3D Gaussian Splatting
3D Gaussian Splatting renders photo-real novel views of captured scenes in real time by optimizing millions of anisotropic 3D Gaussians, replacing slow NeRFs.
Scene = 3D Gaussians with position, anisotropic covariance, opacity and colour, trained with interleaved optimization and density control and a fast visibility-aware rasterizer. Became the dominant 3D representation for capture and the output format of later world models such as Marble.
- Date
- Tuesday, 8 August 2023
- Lab
- Inria / Max Planck (Kerbl et al.)
- Kind
- paper
- Access
- open weights (restricted license)
arXiv v1 2023-08-08 (SIGGRAPH 2023 paper). The reference code is under Inria's non-commercial licence, hence the access label. Authors are Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler and George Drettakis.
Sources
This record was checked against its sources on 6 October 2026. How we check