PhyMotion reward raises 1.3B video model to 14B levels
PhyMotion delivers a structured 3D motion reward that converts generated human videos into SMPL or SMPL-X trajectories and evaluates them inside the MuJoCo physics simulator. The system scores kinematic plausibility, contact consistency, balance, and dynamic feasibility. Reinforcement learning post-training with PhyMotion lifts a 1.3B-parameter video model to performance levels comparable with a 14B-parameter model on human motion benchmarks, replacing weaker 2D pixel-based signals that overlook floating feet and self-penetrations.
PhyMotion
Structured 3D Motion Reward for Physics-Grounded Human Video Generation
paper: https://huggingface.co/papers/2605.14269
AK@_akhaliqPhyMotion Structured 3D Motion Reward for Physics-Grounded Human Video Generation
#1866Jaemin Cho@JMIN__CHOIntroducing PhyMotion! You can now RL-tune your video generator with Real2Sim2Real reward.
PhyMotion is our structured reward for human video generation. It lifts SMPL human motion from video and calculates physical feasibility scores across fine-grained dimensions within MuJoCo.
Experiments show significant gains (e.g., a 1.3B model + PhyMotion matches the performance of a 14B model).
Great work led by Yidong and Zun. Check out more 👇
Yidong Huang@owenhuang117🚨 Excited to introduce PhyMotion🤸: Structured 3D Motion Reward for Physics-Grounded Human Video Generation! ❌ Existing 2D video rewards misleadingly assign high scores to videos with floating feet, self-penetrating limbs, and physics-violating motions. ✅ PhyMotion lifts generated videos into 3D, grounds them in a physics simulator, and scores motion along kinematic / contact / dynamic feasibility. ➡️ RL post-training with PhyMotion improves 1.3B model to match 14B models performance in human prefence. 🧵(1/n)👇