graphdeco-inria/gaussian-splatting

This is the official reference implementation from the paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering," by researchers at Inria, the University of Côte d'Azur, and the Max Planck Institute, published in ACM Transactions on Graphics in 2023. In plain terms, it tackles a problem in computer graphics: given a set of ordinary photos or video frames of a scene from different angles, can a computer figure out the scene's 3D shape and render brand-new views, sweeping the camera through a room you only photographed from a few spots, at real-time frame rates and high resolution? Earlier radiance field methods such as neural-network-based approaches could produce beautiful results but were slow to train and render. This work represents the scene as millions of small fuzzy 3D blobs called Gaussians, each with position, shape, colour and transparency, starting from the sparse point cloud you get from a camera-calibration step. It then iteratively optimises and adds or removes Gaussians so that, when projected onto each input image, they reproduce what was photographed. A custom rendering algorithm splats these Gaussians onto the screen fast enough to display at 1080p and at least 30 frames per second. You would use this as a graphics or vision researcher exploring novel-view synthesis, or as a developer building tools for VR, virtual production, or photogrammetry visualisation. The optimiser is written in Python with PyTorch and CUDA extensions; a real-time viewer is provided in OpenGL, with later updates adding OpenXR support for VR. A CUDA-ready GPU with 24 GB VRAM is recommended for paper-quality training. The full README is longer than what was provided.