Yujing Sun

Diffusion models, guided by vision-language models and geometric constraints, can now reconstruct realistic 3D human interactions from monocular video, even with severe occlusions.

A new point cloud representation, Sparkle, achieves state-of-the-art human motion capture by explicitly disentangling kinematic structure from surface geometry, finally balancing expressiveness and robustness.

Achieve realistic and responsive avatar interactions without expensive full-sequence inference, thanks to a novel modular framework that adapts pre-trained motion priors to diverse interaction scenarios.

A new swarm-based optimization algorithm, inspired by dogfighting but built on kinematic equations, achieves state-of-the-art performance across diverse benchmark and real-world engineering problems.