Quantity over quality? The impact of the number of faces to learn and of their appearance on the robustness of facial representations

For decades, researchers have sought to identify the factors that support robust face learning. Prior work has shown that exposure to varied images improves later recognition. Because image variation can include changes in viewing conditions, pose, expression, or appearance, our research has focused specifically on the role of facial appearance—kept stable or made variable through changes such as hairstyle, hair colour, or makeup. This focus was rooted in the idea that an efficient face‑learning system should refine memory representations only when necessary, after exposure to changes in appearance. Some of our earlier studies showed that memory representations were more robust when participants learned four faces with variable appearances compared to four faces with stable appearances. However, other studies found either no difference between the two conditions or better performance for stable faces when only two faces were learned per condition. To clarify these discrepancies, we conducted a preregistered study manipulating both the number of faces to learn (four, eight; between‑subjects) and appearance (stable, variable; within‑subjects). 236 participants learned faces from the same set of static images. As predicted, recognition performance was higher in the four‑face condition than in the eight‑face condition, indicating a cost of increased learning load. We also found a main effect of appearance, but no significant interaction between appearance and number of faces. Nonetheless, paired comparisons revealed higher recognition performance for variable faces in the eight‑face condition, whereas no difference between appearance conditions was observed when only four faces were learned. These findings suggest that learning load strongly influences recognition performance, and that the effect of appearance may depend on the allocation of limited memory resources in a given learning context. Learning load may therefore help explain inconsistencies in the face‑learning literature and should be considered in future research.