Gaussian Splatting Tools Comparison
This site documents a structured overview and experimental analysis of Gaussian Splatting training pipelines and XR visualization environments.
What's here
The project is organized around two complementary dimensions:
Overview — descriptive documentation of selected Gaussian Splatting tools, covering both commercial and open-source trainers, and a range of visualization and integration tools. No benchmarking — just context.
Analysis — controlled experimental evaluation conducted under fixed, reproducible conditions. Covers open-source training pipelines across indoor and outdoor scenarios, followed by a comparative XR viewer evaluation.
How-To — step-by-step setup and execution guides for each tool included in the experimental analysis.
Reading paths
"I don't know which trainer to use" → Start with Which Trainer to Use? — a one-page decision guide with trade-offs and a bottom-line recommendation
"I want to understand the landscape of tools" → Start with Overview / Trainers and Overview / Viewers
"I want to see benchmark results" → Start with Analysis / Trainer Benchmark Methodology, then Indoor, Outdoor, and Indoor vs Outdoor
"I want to see the XR viewer evaluation" → Start with Analysis / XR Viewer Methodology, then XR Viewers
"I want to reproduce the experiments" → Go directly to How-To
Experimental workflow
The evaluation follows a structured pipeline:
Training → Cleaning → Structural Analysis → XR Visualization → Runtime Evaluation
This separation enables independent assessment of reconstruction structure and density, post-processing impact, immersive rendering stability, and runtime performance behavior.
Methodological scope
The analysis is limited to:
- five selected open-source training pipelines
- two Unity-based XR viewers
All experiments use fixed datasets, consistent optimization settings, a standardized cleaning workflow, identical hardware, and exported .ply files for inspection and XR loading.
Commercial trainers and desktop/web-based viewers are documented at overview level only and are not included in the controlled experimental analysis.
Acknowledgments
Special thanks to Ernesta Maria Sichetti for carrying out the experiments, collecting the results, and producing the initial documentation that forms the foundation of this repository.