[kevin_thibedeau]

The map-to-photo only seems to work because there is a limited training set built with the exact same photographic tile as the vector map. It's reproducing features that are not possible to infer.