I wonder, based on your comment, what the output would look like if the training was inverted. Anything that looks like art would skew towards zero, for instance.