Janusch, AKA Mr.NeRF’s inaugural bounty for his open sourced Gaussian Splatting CUDA repo has finally come to a close and we have a definitive winner! The initial goal was to try and speed up reconstruction times by double, across the standard research benchmark scenes. Well that initial bounty was exceeded by quite a bit. End to end time fell from 49 minutes 50 seconds to 20 minutes 30 seconds, which is roughly two and a half times faster.
The winning submission came from Florian Hahlbohm, who you will no doubt recognize from some of his other work such as INPC and HTGS, or his contributions to the more recent Splatshop and SPaGS.
Florian’s result comes from four primary changes that tighten the training path without sacrificing quality. First, a new rasterizer reduces the number of CUDA kernels compared to the previously used gsplat implementation and condenses the math in the forward and backward passes to only the necessary instructions. It also draws on ideas from prior work, including StopThePop and Taming 3DGS. Second, the MCMC densification strategy is optimized by fusing the operations required during every training iteration into a single kernel. Third, the Adam optimizer is replaced with a fused version that runs significantly faster. Fourth, supporting tweaks clean up overhead so the GPU spends more time on useful work rather than coordination or data loading.
The final bounty came out to $1500, donated by Janusch, Vincent Woo, Mazeyar, Anton, Chris, Julien, and Drew. There were a lot of great entries into the bounty and Florian decided to give $400 of his winnings to second place finisher, melonedo, who also managed to achieve the desired speedup.
Janusch’s code is quickly shaping up as an active competitor to both Nerfstudio’s gsplat and Jawset’s Postshot. Janusch has repeatedly stated that he aims to keep this repository free and open source, with any potential sponsors or donations accessible to keep momentum going.
Janusch has stated that he would like to continue to offer bounties for future work. I personally would love to see a bounty that aims to cut down on as many gaussians as possible through a scene, without a meaningful drop in fidelity.
Congratulations again to Florian for the win on Bounty #1. If you would like to read through the threads or follow along for the subsequent bounties, you can do so from the Mr.NeRF and Brush Discord server. If you would like to use Janusch’s repo, you can access it from his GitHub where it carries a GPL-3.0 license.
