it demands.