I'm actually curious about this one in particular, because it seems to vaguely match either overall shape and color or the type of object something is. I kind of feel like it must be doing some layered thing to composite the images and is only generating parts of it dynamically because of specifically how it is both good and bad at mimicking images.