OpenSplat, the free and open source implementation of 3D gaussian splatting written in C++ has pushed several exciting updates over the last couple of weeks.
When we first covered OpenSplat, we looked at some of the lofty roadmap goals they had in mind. With only a handful of months gone by, they’ve seemingly delivered on quite a few promises. Starting with Mac.
It is now possible to run OpenSplat on Mac, but you will need to build it yourself. However, it’s not that hard to get up and running. The OpenSplat Github further provides step by step documentation on how to use Mac. It also does not matter whether you are using Apple Silicon or on an Intel chip. OpenSplat will work with both.
I trained one of the sample scenes on an M2 yesterday evening and it trained to 2,000 steps in about 12 minutes. While certainly slower than if I were to run on a NVIDIA GPU, it’s really not bad. For 7,000 steps it took roughly 50 minutes. Given that a year ago it would take about a week to run nerfstudio’s nerfacto on a Mac, this feels lightening fast! You can also use the -d option which will use downscaled versions of the input images (for lower resolution outputs), but will allow the process to run faster.
Additionally, some of the other features include that OpenSplat can now be run only on CPU.
You are also now able to process data directly from COLMAP, OpenSfM and ODM, expanding the first stages of pretty much every radiance field method. This grows the possibilities of utilizing more methods, such as SuGaR to pull accurate meshes. This means you no longer need to pass the data through nerfstudio’s ns-process-data pipeline, which was required to use OpenSplat before.
Finally, users can preserve the original reconstruction coordinate reference system. If you’re building an application that displays both point cloud (or mesh) outputs from structure from motion and 3DGS in the same space, you want the models to be in the same coordinate system.
If you’re using Windows, they’ve actually created a single click binary that you can use to build OpenSplat. They’re asking $10 currently for a copy with included updates for the first year. All purchases go back to support the project. Additionally, they do offer a 14 day free trial, so what do you have to lose? It is important to note that this binary only currently works with NVIDIA GPUs.
Getting OpenSplat running on AMD is an ongoing process, but for now you’ll need to stick with NVIDIA. Check out the full Github page here!
New to Gaussian Splatting? Start here
