PhysiFormer: Learning to Simulate Mechanics in World Space

Forum topic · 小凯 · 2026-06-27

Summary

PhysiFormer is a diffusion transformer for generating physically plausible 3D object motion, presented in a paper by Yiming Chen, Yushi Lan, and Andrea Vedaldi (arXiv:2606.27364). 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, along with the object's material type (rigid or elastic), the model samples future vertex trajectories. Rather than relying on ad-hoc latent spaces or explicitly enforcing rigidity and causality as many related neural physics approaches do, PhysiFormer learns physical behavior directly from data in world space. This approach enables physically consistent simulation of 3D object dynamics while maintaining generative flexibility across different material types. The paper falls within the computer vision research area.

Paper Overview

Research area: Computer Vision (CV)
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, PhysiFormer learns physical behavior directly from data in world space.

Key Points

A diffusion transformer designed to generate physically plausible 3D object motion.
Operates in world coordinates on 3D meshes, rather than in view-dependent pixel space like typical video world models.
Takes as input the initial vertex positions and velocities plus the object's material type (rigid or elastic).
Outputs sampled future vertex trajectories describing how the object moves over time.
Contrasts with related neural physics methods that use ad-hoc latent spaces or explicitly enforce rigidity and causality; PhysiFormer instead learns physics from data.

*Auto-collected on 2026-06-27*