gsplat

gsplat is an open source library for CUDA accelerated, differentiable rasterization of Gaussians with Python bindings, built by the team behind Nerfstudio and released under the Apache 2.0 licence. Its repository opened in August 2023, and the Nerfstudio team introduced gsplat in October 2023 as a modular reimplementation of the 3D Gaussian Splatting backend, kept separate from the Nerfstudio application itself. The documentation for version 1.5.3 lists Vickie Ye of UC Berkeley as project lead, Ruilong Li of UC Berkeley as lead for version 1.0, and Angjoo Kanazawa of UC Berkeley and Matthew Tancik of Luma AI as advisors, and the repository lists contributors from UC Berkeley, NVIDIA, ShanghaiTech University, Amazon, Meta, IIIT, Luma AI, SpectacularAI, Aalto University and Carnegie Mellon University.

gsplat is a library rather than an application. A PyTorch front end calls optimized CUDA kernels, and its rasterization function renders colour, accumulated depth or expected depth, alone or together, through pinhole, orthographic, fisheye and F-Theta camera models, with an antialiased mode, batched rendering over several scenes and viewpoints, and distributed rasterization across multiple GPUs. A separate function renders 2D Gaussian Splatting. Densification and pruning sit in interchangeable strategy classes, a default strategy and a 3DGS-MCMC strategy, and a PNG based compression API added in August 2024 shrinks 1 million Gaussians from 236 MB to 16.5 MB for a 0.5 dB drop in PSNR, according to the project documentation.

NVIDIA's 3DGUT was integrated into gsplat in April 2025, letting training run directly on images from distorted pinhole and fisheye cameras and on rolling shutter captures without undistorting them first, and gsplat supports 3DGUT only with the MCMC strategy. The bundled examples train on COLMAP captures, fit 2D images with 3D Gaussians and render large scenes in real time, and the GsplatViewer that updated the older nerfview viewer in May 2025 switches between RGB, depth and alpha render modes and between camera models. The documentation reports that the standalone trainer matches the original Inria implementation on PSNR, SSIM, LPIPS and final Gaussian count while using up to four times less GPU memory and up to 15 percent less training time on Mip-NeRF 360 captures.

gsplat 1.0 shipped in June 2024, followed by version 1.1 in July, 1.2 and 1.3 in August, and 1.4 in September 2024. gsplat 1.5 arrived in April 2025, and version 1.5.3, tagged on July 4, 2025, added batching across multiple scenes and the F-Theta camera model. It remains the latest tagged release and the latest version on PyPI. Development continues on the main branch, most recently on September 19, 2026, and the README groups that unreleased work under v1.6.0, including LiDAR rasterization, an experimental HiGS inference rendering path, sparse and multi-GPU dense rasterization, and a new minimum of PyTorch 2.7.

gsplat needs an NVIDIA GPU with CUDA and an existing PyTorch install. Installing with pip from PyPI compiles the CUDA kernels the first time the library runs, installing from source compiles them during installation, and prebuilt wheels for Linux and Windows cover selected PyTorch and CUDA combinations, with the v1.5.3 wheels, all built for Python 3.10, spanning PyTorch 2.0 to 2.4 and CUDA 11.8, 12.1 and 12.4 on Linux and only PyTorch 2.4 with CUDA 12.4 on Windows. Building on Windows requires Visual Studio Build Tools.

The library is described in a white paper first posted to arXiv in September 2024 and published in the Journal of Machine Learning Research in 2025, which reports up to 10 percent less training time and four times less memory than the original 3D Gaussian Splatting implementation. Nerfstudio itself depends on gsplat for its Splatfacto method, and the Nerfstudio package currently pins gsplat 1.4.0.