LichtFeld Studio

A high-performance C++ and CUDA implementation of 3D Gaussian Splatting

Quick Start • Installation • Usage • Results • Community

Overview

LichtFeld Studio is a high-performance implementation of 3D Gaussian Splatting that leverages modern C++23 and CUDA 12.8+ for optimal performance. Built with a modular architecture, it provides both training and real-time visualization capabilities for neural rendering research and applications.

Key Features

2.4x faster rasterization (winner of first bounty by Florian Hahlbohm)
MCMC optimization strategy for improved convergence
Real-time interactive viewer with OpenGL rendering
Modular architecture with separate core, training, and rendering components
Multiple rendering modes including RGB, depth, and combined views
Bilateral grid appearance modeling for handling per-image variations

Community & Support

Join our growing community for discussions, support, and updates:

Discord Community - Get help, share results, and discuss development
Website - Visit our website for more resources
Awesome 3D Gaussian Splatting - Comprehensive paper list
@janusch_patas - Follow for the latest updates

Active Bounty

Second Bounty: Better 3DGS Initialization and Training without Densification

$2,600 + $500 Bonus Challenge

Details: Issue #284

Previous winner: Florian Hahlbohm (2.4x rasterizer speedup)

Quick Start

# Clone and build (Linux)
git clone https://github.com/MrNeRF/LichtFeld-Studio
cd LichtFeld-Studio

# Download LibTorch
wget https://download.pytorch.org/libtorch/cu128/libtorch-cxx11-abi-shared-with-deps-2.7.0%2Bcu128.zip  
unzip libtorch-cxx11-abi-shared-with-deps-2.7.0+cu128.zip -d external/

# Build
cmake -B build -DCMAKE_BUILD_TYPE=Release -G Ninja
cmake --build build -- -j$(nproc)

# Train on sample data
./build/gaussian_splatting_cuda -d data/garden -o output/garden --eval

Installation

Requirements

Software

OS: Linux (Ubuntu 22.04+) or Windows
CMake: 3.30 or higher
Compiler: C++23 compatible (GCC 14+ or Clang 17+)
CUDA: 12.8 or higher (required)
LibTorch: 2.7.0 (setup instructions below)
vcpkg: For dependency management

Hardware

GPU: NVIDIA GPU with compute capability 7.5+
VRAM: Minimum 8GB recommended
Tested GPUs: RTX 4090, RTX A5000, RTX 3090Ti, A100, RTX 2060 SUPER

Build Instructions

Linux Build

# Set up vcpkg (one-time setup)
git clone https://github.com/microsoft/vcpkg.git
cd vcpkg && ./bootstrap-vcpkg.sh -disableMetrics && cd ..

## If you want you can specify vcpkg locally without globally setting env variable (see -DCMAKE_TOOLCHAIN_FILE version)
export VCPKG_ROOT=/path/to/vcpkg  # Add to ~/.bashrc

# Clone repository
git clone https://github.com/MrNeRF/LichtFeld-Studio
cd LichtFeld-Studio

# Download LibTorch 2.7.0 with CUDA 12.8
wget https://download.pytorch.org/libtorch/cu128/libtorch-cxx11-abi-shared-with-deps-2.7.0%2Bcu128.zip  
unzip libtorch-cxx11-abi-shared-with-deps-2.7.0+cu128.zip -d external/
rm libtorch-cxx11-abi-shared-with-deps-2.7.0+cu128.zip

# Build
cmake -B build -DCMAKE_BUILD_TYPE=Release -G Ninja

## Or if you want you can specify your own vcpkg 
# cmake -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE="<path-to-vcpkg>/scripts/buildsystems/vcpkg.cmake" -G Ninja 

cmake --build build -- -j$(nproc)

Windows Build

note: Detailed instructions here

Run in x64 native tools command prompt for VS:

# Set up vcpkg (one-time setup)
git clone https://github.com/microsoft/vcpkg.git
cd vcpkg && .\bootstrap-vcpkg.bat -disableMetrics && cd ..

