VMetaphor-Bench - Metaphor Logic: leaderboard
Metaphor logic in VMetaphor-Bench (Chen et al., 2026): a Qwen3.5-27B judge rates from 1 to 5 whether the cross-domain mapping in each generated image is conceptually coherent and logically sound. Mean rating; higher is better. Conceptual prompts, which state only the metaphorical idea.
Metric: Judge rating (1-5) (Qwen3.5-27B rates each generated image from 1 to 5 against the metaphor's structured annotation, 1,500 visual metaphors; this board is the dimension metaphor logic: whether the cross-domain mapping between source and target is conceptually coherent and logically sound; 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 | 4.23 | |
| 2 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 4.08 | |
| 3 | Seedream 5.0 Lite | 3.74 | |
| 4 | FLUX.2-dev | 3.46 | |
| 5 | Qwen-Image-2512 | 3.31 | |
| 6 | Z-Image-Turbo | 3.03 | |
| 7 | Janus-Pro-7B | 3.02 | #7341 |
| 8 | Stable Diffusion 3.5 Large | 3.01 | |
| 9 | FLUX.1-dev | 2.87 | |
| 10 | BAGEL-7B-MoT | 2.72 | #5515 |
Interactive version: theaggregate.ai/benchmark?slug=vmetaphor-bench-metaphor-logic · How It Works · Data refreshed daily, snapshot 2026-11-12.