I had some fun with StyleGAN a couple years ago. I should probably have written this up, or something. Some others I showed it too thought I should. But I never took it too serious. I was just bored over a couple weekends when I had access to some beefy GPUs. And y'know, it's quasi-porno.
I opted for dudes not gals. Part personal preference, partly because I happened to get my hands on an enormous dataset of hot shirtless guys taking selfies.
The initial results were rather poor: https://i.imgur.com/npzDcCL.jpg
That's around when I realized how important the training data really is. Too much variation in pose. So I trained a model that detected the face, nipples, and belly button, and aligned them vertically with that. Then I briefly skimmed over the dataset manually, and deleted anything that just varied too much from the median, in a rather arbitrary way.
Results were better: https://i.imgur.com/bS7ERC6.jpg
So I just let it churn away for a week, heating up the basement. Some final results with narrow truncation and style transfer: https://i.imgur.com/rhnWnpK.jpg
One thing I find intriguing is how bland those men are, really. Oh, beautiful, but too classically beautiful. In a sense, StyleGAN is a compression algorithm, and highly truncated outputs are something like a median of the inputs. So the above reflect a sort of "average" idealized image, that contains the traits most commonly found, in the media I trained it with.