Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

huangserva/h3-fight

Media & Design

A Chinese prompt library, validator and ComfyUI runner for MiniMax H3 fight-scene video generation.

104
32
Python
MIT
1 day old

What it is

h3-fight packages a way to write MiniMax H3 fight-scene prompts, aimed at people using AI assistants or writing prompts by hand. It has three layers: a prompt core of 188 shot blocks, phrase parts, move tables, a pitfall list, a six-part writing pattern, 检查.py (checker) and 查重复.py (duplicate checker); an optional ComfyUI-based rendering path; and explicit non-goals — reference-image generation and video upscaling are left to other tools.

The rationale is that free-hand fight descriptions often produce flat choreography, static camera angles, blades passing through bodies, or characters switching sides. The project turns observed failures into pre-submission checks. All shots shown in the demo GIFs were produced with library prompts (C1, k3b, k4d).

Quick start

Only Python 3.8+ is needed to write and check prompts — no GPU. Clone the repo wherever you like (or into ~/.claude/skills/h3-fight / ~/.codex/skills/h3-fight for the AI-skill path), then:

cd <install-dir>
pip install -r requirements.txt
python3 工具/自检.py

Pick shot blocks from 库对照.md, assemble a prompt following SKILL.md, and run the checker:

python3 工具/检查.py my_fight.txt

Rendering requires a ComfyUI with MiniMax H3 support. Build the workflow yourself following the workflow guide, export it as API JSON to 库/模板/三参考图.json (with reference images) and 库/模板/无参考图.json (prompt only), then run:

python3 工具/出片.py my_fight.txt --名字 fight01 --输出 out/ --参考图 hero.png arena.png rival.png

To continue a fight from the last frame of a clip, extract it with ffmpeg and pass --首帧. The repo ships no workflow file; 工具/自检.py cross-checks nodes and model files against ComfyUI's /object_info.

Requirements

The prompt-only path runs on CPU. Rendering was tested on an RTX 4090 48GB modified card with 123GB RAM: 1664×928, 243 frames (~10s) takes roughly 6–8 minutes from submission to download. 24GB cards are unverified — the main model is 21GB int8 and the text encoder 27GB int8, so models must be swapped in and out of VRAM. Reserve about 60GB of disk for models. ffmpeg is needed for stitching and subtitles. On Windows, set git config --global core.quotepath false before cloning because filenames are Chinese.

Workflow components include hybrid fl2va/ref2va int8, turbo LoRA at 0.7, SelfLift TST τ 0.2, and three fight LoRAs at 0.65 / 0.5 / 0.25. Those LoRAs are not distributed here; download them from their authors and follow their licenses. Plugin dependencies are selflift-Avatar, YCNodes-MiniMax-H3 and minimax-h3-audio-T8.

Limits

A clip is at most about 10 seconds (243 frames), and each can hold at most one short line of dialogue. Lightning, magic and energy-effect-heavy shots are not suited to this writing style and tend to go out of control. Written shots and generated footage often diverge, so human review is required; new problems found while watching should go back into the pitfall table. Reference-image and upscaling steps are documented in the reference-image guide and the upscaling note but not implemented.

Licensing

Code under 工具/ is MIT; the prompt library (库/, 库对照.md) and demo GIFs are CC BY-NC 4.0. The prompt samples are original and have had character names, move names and identifying appearance details removed.

HotRepo's take

The repository was created on 2026-11-15 and had 104 stars and 32 forks by 2026-11-16. The only HotRepo snapshot is a new Python-list entry for 2026-11-16 showing 4 stars gained over a 1.5-hour window — a single early record, not enough to read as a trend. What stands out technically is the split between a CPU-only prompt/check workflow and a GPU-heavy render workflow, plus the willingness to leave workflow JSON, LoRAs and reference images outside the repo, which lowers the value of a bare clone for rendering.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. KadoBOT/ComfyUI-H3-Continuity-StudioA ComfyUI custom node pack for planning and rendering multi-shot MiniMax H3 video sequences with visual continuity across clips.Python104
  2. MiniMax-AI/MiniMax-H3-ImageOpen-weights 16-megapixel image generation and editing model with a 33.1B DiT and 6-step sampling.117
  3. Pika-Labs/Pika-H3-FlexGenerate synchronized video and stereo audio locally from prompts, frames, and up to 12 reference files.Python131
  4. Video-Rebirth/hyperflow-h3-flashA 13.7B few-step MiniMax-H3 derivative that generates video with synchronized stereo audio, in INT8 and GGUF formats.Python111
  5. D3voz/openjutsuLocal open-weights video character and object swapping via a Next.js front end over ComfyUI workflows.TypeScript1,771