Gaussian Splatting
Inria's reference code for 3D Gaussian Splatting, turning posed photo sets into real-time scenes.
About
This is the authors' own code for the 2023 ACM Transactions on Graphics paper that made Gaussian splatting the standard way to capture real places in 3D. You supply photos posed with COLMAP (a convert script automates that), and a PyTorch optimizer with custom CUDA rasterization fits a cloud of colored, semi-transparent 3D Gaussians to them, adding and pruning points as it trains. Finished scenes render new views at interactive frame rates in the bundled SIBR tools, a live network viewer for training and an OpenGL viewer for results, with prebuilt Windows binaries. An October 2024 update added faster training, depth regularization, exposure compensation and anti-aliasing. Requirements are steep: a CUDA GPU of compute capability 7.0 or newer, about 24 GB of VRAM to match the paper (less with documented tradeoffs), and a C++ toolchain with CUDA 11, verified on Ubuntu 22.04 and Windows 10. The Inria and Max Planck license allows research and evaluation only, so commercial work needs a separate deal or an independent implementation like the Apache-licensed gsplat. Best for researchers and graphics developers.
Should you use Gaussian Splatting?
Pick it when
Use the original Inria code when you need the reference 3DGS implementation for a paper baseline or to reproduce published results from a multi-photo capture; the authors list 24 GB VRAM for paper-quality training, less with reduced settings.
Look elsewhere when
Skip it for products or new pipelines: it is released for research use, and gsplat offers an Apache-2.0 rasterizer with less training time, lower memory, and more camera models. Nerfstudio adds a viewer and data tools.
Alternatives to Gaussian Splatting
- gsplat
Apache-2.0 for commercial use, up to 4x less GPU memory and faster training, plus fisheye and LiDAR cameras, but it is a library, not a reference app.
- Nerfstudio
Apache-2.0 framework with an interactive web viewer and data-processing utilities around radiance-field training, more turnkey than bare research code.
- Meshroom
Produces a textured mesh you can edit in DCC tools instead of splats, with an easier node UI and MPL-2.0 license, but no real-time radiance field rendering.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- 3D Model Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Advanced (4/5)
- License
- Gaussian-Splatting License (non-commercial)
- Minimum VRAM
- 8 GB
- Added
- Apr 3, 2026