Face processing relies on configural processing, which is thought to be particularly disrupted by inversion. We compared inversion effects in recognition experiments for three types of stimuli, using faces (Experiment 1) and houses (Experiment 2). Stimuli varied by their colour only (colour), by the spatial relations between components (relational), or by the components themselves (eyes, mouths, doors). For faces, relational versions revealed strong inversion effects, component versions moderate, and colour versions no inversion effect. Recognition of houses revealed no inversion effects at all. We suggest that the inversion effects observed for faces in the component condition are due to relational changes, which must accompany any change in components. This proposal may account for the rather inconsistent effects of inversion reported in the literature. Furthermore, we suggest configural processing seems to be somehow face-specific.