This was trained to generate images of twisted and fragmented cityscapes with sideways and upside-down buildings.