## If you want you can specify vcpkg locally without globally setting env variable (see -DCMAKE_TOOLCHAIN_FILE version)
set VCPKG_ROOT=%CD%\vcpkg

# Clone repository
git clone https://github.com/MrNeRF/LichtFeld-Studio
cd LichtFeld-Studio

# Create directories
if not exist external mkdir external
if not exist external\debug mkdir external\debug
if not exist external\release mkdir external\release

# Download LibTorch (Debug)
curl -L -o libtorch-debug.zip https://download.pytorch.org/libtorch/cu128/libtorch-win-shared-with-deps-debug-2.7.0%2Bcu128.zip
tar -xf libtorch-debug.zip -C external\debug
del libtorch-debug.zip

# Download LibTorch (Release)
curl -L -o libtorch-release.zip https://download.pytorch.org/libtorch/cu128/libtorch-win-shared-with-deps-2.7.0%2Bcu128.zip
tar -xf libtorch-release.zip -C external\release
del libtorch-release.zip

# Build 

## Or if you want you can specify your own vcpkg 
# cmake -B build -DCMAKE_BUILD_TYPE=Release -G ninja -DCMAKE_TOOLCHAIN_FILE="<path-to-vcpkg>/scripts/buildsystems/vcpkg.cmake"

# Ninja should be included with Visual Studio installation, 
# if not you can either install it, 
# or ignore this flag and use native generator - Building time might be extended
cmake -B build -DCMAKE_BUILD_TYPE=Release -G Ninja

cmake --build build -j

Docker Build

# Build and start container
./docker/run_docker.sh -bu 12.8.0

# Build without cache
./docker/run_docker.sh -n

# Stop containers
./docker/run_docker.sh -c

Compiler Setup

Ubuntu 24.04+ (GCC 14)

# Install GCC 14
sudo apt update
sudo apt install gcc-14 g++-14 gfortran-14

# Set as default
sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 60
sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 60
sudo update-alternatives --config gcc
sudo update-alternatives --config g++

Ubuntu 22.04 (Build GCC 14 from source)

# Install dependencies
sudo apt install build-essential libmpfr-dev libgmp3-dev libmpc-dev -y

# Download and build GCC
wget http://ftp.gnu.org/gnu/gcc/gcc-14.1.0/gcc-14.1.0.tar.gz
tar -xf gcc-14.1.0.tar.gz
cd gcc-14.1.0

# Configure and build (1-2 hours)
./configure --prefix=/usr/local/gcc-14.1.0 --enable-languages=c,c++ --disable-multilib
make -j$(nproc)
sudo make install

# Set up alternatives
sudo update-alternatives --install /usr/bin/gcc gcc /usr/local/gcc-14.1.0/bin/gcc 14
sudo update-alternatives --install /usr/bin/g++ g++ /usr/local/gcc-14.1.0/bin/g++ 14

Usage

Dataset Preparation

Download and extract the Tanks & Trains dataset:

wget https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/tandt_db.zip
unzip tandt_db.zip -d data/

Training

Basic training:

./build/gaussian_splatting_cuda -d data/garden -o output/garden

Training with evaluation and visualization:

./build/gaussian_splatting_cuda \
    -d data/garden \
    -o output/garden \
    --eval \
    --save-eval-images \
    --render-mode RGB_D \
    -i 30000

MCMC strategy with limited Gaussians:

./build/gaussian_splatting_cuda \
    -d data/garden \
    -o output/garden \
    --strategy mcmc \
    --max-cap 500000

Command-Line Options

Required

-d, --data-path [PATH] - Path to training data with COLMAP reconstruction

Training Configuration

-o, --output-path [PATH] - Output directory (default: ./output)
-i, --iter [NUM] - Training iterations (default: 30000)
-r, --resize_factor [NUM] - Image resolution factor (default: 1)
--strategy [mcmc|default] - Optimization strategy (default: mcmc)
--max-cap [NUM] - Maximum Gaussians for MCMC (default: 1000000)

