nerfstudio Introduces gsplat

We’re about to see quite a few developments from the nerfstudio team and today we’re starting to get a glimpse at their newest addition, gsplat. While it is from the team at nerfstudio, this is separate from the nerfstudio platform.

Please note that this is not a full interactive 3D Gaussian Splatting platform, but it’s very, very close. The idea behind gsplat is to reimplement 3DGS backend code so that you can train gaussian models with it and also to make it a bit more modular and python friendly for other developers to build ontop of. If you can imagine Nerfacto is to nerfstudio’s NeRF method is like gsplat is to Gaussian Splatting in nerfstudio.

gsplat is be a great tool to go hand in hand, with the soon to arrive nerfstudio implementation of 3D Gaussian Splatting. Like so much of nerfstudio, gsplat is meant to empower those by making itself opensource and modular for people to build on top of. For gsplat, comes with making gaussians python friendly for developers.

We’ll have to see what features the nerfstudio implementation launches with, but given how quickly 3DGS has grown, I think it’s great that we have a modular system to build on top of. As soon as the nerfstudio implementation goes live, we’ll be here reporting on it.

This install has been tested on linux with pytorch version 2.0.1 and cudatoolkit 11.8. Compiling on Windows is still a work in progress, but I have been speaking directly to the team who have been hard at work to expand access to Windows users. I will update this article as that changes.

I’ve been noticing a lot of the recent announcements and platforms have been referencing the Renaissance. It feels about right.