In this paper we address the problem of learning image structures directly from sparse codes. We first model images as linear combinations of molecules, which are themselves groups of atoms from a redundant dictionary. We then formulate a new structure learning problem that learns molecules directly from image sparse codes, namely from the image representation in the atom domain. We build on a structural difference function that permits to compare molecules and we derive an algorithm that analyses sparse codes and estimates the most relevant signal structure without reconstructing the images. Experiments on both synthetic and real image datasets confirm the benefits of our new method compared to traditional learning methods.