Nakwon Rim¹, Marc G. Berman¹, Stefan Uddenberg², Shunichi Kasahara³, Wilma A. Bainbridge¹; ¹University of Chicago, ²University of Illinois Urbana-Champaign, ³Sony Computer Science Laboratories

Recent advances in generative artificial intelligence now enable the generation of synthetic human faces that observers cannot reliably distinguish from real ones. These models offer a powerful methodological opportunity for research on face perception and cognition: they are able to produce an arbitrary number of photorealistic faces with rich latent representations that can support systematic sampling and targeted transformation. Here, we evaluated whether such generative models can serve as an effective tool for understanding and augmenting face memorability. We first conducted two analyses on memorability scores for 1,004 StyleGAN2-generated faces to test whether their memorability aligns with that of real faces. In the first analysis, a neural network trained exclusively on the memorability data of real faces successfully predicted memorability in our synthetic set. In addition, we found that facial judgment features previously shown to relate to the memorability of real faces also showed the same relationships with the memorability of synthetic faces. For example, faces judged as more “typical” were less memorable, paralleling previous findings for real faces. Finally, having established this correspondence, we leveraged StyleGAN2’s latent space to test whether memorability could be systematically manipulated. Across two experiments employing distinct manipulation strategies, we were able to consistently increase the memorability of synthetic faces but failed to decrease it reliably. This asymmetry suggests that memorable faces may share a relatively homogeneous set of visual attributes that are readily captured in generative latent spaces, whereas forgettable faces may arise from more heterogeneous, idiosyncratic configurations that are harder to capture. Together, these findings demonstrate that generative models can serve as a valuable tool for studying and influencing face memorability. More broadly, the ability to algorithmically enhance face memorability opens the door to applications such as real-time enhancement of facial recall in communication or educational settings.

Acknowledgements: The research was supported by the National Science Foundation under CAREER Grant No. 2441710.