Research reward that scores human motion in generated video via 3D body recovery and MuJoCo physics simulation.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
PhyMotion is a research method from a team including Yidong Huang, Mohit Bansal and co-authors, titled "PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation" (NeurIPS 2026 per the BibTeX). It is a reward signal for evaluating and improving human motion in generated video, not a standalone video generator.
The method recovers SMPL body meshes from a generated video, retargets them onto a humanoid in the MuJoCo physics simulator, and scores the motion on three axes: kinematic plausibility, contact and balance consistency, and dynamic feasibility. Each axis gives a continuous, interpretable score. The authors report stronger correlation with human judgments than existing 2D-based rewards, and that the three axes complement each other.
Used for RL-based post-training, it is reported to improve motion realism in both autoregressive and bidirectional video generators, in automatic metrics and blind human evaluation. In the project page's Elo table, the authors' 1.3B model ranks above Wan2.2 14B, FastWan 1.3B, Wan 1.3B, EchoMotion 5B and Wan2.2 5B. A creator review says code is released and that some anatomical errors remain. The page provided does not link the code, weights or paper, so those details are unverified here.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Reward system that recovers a 3D human body from generated video and scores it in the MuJoCo physics simulator. This improves anatomical correctness in video models, though some errors remain. Code released.
Closest neighbours by meaning, not just by category.
Model
Monocular video motion capture that combines flow matching with physics-based RL for plausible 3D human motion.
Research
Research framework that turns a single image into a textured 3D asset with articulated parts and physical behavior.
Model
Research method that reconstructs a 4D Gaussian Splatting human from a single casual monocular video.
Model
Research framework that conditions one video world model on force, camera, hand pose and robot pose actions.
Model
Diffusion transformer that generates rigid and elastic 3D mesh dynamics in world coordinates from initial velocities and materials.
Research
Training-free, skeleton-free video motion transfer across subjects using TransPE on a frozen WAN-T2V diffusion transformer.
