3D Gaussian Splatting

3D Gaussian Splatting reconstructs photographed scenes so you can view them from new angles. It's the authors' reference implementation for researchers and developers working on 3D capture, with support for complete scenes and distant backgrounds.

• Scene reconstruction: Build 3D Gaussian models from COLMAP or Synthetic NeRF datasets, with an image-processing tool for preparing your own photos.
• Interactive viewing: Explore trained scenes in real time or watch reconstruction progress through a separate training viewer.
• Rendering controls: Training supports depth regularization, anti-aliasing and exposure compensation.
• Evaluation: Render held-out views and calculate image quality metrics, with pretrained models available for comparison.

The software runs locally on Windows and Linux. Training and the real-time viewer require a CUDA-capable GPU. Training at the paper's evaluation quality calls for 24 GB of VRAM, though smaller configurations are possible; the viewers recommend 4 GB. Prebuilt viewers are available for Windows. A separate OpenXR branch supports VR viewing.

This is the authors’ reference implementation of 3D Gaussian Splatting, distributed through source installation and downloadable pretrained scenes. Its Gaussian-Splatting License permits research and evaluation and prohibits commercial use without prior permission from the licensors.