Load MiniMax H3 Latent Upscaler

Upscaling video is the worst kind of upscaling: every frame has to agree with its neighbors, so you can't just sharpen and move on. MiniMax H3 sidesteps the whole "more pixels over time" mess in a clever way - by upscaling in latent space, before anything is ever decoded to pixels. The model that does that is a small, attention-free 3D learned upscaler, and this node is how you load it.

The trick is that this is a real, separate model - not a script, not a resize node. You put its weights in models/latent_upscale_models/ (the author points at LBH-123-AI/Minimax_h3_latent_Upscaler on HuggingFace), and this loader hands it over with ComfyUI's normal VRAM offload lifecycle, so it sits on disk until it's needed instead of squatting on your card.

How it works

The upscaler is "attention-free" - a 3D convolutional stack that enlarges the video latent spatially without the transformer overhead of the main model, which is what makes it cheap enough to bother with on a Turing card. The loader itself is a thin node: it enumerates files in models/latent_upscale_models/, loads your pick with ComfyUI's managed offload, and hands you a custom upscale_model object that only the pack's MiniMax H3 Latent Upscale node knows how to consume.

The inputs that matter

• model_name - the upscaler file from models/latent_upscale_models/. If you downloaded the HuggingFace weights and they don't show up, they're in the wrong folder.
• precision - auto (default), fp16, bf16, or fp32. Auto picks the efficient native type for your device, and the tooltip says it plainly: FP16 is normally the preferred choice on Turing.

Output: a single upscale_model socket - wire it into MiniMax H3 Latent Upscale and nowhere else; it's a custom type.

How to install

Pack install first, then the weights:

cd ComfyUI/custom_nodes
git clone https://github.com/wjie98/comfyui-svdint4
cd comfyui-svdint4
python -m pip install -v --no-build-isolation -e ./kernel

Then grab Minimax_h3_latent_Upscaler from HuggingFace and drop the weights into ComfyUI/models/latent_upscale_models/. Restart ComfyUI so the folder is scanned.

Where people get burned

Two things, both predictable. The empty model_name dropdown is almost always a weights-folder problem, not a pack problem - check models/latent_upscale_models/, not models/upscale_models/, which is where image upscaler packs go. And second: don't confuse this with the pack's VAE decode path. The upscaler and the VAE are different stages - this node feeds the latent upscale, then the H3 VAE decodes the result. Miss that ordering and you'll be upscaling pixels, which defeats the entire point.