This work investigates ways painters compress the range of luminances present in the natural world into the far smaller range available for painted art. Earlier work suggests that a diverse group of scanned paintings is similar to natural scenes in terms of statistical and modeled neural response properties. While this work suggests paintings recreate basic statistical properties of natural scenes whether subject matter is concrete or abstract, it also suggests artists apply luminance nonlinearities to paintings which resemble nonlinearities in early vision. We tested an expanded set of paintings to determine the extent to which low-level statistics and modeled responses can describe paintings' provenance or type (i.e., landscape, portrait or abstract). Image type experiments were based on forced-choice studies with human judges; judges agreed on classifications for 57 images. Sparseness of modeled retinal responses was similar for all types of paintings but applying a static nonlinearity before filtering, landscapes were more sparse than the other types. Retinal response sparseness was higher for Western provenances but the same for both provenances with the nonlinearity. These regularities could help explain how artists approach the task of compressing the range of natural luminances at different spatial scales, in abstract form, and across cultures.