Recent studies suggest that a stimulus actively maintained in working memory (WM) automatically captures visual attention when subsequently perceived. However, such a WM guidance effect has been so far observed only for stimuli defined by simple features, such as colour or orientation. Here we investigated whether the effect occurs also for naturalistic stimuli, whose identity is defined by multiple features and relations among them, specifically for faces and houses. The experiment comprised two conditions: a WM condition and a mere exposure condition. Subjects (N = 28) either memorized or merely saw a template stimulus, and then performed several dot-probe trials, in which pairs of stimuli were presented laterally as distractors (memorized or seen on the one side, control stimulus on the other) and followed by a target dot. We found that response times were faster when targets followed a memorized stimulus than when they followed a control one, and that the memory-matching stimuli evoked the N2pc ERP component. Further, in an exploratory analysis we found electrophysiological evidence for an early (100-200 ms post stimulus) prioritization specific to the memorized faces. Importantly, neither the RT, nor the electrophysiological effects were observed in the mere exposure condition. In conclusion, showing an attention capture by complex stimuli maintained in WM contributes to identifying the mechanism of WM-based attentional selection and provides novel evidence that complex stimuli can, in principle, guide attention automatically. Finally, we interpret the early prioritization of the memorized faces as evidence supporting the sensory recruitment theory of WM.