Reverse correlation (RC) techniques provide a data-driven approach to model internal representations in an unconstrained way. Here, we used this approach to model social perception of faces. In the RC task, participants repeatedly selected from two face images—created by superimposing randomly generated noise masks on the same face—the face that looked most trustworthy (or, in other conditions: untrustworthy, dominant, or submissive). We calculated classification images (CIs) by averaging all selected images. Trait judgments of independent participants, as well as objective metrics, showed that the CIs visualized the intended traits well. Furthermore, tests of pixel clusters showed that diagnostic information resided mostly in mouth, eye, eyebrow, and hair regions. The current work shows that RC provides an excellent tool to extract psychologically meaningful images that map onto social perception.