PhotoBench: leaderboard
Metric: Set-based F1 (%) of the returned image set against the exhaustive ground-truth set on normal queries (Chinese version), where the backbone LLM drives the authors' ReAct-style retrieval agent with vector-search (RzenEmbed-v2-7B), metadata-filter, face-search and set-composition tools over a personal album; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B | 51.9 | #304 |
| 2 | Claude Opus 4.5 | 50.5 | #79 |
| 3 | O3 | 48 | #121 |
| 4 | Claude Sonnet 4.5 | 47.9 | #138 |
| 5 | GPT-4o | 46.4 | #333 |
| 6 | Qwen 3 32B | 41.4 | #424 |
| 7 | Qwen 3 8B | 41 | #667 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=photobench · How It Works · Data refreshed daily, snapshot 2026-10-11.