A familiar face is instantly recognized in a crowd. This cannot be achieved through a feature by feature comparison of the observed face with either an average face (norm-based model of face recognition) or with a set of similarly constructed faces stored in memory (exemplar-based model of face recognition). A modified norm-based model is thus proposed. Instead of memorizing an average face, the normal variations for each facial feature are used to construct a multidimensional volume of face-space devoid of unusual features, here defined as features whose metrics lie below the 5th or above the 95th percentiles for that feature. A face consisting of 100 independently variable features will thus have, on average, 10 unusual features. Face identification then becomes exception-reporting. It requires only 10 such rare features to render a given face a one in 1013 faces (P=0.0510=9.8×10-14). In a world containing 6.7×109 people, such a face would be unique. Faces remembered in this way can have their unusual features exaggerated or attenuated without loss of identity. This is the basis of caricatures and anti-caricatures. It also means that individuals belonging to a foreign race, possessing several features with modes beyond the "usual range" of the own-race population, will all look alike. Features that render a face unique in the own-race population are now shared by everyone in the foreign race. Average faces are more beautiful than the faces used in the averaging process. This makes evolutionary sense. Natural selection increases the frequency of fit features at the expense of maladaptive features. "Usual features" are therefore fitter than "unusual features", and play an important role in mate selection. Such an existing fundamental sexual attribute could easily have been harnessed for the fast and efficient recognition of individuals in the community. © 2009 Elsevier Ltd. All rights reserved.