TL;DR: Our simulation-rendering framework enables realistic collisions with arbitrary shapes.

Simulation Particles + 2DGS-based Collider

Given multi-view input images, we separately reconstruct the continuum object and the rigid body collider. Particle trajectories are then advected using our Any-Shape Differentiable Material Point Method (AS-DiffMPM). Finally, the updated particle positions are mapped back to rendering primitives for photorealistic and physically plausible animation.

System Identification

Beyond physically plausible animation, the differentiability of our simulation-rendering framework makes it suitable for estimating the physical parameters of objects from video observations (i.e., system identification). Our AS-DiffMPM can be seamlessly integrated with various rendering methods (e.g., 2DGS, DVGO, MDyn-3DGS) for the task at hand.

Select different materials, colliders and samples.

Collision Resolution

Our AS-DiffMPM models collisions with sharp, fine-grained colliders more accurately compared to the GOP-DiffMPM baseline.

Select different colliders.

Citation

The website template was borrowed from PhysDreamer, ReconFusion, Ref-NeRF and DMD.