You should be able to finetune on a dataset that has captions of human faces.