How nv-tlabs/3dgrut works
3DGRUT combines training and rendering for Gaussian scenes, with 3DGRT ray tracing, 3DGUT rasterization, and a hybrid primary-raster/secondary-ray pipeline. The main workflow is dataset ingestion, model training, rendering and evaluation, with trained scenes also exportable and viewable in an interactive playground. The source index supports several internal links; README statements support the higher-level product workflows. Unsampled subsystems are represented at their exact enclosing directories, and their internal wiring is not inferred.
Main entry points: trainer.py Training loop · model.py Gaussian model · render.py Render and evaluation · export Scene export
Components 21
Scene Data
- Dataset loaders
threedgrut/datasets - Batch protocols
threedgrut/datasets/protocols.py
Training
- Training loop
threedgrut/trainer.py - Gaussian model
threedgrut/model/model.py - Scene features
threedgrut/model/features.py - Feature decoder
threedgrut/model/feature_decoder.py - Densification
threedgrut/strategy - Optimizers
threedgrut/optimizers
Rendering
- Render and evaluation
threedgrut/render.py - Render utilities
threedgrut/utils/render.py - 3DGRT tracer
threedgrt_tracer/tracer.py - 3DGUT rasterizer
threedgut_tracer/tracer.py - Hybrid tracer
threedgrut_playground/tracer.py
Export and Tools
- Interactive playground
threedgrut_playground/engine.py - Training viewer
threedgrut/utils/gui.py - Scene export
threedgrut/export - USD and NuRec
threedgrut/export/usd - ISP post-processing
threedgrut/export/usd/post_processing - Exported assets
Other components
- Scene creator
- Evaluation user
Connections 22
- Scene creator → Dataset loaders: selects data
README.md - Dataset loaders → Batch protocols: provides batches
threedgrut/datasets/datasetNcore.py - Training loop → Dataset loaders: loads data
threedgrut/trainer.py - Training loop → Gaussian model: trains
threedgrut/trainer.py - Training loop → Densification: uses strategy
threedgrut/trainer.py - Training loop → Optimizers: uses optimizers
threedgrut/trainer.py - Gaussian model → Scene features: configures features
threedgrut/model/model.py - Gaussian model → 3DGRT tracer: invokes tracing
threedgrut/model/model.py - Gaussian model → 3DGUT rasterizer: invokes rasterizer
threedgrut/model/model.py - Training loop → Render and evaluation: validates renders
threedgrut/trainer.py - Render and evaluation → Gaussian model: renders with
threedgrut/render.py - Render and evaluation → Render utilities: post-processes
threedgrut/render.py - Gaussian model → Feature decoder: decodes features
- Scene creator → Render and evaluation: requests render
README.md - Evaluation user → Render and evaluation: requests evaluation
README.md - Training loop → Training viewer: opens viewer
threedgrut/trainer.py - Scene creator → Interactive playground: inspects scene
README.md - Interactive playground → Hybrid tracer: uses hybrid tracing
threedgrut_playground/engine.py - Scene creator → Scene export: exports scene
README.md - Scene export → USD and NuRec: supports USD
threedgrut/export/__init__.py - USD and NuRec → ISP post-processing: applies processing
threedgrut/export/usd/post_processing/ppisp_controller_writer.py - Scene export → Exported assets: writes assets
README.md