Learning and extracting regularities from the environment is critical for developing a mental model of the world. This process can be either intentionally with one's explicit inference engagement or incidentally without one's awareness of it. However, whether a single or multiple memory mechanisms are involved in these two learning approaches is still under debate. Here, behavioral and electrophysiological evidences from a visual statistical learning task using natural images were presented, and our finding is in favor of an overlapping mechanism which is modulated by subjects' intention-to-learn. In addition to faster response in the intentional condition, response accuracy, ERP components, and EEG theta-gamma phase-amplitude coupling (i.e., an index of working memory load) were all affected by the temporal regularities embedded in the stimulus stream, regardless of subjects' awareness. Specifically, significant repetition priming effect was evident in the facial images with higher transitional probability (contingent as "triplets") in both incidental and intentional conditions. These images were recognized more accurately and elicited lower N170, suggesting that grouping in perceptual representations may contribute to both learning conditions. Moreover, the theta-gamma coupling level was higher in the triplets than stimuli without regularity in the incidental condition, and this pattern was reversed in the intentional condition. Collectively, although working memory may involve in both conditions, it can be allocated more efficiently in searching for regularities under subjects' intention-to-learn which leads to faster response.