My thought is that they basically run gradient descent on the image where the loss is the magnitude of one output plane in one of the layers of the neural network. Probably using gradient descent to push up the magnitude of one of the output plane layers or something like that.