Ideal Facial Proportions in Generative Artificial Intelligence: What Does Artificial Intelligence Consider an Attractive Face?

Abstract

Background: To investigate the facial proportions and physical characteristics deemed ideal and attractive by current popular generative artificial intelligence (AI) models. Methods: "Attractive" and "ideal" face images were generated using five text-to-image models. Facial proportions of the generated images were measured and compared with the neoclassical canons of facial thirds and fifths. Generated facial proportions were compared between AI models and classical facial proportions using analysis of variance and unpaired Student's t-test, respectively. Results: The generated images included 28 (70%) female faces and 29 (75%) Caucasian faces. Mean generated horizontal proportions were 33.7%, 32.0%, and 34.3%. Mean generated vertical proportions were 21.1%, 18.0%, 21.1%, 17.9%, and 21.7%. The middle horizontal segment was significantly smaller in generated "ideal faces," and the lower horizontal segment was larger in generated "attractive faces" compared with classical proportions (p < 0.001 and p = 0.01, respectively); the left and right middle vertical segments were significantly smaller in all generated faces compared with classical proportions (p < 0.001). Conclusion: There are significant differences between the definition of attractive and ideal faces in generative AI models and classical facial proportions. These differences may reflect biases present within the models or may reflect changing cultural perceptions of facial attractiveness.