I’ve been thinking about face generators a lot. I still think about the Face Generator web app a lot. But I also think about This Person Does Not Exist a lot. Those images are generated by an algorithm that has been trained by another one. The first algorithm is given the task of deciding whether a face we give it is a real face or not. Since we know the answer, we can train it. The second algorithm is given the task of creating a face. We obviously know that they’re all fake, so we mix these into the faces for the first algorithm. The goal of the second algorithm is to fool the first. This is a generative adversarial network (GAN) setup. It’s fascinating because it allows us to train a face generator without having to actually make any decisions since we simply train the first one on real faces.
The software exists. The person running This Person Does Not Exist links to three of them. One of them is called Lightweight GAN. It looks fascinating.
The problem is getting the faces. It seems to me that all such collections are ethically flawed. None of the people gave their consent: these are faces scrubbed from social networks, or mugshots, or whatever else. It makes me very uncomfortable.
And yet... And yet! It’s fascinating. With a little post-processing on the command-line, you can get great portraits for your campaign wiki.
True, I wonder about the colour mix, for example. But the fascination is there, for sure.
This is my Makefile:
face: bw.png display bw.png bw.png: source.jpg convert $< -resize 300x300 -monochrome -quality 9 $@ clean: rm *.jpg # get image from the service source.jpg: curl https://thispersondoesnotexist.com/image -o $@