Current mainstream text-to-image models are trained on image-text pairs and are better at recognizing concrete visual descriptors than abstract emotions or metaphors; when encountering a metaphor, they tend to prioritize the most direct visual element for generation.