Latent Upscale Model Loader (MultiGPU)

LTX-2 does a neat trick with resolution: it doesn't render at the size you ask for. It renders internally at half resolution and then upscales - so your "1080p" is really a ~540p generation with an upscale pass on top. That upscale runs in latent space using a separate upscale model, and on a VRAM-strained multi-GPU box it's one more thing ComfyUI wants to load onto card zero. This node loads the latent upscale model and lets you point it at any GPU (or the CPU) instead.

It's the fourth and quietest node in the ComfyUI-LTX2-MultiGPU pack, nathandreamfast's fix for OOMing LTX-2 on a 4090+5090 rig. Like the pack's other nodes, it's a thin adapter - the actual loading is ComfyUI's stock LatentUpscaleModelLoader - with a device dropdown bolted on. If your workflow never touches the upscale pass, you don't need this node at all. If it does, and one of your GPUs is sitting there doing nothing, this is the node that fixes that.

How it works

model_name picks the upscale model from your models/latent_upscale_models folder - a separate download from the checkpoint, part of the LTX-2 release's upscaler files. The node loads it, and the device dropdown decides where the model lives while it does the 2x latent upscale. Same pack-wide mechanism as everything else: the selected device is set for the duration of the load, then ComfyUI's defaults resume.

The dropdown lists every device torch can see - cpu, cuda:0, cuda:1, and so on - and defaults to your first GPU when one exists. On a two-card box that means you can run the upscale on the second GPU while the sampler works the first, or drop it to CPU entirely.

The inputs and outputs that matter

• model_name - required; the latent upscale model file.
• device - optional; the dropdown.

Output: LATENT_UPSCALE_MODEL → into the latent upscale step in your LTX-2 workflow (the node that consumes a latent upscale model), before the final VAE decode.

Installing it

Same one-time pack install as the other three: ComfyUI Manager → search "ComfyUI-LTX2-MultiGPU" → install → restart. Or:

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

Restart. No requirements.txt; ComfyUI's bundled deps suffice. And remember the model file itself - it goes in ComfyUI/models/latent_upscale_models/ and the pack doesn't download it for you.

Common issues & troubleshooting

• Empty model_name dropdown - the upscale model isn't in the latent_upscale_models folder yet. Grab it from the LTX-2 release files and drop it there.
• "LatentUpscaleModelLoader not found in ComfyUI core nodes" - outdated ComfyUI. Update, since this pack wraps core nodes that only exist in recent builds.

That's the whole node, honestly. It's a loader with a device picker - nothing clever, nothing to tune. But in the author's preferred layout (text encoder on CPU, video model and VAE on the cards), having the upscale pass not fight the sampler for VRAM is one less thing to babysit on a generation that takes minutes. Sometimes the boring node is the one that saves you.