Not necessarily, the NN may just concentrate on them in the porn it's trained on, because it's a common factor that's in porn, but not in beach pics for example. (whereas in both you have almost naked people).
Agreed. It's fascinating, really. To me it seems that the algorithm is generating both male/female genitalia, but it's sort of "choosing" based on what "looks best".
That is to say, certain features that in their SFW form already look vaguely like a penis seem to be depicted as a penis (the statues, volcano, etc), whereas other features like the beaches were depicted as vulvas. And the slight red tinge present in some of the canyon/beach images also elicited the vulva, even if the shape of the surface wasn't something "obvious".
Man, ironically this stuff to me is true abstract art. To use a somewhat trite cliche, the [extremely subtle] erotic nature of these environments that are clearly not actually genitalia gives insight into how the human mind works just as well as it does into (convolutional?) neural networks themselves
I had thought these were talking about 'pornography', but this makes it clear that a very large amount of its training set must consist of dick pics.