Jojocodex's Wushu Action v7 Teaches MiniMax H3 Real Martial Arts Moves
A community LoRA trained on 924 curated martial arts clips brings specific wushu technique names into MiniMax H3 video generation, with ComfyUI drop-in support.
Takeaways
- Wushu Action v7 is a style LoRA for MiniMax H3 hitting 14,750 downloads on Hugging Face.
- Trained on 924 curated martial arts clips at 832x480, 24fps for 1000 steps.
- Trigger word
wushu_action, works with fl2va and ref2va H3 pipelines. - Ships a ComfyUI workflow JSON and a full dictionary of Chinese technique tags.
- Requires CFG=1.0, euler/simple sampler, 25 steps, frame counts following 17n+5.
- Stacks with speed LoRAs like HyperFlow and other H3 motion adapters.
Jojocodex has released Wushu Action v7, a low-rank adaptation for MiniMax H3 that steers the open-weights video-and-audio diffusion model toward choreographed Chinese martial arts. Named techniques map to learned motion patterns, while broad prompts such as “kick” or “punch” tend to produce generic movement. At publication, the adapter had about 14,750 downloads.
A LoRA adds a compact set of learned weights alongside the base model without replacing it. This adapter supports first/last-frame-to-video, abbreviated fl2va, and reference-to-video, abbreviated ref2va. It uses the trigger wushu_action and includes a ComfyUI workflow plus a dictionary of tags extracted from its training captions.
| Item | Details |
|---|---|
| Supported pipelines | fl2va and ref2va |
| Trigger | wushu_action |
| Workflow | Bundled ComfyUI JSON using standard nodes |
| Training data | 924 curated wushu clips |
| Training length | 1,000 steps |
| Training format | 832 × 480 pixels at 24 fps |
| License | Available at no charge under custom terms in the repository |
The custom license warrants review before commercial use, redistribution, or integration into a hosted service. The repository’s terms govern those uses even though the weights can be downloaded without payment.
The checkpoint choice matters
Training targeted the full, unpruned int8-convrot diffusion transformer, or DiT, paired with the bf16 Qwen3-VL text encoder. The DiT performs the iterative denoising that turns latent noise into video; the lower-precision formats reduce the memory required to store and run the model. Jojocodex describes this combination as H3’s highest-fidelity quantized configuration.