For every tile it gives the model three references at once: the tile, a 3x3 map of its neighbors, and a thumbnail of the whole image. So each tile knows where it sits before it refines, which means…