3D Gaussian splatting is a scene representation and rendering technique that models a three-dimensional scene as a large set of oriented, semi-transparent 3D Gaussians, each with position, shape, color and opacity. Introduced in the 2023 SIGGRAPH paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" by Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler and George Drettakis of Inria and the Max Planck Institute, it matched or exceeded the visual quality of neural radiance fields while rendering at real-time frame rates, and quickly became the dominant approach to novel-view synthesis.
How it works
Starting from a sparse point cloud produced by classical structure-from-motion, each point becomes a 3D Gaussian whose parameters are optimized by gradient descent so that rasterized renderings match the input photographs. Unlike a NeRF, which stores the scene implicitly in neural network weights and renders by costly ray sampling, Gaussian splatting is an explicit representation rendered by fast GPU rasterization: Gaussians are projected, or "splatted", onto the image plane and alpha-blended in depth order. Densification and pruning during training add detail where needed and remove redundant primitives.
Adoption
Within a year of publication the method spread through visual effects, game engines, drone mapping, real-estate capture and smartphone scanning apps, with support landing in Unreal Engine, Unity and consumer capture tools. Research extensions added dynamic scenes, animatable human avatars, and integration with text-to-3D generation pipelines, where splats serve as the optimized representation. The technique also feeds embodied AI and robotics simulation, producing photoreal environments from casual video capture.
Relation to the field
Gaussian splatting is often described as the practical successor to NeRF: both belong to the radiance-field family born from computer vision and deep learning research, but splatting trades implicit neural compactness for explicit primitives and rendering speed. Active research explores compression, relighting and hybrid neural-explicit models.