2D Gaussian Splatting is a novel-view synthesis and surface reconstruction technique that represents scenes using 2D oriented disks and rasterizes them with perspective-correct differentiable rendering.
The approach addresses the problem of reconstructing geometrically accurate 3D scenes from multi-view images. Rather than using 3D Gaussians, this method represents scenes as surface elements (surfels) - 2D oriented disks - which are rasterized through a custom differentiable rasterization pipeline. The technique includes regularizations designed to enhance reconstruction quality and provides meshing approaches for extracting surfaces from the learned representation, including support for both bounded and unbounded scene reconstruction through adaptive truncated signed distance field methods.
Developers should choose this tool if they need geometrically accurate surface reconstruction alongside novel-view synthesis capabilities. It suits projects requiring high-quality mesh extraction from multi-view data, particularly those working with both bounded indoor scenes and unbounded outdoor environments. The project provides multiple viewing and interaction options through supported viewers including web-based solutions, SIBR integration, and Colab notebooks, making it accessible for both research and practical applications.
The project maintains active development with regular updates addressing performance improvements and feature expansion. Recent work has focused on optimization, with training speed improvements through CUDA operator fusion. The codebase includes a dedicated CUDA rasterization module as a separate repository component, and the project has integrated community contributions including viewer implementations and tutorial resources. Documentation is available through the official gsplat library API reference, and the project benefits from community-contributed resources including comprehensive practitioner guides and multiple viewer implementations.
