OpenSplat

OpenSplat is a free and open source 3D Gaussian Splatting trainer written in C++ by Piero Toffanin. It appeared in February 2024 as a C++ successor to nerfstudio's Splatfacto model, and the code it carries over from Splatfacto remains under Apache 2.0, copyright The Nerfstudio Team. Writing the trainer in C++ rather than Python removes the interpreter lock and gives direct control over memory use and speed, in line with the project's focus on portability, efficiency and speed.

OpenSplat takes camera poses and sparse points from an existing reconstruction and trains a scene from them. It reads COLMAP, OpenSfM, ODX, OpenMVG and nerfstudio project formats, and since March 2024 it has taken COLMAP, OpenSfM and ODM output directly, without nerfstudio's ns-process-data step, while keeping the original coordinate reference system of the reconstruction. Training writes a .ply, .splat, .spz or .rad scene file along with a cameras.json file.

OpenSplat builds on Windows, macOS and Linux. It trains on NVIDIA GPUs through CUDA, on AMD GPUs through ROCm and HIP and on Apple GPUs through Metal, and it also runs on the CPU alone at roughly a hundredth of GPU speed. Apple GPU training arrived in April 2024 through an MPS backend contributed by Andrew Chan, which cut 2,000 training steps on an M2 laptop from about 15 minutes to about 3. The project estimates GPU memory at about 2 GB per million Gaussians and provides Docker builds for CUDA and ROCm.

OpenSplat is licensed under AGPLv3, and the project states that commercial use is allowed and encouraged under those terms. Building from source is free. A digitally signed prebuilt Windows program sells for US$29 with a year of updates and a 14 day money back guarantee, and it supports NVIDIA CUDA GPUs and the CPU but not AMD GPUs. When Mac support landed in March 2024 the same Windows binary cost about $10. The GitHub repository carries no tagged releases, so source builds follow the main branch.