Beyond FID: Human Perceptual Judgments Reveal Systematic Blind Spots in GAN Face Evaluation

Generative Adversarial Networks (GANs) can synthesize highly realistic facial images from random noise
vectors. The Fréchet Inception Distance (FID) is widely used as a standard metric to automatically
evaluate the quality of GAN-generated images. However, it remains unclear to what extent this statistical
measure reflects human perceptual judgments, which ultimately define image realism in practical
applications. To address this, we conducted a psychophysical study in which participants (n = 20)
performed a two-alternative forced-choice task, assessing actual photographs and GAN-generated images as
real or fake. We show that while FID provides a reliable global ordering of image quality, it
systematically fails for localized semantic artifacts (e.g., eyewear and skin texture) that
disproportionately affect human realness judgments. This demonstrates that FID and human perception are
not merely noisy versions of the same signal, but that FID has systematic blind spots for localized
semantic artifacts that disproportionately drive human realism judgments.