graphdeco-inria/gaussian-splatting

gaussian-splatting is the official reference code from a research team at Inria for a 2023 computer-graphics paper, "3D Gaussian Splatting for Real-Time Radiance Field Rendering." Its purpose is "novel-view synthesis": you give it a collection of ordinary photos or video frames of a real-world scene from different angles, and it learns a 3D model that lets you move a virtual camera around and see what the scene would look like from viewpoints it was never shown. The README's abstract explains the approach. Earlier "radiance field" methods produced beautiful images but were costly to train and slow to render, and faster variants tended to sacrifice visual quality. This work represents the scene as a cloud of millions of tiny 3D Gaussians, soft ellipsoids carrying colour and transparency, initialised from sparse 3D points produced by standard camera-calibration software. During training, the system jointly optimises where each Gaussian sits, its shape and orientation (what the authors call anisotropic covariance), and how dense the cloud is, adding and removing Gaussians where needed. A custom visibility-aware rendering algorithm "splats" these Gaussians onto the image plane very quickly, speeding up both training and playback to roughly thirty frames per second at 1080p. You would use this for computer-graphics or computer-vision research on scene reconstruction, for turning phone-shot videos into walkable 3D environments, or for experimenting with VR (an OpenXR viewer has been contributed). The optimiser is written in Python with PyTorch and custom CUDA extensions, there is also an OpenGL-based real-time viewer and a script to prepare datasets from your own images.