How moorethreads/litegs works
LiteGS is a modular 3D Gaussian-splatting training system with Python and CUDA rendering APIs. Its main workflow loads COLMAP cameras and images, initializes clustered Gaussian primitives, trains them through rendering and density control, and persists checkpoints. The README also documents the staged culling-to-rasterization design and flexible Python/CUDA implementations. Some subsystem details come from README or SOURCE INDEX references rather than sampled implementation excerpts; edges reflect only those documented or shown relationships.
Main entry points: trainer.py Training loop · __init__.py Cull and activate · colmap.py COLMAP reader · densify.py Density control
Components 22
Data and persistence
- COLMAP reader
litegs/io_manager/colmap.py - Camera and image data
litegs/data.py - PLY I/O
litegs/io_manager/ply.py - Checkpoint I/O
litegs/io_manager/checkpoint.py
Training and scene
- Run parameters
litegs/arguments.py - Training loop
litegs/training/trainer.py - Gaussian primitives
litegs/scene/point.py - Cluster management
litegs/scene/cluster.py - Density control
litegs/training/densify.py - Sparse optimizer
litegs/training/optimizer.py - Training statistics
litegs/utils/statistic_helper.py
Rendering pipeline
- Cull and activate
litegs/render/__init__.py - Projection and transforms
litegs/utils/wrapper.py - Visibility binning
litegs/utils/wrapper.py - Tile rasterization
litegs/utils/wrapper.py
Rendering compute
- Primitive compaction
litegs/submodules/gaussian_raster/compact.h - CUDA transforms
litegs/submodules/gaussian_raster/transform.h - CUDA binning
litegs/submodules/gaussian_raster/binning.h - CUDA raster kernel
litegs/submodules/gaussian_raster/raster.h - Compacted tensors
litegs/utils/CompactedTensor.py - Spherical harmonics
litegs/utils/spherical_harmonics.py
Other components
- Training user
Connections 23
- Training user → Run parameters: sets options
litegs/arguments.py - Training loop → Run parameters: uses settings
litegs/training/trainer.py - Training loop → COLMAP reader: loads scene
litegs/training/trainer.py - COLMAP reader → Camera and image data: builds frames
litegs/io_manager/colmap.py - Training loop → Camera and image data: loads images
litegs/training/trainer.py - Training loop → Gaussian primitives: initializes points
litegs/training/trainer.py - Training loop → Cluster management: clusters points
litegs/training/trainer.py - Training loop → Cull and activate: renders views
litegs/render/__init__.py - Cull and activate → Cluster management: culls clusters
litegs/render/__init__.py - Cull and activate → Projection and transforms: projects points
litegs/render/__init__.py - Projection and transforms → CUDA transforms: invokes transforms
- Cull and activate → Spherical harmonics: computes color
litegs/render/__init__.py - Cull and activate → Visibility binning: builds visibility
litegs/render/__init__.py - Visibility binning → CUDA binning: invokes binning
- Cull and activate → Tile rasterization: rasterizes tiles
litegs/render/__init__.py - Tile rasterization → CUDA raster kernel: invokes raster
- Cull and activate → Primitive compaction: compacts points
- Training loop → Density control: updates density
litegs/training/trainer.py - Density control → Training statistics: reads statistics
litegs/training/densify.py - Training loop → Sparse optimizer: updates parameters
litegs/training/trainer.py - Sparse optimizer → Compacted tensors: updates compact gradients
litegs/training/optimizer.py - Training loop → Checkpoint I/O: saves state
litegs/training/trainer.py - Training loop → PLY I/O: exports primitives