MiniMax H3 ComfyUI Pack
Ready-to-import ComfyUI API workflows for MiniMax H3 video generation, wired for SageAttention and Spectrum.
Overview
A small pack of API-format ComfyUI workflow JSONs covering the three ways people actually use MiniMax H3 for video: text-to-video, image-to-video and reference-to-video. Each graph is pre-wired with the SageAttention patch and the SpectrumApplyMiniMaxH3 node, so the speed and quality variants can be compared without rebuilding the pipeline by hand.
How it is wired
- Model chain: UNETLoader into PathchSageAttentionKJ into SpectrumApplyMiniMaxH3
- Conditioning: CLIPLoader plus video and audio VAELoaders from the same H3-compatible assets
- Sampling: KSamplerSelect (res_multistep), BasicScheduler (simple, 20), SamplerCustomAdvanced
- Decode path: VAEDecode and VAEDecodeAudio into CreateVideo and SaveVideo
Requirements
You need a SageAttention-compatible ComfyUI setup, the ComfyUI-Spectrum-MiniMax-H3 custom node, and the H3 weights (the pruned int8 UNet and int4 Qwen3-VL text encoder variants). Load any of the three JSONs via the API-format import, run a smoke render, then tune prompts and seeds.
Why it exists
Video workflows get shared as screenshots of node soup more often than as working files. This pack is the boring, useful version: importable JSON, a documented graph pattern, and a checklist to validate the graph before burning GPU time.
Status
Released under MIT on GitHub.