felixwindisch/lodofgaussians

This is the official research code for a paper presented at SIGGRAPH 2026 that tackles a specific challenge in 3D scene reconstruction: how to train and render extremely large scenes, such as entire city blocks or large outdoor environments, on consumer-grade graphics cards that do not have enough memory to hold the whole scene at once. The technique is built on 3D Gaussian Splatting, a method for representing 3D scenes as clouds of small translucent blobs called Gaussians, which can be rendered very quickly. The core innovations are out-of-core streaming, meaning the scene data is stored on disk and only the parts currently needed are loaded into GPU memory, and Level of Detail (LoD), which automatically shows a simplified version of distant parts of the scene and full detail for nearby parts. Together these allow training and real-time rendering of scenes far larger than what would normally fit in GPU memory. In practice you would use this if you have a large collection of photographs of a big outdoor space and want to reconstruct it as a 3D scene you can navigate in real time. The codebase is written in C++ with a Python training interface and requires a CUDA-capable GPU.