Training Backends

Supported Backends

3D Gaussian Splatting

• Input: COLMAP export (sparse/0)
• Preference path: repository containing train.py

Nerfstudio (splatfacto)

• Input: transforms.json
• Preference: conda env name

GS-Lightning

• Input: COLMAP export
• Optional masks supported (use GS-Lightning mask format when needed)
• Preference path: repository containing main.py

gsplat

• Input: COLMAP or transforms.json
• Preference path: gsplat/examples containing simple_trainer.py

Start Training From Blender

1. Open the Training panel.
2. Choose backend.
3. Set training data path and output path.
4. Configure iterations/arguments.
5. Click Start Training.

Common Validation Requirements

• COLMAP backends: sparse/0 must exist
• transforms.json backends: JSON file must exist and match image paths
• Windows: keep output paths short to avoid path-length issues

Post-Training Cleanup

• The addon can run proxy-hull cleanup after training for all built-in backends.
• Cleanup is geometry-based (cleanup/proxy_hulls.json) and does not depend on backend-specific prune behavior.
• Backend-native knobs can still help reduce floaters:
• Nerfstudio: cull_alpha_thresh, continue_cull_post_densification, use_scale_regularization
• gsplat: strategy prune parameters such as prune_opa
• GS-Lightning / 3DGS: regularization and densification tuning