Face Averageness as a Predictor of Perceived Realism and Trustworthiness in Synthetic Faces

Ivory, Matthew and Nightingale, Sophie (2026) Face Averageness as a Predictor of Perceived Realism and Trustworthiness in Synthetic Faces. In: Proceedings of The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 :. IEEE, USA. (In Press)

Text (IvoryNightingale2026_FaceAvgAsPredictorTrust)
IvoryNightingale2026_FaceAvgAsPredictorTrust.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (14MB)

Abstract

Human perception of faces is theorized to occur in face-space, where faces are placed within a multi-dimensional space with dense clusters or regions indicating typicality or increased averageness. It has previously been suggested that the increased realism and trustworthiness of synthetic faces may be a result of their increased averageness compared to real faces. In this paper, the averageness of real and synthetic faces (GAN-generated and Diffusion-generated) was calculated using a generalized Procrustean analysis to assess the impact of averageness on perceived realism and trustworthiness. As averageness increased, no meaningful relationship was observed with the correct classification of real, GAN, or Diffusion faces, revealing a uniform accuracy rate across averageness. The relationship between trustworthiness and averageness was modulated by both gender and race, demonstrating complex relationships between real and synthetic faces as well as between the two synthetic generators. The findings highlight the complex ways synthetic faces occupy human face-space and, as such, have important implications for researchers using synthetic faces instead of real ones, particularly in perceptual studies.