it is still prone to recreate some characters, but I think this is due to lack of diversity of dataset