LichtFeld Studio: a native 3D Gaussian Splatting workstation with a paid Windows build

The gap LichtFeld Studio is trying to close

A typical 3D Gaussian Splatting workflow is a chain of separate programs. One repository trains the splats, a second viewer loads the resulting PLY, a third script performs cleanup, and export to a web-friendly format is somebody's afternoon. Each hop loses context: the training run knows nothing about the edit you are about to make, and the viewer cannot resume a checkpoint.

LichtFeld Studio's README frames the project as "the modular workstation for 3D Gaussian Splatting" and lists the whole chain as one application: train from COLMAP datasets, resume checkpoints, inspect reconstructions in real time, select and transform gaussian subsets and scene nodes with undo/redo, and export to PLY, SOG, SPZ or a standalone HTML viewer. The stated audience is three groups: researchers iterating on reconstruction quality, production teams that need visual inspection and portable delivery without assembling a toolchain, and tool builders who want to drive the app from scripts or agents.

That last group is the differentiator. Most splatting tools stop at a command line or a viewport. LichtFeld Studio exposes a Python plugin system for custom panels, operators, tools and plugin-local dependencies, plus an automation surface described as MCP resources and MCP tools. The repository carries a .mcp.json at the top level and a docs/docs/development/mcp/index.md guide, so the automation story is part of the shipped tree rather than a blog post.

How the pieces fit: training, editing, export, automation

The README describes a single native application rather than a service. Dataset loading and checkpoint resume feed a training loop; the same session can be inspected interactively while training or after convergence. Editing operates on gaussian selections and scene nodes, with history support for undo and redo. Export is the terminal step, with four named targets: PLY, SOG, SPZ and a standalone HTML viewer for sharing.

The implementation is C++23 with CUDA 12.8 or newer, and the repository confirms that shape: CMakeLists.txt, CMakePresets.json, a vcpkg.json manifest, a docker/ directory, and src/ split across the application. Research features named in the README include MCMC optimization, bilateral grid appearance modeling, 3DGUT support for distorted camera models, and timelapse generation. Those are training-side options, not separate binaries.

Extensibility runs in-process. Plugins are Python and can declare their own dependencies, which means a plugin author is not limited to what the host already links. Automation is exposed twice: embedded Python for direct scripting, and MCP resources and tools for external agents and local tools. The repository also ships model weights as separate releases, including LPIPS v0.1 VGG16 weights, SAM 2.1 Hiera base+ weights, and a MoGe-2 ViT-B Normal ONNX model, which suggests the app pulls in segmentation and depth or normal estimation for masking and geometry work rather than relying only on COLMAP input.

Installing LichtFeld Studio on Windows or from source

There is no free prebuilt download on GitHub. The README states that prebuilt Windows binaries are distributed through the LichtFeld Portal, that access is paid, and that you register and make a donation of your choosing to receive the current release plus daily builds. The workflow after that is manual: download the archive, unzip it, run the executable. Binaries are not published on GitHub Releases, so a package manager will not find one.

Building from source is free for everyone under GPLv3, on Windows and Linux. The README points to the project wiki for platform notes and to the repo-local build guide for contributor setup and test commands. The repository also carries a PowerShell helper at the top level:

build_lichtfeld.ps1

The README does not spell out the flags for that script, so read docs/building_and_distribution.md before running it. The hardware floor is explicit: an NVIDIA GPU with compute capability 7.5 or newer (GTX 16-series, RTX 20-series and up) and driver 570 or newer. GTX 10-series and older, AMD and Intel GPUs are not supported.

For a first real use, the documented path is to load a COLMAP dataset, let training run, inspect the reconstruction in the viewport, then export. The README does not print the exact menu path or CLI invocation for that sequence, so treat the wiki as the authority rather than guessing at flags. If you want to script it instead of clicking, the plugin and MCP guides under docs/ are where the automation entry points are documented.

The hardware and licensing constraints are the real adoption cost

The GPU requirement is the hardest boundary in the project. Compute capability 7.5 or newer excludes GTX 10-series cards, all AMD GPUs and all Intel GPUs, and the README states that plainly rather than offering a fallback path. If your workstation is an Apple laptop, there is nothing here for you: one of the recurring search phrases around this project is people asking about macOS, and the README lists Windows as the primary prebuilt distribution target today with source builds on Windows and Linux. No macOS build is described.

