MiniMax H3 NF4 Video VAE Loader / NF4 视频 VAE 加载

H3 decodes its picture through a dedicated video VAE, and if you're running the DiffSynth NF4 weight set you need this loader - the stock VAELoader won't understand the file. video_vae_nf4.safetensors is only 1.61 GB, the second-smallest of the four NF4 files, and this node turns it into a normal VAE output for the H3 sampler and decoder.

What it does

Same pattern as the pack's other NF4 loaders. DiffSynth's checkpoint keeps Linear weights flattened as BitsAndBytes NF4 tensors with the real shapes in .quant_state.bitsandbytes__nf4 sidecars. The loader restores those shapes, rebuilds the VAE's quantized Linear layers, and then does a strict sanity check: it loads into ComfyUI's MiniMax H3 VAE structure and fails if the checkpoint doesn't match - a wrong or foreign VAE file won't silently half-load.

One vae_name combo input, one VAE output. Drop the file in:

| File | Folder | Size | |---|---|---| | video_vae_nf4.safetensors | ComfyUI/models/vae | 1.61 GB |

The NF4 stack context

This is one of the four loaders that together give ComfyUI the whole DiffSynth NF4 FL2VA stack: model (17.16 GB), text encoder (15.32 GB), this video VAE, and the audio VAE (0.28 GB) - 34.4 GB total. Keep the rest of the native H3 workflow unchanged and just swap the four loaders in.

Requires BitsAndBytes and CUDA:

python -m pip install "bitsandbytes>=0.50.0"

And the honest performance note applies across the stack: NF4 saves disk and download size, but weights dequantize for computation, so on the pack's measured 3090 the NF4 stack was slower than the Comfy-Org int8_convrot route (about 32% at the tested resolution). If you're choosing between stacks, size your decision by "will the smaller footprint fit my setup," not "NF4 is faster" - it isn't.

Install

Part of the MiniMax-H3 Tools pack (ComfyUI 0.30.0+). Search "MiniMax H3 Tools" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git

Restart ComfyUI, find MiniMax H3 NF4 Video VAE Loader under MiniMax H3. Fails explicitly on FP4 files, missing sidecars, and incomplete downloads - .incomplete and .aria2log artifacts aren't model files. And as with all H3 weights, the MiniMax H3 Community License applies, which excludes the US, EU, UK and Korea from running the local weights.