gaussian-splatting technique sourced
3D Gaussian Splatting
Scenes are represented as millions of anisotropic 3D Gaussians optimized from photos and rasterized in real time, giving photoreal captured or generated environments at 1080p and 30+ fps.
mechanism
Starting from sparse structure-from-motion points, each Gaussian stores position, anisotropic covariance, opacity, and view-dependent color as spherical harmonics. A differentiable tile-based rasterizer projects, depth-sorts, and alpha-blends the Gaussians; gradients from photo reconstruction error update parameters while adaptive density control splits, clones, and prunes Gaussians. Rendering needs no neural network at runtime, so it runs at real-time rates. Extensions attach physics directly to the Gaussians (PhysGaussian runs a Material Point Method on the kernels), stream level-of-detail on the web, and standardize the format in glTF.
limitations
- Colors store captured radiance, so lighting is baked in; dynamic time of day, player-placed lights, and relighting are not supported by the base method.
- Memory heavy: millions of Gaussians per scene and training peaks above 20 GB.
- No collision geometry or semantics: gameplay needs separate meshes or proxies for physics and pathfinding.
- Editing and stylizing splats is harder than editing meshes; tooling and a ratified interchange standard were still arriving in 2026.
relations
2 stored edgesDecoration and beauty
Photoreal props and scenery can be captured from photos in under an hour of training instead of modeled and textured by hand, enlarging decor catalogs that beauty scores consume; lighting is baked into the capture.
Soft-body deformation
PhysGaussian deforms captured objects with MPM directly on the rendered Gaussians, so a scanned real object can bend, dent, or flow without building a node-beam or finite-element mesh first.
backlinks
computed from other recordsclaims
5 with sourcesKerbl et al. (2023) reached real-time (30 fps or more) novel-view synthesis at 1080p with state-of-the-art visual quality and competitive training times.
A full 3DGS scene trains in about 35 to 45 minutes versus about 48 hours for Mip-NeRF 360, but uses more memory than NeRF methods, with peak training memory over 20 GB, and shows popping and elongated-splat artifacts.
PhysGaussian simulates elastic, metal, non-Newtonian, and granular behavior directly on Gaussian kernels with an MPM solver, avoiding meshing so the simulated and rendered representation are the same.
Khronos released a KHR_gaussian_splatting glTF 2.0 extension as a release candidate on 2026-02-03 with contributors including Autodesk, Cesium, Esri, Niantic Spatial, and NVIDIA, expecting ratification in Q2 2026.
Generated-world products export splats as their highest-fidelity format and render them in the browser (World Labs Marble with the Spark renderer).