Evaluation

--eval - Enable evaluation during training
--save-eval-images - Save evaluation images
--test-every [NUM] - Test/validation split ratio (default: 8)

Visualization

--headless - Run without GUI (terminal-only mode)

Advanced Options

--bilateral-grid - Enable appearance modeling
--steps-scaler [NUM] - Scale training steps for multiple checkpoints
See --help for complete list of options

LPIPS Model Details

The implementation uses weights/lpips_vgg.pt, exported from torchmetrics with:

Network: VGG with ImageNet pretrained weights
Input range: [-1, 1] (conversion handled internally)
Normalization: Included in model

Project Architecture

LichtFeld-Studio/
├── src/
│   ├── core/          # Foundation (data structures, utilities)
│   ├── geometry/      # Geometric operations
│   ├── loader/        # Dataset loading (COLMAP, PLY, Blender)
│   ├── training/      # Training pipeline and strategies
│   ├── rendering/     # CUDA/OpenGL rendering
│   └── visualizer/    # Interactive GUI
├── gsplat/            # Optimized rasterization backend
├── fastgs/            # Fast Gaussian splatting kernels
└── parameter/         # JSON configuration files

Contributing

We welcome contributions! See our Contributing Guidelines.

Getting Started

Check issues labeled good first issue
Join our Discord for discussions
Use the pre-commit hook: cp tools/pre-commit .git/hooks/

Development Requirements

C++23 compatible compiler (GCC 14+ or Clang 17+)
CUDA 12.8+ for GPU development
Apply clang-format for code style

Acknowledgments

This implementation builds upon:

gsplat - Optimized CUDA rasterization backend
3D Gaussian Splatting - Original work by Kerbl et al.

Citation

@software{lichtfeld2025,
  author    = {LichtFeld Studio},
  title     = {A high-performance C++ and CUDA implementation of 3D Gaussian Splatting},
  year      = {2025},
  url       = {https://github.com/MrNeRF/gaussian-splatting-cuda}
}

License

This project is licensed under GPLv3. See LICENSE for details.

Connect with us: Website • Discord • Twitter

Name		Name	Last commit message	Last commit date
Latest commit History 795 Commits
.github		.github
docker		docker
docs		docs
fastgs		fastgs
gsplat		gsplat
include		include
kernels		kernels
nvs-bench		nvs-bench
parameter		parameter
scripts		scripts
src		src
tests		tests
tools		tools
weights		weights
.clang-format		.clang-format
.gitignore		.gitignore
.gitmodules		.gitmodules
CMakeLists.txt		CMakeLists.txt
LICENSE		LICENSE
README.md		README.md
vcpkg.json		vcpkg.json

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

LichtFeld Studio

Overview

Key Features

Community & Support

Active Bounty

Second Bounty: Better 3DGS Initialization and Training without Densification

Quick Start

Installation

Requirements

Software

Hardware

Build Instructions

Compiler Setup

Usage

Dataset Preparation

Training

Command-Line Options

Required

Training Configuration

Evaluation

Visualization

Advanced Options

LPIPS Model Details

Project Architecture

Contributing

Getting Started

Development Requirements

Acknowledgments

Citation

License

About

Uh oh!

Releases

Packages

Languages

License

N-Demir/LichtFeld-Studio

Folders and files

Latest commit

History

Repository files navigation

LichtFeld Studio

Overview

Key Features

Community & Support

Active Bounty

Second Bounty: Better 3DGS Initialization and Training without Densification

Quick Start

Installation

Requirements

Software

Hardware

Build Instructions

Compiler Setup

Usage

Dataset Preparation

Training

Command-Line Options

Required

Training Configuration

Evaluation

Visualization

Advanced Options

LPIPS Model Details

Project Architecture

Contributing

Getting Started

Development Requirements

Acknowledgments

Citation

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages