Load MiniMax H3 Learned Latent Upscaler

Video upscaling normally means decoding to pixels and running an image upscaler frame by frame - slow, memory-hungry, and prone to temporal flicker. MiniMax H3 sidesteps that with a learned latent upscaler: a model trained to upscale in latent space, before you ever decode. VRGDG MiniMax H3 Latent Upscale Model Loader is the boring-but-necessary front half of that pipeline - it loads the upscale model into memory so the actual upscale node can use it.

How it works

This is a loader node, so it's thin by design. You pick a model_name, choose device (cuda default, cpu fallback) and precision (bf16 default, with fp16 and fp32 options), and it hands you a upscale_model of the pack's custom type VRGDG_MINIMAX_H3_LATENT_UPSCALE_MODEL. That output only plugs into the VRGDG MiniMax H3 Learned Latent Upscale node, which is the half that actually does the work.

The device/precision choices are the real dials here. bf16 is the right default on any modern NVIDIA card - it's the precision these models are typically trained in, and it saves the VRAM the joint AV latents are already eating. Drop to fp16 if you hit a precision mismatch on older kernels, and fp32 is the "it's either this or it's broken" escape hatch. cpu is there so the upscale pass can offload when VRAM is the wall, but expect it to be slow.

One detail that catches people: the model_name enum is populated from your registered latent_upscale_models folders. If no models are found, the dropdown literally reads "[no models found in registered latent_upscale_models folders]" - that's the node telling you the file isn't where it looks. The H3 upscaler is a real model download, not something the pack bundles, so this is the step where a missing download surfaces.

Install

Pack-wide: ComfyUI Manager → search vrgamedev (or add https://github.com/vrgamegirl19/comfyui-vrgamedevgirl), restart, hard-refresh. Manual: git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl into custom_nodes, then pip install -r requirements.txt; on Windows portable, install Cython and scikit-build-core first.

Troubleshooting

If the dropdown is empty, you haven't got the model file in a registered latent_upscale_models folder - put it where the loader scans and hard-refresh the node list. If generation runs but the upscale output looks wrong or errors on precision, switch precision down a notch or up to fp32 and re-run before you blame the workflow. And if you're on a tight card, run this at bf16 on cuda and consider letting the upscale pass itself land on CPU - the loader and the work node are separate, so the offload is easy to arrange.