PhysiFormer: Learning to Simulate Mechanics in World Space

arXiv·medium signal

Presents a diffusion transformer that generates physically-plausible 3D object motion by representing objects as 3D meshes in world space, unlike video world models that operate in view-dependent pixel space. The world-space representation yields motion that respects mechanics rather than just visual coherence. Of interest for simulation, robotics, and 3D content generation.