PhotoBench is the first benchmark constructed from authentic, personal albums, designed to shift the paradigm from visual matching to personalized multi-source intent-driven photo retrieval. PhotoBench-Protected is the limited-information release: only pre-computed captions, embeddings, and metadata are provided, so this leaderboard focuses exclusively on agent planning ability.

⚠️ Please confirm you are submitting to the correct leaderboard.

The test sets for PhotoBench-Protected and PhotoBench (full) ↗ are different. For unrestricted retrieval with raw images, please use the full PhotoBench leaderboard ↗. Full dataset download: OneBox ↗.