DCGAN is initialized with random weights, so a random code plugged into your network would produce a completely random image. On the other hand, when you might imagine, the network has a lot of parameters that we could tweak, plus the purpose is to find a environment of those parameters which makes samples produced from random codes appear like t