MiniMax H3 Spatial Physics LoRA: More Realistic Motion in ComfyUI

MiniMax H3 Spatial Physics LoRA is a community adapter for MiniMax H3 that improves the model's spatial reasoning and physics — collisions, stacking, falling, and occlusion between objects — in text-to-video generation. It is trained on the ComfyUI-ready weights (Comfy-Org/MiniMax-H3) and loads with ComfyUI's built-in LoRA loader, with no custom nodes required.

Same prompt, same seed: base MiniMax H3 (left) vs. spatial physics LoRA (right) on a wushu fight scene

What it does

The LoRA focuses on object physics: rigid-body collisions with visible impact and bounce, momentum transfer, gravity-consistent trajectories, and objects that keep clear spatial relationships instead of passing through each other or floating. The model card's side-by-side tests use fight scenes, where the LoRA (strength 0.3) adds believable contact and impact feedback:

Wushu fight scene with the spatial physics LoRA applied (strength 0.3, from the official model card)

The same prompt without the LoRA, for comparison (from the official model card)

How to use it in ComfyUI

The weights are standard safetensors LoRAs, so no custom nodes are needed:

1. Download one of the .safetensors files from the model repository into ComfyUI/models/loras/.
2. Add a LoRA loader to your MiniMax H3 text-to-video workflow and point it at the file. The author suggests a strength around 0.3–0.5 for a subtle enhancement on top of the base model.
3. Describe motion with spatial and physics language — no trigger word is required. Useful phrases include clear spatial layout, gravity-consistent trajectories, objects maintaining spatial relationships, rigid collision with momentum transfer, and visible impact and bounce.

The LoRA is compatible with the MiniMax H3 Turbo acceleration LoRA, so the two can be stacked.

Availability

The model is published on Hugging Face at Jojocodex/minimax-h3-spatial-physics-lora in three variants:

It is trained with ai-toolkit at Rank 16, 512 resolution, 90 frames @ 24 fps. The author describes it as an early training and testing stage release, with the next retraining round already planned, so results may vary — treat it as an enhancement on top of the base model rather than a full replacement.