Run MiniMax-H3 locally on Apple Silicon from ComfyUI (Wild H3C nodes, via h3.c)
Hi everyone — for those on Mac: MiniMax-H3 can now be generated locally on Apple Silicon directly from ComfyUI.
Wild H3C is a ComfyUI custom node package that wraps antirez's h3.c Metal engine:
- 🎬 H3 Text to Video and H3 Reference to Video (Ref2VA) nodes — canvas, duration, steps, presets, seed, first/last-frame anchors, up to 9 reference images + 1 reference audio
- 📈 Live per-phase progress in the UI and clean cancellation
- 🔌 Standard VIDEO output, plugs straight into SaveVideo
- 💾 Works with an existing MiniMax-H3 snapshot (just point the node at it) or a fresh download — both paths are documented
Links:
- ComfyUI Registry: https://registry.comfy.org/nodes/wild-h3c
- Repo: https://github.com/skaiy/wild-h3c-comfyui
Tested on an M-series MacBook Pro with 128 GB unified memory (512×512, 4 steps ≈ 32 s warm). Note that the MiniMax H3 Community License applies to the weights. Feedback, issues and PRs welcome!
我电脑是 Macbook Pro M2 MAX 64GB，Comfyui版本为0.30.0，测试了一下Ref2VA，token_reduction和SSD streaming都关闭，视频全部黑屏，同时打开token_reduction和SSD streaming，结果如下，
480864 30步 Balanced Match 黑屏
480864 26步 Balanced Match 有视频但镜头未遵从提示词，比例基本正常
480864 20步 Balanced Match 有视频但镜头未按提示词，且比例压扁
480864 18步 Balanced Match 有视频基本遵从提示词但比例压扁
480*864 16步 Balanced Match 黑屏
480864 30步 Precise Match 黑屏
480864 20步 Precise Match 黑屏
480*864 16步 Precise Match 黑屏
且生成的视频人物面部特征和参考图人物一致性也不足，请作者看看是哪里出的问题呢