Using a delayed visual search paradigm, we showed that non-rigidly moving faces are better encoded than static faces (Pilz. Thornton and Bülthoff, 2006). In this task, observers learned one dynamic and one static face, and then searched for either target in a static search array. Here, we used high (HSF) and low (LSF) frequency filtered faces during visual search to investigate whether the difference lies in the encoding of different spatial frequencies. In Experiment 1 (N=12), we used a learning procedure which only required observers to rate the targets along different character traits. We found no advantage for dynamically-learned faces, but HSF faces were recognized more accurately (p[[lt]].05). In Experiment 2, we used our previous learning procedure which required observers to assess both targets' personality and facial features using a detailed questionnaire. Observers (N=8) were faster at finding dynamically-learned faces (p[[lt]].05), and more accurate at finding LSF faces (p=0.07). Taken together, these results show that the nature of learning can affect face encoding strategies. Furthermore, the frequency effects suggest that less familiar faces may be recognized more from features than from configural information. In Experiment 3, we tested whether the dynamic advantage was due to the higher familiarity of dynamically-learned faces. Observers (N=8) searched for a colleague and an unfamiliar face, learned with the procedure from Experiment 2. We found that observers were faster (p[[lt]].01) and more accurate (p[[lt]].01) at finding their colleagues, which suggests that the dynamic advantage partly depends on the level of familiarity with the target face.