Instructions to use yichengup/Impact-Flow-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Inference
- Notebooks
- Google Colab
- Kaggle
Impact Flow v2 β H3 motion LoRA
A LoRA adapter for MiniMax H3, trained by yicheng. It targets impact and combat motion: contact weight, directional follow-through, and the camera response to a heavy hit.
- Base: MiniMax H3
- Network: LoRA, rank 16
- Suggested strength: 0.8
Usage
Load with any standard LoRA loader in ComfyUI (for example LoraLoaderModelOnly) on top of a MiniMax H3 checkpoint. Start at 0.8 β drop toward 0.6 if the motion becomes twitchy, raise toward 1.0 if impacts feel weightless.
Technical details
Training settings, tensor coverage and key naming are documented in DETAILS.md.
Licence
Distributed under the MiniMax H3 Community License Agreement β see LICENSE and
NOTICE. Not licensed for generating likenesses of real people without consent, for illegal
content, or for any use prohibited by the upstream licence. You are responsible for what you
generate.
- Downloads last month
- 24