How do GAN's differ?

I am aware variations of GAN’s exist; i.e DCGAN, DiscoGAN, CicyleGAN; Source: https://github.com/eriklindernoren/PyTorch-GAN; My question is, what are the primary constructs that differentiate GANs?
I would imagine that having a specified network architecture acting as the Discriminator creates a variety of a GAN network, but I’m unsure and need clarification. Please let me know if this thinking is correct or not, and in addition please elaborate on what other factors lead to a variant GAN network.
Thank you very much!

Besides changes in the architecture, different GAN variants use different loss functions, connections, inputs etc.
This older paper gives a brief overview. You can also find some blog posts about different GANs, but I’m not sure if there is a good “complete” overview.
If I’m not mistaken, at one point a lot of new GAN architectures were published daily on arxiv, so it was quite hard to follow which variant might be promising.

1 Like

That was helpful! Thanks a lot