I suggested that to deeppomf a while ago (I was thinking of simply using unlabeled anime images from https://gwern.net/Danbooru2017 with random area deletions to simplify the model & training process as much as possible) and his belief is that because genitals are such a small fraction of any images, and the rest of images vary so much while genitals are a fairly small narrow domain, a generic inpainting/denoising CNN will learn to inpaint pretty much anything else possible and neglect genitals specifically.

Presumably if you trained a really big inpainting CNN a lot, it would learn genitals (along with everything else), but it's understandable that he would try a much more targeted approach.