Skip to content

Library Batch Capture (CLI)

Scan a folder of .blend files and run capture, training, and cleanup for every collection.

Quick Start

# Full batch with config file
python tools/library_capture.py --config jobs/example_library.yaml

# Inline config (no file)
python tools/library_capture.py --library-path /assets --output /data

# Preview without executing
python tools/library_capture.py --config lib.yaml --dry-run

# Introspection only (discover + analyze, no capture)
python tools/library_capture.py --config lib.yaml --introspect-only

# Parallel execution on multiple GPUs
python tools/library_capture.py --config lib.yaml --max-workers 4

Key CLI Arguments

Argument Default Description
--config none YAML config file
--library-path none Folder of .blend files to scan
--output none Base output directory
--blender-path blender Path to Blender executable
--granularity COLLECTIONS COLLECTIONS, SCENE, or EACH_OBJECT
--max-workers 1 Parallel workers (1 = sequential)
--dry-run off Preview pipeline without executing
--introspect-only off Analyze collections only, no capture
--adaptive off Auto-detect optimal camera count/resolution
--no-skip-existing off Force re-capture of already-processed items

Workflow

The pipeline runs in five phases:

  1. Discovery — scans library_path for .blend files, applies include/exclude filters.
  2. Introspection — launches Blender headless per file to discover collections, mesh counts, vertex totals.
  3. Job Generation — creates a pipeline config per collection, respecting filters and skip-existing logic.
  4. Execution — runs capture/train/cleanup jobs (optionally in parallel with GPU load-balancing).
  5. Reporting — writes library_report.json with per-file status, timings, and error messages.

Config File

See jobs/example_library.yaml for the full config schema, including filtering, adaptive analysis, and capture/training/cleanup settings.

Notes

  • Use --introspect-only to preview what would be captured before committing to a long batch.
  • skip_existing: true (default) avoids re-processing items that already have a headless_capture_summary.json.
  • GPU assignment uses round-robin across gpu_ids for parallel workers.