I can't help but think that DiscoGAN (https://arxiv.org/abs/1703.05192 "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks", Kim et al 2017) would be perfect for this. Simply feed it a ton of photos of regular clothed people and other naked people (don't have to be the same people), and it'll learn a mapping on its own. Scale that up, refine it for a few years...