lichtfeld-node

Windows GPU splat processing node for MobileScannerPhotogrammetry using LichtFeld Studio.

Runs natively on Windows with real-time training visualization. Polls Firestore for scan tickets, runs COLMAP + LichtFeld Studio training, aligns to ARKit, uploads results.

Quick Start

1. Install Prerequisites

• NVIDIA Driver 570+ (download)
• COLMAP with GPU support (download)
• LichtFeld Studio (download binary) — unzip to C:\tools\LichtFeld-Studio\
• Python 3.10+ (download)

Make sure colmap is on your PATH, or pass --colmap "C:\path\to\colmap.exe".

2. Clone & Setup

git clone https://github.com/theLodgeBots/lichtfeld-node.git
cd lichtfeld-node

Copy service-account.json from the MobileScannerPhotogrammetry project into this directory.

Double-click setup.bat — this creates a Python virtual environment and installs all dependencies.

3. Run

Double-click run.bat or:

.venv\Scripts\activate
python lichtfeld_node.py --lfs "C:\tools\LichtFeld-Studio\bin\LichtFeld-Studio.exe"

The LichtFeld Studio viewer window will open during training so you can watch gaussians form in real time.

Dashboard at http://localhost:8788

What It Does

1. Watches Firestore for splat_status=300 (requested) tickets
2. Downloads dataset from Firebase Storage
3. Converts HEIC → JPG (if needed)
4. Runs COLMAP SfM (GPU-accelerated SIFT, exhaustive matching, undistortion)
5. Trains gaussian splats using LichtFeld Studio with MCMC + bilateral grid + mip filtering
6. Aligns to ARKit coordinate space (Procrustes)
7. Exports PLY → .splat format
8. Uploads to Firebase Storage
9. Updates ticket status to 500 (success)

Configuration

--lfs PATH          Path to LichtFeld-Studio executable (required)
--port 8788         Dashboard port
--data-dir ./data   Working directory
--max-steps 30000   Training iterations
--downsample 1      Image downsample factor (1 = full res)
--strategy mcmc     Training strategy: mcmc (default) or default (ADC)
--colmap colmap     Path to COLMAP executable
--headless          Disable LichtFeld viewer (run without GUI)

You can pass extra args through run.bat, e.g.:

run.bat --max-steps 50000 --headless

LichtFeld Studio Features Used

• MCMC strategy — better gaussian placement with fixed budget
• Bilateral grid — per-image appearance/exposure correction
• Mip filtering — anti-aliasing for multi-scale viewing
• Real-time viewer — watch training progress live (disable with --headless)

Comparison with gsplat-node

Feature	gsplat-node	lichtfeld-node
Platform	Docker (Linux)	Native Windows
Trainer	gsplat (Python/CUDA)	LichtFeld Studio (C++/CUDA)
Rasterization	gsplat	2.4x faster custom CUDA
Real-time viewer	No	Yes
Strategy	default or mcmc	mcmc (default) or ADC
Appearance model	bilateral grid	bilateral grid
License	Apache 2.0	GPLv3

Requirements

• Windows 10/11
• NVIDIA GPU with compute capability 7.5+ (RTX 2060+, tested: RTX 4080)
• NVIDIA driver 570+
• 8GB+ VRAM recommended

License

GPLv3 (due to LichtFeld Studio dependency)