gsplat is a high-performance differentiable rasterization engine for 3D Gaussian splatting, designed for real-time novel view synthesis from 2D images. It provides a complete pipeline for reconstructing 3D scenes by optimizing differentiable Gaussian representations, training models from COLMAP-processed captures or proprietary device files, and generating new viewpoints through a CUDA-accelerated rendering backend.
The framework distinguishes itself through memory-optimized CUDA kernels that reduce training memory usage by up to 4x compared to standard implementations while matching published quality metrics. It supports large-scale scene reconstruction with millions of Gaussians through gradient-based densification and pruning strategies, multi-GPU distributed rendering for handling massive scenes, and antialiased rendering to minimize visual artifacts. Additional capabilities include rendering arbitrary-dimensional feature vectors beyond RGB, compressing scene representations for reduced storage, and providing a browser-based interactive viewer for real-time exploration of trained models.
The project covers the full workflow from scene optimization and rendering to performance profiling and quality evaluation, with automated benchmark execution for reproducing standard metrics on common datasets.