PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions

Researchers introduced PhyGenHOI, a framework for generating dynamic human-object interactions from text with stronger physical realism.

It pairs a motion diffusion model for human motion with a physics-based object simulation, using attraction and contact-triggered resimulation to better capture timing and impact.

A masked video prior and 3D Gaussian splats further improve visual fidelity, and the method reportedly outperforms prior baselines.

The work could advance 4D scene generation for more realistic simulations, videos, and interactive content.