LichtFeld Studio

LichtFeld Studio is an open source 3D Gaussian Splatting application started by MrNeRF and licensed GPL-3.0. It trains scenes from COLMAP datasets, resumes from checkpoints, inspects reconstructions in real time, edits gaussian selections and scene nodes with undo and redo, and exports the result, all from a single native desktop application rather than a training script paired with a separate viewer. Since version 0.5.4 a .licht project bundles the dataset, checkpoints, trained models, scene edits, camera paths and workspace layout, and a Project Manager browses, searches, filters, previews, renames and organizes both local and external projects. The repository carries roughly 3,800 stars, and the project is based in Switzerland.

The application is written in C++23 against CUDA 12.8 or newer and runs only on NVIDIA hardware. It requires a GPU at compute capability 7.5 or newer, which means a GTX 16 series card, an RTX 20 series card or anything later, and a pre-flight check refuses to start on parts below that line. GTX 10 series and older NVIDIA cards, AMD cards and Intel cards are not supported at all. Version 0.5.3 replaced the CUDA renderer with a full Vulkan viewer, which changed how a scene is drawn without loosening the CUDA requirement for training. Version 0.5.4 lowers training VRAM with better control of peak memory, reduces viewer memory, releases cached images under memory pressure and reports VRAM use more accurately. The 0.5.4 release cuts training VRAM by 30 percent, speeds up training by 9 percent through an edge-map fix and by 6 percent on high-count scenes, and makes GUT training 25 percent faster, while its larger gains are in file handling, with .licht projects opening up to five times faster, COLMAP import almost five times faster and PLY export 3.5 times faster.

Distribution splits by platform. Prebuilt Windows archives come from the Lichtfeld Portal at portal.lichtfeld.io, which asks for registration and a donation and in return serves daily builds alongside tagged releases. Windows is the only prebuilt target. Linux is supported by building from source, with a Dockerfile in the repository and platform notes in the project wiki. The current tag is v0.5.4.1, tagged on October 18, 2026, which followed v0.5.4 on October 9 and v0.5.3 from June 24, and v0.5.2 shipped in April 2026.

File support covers PLY, SOG, SSOG, SPZ, USD, glTF and GLB, plus a standalone HTML viewer export that packages a scene for sharing without a server and carries labels, multiple measurements, screenshots and VR oriented interaction. Projects also publish directly to the Gallery on lichtfeld.io with a title, description, thumbnail, camera view, field of view, background and presentation settings, and changes made on the Gallery side sync back to the project. Research facing features include MCMC optimization, bilateral grid appearance modeling, 3DGUT support for distorted camera models, and timelapse generation. Version 0.5.4 adds normal guided training, which generates normal maps automatically and supervises geometry alongside color through a normal loss, user defined masks and normal and depth priors. The MRNF strategy gains far field reconstruction and improved opacity handling, pruning, initialization and resume, the GUT path improves exposure correction and undistortion, and camera models now cover fisheye lenses and more complex distortion. The interface has moved off Dear ImGui and from GLFW to SDL3, and headless operation is supported for batch work.

A Python plugin system with a Plugin Manager lets third parties extend the application, and an MCP surface alongside embedded Python exposes the same operations to external scripts and agents. Moshpit360 brought one click 360 video ingestion to the application in May 2026, and Alex Gee's 360 Plugin, licensed GPL-3.0-or-later, adds a six stage pipeline that demuxes DJI .osv and Insta360 .insv containers, masks the camera operator with Meta's SAM 3, and aligns frames through COLMAP 4.1 using native equirectangular and fisheye camera models. That plugin requires LichtFeld Studio 0.5.0 or newer, and its first install pulls roughly 4.2 GB of CUDA dependencies.

Production work already runs through it. A gem bouquet from the Natural History Museum Vienna was reconstructed as a 3D Gaussian splat by Geofront in May 2026. LichtFeld Studio is a free download.