PhysLDM: Latent Diffusion for High-Fidelity Deformable Simulation

arXiv cs.GR details PhysLDM, a new latent-diffusion framework for one-shot volumetric deformable simulation. The pitch is straightforward: high-resolution deformable bodies are expensive to predict directly, and autoregressive rollouts tend to drift, so the authors compress motion into a compact latent space before generating trajectories.

The core technical piece is a holistic spatiotemporal VAE designed to avoid the “staircase” artifacts common in temporal compression schemes. The paper says it reaches about 2.48 mm reconstruction precision on meter-scale scenes while compressing tokens by up to 78x, which is the kind of efficiency gain that matters if you want neural simulation to scale beyond toy examples.

The authors also compare deterministic regression against diffusion in this latent space and argue that chaotic deformable dynamics are better handled as a distribution than as a single best guess. In their setup, diffusion produces more physically plausible trajectories, while regression tends to collapse toward averaged, non-physical motion.

Trained purely on kinematic data from an Objaverse-scale dataset, PhysLDM reportedly generalizes zero-shot to GSO and Toys4K. Because the model is differentiable, it also opens the door to inverse problems and higher-order design optimization, which makes it relevant not just for simulation research but for tooling around physics-aware content creation.