Automation Pipeline (CLI)¶

End-to-end orchestrator that takes a .blend file through capture, GS training, and splat cleanup in one command.

Quick Start¶

# Single job from config
python tools/pipeline.py --config jobs/my_job.yaml

# Multiple jobs
python tools/pipeline.py --config jobs/a.yaml jobs/b.yaml

# Preview without executing
python tools/pipeline.py --config jobs/my_job.yaml --dry-run

Three Phases¶

Capture — launches Blender with headless_capture.py to render multi-view images and export COLMAP/transforms.json.
Training — runs a Gaussian Splatting backend (3DGS, Nerfstudio, gsplat, etc.) on the captured images.
Cleanup — filters the trained .ply model using proxy hull volumes to remove floating splats.

Each phase is optional and can be enabled/disabled in the config.

Config File¶

See jobs/example_capture.yaml for the full config schema. Key sections:

blender_path: "blender"
scene: "scene.blend"
output_base: "/output/path"

capture:
  cameras: 150
  preset: "3dgs"
  engine: EEVEE

training:
  enabled: true
  backend: "original"
  iterations: 30000

cleanup:
  enabled: true
  proxy_hulls: true

Pipeline Report¶

Each run writes pipeline_report.json to the output directory with per-phase timings, success/failure status, and error details.

Relationship to Other Tools¶

Tool	Purpose
`tools/pipeline.py`	Orchestrates a single .blend through all phases
`tools/headless_capture.py`	Capture-only (called by pipeline.py)
`tools/library_capture.py`	Batch-processes a folder of .blend files using pipeline.py

For batch processing many .blend files, use Library Batch Capture instead.

Notes¶

Blender is auto-detected if on PATH, or set blender_path in config.
The pipeline runs outside Blender (standalone Python).
Training requires the chosen backend to be installed in the active Python environment.