gaussian-splatting: Reconstruct and Render 3D Scenes in Real Time

Overview

This repository is the authors’ reference implementation of 3D Gaussian Splatting for reconstructing scenes from multiple images and rendering new viewpoints. It represents a scene with optimized 3D Gaussians and combines training tools with interactive viewers, addressing the cost of producing high-quality, real-time novel-view renderings with earlier radiance-field methods.

It is aimed at researchers and developers working in computer graphics or computer vision, rather than users looking for a ready-made viewer for arbitrary media. You provide suitable image and camera data, train a scene model, then render or inspect it. The project page and paper are linked from the official repository and project website.

Key Features

• Train Gaussian scene representations from COLMAP or synthetic NeRF datasets.
• Render training and test views and calculate image-quality metrics.
• Use a network viewer to inspect optimization while training.
• Explore trained models in an OpenGL-based real-time viewer.
• Convert image collections into COLMAP-style inputs with the included script.
• Enable optional depth regularization, exposure compensation, and anti-aliasing.
• Use an accelerated rasterizer and sparse Adam optimizer through the documented training-speed options.

Use Cases

• Graphics researchers can reproduce and evaluate the method described in the associated paper.
• Computer-vision practitioners can reconstruct a photographed scene for interactive novel-view inspection.
• Developers can use the included differentiable Gaussian rasterizer as a component in their own project.
• Teams comparing reconstruction quality can train, render, and compute metrics on supported evaluation datasets.

Project Facts

• Language: Python
• License: NOASSERTION
• Stars: 24.1k
• Forks: 3.5k
• Topics: computer-graphics, computer-vision, radiance-field
• Archived: false

Getting Started

Clone the repository and its submodules, then create the Conda environment and train on a prepared dataset:

git clone https://github.com/graphdeco-inria/gaussian-splatting --recursive
cd gaussian-splatting
conda env create --file environment.yml
conda activate gaussian_splatting
python train.py -s <path to COLMAP or NeRF Synthetic dataset>

See the README for dataset preparation, platform setup, evaluation, and viewer instructions.

Alternatives

• LAM: LAM creates and animates a Gaussian head avatar from a single image, while gaussian-splatting reconstructs general scenes from posed images.

Considerations

• The documented training setup requires a CUDA-capable GPU with Compute Capability 7.0 or higher. The README recommends 24 GB of VRAM for training to the paper’s evaluation quality.
• Building the Python CUDA extensions and SIBR viewers requires compatible compiler and CUDA tooling. The documented setup uses CUDA SDK 11, and Windows builds may need additional troubleshooting.
• Input images need camera and structure-from-motion data in a supported format. The supplied conversion workflow relies on COLMAP, and rasterization expects SIMPLE_PINHOLE or PINHOLE cameras.
• The authors note that maintenance and feature work are constrained by limited resources. Some newer features are optional additions, and results or resource use can vary with configuration and dataset.
• The repository metadata lists its license as NOASSERTION, so check the repository for licensing terms before adopting it in a project.