We consider one of the most basic questions in computer vision, that of finding a low-level image representation that could be used to seed diverse, subsequent computations of image understanding. Can we define a relatively general purpose image representation which would serve as the syntax for diverse needs of image understanding? What makes good image syntax? How do we evaluate it? We pose a series of such questions and evolve a set of answers to them, which in turn help evolve an image representation. For concreteness, we first perform this exercise in the specific context of the following problem.