gaussian-splatting-lightning
v0.1.0PyTorch Lightning-based 3D Gaussian Splatting framework supporting multiple paper implementations and interactive web viewing
Development Activity
Sample Renders
Overview
Best for
Researchers comparing multiple Gaussian Splatting variants in a unified PyTorch Lightning framework with structured training, logging, and multi-GPU support
Not ideal for
Production deployment, non-NVIDIA hardware, users seeking a polished end-user application, or those needing NeRF support alongside Gaussian Splatting
Strengths
- PyTorch Lightning integration provides structured training with built-in logging, checkpointing, and multi-GPU support, making experiments more reproducible than ad-hoc scripts
- Implements multiple 3DGS paper variants in a unified framework, allowing fair comparison between methods without reimplementation effort
- Interactive web-based viewer enables quick qualitative assessment of trained models without installing separate viewer software
- MIT license allows unrestricted use in both academic research and commercial applications
- Modular architecture makes it relatively straightforward to add new Gaussian Splatting variants as they are published
Limitations
- Requires NVIDIA GPU with CUDA — no CPU fallback or support for AMD or Intel GPUs, limiting accessibility
- Less established than Nerfstudio as a Gaussian Splatting platform — smaller community, fewer tutorials, and less comprehensive documentation
- Fast-moving codebase may have breaking changes between versions as the 3DGS research landscape evolves rapidly
- Training quality and speed may lag behind specialized implementations like gsplat that focus on optimizing a single method rather than supporting multiple variants
- Input pipeline assumes COLMAP-style camera pose estimation — integrating with other structure-from-motion tools or custom camera rigs requires additional effort
Background
gaussian-splatting-lightning is a research framework built on PyTorch Lightning that provides a unified training pipeline for multiple 3D Gaussian Splatting variants. Created after the landmark INRIA 3DGS paper in August 2023, it leverages PyTorch Lightning's structured training paradigm — with built-in logging, checkpointing, multi-GPU support, and experiment tracking — to make Gaussian Splatting research more reproducible and extensible than ad-hoc training scripts.
The framework implements the original 3DGS algorithm alongside several published extensions and variants, allowing researchers to train, evaluate, and compare different methods within a consistent codebase. It accepts COLMAP-style input (camera poses and images), outputs trained Gaussian models in PLY format, and includes an interactive web-based viewer for qualitative assessment of trained models. The training pipeline supports configurable hyperparameters, densification strategies, and regularization techniques.
Positioned between the original INRIA reference implementation (a single-purpose research code) and Nerfstudio (a full-featured platform spanning both NeRF and Gaussian Splatting), gaussian-splatting-lightning focuses specifically on structured, reproducible Gaussian Splatting experiments. Its PyTorch Lightning foundation appeals to researchers already familiar with that ecosystem. However, as a relatively young project in the rapidly evolving 3DGS landscape, it has a smaller community and less documentation than more established tools.
Quick Start
pip install git+https://github.com/yzslab/gaussian-splatting-lightningRelated Renderers
Community & Resources
Tutorials & Resources
Performance Benchmarks
No benchmark data available for gaussian-splatting-lightning yet.
Benchmarks will be added as more renderers are tested across our standard scene suite.
Learn about our methodology