This study proposes a facial expression generation method that comprehensively covers the emotional spectrum defined by the Russell Circumplex Model of Affect. We employ StyleGAN to synthesize face images tailored to specific ages and genders, and apply GANimation to generate 49 expressions per face, guided by a rule set that combines facial muscle movements-Action Units (AUs)-according to their patterns of co-activation and mutual exclusivity. Expressions were generated for four synthetic face models, and subjective impression ratings were collected via a Visual Analog Scale (VAS). Our method generated all 49 expressions without artifacts; however, impression ratings varied with the age and gender of the target face. Further analysis revealed that while some AUs consistently influenced impressions across models, others varied depending on the face. The proposed method enables flexible and culturally adaptive facial expression design, with potential applications in personalized affective computing and multicultural human-computer interface design.
ZHANG et al. (Thu,) studied this question.