Are GANs magic , or am I missing something ?(Project Idea)

It seems that you can give a GAN anything , and tell it to do anything.
(for eg. from noise it can generate faces, provided you have a few thousand faces to allow the discriminator to do its job)

I was thinking of giving a GAN a patch of a person’s face (probably the top area ,and everything else blacked out), and the model’s task is to complete the rest of the face.

My understanding of GANs is not very good , but isnt this task exactly what Jeremy should us for the CIFAR using WGAN ?
The only thing that will change is inputs and outputs. Am I missing something or is this really just it ?

You should check out this new paper from Deepmind on a change in implementation in GAN’s. They actually use this exact example in their paper to show its efficacy.

1 Like