gsplat

Born out of the Nerfstudio project, gsplat packages CUDA-accelerated rasterization of 3D Gaussians behind clean Python bindings and has become the standard open library for Gaussian splatting research and production pipelines. Compared with the original Inria implementation it trains with up to 15 percent less time and up to 4x less GPU memory, and it layers on features the reference code never had: pinhole, fisheye, and LiDAR camera models, NVIDIA 3DGUT integration, batched rendering of arbitrary scene sets, TorchScript deployment, and experimental HiGS inference rendering. Installation is pip install gsplat with just-in-time kernel compilation on first run, or prebuilt wheels for common Python, PyTorch, and CUDA combinations; an NVIDIA GPU is required. Contributors from UC Berkeley, NVIDIA, ShanghaiTech, Amazon, and Meta maintain the library, with documentation at docs.gsplat.studio and version 1.5.3 current as of May 2025 and active commits into 2026. Everything ships under Apache-2.0, so commercial use is unrestricted, which is why many downstream splatting toolkits build on it; the repo holds about 5.5k stars.