1

I have seen many examples of NVIDIA StyleGAN. They generate endless human faces, anime faces, cats, dogs. I wonder is it possible to generate multiple images of the same human, anime, cat or dog?

pouya
  • 3,400
  • 6
  • 38
  • 53

1 Answers1

1

StyleGAN doesn't have any concept or representation of a specific subject (person, animal or what have you). It only has "styles" at different levels of scale (coarse, medium, fine). It also has noise inputs at each level. If you vary the styles you're probably going to get images that look like different people (different shapes of the face if you change coarse style, different skin color and hair if you change finer layers). If you keep the style inputs the same and let the noise inputs vary, you will technically get different images of what will probably look like the same subject, but they won't be different in any interesting way.

I think there are more recent versions of StyleGAN that allow you to change the pose of a face while keeping the identity more or less the same, but the original StyleGAN makes no such guarantees.

Kristoffer Sjöö
  • 286
  • 1
  • 9
  • So i guess it is possible but not for now. This could be particularly the Fintech world nightmare now that the world is moving towards virtual banking fast (things like COVID-19 is accelerating this even more) and these companies are competing over account opening time. How many of these accounts will be genuine? You can literally fake everything – pouya Aug 24 '20 at 14:45
  • @pouya In the StyleGAN2 paper (see https://github.com/NVlabs/stylegan2) the authors point out that with their newer version makes it easier to detect whether a given image has been generated by the model or not. So they're certainly aware of the issue. – Kristoffer Sjöö Aug 25 '20 at 16:54