LTXV2 Audio VAE Loader (MultiGPU)

Most LTX-2 OOM stories are about the text encoder, but the audio half of the model has a claim on your VRAM too. LTX-2's whole trick is synchronized audio and video, and the audio runs through its own VAE, loaded out of the same checkpoint. This node is that audio-VAE loader from the ComfyUI-LTX2-MultiGPU pack, with the pack's signature addition: a device dropdown, so you decide which GPU (or the CPU) holds it instead of letting ComfyUI stack everything on card zero.

It's the least glamorous node in the pack, and it's honest about it - the code is a thin adapter around ComfyUI's built-in LTXVAudioVAELoader. What it adds is placement control, which is the whole point of the pack by nathandreamfast (built after OOMing a 4090+5090 box on LTX-2 within days of the January 2026 release). If you've ever watched one GPU fill to the brim while the other idles, you know the pitch already.

How it works

One required input, one dropdown. ckpt_name points at your LTX-2 checkpoint - the audio VAE doesn't ship as its own file, it lives inside the checkpoint alongside the video model and the video VAE. The node pulls it out with ComfyUI's core LTXVAudioVAELoader, and the device dropdown decides where the loaded weights land. Same mechanism as every node in this pack: during the load, ComfyUI's model management is told to use your selected device, then the setting reverts.

If you have ComfyUI-MultiGPU installed, the pack detects it and uses that pack's device management rather than patching comfy.model_management itself - the startup log will say which mode you're in. The dropdown enumerates what torch can actually see: cpu plus every cuda:N on the box.

The inputs and outputs that matter

• ckpt_name - required; your LTX-2 checkpoint (the one with the audio VAE inside).
• device - optional; the dropdown. cpu / cuda:0 / cuda:1 ...

Output: audio_vae (VAE) → into the audio side of the pipeline, ComfyUI's LTXV audio encode node, before the audio latent gets combined with the video latent for the synchronized generation.

Installing it

Pack-level install, once for all four nodes: ComfyUI Manager → search "ComfyUI-LTX2-MultiGPU" → install → restart. Or:

cd ComfyUI/custom_nodes/
git clone https://github.com/dreamfast/ComfyUI-LTX2-MultiGPU

Restart ComfyUI. No requirements.txt - it uses only ComfyUI's bundled torch and model management. Real prerequisites: a ComfyUI with native LTX-2 support (the core node this wraps only exists there) and the checkpoint downloaded.

Common issues & troubleshooting

• "LTXVAudioVAELoader not found in ComfyUI core nodes" - same story as the rest of the pack: update ComfyUI. This node is glue over a core node that only exists in recent builds.
• Audio encode errors later in the graph - the author hit one himself, and it cleared up with the Video Combine node plus a current ffmpeg. If the preview shows nothing, right-click Video Combine → Open Preview; that preview is buggy on some builds.
• Distilled-checkpoint audio weirdness - the first LTX-2 distilled build shipped with a bad VAE, and Lightricks replaced it mid-January. If the audio VAE loads but output sounds off, refresh the checkpoint.

If you're already happy with ComfyUI's stock LTXVAudioVAELoader, this node adds nothing except placement control. That's the honest take: reach for it when your second GPU is sitting idle or the audio VAE is crowding the video model off your main card. For a one-input loader, that's a reasonable job.