SplatReady

https://www.youtube.com/watch?v=5uP29nmTIus

Convert video to COLMAP datasets ready for Gaussian splat training — a plugin for LichtFeld Studio.

SplatReady automates the full pipeline from drone/camera video to a training-ready dataset:

1. Frame Extraction — Extract frames from video files using FFmpeg, with automatic GPS embedding from SRT files
2. 3D Reconstruction — Align images and produce a COLMAP-compatible dataset using your choice of reconstruction tool
3. Import — One-click import into LichtFeld Studio to start training

Requirements

• LichtFeld Studio v0.4.2 or later
• FFmpeg — must be on your system PATH (used for frame extraction)
• One of the following reconstruction tools:

COLMAP (free, open source)

• Download from colmap.github.io
• Windows: download the pre-built binary, the executable is at colmap-x64-windows-cuda/bin/colmap.exe

Agisoft Metashape (commercial)

• Download from agisoft.com
• Default install path: C:\Program Files\Agisoft\Metashape Pro\metashape.exe
• Requires a Professional license for headless/scripted processing

RealityScan (free from Epic Games)

• Download via the Epic Games Launcher — search for "RealityScan"
• Default install path: C:\Program Files\Epic Games\RealityScan_2.0\RealityScan.exe

Installation

Copy the splat_ready folder into your LichtFeld Studio plugins directory:

~/.lichtfeld/plugins/splat_ready/

Or install from this repo:

git clone https://github.com/jacobvanbeets/SplatReady.git ~/.lichtfeld/plugins/splat_ready

LichtFeld Studio will automatically install the plugin dependencies (Pillow, piexif) on first load.

Usage

1. Open LichtFeld Studio — the SplatReady tab appears in the main panel
2. Stage 1: Frame Extraction
   • Select a video file (or folder of videos) and an output folder
   • Set the frame rate (e.g., 0.5 fps for drone footage)
   • If you have DJI SRT files alongside your videos, GPS data will be embedded into the extracted frames automatically
3. Stage 2: 3D Reconstruction
   • Choose your reconstruction method (COLMAP, Metashape, or RealityScan)
   • Browse to the executable
   • For COLMAP: optionally enable fisheye mode and adjust max image size / min scale
4. Click Run Pipeline
   • A progress bar shows the current stage
   • When finished, click Import Dataset to load directly into LFS for training

Tips

• You can skip Stage 1 if you already have extracted frames — just check "Skip" and point to your frames folder
• You can skip Stage 2 if you already have a COLMAP dataset — just check "Skip"
• All settings are saved between sessions

Output Structure

The plugin produces a standard COLMAP dataset layout:

[Base Output]/
├── frames/
│   └── [VideoName]/          # Extracted JPEG frames with GPS EXIF
└── colmap/
    └── undistorted/
        ├── images/           # Undistorted images
        └── sparse/
            └── 0/
                ├── cameras.txt
                ├── images.txt
                └── points3D.txt

License

MIT

Credits

Built for LichtFeld Studio by Jacob van Beets.

Co-Authored-By: Oz oz-agent@warp.dev