In daily life, face identification requires that the observer select a single representation from hundreds if not thousands in memory. This breadth of choice is nearly impossible to replicate in the laboratory using newly learned faces, especially in the context of a Bubbles experiment (Gosselin & Schyns, 2001). In this study, we obviated this concern by studying the performance of observers in a face-naming task using 210 faces of celebrities. On each trial, we presented a face randomly sampled with Bubbles. We performed least-square multiple linear regressions on the location of the samples and on accuracy to pinpoint the facial features that were used effectively in this task. Correct face identification relied primarily on the eye areas in spatial frequency bands ranging from 4.37 to 70 cycles per face (cpf) and on the mouth and the nose in a spatial frequency band ranging from 8.75 to 17.5 cpf. A comparison with other studies (Caldara et al., 2005; Schyns, Bonnar, & Gosselin, 2002) in which Bubbles was used with a set of 10 newly learned faces revealed that although the eye areas were useful across studies, the mouth area and higher spatial frequencies gained in importance when few newly learned faces were used.