Skip to content

LichtFeld Studio — How-To (Training, Editing, and Export)

This document describes the workflow used for the LichtFeld Studio implementation, including:

  • application setup (Windows prebuilt)
  • dataset loading
  • training
  • editing and inspection
  • export for downstream processing

References

Important Note

For the experiments, a precompiled Windows binary was used.

Instructions for building the software from source are available in the documentation and in the tutorial video.
The procedure described below refers to the precompiled version.

1. Overview

LichtFeld Studio is an implementation of 3D Gaussian Splatting that provides both training and real-time interactive visualization.

It operates on a calibrated multi-view reconstruction and does not estimate camera poses.

The workflow consists of:

  1. Provide a calibrated dataset
  2. Run Gaussian optimization inside the application
  3. Inspect the reconstruction
  4. Export the final scene

2. Requirements

Hardware requirements:

  • NVIDIA GPU (Compute Capability ≥ 7.5)
  • Recommended VRAM ≥ 8GB

3. Installation (Prebuilt Version Used)

For the experiments a precompiled Windows binary was used.

Download from:

https://github.com/MrNeRF/LichtFeld-Studio/releases

After download:

  1. Extract the archive
  2. Open the bin folder
  3. Run the executable

4. Input Dataset

LichtFeld Studio requires a precomputed SfM reconstruction. It does not compute Structure-from-Motion internally but trains Gaussian splats from an existing reconstruction.

The dataset must already contain camera poses and intrinsics (e.g. generated by COLMAP, nerfstudio, Metashape, RealityCapture, OpenSfM).

The preferred way to use LichtFeld Studio is to import input data (undistorted images + pointcloud + camera locations) in COLMAP format:

<location>
|---images
|---sparse
    |---0
        |---cameras.bin
        |---images.bin
        |---points3D.bin

5. Training

The training was performed using the GUI.

  1. Open LichtFeld Studio
  2. Create a new project
  3. Import dataset
  4. Start training

Training runs interactively inside the application.

During optimization the scene can be navigated in real time while Gaussian primitives progressively appear and refine. The interface provides live feedback on the reconstruction and allows monitoring convergence visually.

The training panel exposes parameters such as maximum number of Gaussians, spherical harmonics degree, and optimization settings.

The user can pause, restart, or clear the optimization at any time and immediately evaluate the impact of parameter changes on the reconstruction quality.

The training can be performed also by command line, as shown in the documentation.

6. Export

The trained scene can be exported as a .ply file directly from the application.