DCGAN is initialized with random weights, so a random code plugged into the network would deliver a totally random image. On the other hand, as you may think, the network has numerous parameters that we can easily tweak, along with the goal is to find a location of these parameters which makes samples created from random codes seem like the teachin