Two ComfyUI graphs ship with this episode, both built on MiniMax H3 and aimed at martial-arts motion.
They are supplied files, not benchmark runs. No clip was rendered for this article, no timing was
taken, and nothing here is a quality comparison.
What is actually in the box
| MiniMax H3 Martial Arts Workflow | MiniMax H3 Martial Arts Workflow (Multi-Keyframe) | |
|---|---|---|
| Nodes / links | 119 / 111 | 138 / 152 |
| Groups | 24 | 25 |
| Node types | 46 | 49 |
| Timeline math nodes | 2 | 8 |
Both graphs share the same LoRA stack, so learn it once:
-
Minimax H3真实电影质感V0.1_FeiHouRemix_v0.6_compat_v001_T8.safetensorsat 0.5 model / 1.0 clip -
camera_motion_h3_lora_v1_3000_pruned.safetensorsat 0.3 model / 1.0 clip -
wushu_spatial_physics_clean_2500.safetensorsat 0.3 model / 1.0 clip -
wushu_action_fl2va_v7_2000step_no_adaln.safetensorsat 0.4 model / 1.0 clip
The first three are what the video means by "three-LoRA mixing": wushu action, wushu spatial physics,
camera motion. The turbo LoRA and the film-grain LoRA sit alongside them. Those strengths are saved
widget values, not tuned recommendations — treat them as a starting point to test on your own footage.
The single-frame-pair graph
The graph leans on SetNode/GetNode reroutes: conditioning is published once and consumed by several groups, which keeps the canvas readable at 119 nodes. Audio is loaded and cropped inside the same graph, so the render and the soundtrack stay in one file. A first frame and a last frame pin the motion; the sampler interpolates between them.
Saved sampler state: seed 755708092963751 (fixed), 20 steps, CFG 1, euler with the simple scheduler, denoise 1.
Check this before you trust the defaults. The stack loads the 8-step turbo LoRA at 0.8, yet the
saved sampler array says 20 steps. Either the running interface overrides the step count, or
the file ships a mismatch. Confirm it in your own UI rather than assuming the stored widget executes.
The multi-keyframe variant
138 nodes and 152 links — bigger, because the timeline is computed instead of typed
in: 8 math and integer nodes assemble keyframe positions. That is the practical
difference between the two files. If you want one motion beat, the first graph is enough; if you need
several keyframes on one path, the second one carries the arithmetic for you.
What none of this proves
- No inference, no rendered output, no speed numbers. If a video claims motion quality, that is the author's claim about their own run.
- The weights are not bundled. Check the licence of MiniMax H3, each LoRA and any input footage separately before publishing anything you make.
- Covers are AI-generated promo frames with a platform watermark; they are not frames from these graphs.
Get the files
- Hugging Face:
AI-KSK/minimax-h3-martial-arts-two-graphs-301— both JSONs plus covers and a hash list. - ModelScope:
Sunjian520/minimax-h3-martial-arts-two-graphs-301— same package, Chinese README. - RunningHub: single pair · multi-keyframe
- Civitai: the two models built from the same files.
Source video: https://www.youtube.com/watch?v=3Ak4mkrhOrY · Chinese mirror: https://www.bilibili.com/video/BV1rBty6AENs
Top comments (0)