0

It seems to me that the Generative Adversarial Networks have a practical issue when trying to reproduce some of their output images

For example, as you can see https://www.youtube.com/watch?v=oIzwe_MOeQI&t=1057s seems to generate shape and background changes as it changes rotations.

Can any of the StyleGANs be convinced to take an image of a car, along with some sort of key that says how it is rotated then consider that when the trained model is generating?

Obviously, it might be a substantial effort to label the data, but I'm curious if it's been done in a well-published fashion.

Mariusmarten
  • 383
  • 1
  • 16
user1833028
  • 101
  • 3

1 Answers1

2

The idea of taking an additional label as input sounds like a conditional generative adversarial network (cGAN). Within a conditional GAN both the discriminator and the generator are conditioned on extra information provided through a label. The idea was first introduced in this paper.

This image from this paper nicely highlights how a conditional GAN differs from a standard GAN:

GAN_comparison

Such an architecture would likely also be suitable to be conditioned on image rotations.

Mariusmarten
  • 383
  • 1
  • 16