To generate realistic looking headshots, you need high quality training data (i.e. high quality headshots) of the person. Then what's stopping them from using stable diffusion and stylize those existing photos in the first place?