Apart from its traditional definition as an autobiographical and temporally dated experience that can be consciously recollected, episodic memory can also refer to a trace of a momentary sensory input—a snippet of information that may serve either as a building block for developing more abstract representations or as a subconsciously accessed piece of memory. In both roles—as subconscious snippets and as components of abstracted knowledge—this kind of information requires consolidation for long-term retention. However, behavioral measures that clearly distinguish the consolidation of such snippets from that of abstractions, as well as the consequences of each type of consolidation, remain rare in the literature. We present a study aimed at achieving such a separation, based on the phenomenon of transfer learning, which involves the re-application of previously learned higher-level regularities to novel input. Previous empirical studies have investigated human transfer learning in supervised or reinforcement learning settings, typically focusing on explicit knowledge. Consequently, it remains unknown whether such transfer occurs naturally during the more common type of implicit and unsupervised learning—and, if so, how it relates to the consolidation of nonspecific, unconscious memory across different levels of abstraction. We compared the transfer of newly acquired abstract knowledge—ranging from somewhat explicit to fully implicit—during unsupervised learning by extending a visual statistical learning paradigm to a transfer learning context. The visual statistical learning paradigm exposes observers—without any task—to a large set of compound images composed of separate shapes and subsequently measures sensitivity to hidden structures, such as pairs of shapes that consistently repeat within the stream. We introduced higher-level features into the paradigm by biasing the dominant orientation of the hidden pairs to be either horizontal or vertical. In the transfer phase, all shapes were replaced with a new, never-before-seen set, and we measured how the horizontal or vertical structure embedded in the first phase influenced the implicit learning of new pairs of any orientation in the second phase. Using this method, we found evidence of transfer during unsupervised learning, but with important differences depending on the explicitness or implicitness of the acquired knowledge. Observers who acquired more explicit knowledge of pair associations (without awareness of the dominant orientation structure) during the initial learning phase were able to immediately transfer the general orientation information to the second phase by learning similarly oriented pairs better. In contrast, observers with the same amount of implicit knowledge showed the opposite effect—structural interference during transfer. Importantly, when sleep occurred between the learning phases, these implicit observers—while still remaining unaware of structures—shifted their behavior and exhibited the same pattern of transfer as the explicit group. This effect was specific to sleep and did not occur after a comparable period of wakeful consolidation. Our results highlight both the similarities and differences between explicit and implicit learning, as well as the influence of available explicit and implicit knowledge on the acquisition of generalizable higher-level knowledge. They also underscore the complex role of consolidation in restructuring internal representations.