SADL - Open Detection: leaderboard
The open detection protocol of SADL (2026): from a photograph and its intended subject alone, a vision-language model names the distractors, which are grounded to masks by SAM 3 and matched to the annotations. F1 in percent; higher is better.
Metric: Open detection F1 (%, times 100): the model lists distractors from the image and subject description alone, the labels are grounded to masks by SAM 3 and matched to annotated distractors at IoU 0.5, over the 14,617 annotated candidates of SADL (1,800 subject-aware cases on 1,000 photographs, including 1,938 hard negatives), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 40.6 | #137 |
| 2 | Qwen 3 VL 235B A22B | 36.2 | #479 |
| 3 | GPT-4.1 | 32.5 | #495 |
| 4 | Pixtral-12B | 29.8 | #2476 |
| 5 | GPT-4o | 27.6 | #741 |
| 6 | Molmo2-8B | 27.5 | #1483 |
| 7 | Llama 4 Scout | 26.5 | #1776 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sadl-open-detection · How It Works · Data refreshed daily, snapshot 2026-11-13.