openclaw-vision-storyboard

Sample the minimum frames needed per scene, resize them to the smallest resolution that preserves evidence, and run vision inference to identify caption-safe regions. The output is scene-indexed observations with concrete revisions for caption placement.

Run exactly one admitted vision workload at a time and never keep a vision model resident beside an active 7B model on constrained hosts. Release the vision model before starting another heavyweight model stage.

Yes, frame inspection extracts OCR text as part of structured observations covering composition, continuity, and visual defects. When OCR confidence is low, the skill reports insufficient evidence rather than inventing text.

Frame sampling trades coverage for resource safety, so issues between sampled frames may be missed. The workflow stops when each sampled scene has enough evidence for a deterministic revision or an explicit insufficient-evidence result.

Avoid it when frames cannot be sampled at a resolution preserving required evidence, or when a vision model cannot be loaded without displacing an active heavyweight model. In those cases, defer inspection until resources free up.