DCGAN is initialized with random weights, so a random code plugged to the network would generate a totally random graphic. Nonetheless, when you might imagine, the network has millions of parameters that we could tweak, and also the aim is to find a location of these parameters which makes samples g