What are the boundaries of implicit learning of contingencies? We could show in previous work that visual features, such as color and shape of distractors, can be learned to predict a target location in a visual search task. In the present study, in an adapted contextual cueing paradigm, we tested whether participants can learn such contingencies in complex, real-life scenes. In a first experiment, the contingency was between predominant color of a scene, and target location. In a second experiment, the scene category (bathroom, kitchen, bedroom, living room) predicted the target location. Participants were either explicitly instructed to learn these contingencies (explicit learning condition) or not informed about any contingencies in the task (implicit learning condition). We found evidence for implicit learning of low-level feature cues, but for semantic cues, learning was only evident in the explicit learning condition. In the implicit learning conditions, we found no evidence for explicit knowledge of the contingencies. We discuss our findings against the background of theories of consciousness and frameworks of implicit learning mechanisms.