Text-to-image models reveal specific color-emotion associations

Abstract

Text-to-image models learn associations between human-provided image tags and image features over billions of examples. As a result, such models provide a powerful mean to study the psychological relationships between colors and emotions. We generated images for different emotions descriptions varying in valence, arousal and dominance across several subjects and then extracted color features (chroma and L*a*b* values) from the resultant images to find color-emotion associations. Results show a joint effect of red and chroma to generate effects of joy, rage and negative powerless. In addition, lightness is key in generating effects of serenity, threat and a relief/stress divergence. Dominance emerged as an important dimension to understand interactions and nuances in color-emotion associations. The study highlights that specific combination of color elements convey emotions, rather than and beyond simple associations such as red-anger or lightness-valence.