SAX MiniMax H3 Loader

What this actually is

MiniMax H3 is the 33B omni-modal video model MiniMax - the lab behind the closed Hailuo service - opened the weights to in August 2026. Four to fifteen second clips, up to 2K at 24fps, with stereo audio generated with the picture rather than glued on afterwards. It's the first open-weight answer to the one thing Veo had and local video didn't.

One thing to clear before you spend an evening on it: the H3 Community License carves the US, EU, UK and South Korea out of its Applicable Territory. The weights download fine; the grant just doesn't cover those regions.

Now the reason this node exists. H3 isn't a checkpoint. Running it locally needs a diffusion model, a Qwen3-VL 32B text encoder, a video VAE, an audio VAE - and a second diffusion model if you want reference image/video/audio conditioning. The canonical graph feeds all of that into MiniMaxH3-Director (a separate pack, seesee75-commits/ComfyUI-MiniMaxH3-Director), which builds the prompt, latent and sigma shift - five loaders plus a LoRA branch before you've generated a single frame.

SAX MiniMax H3 Loader collapses all of it into one node, wires the outputs to the Director's input names one-for-one, and stuffs the sampling settings into a PIPE so the sampler downstream doesn't need them repeated.

How it works

Under the hood it's doing the boring, correct thing: the diffusion model comes from diffusion_models via ComfyUI's own loader, the text encoder from text_encoders, both VAEs from vae. The text encoder type is set for you - H3 needs CLIPType.MINIMAX, which is exactly the sort of dropdown nobody finds on the first try. If your ComfyUI predates H3 support you get a named error instead of a mystery conditioning failure.

Two model slots, not one:

• unet_name - the FL2VA checkpoint: text-to-video and first/last-frame image-to-video. Director with Refs off.
• ref_unet_name - the REF2VA checkpoint: reference images, videos, audio. Director with Refs on. Selecting it loads a second ~20GB model, so leave it on None until you actually need references.

Either slot alone is fine; both empty is an error. The two LoRA slots are equally strict - lora_name applies to FL2VA and ref_lora_name to REF2VA, lora_strength covers both, and a LoRA attached to a slot whose model is None throws rather than silently doing nothing. That strictness is the good kind: a wrong-type LoRA on the other model is a silent no-op, so the pack at least catches the version of the mistake it can see.

On defaults: each file dropdown pre-selects the first file whose name contains fl2va, qwen3vl, video_vae / audio_vae, so name your downloads sensibly and you can mostly ignore the dropdowns. steps defaults to 20 and that's the reference value - below about 15 quality falls off a cliff. res_multistep with the simple scheduler is the official pairing; beta or normal can suit reference-heavy prompts.

Outputs: PIPE, MODEL, MODEL_REF2VA, CLIP, VAE, AUDIO_VAE. The five model-ish outputs go straight into the Director. PIPE goes to SAX MiniMax H3 Sampler.

One trap worth internalising: vae_name and audio_vae_name are both plain VAE dropdowns over the same folder. Swap them and you get noisy video with no error message. Audio VAE first, then the video VAE.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/so16tm/SAX_Bridge
# restart ComfyUI

Or in ComfyUI Manager, install via Git URL with the same address (searching "SAX_Bridge" also finds it). Nothing to pip install for these two nodes - the pack's optional dependencies (sam3 plus triton on Windows, Impact Pack, pysssss custom-scripts) belong to other nodes in the pack and won't block H3.

You do need to place the weights: H3 diffusion models in ComfyUI/models/diffusion_models/, the Qwen3-VL 32B text encoder in models/text_encoders/, video and audio VAEs in models/vae/. Add the REF2VA checkpoint only if you're using references.

ComfyUI 0.30.0 or later is a hard requirement - H3 landed with day-zero core support because ComfyUI had the weights early, and both of these nodes lean on that core.

Where people get burned

• "Requires a ComfyUI version with MiniMax H3 support (CLIPType.MINIMAX)." Update ComfyUI. Not a pack bug.
• Reaching for ref_unet_name "just in case." It's a second ~20GB load. Use it when you're using references and the Director has Refs ON, not before.
• Hardware honesty. H3 is 33B, ~42.5GB of weights, on top of a 32B text encoder - and ~20GB more for REF2VA. The launch threads had people discovering their 16GB card wasn't going to hold the resident set; quantised community builds (int8/int4) are where that crowd ended up. ComfyUI offloads, but this is the heaviest thing in the video folder and it behaves like it.
• Looking for width, height or frame count here. There aren't any, deliberately. Resolution, clip length, latent and conditioning are the Director's job; this node loads models and carries settings.