Licon VBVR LoRA helps MiniMax H3 keep objects consistent and actions in order
Licon Studio's free VBVR LoRA for MiniMax H3 image-to-video targets three common failures: objects that suddenly duplicate, bodies clipping through props, and actions that happen in the wrong order. It is a 149 MB file that loads in ComfyUI's stock LoRA loader with any H3 first-frame workflow.
What it does
It was trained on VBVR video-reasoning clips, where each clip is a checkable task such as tracking an object or doing steps in a stated order, rebuilt for H3's training pipeline: 434,894 items over 7,000 steps at 1280×720, rank 16. Training was video only, so the sound H3 generates is unchanged.
Results and limits
The release has six validation pairs rendered from the same seed with and without the LoRA, for example rinsing one green pepper, cutting it and moving the pieces into a pan, or moving a magazine and then a remote in a set order (864×480, 12.25-second clips). The comparisons come from the author, with no benchmark and no test of lower LoRA strengths.
Worth knowing
Put H3-Licon-VBVR-I2V.safetensors in ComfyUI/models/loras and pick it in LoraLoaderModelOnly; it belongs on the first-frame (FL2VA) H3 base, not text-to-video. No workflow file is published, the validation prompts describe the action step by step and forbid extra objects, and the weights are Apache 2.0.