This is the authors' abstract. We don't add key points for this paper.
AI-generated faces are increasingly used as substitutes for real facial photographs in various domains of science and in practical applications, but their correspondence to natural human faces remains largely untested. We have compared real facial photographs of Cameroonian women and men with faces generated by six widely used AI models (DALL-E 3, Midjourney, Stable Diffusion, Gemini, Qwen Image, and FLUX-1). Using geometric morphometrics, we have then quantified facial morphological disparity, sexual shape dimorphism, and facial asymmetry. In all models, AI-generated faces showed substantively lower morphological disparity than real faces, which indicates a systematic compression of facial shape variation. Sexual shape dimorphism was markedly stronger in AI-generated samples than in the real population, which suggests exaggerated or unrealistic sex differentiation. Facial asymmetry did not differ between AI and real faces overall, although we did observe a small reduction in AI-generated female faces. These findings show that, despite their possible perceptual realism, AI-generated faces are not morphologically equivalent to real Cameroonian faces. The consistent reduction of facial variability and distortion of population-specific sexual dimorphism underscore the need for quantitative morphometric validation of synthetic face datasets prior to their use in scientific research or applied settings.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Budnik et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: