Would it be possible to have an algorithm that produces the images from the training set whose parts are most similar to the produced output? This looks super impressive but I still wonder how much the network just recycles parts of images it has seen before.