- Unlike vanilla GANs, CGAN, is a condition to both the generator and the discriminator. This condition tells the GAN what image we are expecting our generator to generate. So, both of our components—the discriminator and the generator—act upon this condition.
- The code, c, is basically interpretable disentangled information. Assuming we have some MNIST data, then, code, c1, implies the digit label, code, c2, implies the width, c3, implies the stroke of the digit, and so on. We collectively represent them by the term c.
-
Mutual information between two random variables tells us the amount of information we can obtain from one random variable through another. Mutual information between two random variables x and y can be given as follows:
It is basically the difference between the entropy of y and the conditional entropy of y given x.
- The code, c, gives us the interpretable disentangled information about the image. So, we try to find c given the image? However, we can't do this easily since we don't know the posterior
, so, we use an auxiliary distribution 
to learn c.
-
The adversarial loss alone does not ensure the proper mapping of the images. For instance, a generator can map the images from the source domain to a random permutation of images in the target domain that matches the target distribution. So, to avoid this, we introduce an additional loss called cycle consistent loss. It enforces both of the generators G and F to be cycle consistent.
- We have two generators:
and 
. The role of 
is to learn the mapping from 
to 
and the role of the generator, 
, is to learn the mapping from y to 
. - Stack GANs convert text descriptions into pictures in two stages. In the first stage, artists draw primitive shapes and create a basic outline that forms an initial version of the image. In the next stage, they enhance the image by making it more realistic and appealing.