VMetaphor-Bench - Structure Type: leaderboard
Structure type in VMetaphor-Bench (Chen et al., 2026): one multiple-choice question per generated image on whether the two domains are combined as intended (replacement, fusion or juxtaposition), answered by a Qwen3.5-27B judge. Accuracy in percent; higher is better. Conceptual prompts, which state only the metaphorical idea.
Metric: MCQ accuracy (%; one question per image on how the two domains are combined: replacement, fusion or juxtaposition; the generated image is judged by Qwen3.5-27B on multiple-choice questions (four options including cannot be determined) built from each metaphor's structured annotation; 1,500 visual metaphors curated from real creative imagery; conceptual prompts that state only the metaphorical idea). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT Image 1.5 | 78.1 | |
| 2 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 76.3 | |
| 3 | Qwen-Image-2512 | 66.1 | |
| 4 | Seedream 5.0 Lite | 65.6 | |
| 5 | FLUX.2-dev | 63.7 | |
| 6 | Stable Diffusion 3.5 Large | 60.1 | |
| 7 | Z-Image-Turbo | 59.6 | |
| 8 | Janus-Pro-7B | 59.1 | #7341 |
| 9 | FLUX.1-dev | 54.9 | |
| 10 | BAGEL-7B-MoT | 54.8 | #5515 |
Interactive version: theaggregate.ai/benchmark?slug=vmetaphor-bench-structure-type · How It Works · Data refreshed daily, snapshot 2026-11-12.