The original 3D Gaussian Splatting repository continues to receive updates, albeit infrequently. Just today, the development team merged several exciting updates into the repository, enhancing both performance and usability.
Training speeds are getting a big boost from integrating drop-in replacement from Taming-3DGS along with fused SSIM into the original codebase. The optimized rasterizer delivers:
1.6× speedup using
--optimizer_type default2.7× speedup using
--optimizer_type sparse_adam
To benefit from faster training times, install the accelerated rasterizer in your environment:
Additionally, the SIBR Viewer is getting a few updates and bug fixes with added features in the Top View, allowing visualization of input and user cameras.
After installation, use the sparse Adam optimizer when running train.py by adding:
Compatibility with Previous Updates
The newest update is also fully compatible with previous enhancements, including:
Depth Regularization: Utilizes depth maps as priors during optimization with each input image. This technique is particularly effective on untextured areas (e.g., roads) and helps eliminate floaters. Similar approaches have been adopted in:
The depth regularization integrated here is based on their Hierarchical 3DGS paper, applied to the original 3DGS. While it significantly improves quality in some scenes (e.g., DeepBlending scenes), the impact varies and can sometimes be negligible or negative. Additionally, it is currently not possible to use depth regularization with the training speed acceleration since they use different rasterizer versions.
Anti-Aliasing: Reduces aliasing artifacts to improve image quality.
Exposure Compensation: Adjusts exposure levels for consistent image brightness.
Please be aware that this original implementation is not permitted for commercial use without obtaining a license. For more details, refer to the GitHub repository.
New to Gaussian Splatting? Start here
