Skip to main content

Lichtfeld Studio

TOOL
📖5 min read
Latest Version: 0.5.3
Lichtfeld Studio: The Best Open-Source 3D Gaussian Splatting Software? (Full Tutorial) by Pixel Reconstruct

Open Source

The application is GPLv3 open source. A third party can read the code and build it, which matters when the pipeline itself is part of your record.

Native Training

It trains directly from a COLMAP dataset, previews the reconstruction in real time, and resumes training from a checkpoint.

Scriptable

Python plugins and MCP automation run the workflow headless. A run is a recorded script, not a series of mouse clicks.

Introduction​

LichtFeld Studio is a native desktop workstation for 3D Gaussian Splatting from MrNeRF. It covers the whole splat lifecycle in one application: train a scene from a photo set, inspect it in real time, edit the splats with selection and masking, and export the result. It runs on Windows and Linux.

Pick LichtFeld Studio when the splat is the working format: you train, clean, and export it here. Pair it with COLMAP for the camera solve, and with the Gaussian Splatting technique for the method, its limits, and the investigation rules. For a mesh-based reconstruction, use the photogrammetry tools instead.

Installation​

  1. Get the application. Two routes:
    • Prebuilt Windows binary from the LichtFeld Portal. Portal access is paid: you register and make a donation of your choice, which funds development.
    • Build from source on Windows or Linux from the GitHub repository. The code is GPLv3 and the build is free.
  2. Verify the version in the application and record it. The current release is 0.5.3 (June 2026).
  3. For the source route: the repository README covers the build requirements (CMake, a CUDA-capable GPU).

Usage​

The standard route solves the cameras in COLMAP first, then trains and cleans the splat in LichtFeld Studio.

Solve the cameras in COLMAP· sparse reconstruction

Run the COLMAP pipeline on the image set. What goes in: photos or extracted frames. What comes out: the image directory plus the COLMAP database and sparse model, which is the input dataset for LichtFeld Studio.

Import the dataset· dataset ready

Load the dataset into LichtFeld Studio. The application reads the images, the camera positions, and the sparse point cloud. Check the camera frustums in the 3D view before you train: a broken alignment is a broken splat.

Train the splat· live preview

Start the training. The viewport shows the scene forming while the GPU works. Record the training settings (iterations, quality mode, filtering). If the run must stop, the checkpoint resumes where it left off.

Inspect and clean· selection and masks

Orbit novel views and check them against the source footage. Select and remove problem regions: sky, background, floaters, ghosts of moving objects. Masks keep excluded regions out of the reconstruction.

Export· splat file

Export the cleaned splat. What comes out: the splat file for Blender (via SkySplat or BlendSplat) and the training settings for the case record.

In an Investigation​

Use the block below to document how you ran LichtFeld Studio on the evidence.

Documentation: how to document LichtFeld Studio

In Your Method Section​

Guidance: Describe the purpose, the tool version, and the input data.

  • Investigative Objective: To reconstruct a navigable 3D scene of the capture for spatial analysis and visual presentation.
  • Software Version: LichtFeld Studio [Version] (e.g., 0.5.3), build from [prebuilt binary / source commit], plus the COLMAP version used for the camera solve.
  • Input Data: [Image set name/ID], [N] images derived from [source footage/document], Source Hash: [First 6 digits].

In Your Decision Log​

Guidance: Record the specific procedural parameters and integrity checks.

  • Specific Settings: Frame selection rationale, training settings (iterations, quality mode, filtering), cleanup operations (selections, masks, outlier removal) and what each removed, export format.
  • Assumptions: The scene was static during the capture; the COLMAP alignment is correct (state the registration ratio).
  • Integrity Check: Analysis ran on a working copy, never the original evidence file.
  • Date Performed: [YYYY-MM-DD].

Verification & Mitigation​

Guidance: How did you verify accuracy and mitigate bias?

  • Verification: Compared novel views against source frames from matching viewpoints. Checked known feature positions in the scene.
  • Bias Mitigation: Treated the splat as a visualization. Routed every measurement to a mesh or the photogrammetry pipeline.

Example Methodology Statement​

"A Gaussian Splat scene was reconstructed with LichtFeld Studio 0.5.3 (build: [route]) from [N] images of [source] (Hash: A1B2C3). Cameras were solved in COLMAP [version]; [N] of [N] images registered. Cleanup removed [regions] with selection and masks. Novel views were verified against the source frames; no dimensions were taken from the splat itself."

Into Blender​

LichtFeld Studio exports the splat for Blender import.

  • Splat file: import with SkySplat or BlendSplat. For rendering-focused review, KIRI 3DGS Render is the third option in the KB.
  • Scale: if the COLMAP dataset was anchored with a known dimension or ground control points, the splat carries that scale. Otherwise the scale is arbitrary: anchor it before you place other objects.
  • Check the orientation of the imported scene and rotate it onto the Blender ground plane.
  • For georeferenced scenes: place the reconstruction relative to the site with the QGIS baseline (custom CRS and origin point).

Limitations​

  • The splat is a rendering approximation. It carries no solid geometry, so no dimension taken from it enters a finding. Derive a mesh or use photogrammetry for measurements.
  • Prebuilt binaries are distributed through the paid portal. The source route is free but needs a build, which is a barrier for a third party who wants to rerun you.
  • It trains from a solved dataset. The camera solve (COLMAP) is still the weak link: bad alignment in, bad splat out.
  • The same capture weaknesses apply: reflective and transparent surfaces, and moving objects, degrade the reconstruction.

Resources​

Sources​

  1. LichtFeld Studio, official site
  2. LichtFeld Studio on GitHub, source, GPLv3 license, and build instructions
  3. LichtFeld releases on GitHub, current version 0.5.3 (2026-06-24)
  4. LichtFeld Portal, prebuilt binary distribution