Two experiments were designed to isolate the facial information utilised in the learning of new faces. In experiments 1 and 2, two groups of subjects were each trained on different groups of faces by means of a dynamic video presentation. They were then shown both trained and novel faces in a same-different decision task, where 'different' trials included manipulations of internal and external facial features, and the task was to decide whether two images were identical or had a difference in one or more features. Both experiments showed that hair change was most easily detected in untrained (unfamiliar) faces. When faces had been trained (familiar), detection of eye changes was selectively enhanced and sensitivity to hair changes was maintained. While previous studies have suggested that familiar face representations are weighted towards their internal features, our experimental results show that this is due to selective enhancement of sensitivity to changes made to the eyes alone, with no reduction in the salience of the hair. Moreover, within the limits of the familiarisation used here, there was no enhancement of the representation of the other internal face features examined.