What are GANs that take images as input instead of latent vectors?

Question

Here I am a newcomer to the field of GAN. I know the original GANs take latent vectors as input. But if I want to complete tasks as style converting and watermark removal, the input should possibly be an image.

Then it leads me to think that I probably need an autoencoder to translate an image to latent vector if I want to do such work based on original GAN architectures. Is it a legit idea?

Now I know Pix2pix is likely what I need. But what are the early-era GAN architectures to accomplish this 'image converting' task?

Many thanks.

Answer 1

There is VAE-GAN, which likely can achieve what you want, you likely don't even need the "variational" part, You might also want to look into CycleGAN.