/ComfyUI_RH_MinMaxH3MiniMax-H3 native workflows for ComfyUI — finally repaired, fully English, and ready for 0.30.1+Forked from the RunningHub pack and heavily expanded. All validated graphs now use ComfyUI’s native sampling path.
10 ready-to-load workflows covering:
• T2VA / FL2VA / Ref2VA
• First-frame, last-frame & first+last-frame control
• Image, image+audio, video+audio & custom audio references
• Comfy-native INT8-CONVROT (default) + scaled-FP8 options
• NVFP4/AWQ Qwen3-VL text encoder
• Single-file VAEs in standard folders
• Stock res_multistep sampling
• Bundled SageAttention + experimental Sol-style attention (no extra packs)
Older Direct runtime & Dual Sigma nodes kept for compatibility/research.
Repo + all workflows:
Description
The native MiniMax-H3 nodes+workflow set for ComfyUI is now repaired, English throughout, and ready for ComfyUI 0.30.1+.
Forked from the original RunningHub node pack and substantially expanded.
The reliable examples use ComfyUI’s native sampling path. The older Direct runtime and Dual Sigma node remain available for compatibility and research, but are not used by the validated workflows.
Repo:
https://github.com/exportAnything/ComfyUI_RH_MinMaxH3
Additional speed-focused nodes are planned as separate, ground-up work.
FAQ
Comments (4)
Looks like we're finally surpassing the strengths of both LTX and Wan 2.2.
Prompting is the biggest hurdle for each. They all have a way they like to be spoken to.
@exportAnything Yes. I'm on board with learning that. The biggest problem is when they don't have the vocabulary for certain needed camera/scene actions. But H3 seems pretty complex in that department, so I hope it can be told to do more complex scene/shot control, even if it's harder to learn.
Ah. Yeah. Un-be-freakin-lievable.