FL MiniMax H3 Load Latent Upscaler

If you load an H3 neural-upscale workflow and wonder why there's a loader node nobody ever touches sitting in front of the interesting one - that's this. FL MiniMax H3 Load Latent Upscaler is the boring-but-necessary half of the FL MiniMax H3 neural upscale pair. It points at a checkpoint in your models/latent_upscale_models folder, picks a precision, and hands the result to either Neural Latent Upscale 2D or 3D over a single upscale_model socket.

What makes it more than a plain file picker is that MiniMax H3 ships two different upscaler architectures, and this node knows which one it just loaded. Open the code and you can see it inspecting the checkpoint's keys - whether it finds a resizer layout with temporal conv blocks or a plain 2D resizer decides which family it belongs to. So the same dropdown serves both the fast spatial-only 2D checkpoint and the heavier 3D spatiotemporal one. You just pick the file; the node figures out what it is, and the downstream 2D or 3D node errors out if you've mismatched them.

The other thing it does properly is hand memory management to ComfyUI. Instead of keeping its own copy of the weights pinned on the GPU, it loads the checkpoint through ComfyUI's model-management path (it builds a core model patcher over the upscaler), so it offloads to CPU/disk when nothing's using it. Given these upscalers sit on top of an already-VRAM-hungry 33B video model, that's not a nicety - that's the difference between the upscale running and an OOM you have to work around.

Inputs are the whole story here:

• model_name - a dropdown of .safetensors/.pth files in ComfyUI/models/latent_upscale_models/. Empty? You haven't dropped the weights in yet (or haven't restarted after adding them).
• precision - auto is the sane default; ComfyUI picks fp16/bf16/fp32 by what your device supports. Only reach for a manual setting when you know why.

Output: a single upscale_model, wired straight into the 2D or 3D node's upscale_model input.

Getting the weights. The pack doesn't bundle or download them. The community-trained checkpoints live on HuggingFace at LBH-123-AI/Minimax_h3_latent_Upscaler. Drop the files you want into the folder and restart:

# your ComfyUI install dir
mkdir -p models/latent_upscale_models
# put the downloaded 2D and/or 3D checkpoint(s) in there, then restart ComfyUI

Installing the pack. Easiest is ComfyUI Manager - search FL MiniMax H3 and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxH3.git

Then restart. This is from filliptm (the ComfyUI Fill Nodes author), Apache-2.0 licensed, and it needs a current ComfyUI with MiniMax H3 support - the H3 model, text encoder and VAEs are yours to supply, not the pack's. The only Python dependency beyond ComfyUI is huggingface_hub, which the loader itself doesn't even use (that's the VDN node's download).

Where people get burned: pairing a 2D checkpoint into the 3D node or vice versa - both error with a clear "requires a 2D/3D checkpoint" message, so read it before assuming the loader is broken. And remember the loader is the same one both architectures share, so if you're A/B-testing 2D vs 3D, you don't need two loaders - just swap the file.