fastgs/FastGS — reverse-engineered prompt

Build me a clean, usable Python project for FastGS, a tool that trains 3D Gaussian Splatting scenes much faster while keeping good rendering quality.

I want it to let someone set up the conda environment, point it at a COLMAP or NeRF style dataset, choose a normal fast training run or a bigger higher quality run, then render and evaluate the trained scene. Keep the workflow simple, with clear scripts for training, rendering, metrics, and full evaluation. Include sensible defaults so a user can get started without understanding every research setting, but still expose the important training options for advanced users.

It should work with CUDA GPUs and PyTorch, use the optimized Gaussian rendering pieces, and explain the expected dataset layout for MipNeRF360, Deep Blending, and Tanks and Temples style data. Please include a short README with setup, hardware notes, example commands, and where outputs go. Look up current docs online if you need to.