Training Reference

3D Gaussian Splatting by INRIA

The original GraphDeco-INRIA implementation remains the reference point for 3D Gaussian Splatting research, datasets, metrics, and many downstream tools.

Stage: TrainingBaseline qualityResearch comparisonLearning internals

Open official linkGitHub Repository->

What It Does

This repository is the source many other trainers, viewers, and file conventions grew from. It includes a PyTorch optimizer, conversion script, network viewer, and SIBR real-time viewer.

Use it when you need a trusted baseline or want to understand how densification, spherical harmonics, opacity, and scale optimization work in the original method.

How To Use It In 3DGS

Prepare a COLMAP dataset with images and sparse reconstruction.
Create the conda environment and clone submodules exactly as documented.
Train with the provided scripts, then inspect the result in the viewer or export to a browser-friendly format.
Use its metrics and evaluation flow when comparing new methods against the original paper.

Things To Watch

The setup is more demanding than modern app-style tools, especially around CUDA, compilers, and submodules.
The default output can be large, so convert or compress before web delivery.
It is a reference implementation, not always the most ergonomic production pipeline.

Official And Related Links

GitHub Repository Paper Page COLMAP

Why It Still Matters

The repository defines many conventions that downstream tools still understand: COLMAP input layout, PLY Gaussian attributes, SIBR viewers, and evaluation scripts.
Recent updates added training speed acceleration, depth regularization, anti-aliasing, exposure compensation, and viewer improvements, so it is not just a frozen 2023 artifact.
Use it as the baseline when evaluating newer tools like FastGS, gsplat-based trainers, or productized desktop apps.