LTXV2 Checkpoint Loader (MultiGPU)

LTX-2, Lightricks' 19B audio-video model, isn't one file - it's three things riding inside one checkpoint: the video diffusion model, the video VAE, and a separate audio VAE. Stock ComfyUI gives you separate loaders for each and dumps them all on your main GPU, which is exactly how people running a 4090 and a 5090 still managed to run out of VRAM. This node is the fix: a combined loader that reads all three pieces out of a single .safetensors and pins the whole bundle to whichever device you pick in the dropdown.

It's the flagship node of the ComfyUI-LTX2-MultiGPU pack by nathandreamfast. He hit OOM errors on a 4090+5090 box within days of LTX-2's open-weights release in January 2026, made a PR to pollockjj's ComfyUI-MultiGPU, and shipped this pack in the meantime so people didn't have to wait. The pack's whole thesis: on a multi-GPU machine, ComfyUI's default is "everything on card zero." These nodes let you say otherwise.

How it works

It's an adapter, not a new loader. When you run it, the node calls ComfyUI's built-in CheckpointLoaderSimple to pull out the model and video VAE, then calls the core LTXVAudioVAELoader on the same checkpoint to grab the audio VAE - three objects, one file read.

The "MultiGPU" part is a device dropdown injected into every loader in the pack. During the load, ComfyUI's model management is told to hand back your selected device instead of its default, everything loads there, and the setting reverts when the node is done. The dropdown lists whatever torch can actually see - cpu plus each cuda:N on the box (and XPU/MPS/DirectML if you're on one of those builds), defaulting to the first GPU when one exists.

If you also have ComfyUI-MultiGPU installed, the pack detects it at startup and borrows its device management so the two don't double-patch. Otherwise it patches comfy.model_management itself. Either way, the startup log tells you which mode you're in: [LTXV2 MultiGPU] ComfyUI-MultiGPU detected, using its device management or [LTXV2 MultiGPU] Running in standalone mode.

The inputs and outputs that matter

• ckpt_name - the LTX-2 checkpoint (e.g. ltx-2-19b-dev-fp8.safetensors) from your checkpoints folder. This one file has to contain all three components.
• device - the dropdown. cpu / cuda:0 / cuda:1 ...

Three outputs:

• model (MODEL) → into the sampler.
• vae (VAE) → the video VAE, into VAE Decode.
• audio_vae (VAE) → into the audio encode path, before the audio latent gets combined with the video latent for synchronized generation.

Installing it

ComfyUI Manager → search "ComfyUI-LTX2-MultiGPU" → install → restart. Or the manual route:

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

Restart ComfyUI. There's no requirements.txt - the pack only uses what ComfyUI already ships (torch and comfy.model_management), so nothing extra to pip install. What you do need: a recent-enough ComfyUI with native LTX-2 support, PyTorch built with CUDA for multi-GPU to mean anything, and the model file itself (~25GB fp8 dev, ~20GB fp4, or the distilled version plus its LoRA). The bundled workflow in workflows/multi-gpu-ltx2-t2v.json shows the intended setup and even throws in an ollama prompt enhancer, since LTX rewards the long, detailed prompts the community swears by.

Common issues & troubleshooting

• "LTXVAudioVAELoader not found in ComfyUI core nodes" - your ComfyUI predates native LTX-2 support (merged early January 2026). Update ComfyUI first; this node is glue over core nodes that only exist in recent builds.
• Audio that sounds wrong or garbled - the first distilled LTX-2 checkpoint shipped with a broken VAE; Lightricks pushed a corrected file mid-January. If you grabbed an early build, re-download the checkpoint.
• Audio encoding errors further down the graph - the author hit one himself, and it cleared up with the Video Combine node plus a current ffmpeg. And if the preview comes up blank, right-click Video Combine → Open Preview; that preview is buggy on some builds.

The real win is the division of labor: text encoder on CPU (that's the pack's other headline node), video model and video VAE on cuda:1, decode on cuda:0. The author's notes say that exact arrangement produced 10-second 1080p clips where a single-GPU setup OOMed. Multi-GPU is genuinely fiddly to configure - this is the least fiddly part of it.