When LichtFeld Studio first emerged from Mr.NeRF, it stood out for being an easy and free way to process gaussian splatting locally. However, over the following months and several bounties, several improvements hit the platform, leading to a recent culmination of nightly builds.
It feels like the pace is picking up yet again, with the release of v0.4. There are several new updates in this release, compared to the nightly builds, starting with depth aware training. The new depth based grid introduces a stronger geometric backbone to the optimization process. Depth cues help stabilize training behavior and you feel it most in difficult datasets such as interiors, thin structures, and scenes with layered occlusion.
Training configurations can now be imported and exported. Parameters can be versioned, shared, reused, and resumed. The interface has been reworked to support this shift.
Editing reconstructions has also become more intentional. The addition of a mirror tool allows users to quickly rotate a splat and see the other side. Crop tools are also more precise and transform handling has been improved. Orbit navigation no longer collapses into gimbal lock and there are even subtle fixes, like ensuring deleted splats don’t silently persist in exported PLYs.
LichtFeld literally now speaks more languages. A full localization system has been introduced, with support expanding rapidly across Japanese, Chinese, Korean, and much of Europe. Just as importantly, technical parameter names remain consistent in English. That localization effort forced the project to confront Unicode path handling. In previous versions, datasets stored in non ASCII directories could crash the application at import, export, logging, or checkpoint save. v0.4 systematically eliminates those failure points by handling Unicode paths end to end.
Additionally the release comes with infrastructural changes. Portable builds are more reliable and Windows CUDA compilation paths are less brittle. Dependency baselines have been updated. Test coverage has been repaired and expanded.
Performance tuning continues in parallel. CUDA streams are used more effectively, with the Viewer VRAM usage reduced, and training memory pressure has been eased through both architectural changes and practical controls like render resolution scaling.
There is still a ton of progress occurring daily within Lichtfeld Studio. The platform will remain GPL-3.0 licensed and can be downloaded from their GitHub.
New to Gaussian Splatting? Start here
