LiconStudio/Minimax-H3-VBVR

Overview

Licon VBVR is an image-to-video LoRA for MiniMax-H3, trained on VBVR data rebuilt and reprocessed specifically for H3.

It improves object consistency and character action logic, reducing unintended object duplication, clipping, and interpenetration during motion and interactions.

Key Improvements

• Reduced object duplication: fewer unintended copies of objects during generation.
• Reduced clipping and interpenetration: improved spatial interactions between characters, objects, and surfaces.
• Improved character action logic: more coherent movement sequences and interactions.

Training Data

The training data is derived from VBVR-Pro-SFT-Video.

I rebuilt the VBVR training data around H3's characteristics, adapting data preparation and preprocessing to its training pipeline. The goal is to transfer structured motion and interaction patterns into practical video generation, with a focus on object consistency, spatial relationships, and action logic.

Video Comparisons

Usage

Download H3-VBVR-I2V.safetensors and load it with a compatible MiniMax-H3 image-to-video workflow.

For ComfyUI, place the file in ComfyUI/models/loras/ and select it in the LoRA loader used by your H3 workflow.

Acknowledgments

Thanks to MiniMax for the base model and the VBVR team for the source dataset.