How bchao1/textured_gaussians works
Textured Gaussians is a PyTorch/GPU library and example workflow for fitting 2DGS scenes, refining them with spatially varying texture and alpha, and rendering the resulting Gaussian scene. The documented path is dataset loading, 2DGS pre-training, then textured refinement; the trainer also computes image-quality metrics. Runtime capabilities include 2DGS/3DGS rasterization, densification strategies, distributed rendering, selective optimization, and PNG splat compression. Training and dataset internals were not excerpted, so those links rely on the README; other links are limited to relationships visible in sampled source or the SOURCE INDEX.
Main entry points: simple_trainer_textured_gaussians.py Example trainer · _wrapper.py CUDA operation wrapper · rendering.py Rasterization APIs · ops.py Gaussian edits
Components 18
Training workflow
- Dataset loaders
examples/datasets - Example trainer
examples/simple_trainer_textured_gaussians.py - Image metrics
examples/simple_trainer_textured_gaussians.py
Rendering engine
- Rasterization APIs
textured_gaussians/rendering.py - Textured 2DGS kernels
textured_gaussians/cuda/csrc - CUDA operation wrapper
textured_gaussians/cuda/_wrapper.py - CUDA backend
textured_gaussians/cuda/_backend.py - PyTorch raster ops
textured_gaussians/cuda/_torch_impl.py - 2DGS PyTorch ops
textured_gaussians/cuda/_torch_impl_2dgs.py
Optimization and adaptation
- Default strategy
textured_gaussians/strategy/default.py - MCMC strategy
textured_gaussians/strategy/mcmc.py - Gaussian edits
textured_gaussians/strategy/ops.py - Gaussian relocation
textured_gaussians/relocation.py - Selective Adam
textured_gaussians/optimizers/selective_adam.py
Runtime support
- Distributed collectives
textured_gaussians/distributed.py
Splat assets
- PNG compression
textured_gaussians/compression/png_compression.py - Compression sorting
textured_gaussians/compression/sort.py
Other components
- Researcher
Connections 18
- Researcher → Dataset loaders: provides data
README.md - Dataset loaders → Example trainer: loads scenes
README.md - Researcher → Example trainer: configures stages
README.md - Example trainer → Rasterization APIs: renders training views
README.md - Example trainer → Default strategy: may refine
- Example trainer → MCMC strategy: uses MCMC
README.md - Example trainer → Image metrics: computes metrics
README.md - Rasterization APIs → CUDA operation wrapper: calls operations
textured_gaussians/rendering.py - Rasterization APIs → Distributed collectives: uses collectives
textured_gaussians/rendering.py - CUDA operation wrapper → CUDA backend: dispatches kernels
textured_gaussians/cuda/_wrapper.py - CUDA operation wrapper → PyTorch raster ops: uses implementations
textured_gaussians/cuda/_torch_impl.py - CUDA operation wrapper → 2DGS PyTorch ops: uses 2DGS ops
textured_gaussians/cuda/_torch_impl_2dgs.py - Default strategy → Gaussian edits: edits Gaussians
textured_gaussians/strategy/default.py - MCMC strategy → Gaussian edits: edits Gaussians
textured_gaussians/strategy/mcmc.py - Gaussian edits → Gaussian relocation: relocates splats
textured_gaussians/strategy/ops.py - Gaussian relocation → CUDA operation wrapper: calls operations
textured_gaussians/relocation.py - Selective Adam → CUDA operation wrapper: updates parameters
textured_gaussians/optimizers/selective_adam.py - PNG compression → Compression sorting: uses sorting
textured_gaussians/compression/png_compression.py