Color-taste associations for food in text-to-image models

This study explores color-taste correspondences using text-to-image models within food (a context-specific domain) and provides insights into the models' sensitivity to such associations. Results confirmed the association of bitterness with black and saltiness with blue, while links between sweetness and red or sourness and yellow were inconsistent across food items. However, sweet foods described as bitter showed a reduction in red hues, being informative about the difference between expected and depicted taste. Findings suggest that emotional and semantic associations play an important role across a diverse range of food items. Extrinsic visual cues were frequently found in generated images, highlighting implications for food marketing and packaging. • Text-to-image models enable research on color-taste associations across food items. • Bitterness links to black and saltiness to blue across foods in generated images. • Red-sweetness and yellow-sourness associations are relatively inconsistent. • Text-to-image models capture emotional and semantic crossmodal correspondences. • Prompting the model with sweet food as bitter reduced image redness.