Edits images from one to three reference images and a natural-language instruction, preserving key details such as facial features and identity while applying the requested changes