LTX2_SM_AUDIO_VAE

LTX-2.3 is the only open model that makes the picture and the sound in one generation pass, and this node is how that second half gets in and out of latent space. LTX2_SM_AUDIO_VAE loads the audio VAE and, like the video VAE node, hands you two objects from one file: an a_decoder that turns audio latents into a waveform, and an a_encoder that turns a real audio file into the latent conditioning the model needs for audio-to-video (A2V).

What it feeds

• a_decoder → LTX2_DECO_AUDIO. After the KSampler outputs audio_latents, this decodes them to an AUDIO dict you can save or preview.
• a_encoder → LTX2_LATENTS. Feed it an audio file there and you're doing A2V - the model generates video that follows the audio, which is the trick behind the "make your character sing what you typed" workflows and the genuinely good native lip-sync LTX is known for.

The dropdown reads ComfyUI/models/vae/ and expects ltx-2.3-22b-distilled_audio_vae.safetensors, from the vae folder of unsloth/LTX-2.3-GGUF (or the pack author's mirror). It's the sibling of the video VAE file - same naming pattern, don't grab the wrong one.

Gotchas worth knowing

• It's not the video VAE. There's no type confusion at the socket level (both are VAE in this pack, so the wire is compatible) but they're different files and different models. If your generated video has no audio and you expected some, check that LTX2_SM_AUDIO_VAE is actually in the graph - the video path runs happily without it, you just get silence.
• A2V vs video-to-audio are different sampling modes. Feeding audio into LTX2_LATENTS conditions the generation; the sampling_mode you picked on LTX2_SM_Model (audio2v vs two_stages etc.) decides how that conditioning is used. If audio input is being ignored, the mode is usually the culprit.
• Expect a reload tax. Like every node in this pack, executing clears Comfy's model cache. On small cards that's the point - but it means each run re-streams the VAE.

Installing

Same as 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

That installs diffusers, accelerate, transformers, tokenizers, gguf, and omegaconf. Restart, drop the VAE file in place, and the dropdown fills in. The audio comes out via LTX2_DECO_AUDIO, then a save/preview node of your choice - this node itself produces nothing you can hear.