nicolasvonluetzow/GaussianGPT ? reverse-engineered prompt

Build me the full project for autoregressive 3D Gaussian scene generation, where I can train a scene compressor, turn scenes into tokens, train a transformer on those tokens, and then sample new scenes back into renderable 3D Gaussians.

I want one clear pipeline that can do generation, completion, and outpainting from the same model, with a simple way to run each step from the command line. Please include training scripts, tokenization, decoding, and a couple of rendering helpers so I can preview the results from orbit and top down views. Use the existing config driven setup, and make it easy to change settings from the command line.

If you need to, look up the current docs for any tricky dependencies or GPU setup details, and make the code runnable as a real research project rather than just a demo.