DiffSynth: MiniMax H3 Loader

MiniMax H3 is the model that made people pay attention to open video again: a 33B omni-modal system that treats text, image, video and audio as one context and produces 4–15 second clips with synced stereo audio, landing at the top of the video-editing arena on release. It dropped on HuggingFace in August 2026. It is also, by reported weight size, about 42.5 GB before you've loaded anything.

This loader is the front door. Budget disk space before you queue anything.

Required inputs

• model_configs - the merged list. H3's repo is organised by variant, so the pattern matters: FL2VA/text_encoder/model*.safetensors, FL2VA/transformer/model*.safetensors, FL2VA/video_vae/source/model.safetensors, FL2VA/audio_vae/model.safetensors. The Ref2VA graphs are identical with a different folder prefix. The pack's NF4 template swaps these for prequantized files from DiffSynth-Studio/MiniMax-H3-NF4 instead - which is the sane path on consumer VRAM.
• torch_dtype / device - bfloat16 and cuda as usual.

The optional inputs

processor_config - the repo's FL2VA/processor/ folder, wired as its own ModelConfig. Text preprocessing lives there.

redirect_common_files (default true) - the one worth understanding. H3's FL2VA and Ref2VA variant folders each ship their own copy of components that are functionally shared, and this flag redirects the load so the pack uses one common set instead of pulling and holding duplicates. On a model this size that's tens of gigabytes of disk and a meaningful chunk of load time. Leave it on unless you're deliberately testing one variant's own copy of a component against the other's.

vram_limit - from the VRAM Limit node. With 33B and no verified consumer floor at release, giving the loader an explicit budget is not optional in spirit.

Read this before you build anything on it

The H3 Community License grants broad rights below $20M annual revenue and MiniMax claims nothing from your outputs - but the grant is scoped to an "Applicable Territory" that excludes the European Union, the United Kingdom, the Republic of Korea and the United States. The exclusion reaches outputs, not just weights. This is a paper restriction with no technical enforcement, and MiniMax's hosted Hailuo API is not geofenced, but if you're in one of those regions the local-weights path is the one the license doesn't cover. Know that going in rather than discovering it in a README thread.

Install

ComfyUI Manager, search DiffSynth-ComfyUI, install, restart. Or:

pip install git+https://github.com/modelscope/DiffSynth-Studio.git
cd ComfyUI/custom_nodes
git clone https://github.com/modelscope/DiffSynth-ComfyUI.git

Then open general_minimax_h3_fl2va (full precision) or featured_minimax_h3_prund_nf4 (the quantized 4-step one) from the Template tab. If you have less than 24GB of VRAM, start with the NF4 one and save yourself the disappointment.

Where people get burned

Downloading the full-precision weights when you needed the quantized ones. Both are 30GB-plus categories of decision. The NF4 repo plus a Quantization Config with load_prequantized on is what the featured template does, and it's why that graph also pairs a 4-step LoRA with 4 inference steps instead of 50.

Wiring FL2VA files into a Ref2VA graph. The two variants have separate folder trees; mixing them fails at load after the downloads.

Missing the audio VAE. H3's native audio is not a bolt-on, and a graph wired without it won't produce sound.

Judging it before you've re-run it with fewer frames. The defaults are 50 steps at 768×1344 with 124 frames, which is a serious ask. Drop the frame count first.