Wushu Action is a community martial-arts adapter collection for MiniMax H3. The search term wushu_action leads to several files and two substantially different descriptions: V7 focuses on fighting footage, while V8 describes motion-reference animation made from human models. Choosing the file is the first creative decision.
This guide follows the Jojocodex model card and repository file list, checked October 18, 2026 at revision 4808de9b. It explains the author's published configuration and its unresolved inconsistencies. We have not run these adapters or verified the quality claims. The collection's latest recorded modification predates today's review; this is not a new-release announcement.
Distinguish the files before downloading
The collection name still contains V7, but its tree also includes V8 adapters. Do not use the repository name as the only version selector.
| File family | What the published material establishes | What still needs care |
|---|---|---|
| V7 FL2VA, 1000-step filename | Listed as the E3 main file with AdaLN calibration | The card's English and Chinese training-step descriptions disagree |
| V7 FL2VA, 2000-step filename | A distinct file exists in the current tree | Existence alone does not establish that every E3 instruction applies unchanged |
wushu_h3_r2v_v7_2000.safetensors | A separate R2V-named artifact exists | Do not infer a complete Ref2VA recipe from the FL2VA section |
| V8, 2000-step and 5000-step filenames | Two files exist; V8 has its own motion-focused guidance | The card does not provide a controlled comparison between the two checkpoints |
The E3 section describes an unpruned full INT8-convrot diffusion base and BF16 text encoder during training. Its suggested loading table also mentions an NVFP4 encoder option. That is a usage suggestion, not evidence that all precision choices reproduce the training setup.
The card describes 924 selected action clips for the earlier section, but gives conflicting 1000- and 2000-step wording. Preserve the filename and revision in your project notes rather than combining those statements into one invented training history.
Match the base and workflow
For the E3 path, the author names minimax_h3_fl2va_int8_convrot.safetensors, emphasizing the full, unpruned model. Put it in ComfyUI's diffusion-model directory, the Qwen3-VL encoder in the text-encoder directory, both video and audio VAEs in the VAE directory, and the selected adapter under LoRAs.
Import the supplied multi-reference fighting workflow only after checking the models and nodes it actually requests. A file named R2V and a workflow using references are different pieces of evidence; neither proves that any FL2VA adapter can be substituted into any Ref2VA graph.
The author recommends Euler sampling, the simple schedule, 25 steps, and CFG 1.0 for the documented baseline. The table also mentions lower steps with Turbo bases while separately warning that low-step output can be blurry. Start from one internally consistent recipe before adding a speed adapter. An adapter stack changes more than rendering time.
For dimensions, the guide uses 832×480 and a frame grid of 17n + 5. Its 124-frame example at 24fps is approximately 5.17 seconds. Width and height must follow the workflow's alignment requirements. Use the current imported graph to validate those values instead of assuming a generic five-second preset has identical timing.
Use the trigger and describe a readable action
The author places wushu_action at the beginning of the prompt. The card favors specific motion names and short action phrases over a generic request to fight. Keep appearance consistent, then describe preparation, execution, and recovery.
An editorial planning example is: one blue-uniformed practitioner in an empty courtyard lowers into a guarded stance, performs one low sweeping kick, and recovers balance in a stable guard. Use a full-body view, a clear floor, and cloth and footstep sounds. This is a suggested evaluation scene, not a tested prompt or promised movement.
Begin with one practitioner and one action. Inspect foot contact, center of mass, limb continuity, and whether recovery follows the sweep. Only then try a second character or more complex choreography. The V8 section explicitly says its training clips are single-person and short; two-person fights are an extrapolation with weaker stability.
Understand V8's white-model and black-background tendency
The V8 section describes human-model animation and warns that white untextured figures against black backgrounds can appear by default. That visual tendency reflects the training material. It should not be presented as a general cinematic realism improvement over V7.
Its recommended range distinguishes motion-reference creation from ordinary scenes: approximately 0.9–1.0 for the former and 0.6–0.8 as a starting point for the latter. Those are author recommendations, not results from our testing. Lowering strength may reduce the unwanted training look while also weakening action guidance.
The collection includes a V8 move vocabulary. Use that resource to choose a supported movement, then keep the character, environment, and camera simple enough to assess it. A long sequence with many cuts can conceal whether the adapter controls the actual body motion.
Keep installation success separate from motion quality
The author suggests comparing adapter strength zero and one when diagnosing whether it loaded, and notes that key remapping may be needed if no effect appears. A visible difference confirms influence, not anatomical correctness or a successful motion transfer.
Retain the base model, adapter filename, revision, strength, sampler, steps, seed, canvas, and exported audio/video with each test. Compare short matched clips before stacking Combat Base, Motion Continuity Repair, or physics adapters suggested in the card. Those combinations have separate identities and can introduce additional variables.
For continuous two-person choreography, the Fight Flow guide covers another named resource. The H3 LoRA page explains the broader workflow. H3Vid has not registered these Wushu files as hosted presets; the published local setup is not evidence of compatibility with a generic provider LoRA endpoint.