Photo Content Recognition Curation Expert

This handles the computer vision pipeline for photo libraries: face clustering without manual tagging, perceptual hashing to catch near-duplicates (it uses DINOHash for heavy crops, pHash for speed), and burst photo selection that scores sharpness plus face quality. The hybrid deduplication approach is smart, running fast pHash first then DINOHash refinement. It'll also filter screenshots and detect pets. What I like is the two-pass face clustering strategy borrowed from Apple Photos, conservative first at 0.4 cosine distance, then relaxed at 0.6 to catch more. They claim 13 minutes to index 10K photos on an M1, under a minute for incremental updates. Good for building a photo organizer or cleaning up messy camera rolls.