PHYSIFORMER: Learning to Simulate Mechanics in World Space

Traditional physics simulation is computationally expensive, while autoregressive neural models often suffer from error accumulation and drift over long sequences. The authors ask whether a diffusion-based transformer can learn to simulate 3D mechanical dynamics directly in world space, avoiding the view-dependent ambiguities of pixel-based video models and the limitations of autoregressive rollouts.

PhysiFormer treats physical simulation as a generative denoising task. Instead of predicting the next state step-by-step, the model generates the entire future trajectory of 3D mesh vertices in a single forward pass. The architecture is built on a Diffusion Transformer (DiT) backbone, modified with factorized attention mechanisms that operate across time, space, and objects. This allows the model to process multiple objects and materials (rigid or elastic) without needing explicit object identifiers or complex latent space encodings. The model is trained on 100k simulated trajectories, learning to map initial vertex positions and velocities to physically plausible future sequences.

PhysiFormer demonstrates that coordinate-space diffusion is highly effective for mechanical simulation. By generating trajectories holistically, it avoids the cumulative errors that plague autoregressive baselines, leading to better preservation of object shape and momentum. The model generalizes well to unseen geometries, mixed-material configurations, and varying object counts. Furthermore, the probabilistic nature of the diffusion process allows the model to capture uncertainty in dynamics, enabling the generation of multiple plausible futures from a single initial state.