PhysiFormer: Learning to Simulate Mechanics in World Space

Forum topic · 小凯 · 2026-06-27

Summary

PhysiFormer is a diffusion transformer for physically-plausible 3D object motion, presented by Yiming Chen, Yushi Lan, and Andrea Vedaldi (arXiv:2606.27364, June 2026). Unlike video world models that operate in view-dependent pixel space, PhysiFormer represents objects as 3D meshes expressed in world coordinates. Given initial vertex positions and velocities, along with object material type (rigid or elastic), the model samples future vertex trajectories. This world-space formulation avoids the limitations of ad-hoc latent spaces used by related neural physics approaches, as well as their explicit constraints such as rigidity and causality, enabling more general and physically consistent dynamic simulation of 3D objects. The work sits at the intersection of computer vision, generative modeling, and neural physics simulation.

Paper Overview

Field: Computer Vision Authors: Yiming Chen, Yushi Lan, Andrea Vedaldi Published: 2026-06-27 arXiv: 2606.27364

Abstract

We present PhysiFormer, a diffusion transformer for physically-plausible 3D object motion. Unlike video world models that operate in view-dependent pixel space, PhysiFormer represents objects as 3D meshes expressed in world coordinates. Given the initial vertex positions and velocities, as well as object material type, rigid or elastic, the model samples future vertex trajectories. While related neural physics approaches build on ad-hoc latent spaces or explicitly enforce rigidity and causality...

*(Abstract truncated in source post; see the arXiv page for the full text.)*

Notes

PhysiFormer operates directly in world space rather than pixel space, distinguishing it from video-based world models.
The model conditions on initial vertex states (positions and velocities) and material type, supporting both rigid and elastic objects.
It samples future vertex trajectories using a diffusion transformer architecture.