6kplus/PhyMotion-CausalForcing-1.3B

PhyMotion — Causal Forcing 1.3B

LoRA adapter for Causal Forcing 1.3B (the autoregressive distilled version of Wan2.1 T2V-1.3B), post-trained with RL using the PhyMotion reward — a structured 3D motion reward grounded in SMPL recovery and MuJoCo inverse dynamics.

• Paper: PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation
• Project Page: https://phy-motion.github.io
• Repository: https://github.com/h6kplus/PhyMotion
• Prompt Dataset: 6kplus/PhyMotion-MotionX-Prompts

Description

Generating realistic human motion is a central yet unsolved challenge in video generation. PhyMotion is a structured, fine-grained motion reward that grounds recovered 3D human trajectories in a physics simulator and evaluates motion quality along multiple dimensions of physical feasibility: kinematic plausibility, contact and balance consistency, and dynamic feasibility.

What's in this repo

File	Description
adapter_model.bin	PEFT LoRA weights (rank 256, targets CausalWanAttentionBlock)
adapter_config.json	LoRA configuration

Usage

To use this LoRA adapter, clone the PhyMotion repository, place the base model checkpoint and this LoRA, then run inference (full instructions available in the repository README):

git clone https://github.com/h6kplus/PhyMotion.git
cd PhyMotion

# Download this LoRA adapter
huggingface-cli download 6kplus/PhyMotion-CausalForcing-1.3B \
  --local-dir checkpoints/phymotion-causalforcing

# Download MotionX prompts (train + test)
huggingface-cli download 6kplus/PhyMotion-MotionX-Prompts \
  --repo-type dataset --local-dir dataset/motionx

# Inference
# Note: You still need the base Causal Forcing 1.3B checkpoint (causal_forcing.pt)
torchrun --nproc_per_node=1 scripts/inference_wan.py \
  --base_model checkpoints/causalforcing/chunkwise/causal_forcing.pt \
  --lora_path  checkpoints/phymotion-causalforcing \
  --prompt_file dataset/motionx/test.txt \
  --output_dir outputs/test \
  --num_frames 45 --height 480 --width 832 \
  --guidance_scale 3.0 \
  --denoising_steps "1000,750,500,250" \
  --num_frame_per_block 3 \
  --mixed_precision bf16 --seed 42

Citation

@article{huang2026phymotion,
  title   = {PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation},
  author  = {Huang, Yidong and Wang, Zun and Lin, Han and Kim, Dong-Ki and
             Omidshafiei, Shayegan and Yoon, Jaehong and Cho, Jaemin and
             Zhang, Yue and Bansal, Mohit},
  journal = {arXiv preprint arXiv:2605.14269},
  year    = {2026}
}

License

Apache 2.0. The base Wan2.1 / Causal Forcing weights retain their original license.