Because we are learning a lot of images in mannequin state, maybe we can't output them well.