The role of variability in appearance, exposure and learning procedure in dynamic face learning

Research suggest that humans recognize familiar faces reliably but struggle to learn new ones. Variations—in facial appearance and viewing conditions—facilitate learning new faces in lab-based studies. In contrast, stability in appearance seems to support early learning stages in the real-world and changes in appearance disrupt recognition regardless of familiarity levels. To reconcile these findings, we have proposed a cost-efficient learning mechanism: stable appearances help form initial coarse representations, while variability yields refinement over time. We tested this through five pre-registered experiments using a strictly controlled but ecological video database. A total of 922 participants (over years 2024 and 2025) learned two stable and two variable faces across different exposure levels (3, 6, 9, or 12 videos). Learning occurred either through spread learning episodes (Experiments 1, 2 and 5) or grouped blocks (Experiments 3 and 4). The similarity between test and learning materials varied across experiments. We found that stable faces were recognized better than variable ones, but only when test images closely resembled stable learning material. Recognition only improved with additional exposure in the spread learning condition. These findings help bridge real-world and lab-based face learning research, refining our understanding of how humans may encode facial information efficiently.