OPEN LINK ↗
// 1h agoOPENSOURCE RELEASE
PhyMotion uses physics simulators to fix AI video glitches
PhyMotion is an open-source framework that integrates the MuJoCo physics simulator into the video generation pipeline to evaluate and reward physical plausibility. By recovering 3D human body meshes and checking for balance, ground contact, and joint limits, it provides a structured 3D reward signal that eliminates common artifacts like "foot sliding" and floating characters in generated human movement.
// ANALYSIS
Video models are finally getting a "sanity check" from real-world physics, moving beyond pure 2D pixel-matching to understand actual body dynamics.
- –Integrates MuJoCo physics simulator directly into RL-based post-training rewards for video models
- –Evaluates kinematic feasibility, contact consistency, and dynamic forces using SMPL body meshes
- –Achieves a +68 Elo gain in human preference evaluations compared to state-of-the-art 2D-based rewards
- –Open-source release includes training code, datasets, and fine-tuned model weights for Causal Forcing and FastWan
- –Effectively solves "phantom" movements and physical impossibilities that break immersion in character-heavy AI videos
// TAGS
phy-motionvideo-gen3d-genreinforcement-learningphysics-simulationmujocosmplopen-source
DISCOVERED
1h ago
2026-05-17
PUBLISHED
1h ago
2026-05-17
RELEVANCE
8/ 10
AUTHOR
AI Search