How "far" from the training art is the resulting generated art? Do the generated cards have features which are recognizably lifted from the training data?