This started as a $1700 bounty to optimize the CUDA Gaussian Splatting renderer. The PR delivers impressive performance improvements with a 2.4× speedup. Training time for 7 scenes from the mipnerf360 dataset dropped from 50 minutes to ~20 minutes without impacting final splat quality.
This is the new sota and probably the fastest implementation out there. I would love to see if someone could improve here and get another factor. More details are found in the repo.
Open source software deserves celebration. Its impact on the world is profound and transformative. Recently, I began porting the 3D Gaussian splatting paper to C++ and CUDA with an aim to optimize the training speed of such models. This endeavor was only possible because the original authors open-sourced their work. The feedback has been encouraging, and I feel that I'm making a difference—perhaps even more than in my regular job. That said, it's essential to remember that contributing to open source, while rewarding, can be demanding if it's not your full-time occupation. Yet, through such efforts, we ensure these tools remain freely accessible and aren't monopolized by corporations.
I provide you the link to this GitHub project in case you are interested.
https://github.com/MrNeRF/gaussian-splatting-cuda
Original bounty: https://github.com/MrNeRF/gaussian-splatting-cuda/issues/135 PR: https://github.com/MrNeRF/gaussian-splatting-cuda/pull/245