“3D Gaussian Splatting for Real-Time Radiance Field Rendering,” submitted to arXiv on August 8, 2023 by Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis of INRIA, changed how reconstructed 3D scenes are represented and rendered. Neural radiance fields (NeRF) had shown you could capture a real scene from a set of photos and render it from new viewpoints, but doing so required querying a neural network millions of times per frame, which was far too slow for real-time use.
Gaussian splatting drops the neural network for the scene representation. A scene is described instead as a large collection of 3D Gaussians - fuzzy, ellipsoidal blobs each with a position, shape, color, and transparency. Starting from the sparse 3D points that a structure-from-motion step recovers from the input photos, the method optimizes these Gaussians and adaptively controls their density, splitting and pruning them so detail concentrates where it is needed. A fast, visibility-aware rasterizer then “splats” the anisotropic Gaussians onto the screen.
The result matched or beat the visual quality of the best NeRF methods while rendering complete scenes at 1080p resolution at 30 frames per second or more, with competitive training times measured in minutes. The paper won a best-paper award at SIGGRAPH 2023 and was adopted with unusual speed across graphics, virtual reality, robotics, and visual-effects work.
Why business readers should care: Gaussian splatting made photorealistic, explorable 3D capture fast and cheap enough for real products - virtual tours, e-commerce, mapping, and immersive media. It is a reminder that the right data structure can beat a neural network on its own turf.