This project implements the 3D Gaussian Splatting paper in Taichi language with PyTorch. The algorithm takes multi-view images, sparse point clouds, and camera poses as input, then uses a differentiable rasterizer to train a dense point cloud with covariance, color, and other features. For inference, it renders novel views from any camera pose using the same rasterizer. Unlike NeRF-based approaches, it leverages the point cloud representation for efficient real-time rendering and easier scene merging.