The second constraint is the distribution model. GPLv3 covers the source, and building from source is free, but the convenient path, a working Windows binary, is gated behind a donation through the portal. That is a legitimate funding choice for a project that depends on CUDA toolchains and GPU testing, and it is also a procurement problem for teams that need reproducible installs in CI or an auditable artifact chain. A team that cannot or will not build C++23 and CUDA 12.8 from source is effectively buying access.

The third is that the README is thin on operational detail. It does not document rollback, does not list which CUDA minor versions have been validated beyond the 12.8+ floor, and does not describe how checkpoint compatibility is handled across releases. The repository has tests/ and eval/ directories, so there is some verification infrastructure, but the README does not say what the test suite covers.

Where LichtFeld Studio sits against Postshot and gsplat

Two comparisons come up repeatedly in searches around this project: Postshot and gsplat. They are not the same kind of tool, and the difference matters more than any feature checklist.

Postshot is a commercial Windows application for Gaussian Splatting. The practical difference is distribution and source: LichtFeld Studio's source is available under GPL-3.0 and can be built and modified, while Postshot is a closed product. If your reason for choosing a tool is that you need to read, patch or redistribute the code, that distinction decides the question before anything else does. If you only want a working binary with support, the calculus runs the other way, and LichtFeld Studio's paid portal is the closest thing it has to that model.

gsplat is a research codebase from the Nerfstudio ecosystem, Python and PyTorch based. The difference in approach is architectural: gsplat is a library and training implementation you import into your own Python pipeline, while LichtFeld Studio is an application with a viewport, an editing model and an export step built in. A team that already has a PyTorch training pipeline and wants a differentiable rasterizer to call will find gsplat fits without imposing a GUI. A team that wants to load a COLMAP dataset, watch it converge, fix a region by hand and hand someone an HTML viewer has to assemble that themselves around gsplat, and LichtFeld Studio's README claims to provide it in one place. The trade is that you accept a native C++ application and its GPU requirements instead of a Python dependency.

Maintenance, upgrades and what GPL-3.0 means for your build

The repository is not archived, and the last push was on 2026-09-15, so the codebase is being touched. Model weights are also versioned as separate releases: LPIPS v0.1 VGG16 weights on 2026-09-03, SAM 2.1 Hiera base+ weights on 2026-08-28, and a MoGe-2 ViT-B Normal ONNX model on 2026-07-05. Those releases move independently of the application, which means an upgrade can involve two things: a new app build and new weight files, and the README does not describe how the app resolves which weight version to use.

Upgrade cost for source builders is the usual C++ story. The README specifies modern C++23 and CUDA 12.8+ toolchains, a vcpkg manifest for dependencies, and CMake presets, so a toolchain bump on your machine can break a build that worked before. There is no documented migration path for checkpoints between versions, and no rollback procedure in the README.

On licensing: the project is GPL-3.0, and the repository ships a THIRD_PARTY_LICENSES.md alongside LICENSE, which is where you should look for the terms attached to bundled dependencies. GPL-3.0 is a copyleft licence, so the obligations attach to distribution of derivative works, and how that interacts with your own product is a question for your legal counsel, not for this article. One concrete point worth noting: the model weight releases are separate artifacts from the GPL-3.0 source tree, so their terms are not automatically the same, and the README does not state them.

Editorial conclusion

Adopt LichtFeld Studio if you have an NVIDIA GPU of compute capability 7.5 or newer and you want training, inspection, gaussian editing and export in one window: researchers iterating on reconstruction quality, production teams delivering PLY, SOG, SPZ or an HTML viewer, and tool builders wiring scenes into pipelines through Python plugins or MCP. Do not adopt it if you are on macOS, AMD or Intel graphics, or a GTX 10-series card, and do not expect to download a free binary from GitHub Releases: the README states binaries are not published there and that Windows builds come through the LichtFeld Portal behind a donation. Before committing, verify three things: that your driver is 570 or newer, that docs/building_and_distribution.md builds cleanly on your toolchain, and that GPL-3.0 is acceptable for how you plan to ship whatever you produce with it.

Frequently asked questions

Official sources

1. License: GPL-3.0
2. MrNeRF/LichtFeld-Studio on GitHub
3. Project website
4. README
5. Releases