The above paper uses two methods to construct the same image, one through SGD with the loss against pixels and the other through an "evolutionary" method. Although both methods construct the image, the evolutionary method stores better representations of how to construct the image in its weights.