Look up "adversarial networks". If you have a classifier that tells when an image is faked, then you can use that classifier to train a network to generate images which make the classifier say the images are real.