PhysiFormer

PhysiFormer (pronounced "fizzy former") is a research model for learning mechanics in world space. It is a diffusion transformer that represents objects as 3D meshes in world coordinates rather than view-dependent pixels. Given initial per-vertex positions and velocities and a material type (rigid or elastic), it samples full-horizon vertex trajectories, with the input mesh topology imposed at inference.

The model uses attention factorised over time, space and objects, which allows permutation-invariant multi-object reasoning without explicit object encoding. It was trained on over 100k simulated trajectories, on scenes with at most 10 objects and 356 vertices. The project page reports generalisation to mixed materials, including per-vertex material conditioning within one object, unseen real-world geometries, and larger scenes (15 and 50 objects). Because it is probabilistic, it can produce several plausible futures from the same initial conditions. The authors report that it outperforms autoregressive baselines in trajectory accuracy, rigidity preservation and momentum-based physical consistency.

It is trained to generate 49-frame sequences, and chunked rollout can extend this to 193 frames. Reported limits include penetration and spurious contacts in very dense scenes. Training objects have uniform density, so the learned dynamics assume mass is set by volume. A YouTube creator's summary says code has been released and the dataset is still to come. This is a research project, not a hosted product, and the page text provided did not include a repository link, a license or a paper link.