Spaces:
Running
title: Gaussian Physics
emoji: 🧊
colorFrom: blue
colorTo: green
sdk: static
pinned: false
license: apache-2.0
Gaussian Physics — the kernels are the simulation
A 3D Gaussian Splat is usually something you look at. Here the gaussian kernels are the particles of a continuum simulation, so the same representation that renders the object also carries its physics — no mesh, no tetrahedralisation, no round trip between two representations.
Each kernel holds a deformation gradient F, and its covariance is transported by it:
Σ′ = F Σ Fᵀ
That single line is why a squashed region renders as flattened kernels rather than as kernels that merely moved somewhere else.
What is running
- MLS-MPM transfers (Hu et al. 2018): particle → grid, solve, grid → particle. The APIC affine velocity doubles as the velocity gradient, so no separate gradient estimation is needed.
- Fixed corotated elasticity with a polar decomposition per kernel per substep.
- Six materials. Foam, jelly and plasticine differ by stiffness and yield; snow and sand get their character from clamping the stretch — which is also what makes them stop springing back. Drop a jelly and it recovers; drop sand and it stays compacted.
Everything runs on the CPU in your browser. No GPU, no server, nothing uploaded.
Provenance
Independent implementation of the method described in PhysGaussian (arXiv 2311.12198). That repository ships no licence, so no code from it is used here, and the test objects are generated procedurally rather than taken from any splat dataset — published splat reconstructions carry research-only or unstated terms.
Built by VIDraft.