Load LTXV Audio VAE

You came here because you want a video that talks, and this is the node that turns LTX-2's audio latents into actual sound. LTX-2 is Lightricks' audio-plus-video model - the 19B one from January 2026 that finally made synchronized sound a local, open-weights thing - and the Audio VAE is the codec that sits at the audio end of the pipeline, exactly like a normal VAE sits at the image end. Encode sound into latent space, let the model generate alongside it, decode back into a waveform.

It's a deliberately boring node and that's fine. One input, one output, nothing to tune. But it's easy to misread the whole audio side of an LTX-2 workflow if you don't know what you're looking at, so let's map it.

How it works

The ckpt_name dropdown lists files in ComfyUI/models/checkpoints, and here's the part that confuses people: you point it at the LTX-2 checkpoint itself, not a separate audio VAE file. The full LTX-2 model file carries an audio_vae section and a vocoder inside it. This node loads that checkpoint, pulls out those weights, renames the prefix into what ComfyUI's VAE machinery expects, and hands you a ready-to-use VAE object.

You then feed the VAE output into LTXV Audio VAE Encode or Decode, alongside nodes like LTXV Empty Latent Audio (to create a blank audio latent) and LTXV Concat AV Latent (to bind audio and video latents together so the model generates them as one clip). Load the Audio VAE, load the text encoder, load the model, and the audio side of your graph is complete.

The output type is literally VAE - the same type your image VAE uses - which is the second common confusion. It is not a drop-in for the video VAE. LTX-2 has separate VAEs for pixels and for audio, and they don't cross. Your Save Audio / VAE Decode node for audio wants this one; your video decode wants the other.

What trips people up

The audio pipeline is the part of LTX-2 that burns the most new users, and the failure modes are mostly upstream of this node:

• You need the full checkpoint, not a stripped DiT-only build. If the file you selected doesn't contain an audio_vae section, this node fails on load. The quantized GGUFs from the community usually carry it; a bare transformer file does not.
• Audio VAE + text encoder + model must agree. The node description for its sibling Load LTXV Audio Text Encoder is explicit: ltxav: gemma 3 12B. Mixing the audio encoder from a different LTX-2 build with this VAE gets you mumbling or silence.
• VRAM is the real enemy. LTX-2 wants 32GB+ in the official docs (realistic low end is more like 16GB with quantization and offloading), and the Gemma text encoder is the single biggest chunk of that. If you're OOMing before a single frame renders, don't blame the VAE loader - look at your encoder and the --reserve-vram launch flag.
• Audio quality has a reputation for being tinny. That's the model and the vocoder, not this loader, and Lightricks' 2.3 update (rebuilt VAE, new vocoder) is the fix people actually use.

It's a core node, shipped with ComfyUI since LTX-2 got native support in January 2026. The one thing you actually have to get right is picking the checkpoint - and remembering that this tiny node is load-bearing: no Audio VAE, no sound, no matter how good your prompt is.