LTX2_DECO_AUDIO

This is the smallest node in the pack, and the one people forget - until they generate a clip with no sound and realize they never hooked it up. LTX2_DECO_AUDIO decodes the audio_latents output of LTX2_SM_KSampler into an actual AUDIO dict using the audio VAE's decoder. Two inputs, one output, done. But it's the difference between a silent animation and the thing LTX-2.3 exists for.

What it is

Every LTX-2.3 generation produces two latent streams: the video latent and the audio latent, both out of the same pass. The video decode goes through LTX2_DECO_VIDEO. The audio half comes here. Feed it:

• a_decoder - from LTX2_SM_AUDIO_VAE (the audio VAE's decoder, not the video VAE's).
• audio_latents - the second output of the KSampler.

It returns an audio object with a waveform and a sample rate, in the standard ComfyUI AUDIO format. From there it plugs into any audio-capable save or preview node in your setup - VHS save audio, a waveform preview, whatever you normally use. That's the entire node. It decodes audio_latents["samples"] through the AudioDecoder and hands you the result.

The mistakes people actually make

• Feeding it the video latent. The KSampler's latent output is the video stream; this node wants audio_latents. Wrong input, garbage (or a type error). The audio decode only makes sense with audio latents.
• Using the video VAE's decoder. a_decoder must come from LTX2_SM_AUDIO_VAE, which loads the separate ltx-2.3-22b-distilled_audio_vae.safetensors file. Cross-wiring the two VAEs is the classic silent-failure setup.
• Forgetting the node exists. The video path runs happily without the audio decoder wired - you just get a clip with no sound. If your generated singing character is mysteriously mute, this is the first thing to check, not the sampler.

Install

Identical to the rest of the pack - ComfyUI Manager (search "ComfyUI_LTX2_SM") or:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_LTX2_SM.git
cd ComfyUI_LTX2_SM && pip install -r requirements.txt

Restart, make sure the audio VAE is in ComfyUI/models/vae/ and the dropdown on LTX2_SM_AUDIO_VAE picks it. Like every node here, it runs on the pack's vendored LTX-2.3 pipeline (diffusers/transformers stack plus gguf and omegaconf), and it clears Comfy's model cache on execute - a reload tax per run, but the audio decode itself is fast once everything's loaded.