Anyone willing to dig through the code or papers for the actual algorithm? It looks like the GitHub and papers have not been optimized for communication.