About nerfstudio-project/gsplat

The open, permissive CUDA library for Gaussian splatting maintained by the Nerfstudio team, accompanied by the mathematical write-up Mathematical Supplement for the gsplat Library (Ye et al., 2023). See the paper page for context.

gsplat re-implements the core differentiable rasterization of 3D Gaussians with explicit, documented gradients suitable for research extensions. It is the backbone used by Nerfstudio's 3DGS training, and downstream methods such as MVSplat, Scaffold-GS, and Feature 3DGS plug into its CUDA kernels. Mirrored here under Apache-2.0; only the README is included.