A from-scratch PyTorch implementation of 3D Gaussian Splatting (3DGS) for novel view synthesis. Built to understand the math behind the original paper and the engineering choices in the reference CUDA implementation.

Key Features

• Pure PyTorch implementation — rendering and adaptive control in <500 lines of Python code for readability

[!TIP] Technical Deep Dive: For a derivation of the EWA projection math, the adaptive control logic, and learnings when reading and comparing with the reference implementation, see my Learning Notes →