To investigate temporal integration in face recognition, top and bottom halves of pictures of famous people were presented sequentially, either upright or inverted, with varying temporal intervals between the two halves. The inversion effect, a marker of configural processing, was comparable across 0-400 ms intervals, but decreased at intervals exceeding 400 ms (Exp. 1). When an interfering stimulus appeared during the interval between the two face parts (Exp. 2), it disrupted the integration of the parts but not their perception. This is the first report of such an effect. Thus, performance equalled the combined accuracy of each part when presented alone, which in turn was worse than when they were integrated. Our findings indicate that (a) configural processing of faces depends on integration of face parts that are maintained temporarily in a visual buffer; (b) without integration, identification depends on recognition of individual parts whose contributions are additive; and (c) an interfering visual stimulus can obstruct integration, but leaves perception of individual parts intact. The ability to integrate temporally separated face parts into a unified representation is discussed in light of theories of face perception and temporal integration.