DLEBench - Instruction Following: leaderboard
Metric: Instruction-following score (% of the maximum) on a 4-level rubric (localization failure, wrong action, over-modification, flawless), normalized to 100, averaged over the seven instruction types with equal weight; 1,889 small-object editing samples (target 1-10% of the image area) converted from V*-Bench, MME-RealWorld and Pixel-Reasoner, seven instruction types; Oracle-guided evaluation with a Gemini-3-Pro judge on crops around human-annotated target boxes and reference edits; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana Pro (Gemini 3 Pro Image) | 48.97 | |
| 2 | GPT-Image-1 | 45.32 | |
| 3 | Step1X-Edit (DLEBench checkpoint unspecified) | 42.65 | |
| 4 | UniWorld-V2 | 42.32 | |
| 5 | UniREdit-Bagel | 38.03 | |
| 6 | BAGEL-7B-MoT (Thinking) | 35.55 | |
| 7 | OmniGen2 | 24.32 | |
| 8 | Qwen-Image-Edit (DLEBench checkpoint unspecified) | 17.03 | |
| 9 | MagicBrush | 15.34 | |
| 10 | UniWorld-V1 | 13.84 |
Interactive version: theaggregate.ai/benchmark?slug=dlebench-instruction-following · How It Works · Data refreshed daily, snapshot 2026-10-11.