PhysiFormer: Learning to Simulate Mechanics in World Space

TL;DR

A model that directly applies a diffusion process in world coordinates to generate physically plausible 3D object motion.

Problem

Existing video world models operate in view-dependent pixel space, making them geometry-unaware, while neural physics approaches rely on inductive biases or latent spaces.

Approach

It takes initial vertex positions, velocities, and material type as input, framing vertex trajectory prediction as a single denoising diffusion process directly in world coordinates. It uses attention factorized over time, space, and objects for efficiency.

Results & Contribution

Trained on over 100k simulated trajectories, it generates rigid and elastic mechanics and generalizes to unseen geometries and mixed materials. It substantially outperforms autoregressive baselines in trajectory accuracy and physical consistency.