The gaussian-splatting repository is the official reference implementation of the paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" by researchers from INRIA's GRAPHDECO team, including Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. The project addresses a fundamental challenge in novel-view synthesis: achieving high-quality photorealistic rendering of scenes captured from multiple photographs or videos while maintaining real-time performance at 1080p resolution and above.
The core innovation of this work involves three key technical contributions. First, the method represents scenes using 3D Gaussians initialized from sparse points generated during camera calibration, which preserves the beneficial properties of continuous volumetric radiance fields while avoiding unnecessary computation in empty space. Second, the approach employs interleaved optimization and density control of the 3D Gaussians, with particular emphasis on optimizing anisotropic covariance to achieve accurate scene representation. Third, the authors developed a fast visibility-aware rendering algorithm that supports anisotropic splatting, simultaneously accelerating training and enabling real-time rendering capabilities.
The codebase is structured around four main components: a PyTorch-based optimizer that produces 3D Gaussian models from Structure-from-Motion inputs, a network viewer for visualizing the optimization process, an OpenGL-based real-time viewer for rendering trained models, and a utility script for converting personal image collections into optimization-ready datasets. The optimizer requires a CUDA-capable GPU with Compute Capability 7.0 or higher and 24 GB of VRAM for training to paper evaluation quality, along with CUDA SDK 11 and appropriate C++ compiler support.
Recent updates documented in the repository include training speed acceleration integrated in October 2024, made compatible with depth regularization, anti-aliasing, and exposure compensation features. The same update enhanced the SIBR real-time viewer with bug fixes and new functionality in the Top View feature for visualizing input and user cameras. An earlier update in Spring 2024 added OpenXR support for VR viewing, contributed by Orange Labs. The project provides substantial resources for users including pre-trained models totaling 14 GB, evaluation images of 7 GB, Windows viewers of 60 MB, and the T&T+DB COLMAP dataset at 650 MB. The research was funded by the ERC Advanced grant FUNGRAPH No 788065 and supported by computational resources from Université Côte d'Azur and GENCI-IDRIS.
