Difference between PatchGAN and Fully Convolutional Layer


I’m reading this post to get to know PatchGAN more.
And I saw this pic.

How different of between PatchGAN and Fully Convolutional Layer?
Those are looks so same.

Hi! I am the author of this blog. I can answer this question.
See, whatever you have heard all about ConvNet like ResNet, U-Net, etc are like usual. Whereas PatchGAN is special case for ConvNet especially Discriminator in GAN theory. In PatchGAN, the output of the architecture only infer you whether it is fake or real. That’s IT!!

Now see the image below and let say,
if each pixel close to ‘0’ means fake
if each pixel close to ‘1’ means real.
(As I already explained in the blog that: each pixel from output comes from 70x70 image patch from Input image)

Now what does its infering you? Its telling to the Discriminator that: there are some image patches from an input image which contain fake and other image patches which contain real.

Now, If you are familiar in GAN theory, you should know they were trying to achieve their own the objective of generator and discriminator model. So, PatchGAN comes under Discriminator.

Hope you got it.


Oh, I like your post.