ComfyUI-KinetiX-Bridge

MiniMax-H3 5120-d Combat Semantic Bridge for High-Impact Martial Arts and Combat Action Video Generation

By luluka-code·Created 2 days ago·Updated about 12 hours ago· 0

luluka-code/ComfyUI-KinetiX-Bridge

Nodes—

On cloudLocal install

Stars0

Updatedabout 12 hours ago

Readme

ComfyUI-KinetiX-Bridge

MiniMax-H3 5120-d Combat Semantic Bridge (Pure Inference Runtime)

A lightweight runtime node for MiniMax-H3 that injects authentic physical combat causality, heavy impact recoil, and contact alignment into action video generation without facial deformation or style drift.

🚀 Quick Start

1. Installation

cd ComfyUI/custom_nodes
git clone https://github.com/luluka-code/ComfyUI-KinetiX-Bridge.git

Restart ComfyUI after installation.

2. Model Loading (Auto Download or Manual)

Option A: Auto Download (Recommended)
Simply leave the bridge dropdown on:
kinetix_combat_bridge_v1.safetensors (Auto Download)
The node will automatically download the official weights (~16 MB) from Hugging Face (kinetixfx/Kinetix_Combat_Bridge) directly into ComfyUI/models/kinetix_bridge/. Built-in mirror fallback is included.
Option B: Manual Download
Download kinetix_combat_bridge_v1.safetensors manually and place it into any of the scanned directories:
- ComfyUI/models/kinetix_bridge/ (Recommended standard)
- ComfyUI/models/loras/
- ComfyUI/models/wushu_bridge/
- ComfyUI/custom_nodes/ComfyUI-KinetiX-Bridge/models/

🎛️ Recommended Presets

Mode A: Text-to-Video (T2V) — Maximum Action Impact

alpha: 0.18 ~ 0.25 (intense sparks, blade clashes, and staggered knockbacks)
magnitude_match: per_token (prevents color shifts)
token_span: all

Mode B: Reference-to-Video (Ref2VA) — Face Identity Guard

alpha: 0.10 ~ 0.12 (smooth combat flow)
magnitude_match: per_token
token_span: tail ⭐️ (Freezes reference image tokens to protect face & costume identity)
tail_ratio: 0.5

📦 Node Menu

Found under: MiniMax H3/KinetiX Bridge -> KinetiX Combat Semantic Bridge

📄 License

Apache-2.0