Gaussian Splatting

A novel view synthesis method: it models a scene with millions of explicit 3D Gaussians, enabling photorealistic, real-time radiance field rendering at $\ge 100$ FPS.

Gaussian Splatting (3DGS) is a 2023 breakthrough for 3D reconstruction and novel view synthesis, introduced by Bernhard Kerbl et al.. It replaces the slow, implicit neural network of NeRF with an explicit, high-fidelity representation: a large collection of 3D Gaussians (ellipsoids). Each Gaussian is defined by parameters like position, anisotropic covariance, and color, and is optimized using a fast, differentiable Gaussian rasterizer. This explicit structure delivers massive performance gains: training is up to 50 times faster than NeRF, and rendering achieves true real-time rates, consistently hitting $\ge 100$ FPS at 1080p resolution.

https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/