For what it's worth, we're using a GAN to generate fake user avatars for our products. They're real enough that we can use them in advertising, and since there's no actual person whose photo was taken, we don't require a signed model release form - something that was really difficult to get out of modeling studios since we wanted to open source the images after. GANs really solved our problem in this sense.
It'd be really interesting to try that. I think anything organic it'd work well with - people, landscapes, etc. But I've noticed that a lot of the cars that come out of stylegan are pretty abstract - it seems to have trouble with machined objects. Lines don't always line up, which is fine on something organic since there's already an aspect of randomness in our bodies, but on mechanical objects it feels foreign.
Once we open source this I'd be interesting in doing gan placeholders as a service though, completely license free.
