Plus porn seems like one of the most optimizable problem spaces. There's heaps of material, lots of images that come in series and are just slightly different,

Plus porn seems like one of the most optimizable problem spaces. There's heaps of material, lots of images that come in series and are just slightly different, lots of pre-existing collections, categories, and tags for all of it. Everything has some specific name for it too. Porn is probably one of the easiest datasets to train on. You can also easily seek user feedback: ask people to rate images, point out defects, etc. You'll have your audience fine tuning it for free.