Seems the less accurate images occur at smaller training set sizes (at least that's how I read those image captions about kimg).