MiniMax-H3 ONNX VAE Loader

The MiniMax-H3 VAE is the autoencoder inside the 33B omni-modal H3 video model - the thing that turns the sampler's 24-channel latents into actual pixels and back. It works, but in plain PyTorch it's a slow chunk of your video pipeline. This node is the "just run it" option: it loads the ONNX export of the H3 VAE directly, no TensorRT, no compile step, no NVIDIA-only toolchain. You pick the encoder and decoder, you get a VAE, you wire it into your H3 workflow.

The catch is in the pack README's own words: "You can run ONNX models directly, but this is usually slow." That's not a lie - ONNX is a portable format, not a fast one, and this loader exists so you can sanity-check that your model files downloaded correctly before you invest in the compile path the pack is really built around (the MiniMax-H3 TRT Compiler + MiniMax-H3 TRT VAE Loader pair). Think of it as the try-before-you-buy option. If your H3 videos decode in a tolerable time, you may never need TensorRT at all.

What you feed it

Just two dropdowns, both required:

• decoder and encoder - ComfyUI scans your models/vae folder for .onnx files and lists them here. Select the decoder ONNX in one and the encoder ONNX in the other.

That's the entire input surface. The single output is a standard VAE, which plugs straight into the VAE Decode / VAE Encode nodes in any MiniMax-H3 workflow. ComfyUI treats it like any other VAE loader, so nothing downstream knows or cares that the guts are ONNX.

Installing and the files it needs

Install the pack the usual way - ComfyUI Manager (search ComfyUI-H3VAE_TRT) or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT
python -m pip install -r ComfyUI-H3VAE_TRT/requirements.txt

Then grab all three files from lihaoyun6/MiniMax-H3-VAE-ONNX and drop them into ComfyUI/models/vae:

• minimax_h3_vae_decoder.onnx
• minimax_h3_vae_decoder.onnx.data - the decoder's actual weights live here, and this file must sit next to the .onnx
• minimax_h3_vae_encoder.onnx

One honest caveat: the current shipped build of this pack registers only the two TensorRT nodes in its __init__.py, so this ONNX loader may not show up in your menu after an update - the README still describes direct ONNX as an option, but the maintainer's momentum is clearly behind the TRT path. If it's missing, don't fight it; run the TRT Compiler once and use the TRT loader instead.

What usually trips people

Selecting only one model errors out - the node wants both encoder and decoder. And if you skipped the .onnx.data file (easy to do, since HF lists it as a separate download), the decoder won't load and you'll get a parse or missing-weight failure. The export is Apache-2.0, by the way, so it's cleanly usable even though the full H3 weights carry MiniMax's territory-restricted community license.