Assuming you don’t want to go to effort of training for a single image, I’d try inputting the image of a face in img2img mode with some Controlnet options to selectively retain elements of the image.

I'm pretty sure Inpainting can deal with that. For the background, for example, you keep the background as original and set the impact on a low end and it should keep it similar but different.

There's a specific "inpaint" model that supposedly gives better results for such things, but I'm no expert.