We are very familiar with certain objects; we can quickly recognize our cars, friends and collaborators despite heavy occlusion, unusual lighting, or extreme viewing angles. We can also determine if two very different views of a stranger are indeed of the same person. How can we recognize familiar objects quickly, while performing deliberate, perceptual inference on unfamiliar objects? We describe a model combining an identity classification network for familiar faces with an analysis by synthesis approach for unfamiliar faces to make rich inferences about any observed face. We additionally develop an online non-parametric clustering algorithm for recognition of repeatedly experienced unfamiliar faces, and show how new faces can become familiar by being consolidated into the identity recognition network. Finally, we show that this model predicts human behavior in viewpoint generalization and identity clustering tasks, and predicts processing time differences between familiar and unfamiliar faces.