OpenSplat is a command-line C++ implementation of 3D gaussian splatting. It reads camera poses and sparse points from ODX, OpenSfM, COLMAP, OpenMVG or nerfstudio projects and writes a scene as .ply, .splat, .spz or .rad. It trains on NVIDIA (CUDA), AMD (ROCm/HIP) and Apple (Metal) GPUs, or on CPU at roughly 100x slower.
Who it is for: Photogrammetry and 3D capture users training gaussian splats from SfM output
Strengths
Runs on CUDA, ROCm/HIP, Metal and plain CPU from one codebase
Reads ODX, OpenSfM, COLMAP, OpenMVG and nerfstudio projects directly
Supports image masks, resuming from a .ply and compressed .splat output
Documents GPU memory use: about 2 GB per million gaussians
Weaknesses
Requires sparse points in the input; random initialization is not supported
Source build needs OpenCV and a matching libtorch, which takes manual setup
Pre-built Windows binary is sold; free route is building from source
AGPL-3.0 license constrains use in closed-source products
What it needs
GPU optional
Docker
Needs libtorch, OpenCV, CUDA, ROCm, Xcode Metal toolchain, CMake