taichi_3d_gaussian_splatting

Why we included this project

Novel-view synthesis and photogrammetry teams often hit a wall with the official 3D Gaussian Splatting code, which leans heavily on CUDA and is hard to follow. This project reimplements the same radiance-field method in pure Python on the Taichi language, keeping the code readable while still rendering new views in real time. Training takes multi-view images, camera poses, and a sparse point cloud, runs them through a differentiable rasterizer, and produces a dense point cloud with per-point covariance and color. Because the scene lives in a point cloud rather than a neural volume, merging separate scenes or objects is straightforward, something classic NeRF pipelines struggle with. It is a solid reference for studying the algorithm internals or tweaking the training loop, though it currently targets the CUDA backend and runs slower than the official build, so treat it as a learning and prototyping base rather than a production renderer.