Original title: Blending Concepts: Benchmarking Visual Metaphor Generation in Text-to-Image Models
Chuer Chen, Zichen Wang, Yi He, Zhengxi Yu +1 · September 2026
Image generators can easily draw simple objects, but can they create visual metaphors? A visual metaphor is a picture that expresses an abstract idea by merging elements from two different concepts, like fusing an hourglass with an iceberg to show melting time. Researchers wanted to test this skill, so they created a test called VMetaphor-Bench. They collected creative metaphors and tested eleven image-making systems using two evaluation methods: a multiple-choice quiz and a scoring system that judges visual qualities. They found that even the most powerful paid systems struggle to combine concepts and connect different domains together properly. Making abstract, metaphorical images is still a major challenge for these systems.
Imagine trying to explain 'loneliness' by drawing a single tree in a vast desert. The challenge is not just drawing the tree and desert separately, but making the image feel like it represents the abstract feeling. In this paper, the benchmark tests whether AI models can similarly blend concepts from different domains to create such evocative images, and finds they often fail to structure the composition properly.
Artists and designers relying on AI image tools cannot yet produce visual metaphors, which are essential for creative expression and communication. This benchmark shows where current models fall short, guiding future improvements for more nuanced AI-assisted art.
The abstract does not explicitly state limitations, but the benchmark is based on 1,500 curated metaphors and evaluates only 11 models, which may not cover all possible metaphor types or systems.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Blending Concepts: Benchmarking Visual Metaphor Generation in Text-to-Image Models — Chuer Chen, Zichen Wang, Yi He et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.