How graphdeco-inria/gaussian-splatting works
This repository implements 3D Gaussian Splatting: prepare or supply multi-view images and SfM data, optimize a Gaussian scene representation, then render novel views and evaluate image quality. Training also supports depth regularization and a network-connected interactive viewer; the README describes a separate SIBR real-time viewer. The sampled source confirms the training, scene-loading, rasterization, and GUI paths. Evaluation scripts and SIBR internals were not sampled, so their detailed wiring is not inferred.
Main entry points: train.py Training loop · __init__.py Gaussian renderer · convert.py SfM converter · metrics.py Image metrics
Components 23
Dataset preparation
- SfM converter
convert.py - Scene loading
scene/dataset_readers.py - COLMAP model I/O
utils/read_write_model.py - Depth calibration
utils/make_depth_scale.py - Scene assembly
scene/__init__.py
Optimization
- Training loop
train.py - Training parameters
arguments/__init__.py - Camera model
scene/cameras.py - Gaussian model
scene/gaussian_model.py
Rendering and viewing
- Gaussian renderer
gaussian_renderer/__init__.py - CUDA rasterizer
submodules/diff-gaussian-rasterization - Training network GUI
gaussian_renderer/network_gui.py - Novel-view rendering
render.py - SIBR real-time viewer
SIBR_viewers
Evaluation
- Evaluation runner
full_eval.py - Image metrics
metrics.py - LPIPS metric
lpipsPyTorch/__init__.py
Other components
- Researcher
- Input images
- COLMAP
- Trained Gaussian model
- Rendered views
- Reference images
Connections 33
- Researcher → SfM converter: prepares dataset
- Input images → SfM converter: supplies images
- SfM converter → COLMAP: runs SfM
- COLMAP → SfM converter: returns reconstruction
- SfM converter → Scene loading: writes dataset
- Scene loading → COLMAP model I/O: reads model
- Scene loading → Scene assembly: builds scene data
- Researcher → Training loop: starts training
- Training loop → Training parameters: loads settings
- Training loop → Scene assembly: loads scene
- Training loop → Gaussian model: optimizes
- Training loop → Gaussian renderer: renders training views
- Gaussian renderer → CUDA rasterizer: rasterizes Gaussians
- Gaussian renderer → Gaussian model: reads parameters
- Gaussian renderer → Camera model: uses camera transforms
- Training loop → Training network GUI: serves viewer requests
- Training network GUI → Gaussian renderer: requests custom render
- Training loop → Trained Gaussian model: saves checkpoints
- Researcher → Novel-view rendering: requests views
- Novel-view rendering → Gaussian renderer: renders camera views
- Novel-view rendering → Trained Gaussian model: loads model
- Gaussian renderer → Rendered views: produces images
- Researcher → SIBR real-time viewer: views scene
- Trained Gaussian model → SIBR real-time viewer: loads for viewing
- Researcher → Evaluation runner: runs evaluation
- Evaluation runner → Novel-view rendering: requests renders
- Evaluation runner → Image metrics: computes scores
- Image metrics → LPIPS metric: computes perceptual distance
- Image metrics → Rendered views: compares renders
- Image metrics → Reference images: compares references
- Researcher → Depth calibration: calibrates depth
- Depth calibration → COLMAP model I/O: reads SfM model
- Depth calibration → Scene loading: writes depth parameters