In this paper a model is presented in which face recognition is analysed into several stages, each of which may be independently lateralized. Evidence is reviewed which suggests that lateralization is important at all stages of processing a face. Early visuospatial processing, and the creation and comparison of facial representations, appear to be carried out more efficiently by the right hemisphere. Comparisons based on discrete, namable features of faces may yield a left hemisphere advantage. It is also proposed that faces may activate semantic information, including names, more efficiently in the left hemisphere. The model is useful in resolving inconsistencies in the degree and direction of asymmetries found in face-recognition tasks. Suggestions are also made for future research.