The graphdeco-inria/gaussian-splatting repository provides the official reference implementation for the method described in the paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering." This technique addresses the challenge of real-time, high-quality novel-view synthesis from multi-view images or videos, a domain previously dominated by neural radiance field (NeRF) methods that are computationally intensive and slow to render. The repository aims to deliver state-of-the-art visual quality while enabling real-time rendering at 1080p resolution, making it suitable for both research and practical applications in computer graphics and vision.
The core innovation of the method is the representation of a scene using 3D Gaussians, which are optimized from sparse points generated during camera calibration. This approach preserves the desirable properties of continuous volumetric radiance fields but avoids unnecessary computation in empty space, resulting in efficient scene optimization. The method further incorporates interleaved optimization and density control of the 3D Gaussians, including anisotropic covariance optimization for accurate scene representation. A fast, visibility-aware rendering algorithm supports anisotropic splatting, which accelerates both training and real-time rendering.
The repository is structured around four main components: a PyTorch-based optimizer for generating 3D Gaussian models from Structure-from-Motion (SfM) inputs, a network viewer for visualizing the optimization process, an OpenGL-based real-time viewer for rendering trained models, and a script to convert user images into optimization-ready SfM datasets. The codebase is compatible with Windows 10 and Ubuntu Linux 22.04, and includes detailed instructions for hardware and software setup. Training requires a CUDA-ready GPU with at least 24 GB VRAM for evaluation-quality results, and the recommended setup uses Conda for environment management.
Users can train models using their own datasets or provided sample datasets, such as Tanks&Temples and Deep Blending, and evaluate performance using built-in scripts. The repository includes pre-trained models, evaluation images, and viewers for Windows, facilitating both experimentation and benchmarking. The training process is highly configurable, with numerous command-line arguments allowing fine-tuning of learning rates, resolution, evaluation splits, and other parameters. Evaluation scripts enable users to render images and compute error metrics, supporting rigorous comparison with other methods.
Recent updates have introduced features such as training speed acceleration, depth regularization, anti-aliasing, exposure compensation, and improved visualization tools. OpenXR support for VR viewing has also been added, broadening the applicability of the method. The repository provides pre-built binaries for Windows, and the SIBR framework underpins the interactive viewers, allowing real-time exploration of trained models.
Overall, the graphdeco-inria/gaussian-splatting repository is a comprehensive resource for researchers and practitioners interested in real-time radiance field rendering. It offers a robust, efficient, and extensible implementation of 3D Gaussian Splatting, complete with tools for training, evaluation, and visualization, and is backed by extensive documentation and community support.
