VMetaphor-Bench: leaderboard
VMetaphor-Bench (Chen et al., 2026) tests whether text-to-image models can generate visual metaphors that combine a source and a target domain, over 1,500 metaphors curated from real creative imagery. A Qwen3.5-27B judge answers 9,594 multiple-choice questions about the generated images (domain presence, structure type, meaning and element mapping), and the board is the pooled accuracy in percent; higher is better. Conceptual prompts, which state only the metaphorical idea.
Metric: MCQ accuracy (%; pooled over all 9,594 fidelity questions: two domain-presence questions per image, one structure-type, one meaning and the element-mapping questions; 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 | 85.4 | |
| 2 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 84.8 | |
| 3 | Seedream 5.0 Lite | 78.5 | |
| 4 | FLUX.2-dev | 76 | |
| 5 | Qwen-Image-2512 | 74.9 | |
| 6 | Janus-Pro-7B | 73 | #7341 |
| 7 | Z-Image-Turbo | 72.4 | |
| 8 | Stable Diffusion 3.5 Large | 70.8 | |
| 9 | FLUX.1-dev | 68.4 | |
| 10 | BAGEL-7B-MoT | 67 | #5515 |
Interactive version: theaggregate.ai/benchmark?slug=vmetaphor-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.