Stable Diffusion, refined with img2img and rearranging keywords in the prompt across runs for emphasis on certain facial features.