The overall architecture of the GAN is shown in the following figure. The generator and discriminator, which are each separate deep neural networks, are oversimplified as a black box for the sake of easy consumption. We will get to their individual architectures shortly, but first I want to focus on how they interact:

The generator is given a vector of random noise (z) and creates an output G(z) (an image in the case of a DCGAN) that it hopes will trick the discriminator.

The discriminator is given both real training data (X) and generator output G(z). It's job is to determine the probability that it's input is actually real P(X). 

The discriminator and generator are both trained together, in a stack. As one improves, the other also improves until hopefully the generator produces such good output that the discriminator is no longer able to identify the difference between that output and the training data.

Of course, there are a few more details we should cover before you'll be ready to build your own GAN. Next, let's take a deeper look at the generator.