Subjects studied faces in a full- or a divided-attention condition and then received a recognition test that included old faces, new faces constructed by combining facial features from previously studied faces (“conjunction faces”), and partly or completely new faces. Full- but not divided attention subjects responded “old” more often to old than to conjunction faces; all subjects responded “old” to these faces more often than to partially or completely new faces. Thus it is less attentionally demanding to encode facial features than it is to encode their interrelations. Dividing attention had identical effects on an incidental and an intentional learning group. Experiment 3 demonstrated that dividing attention primarily affected explicit recollection rather than stimulus familiarity.