Beyond Content: How Grammatical Gender Shapes Visual Representation in Text-to-Image Models