In a previous post I tried making dragons with not much success. This time I’m just using an existing model for creating people (see thispersondoesnotexist.com). As you can see the model produces very photorealistic images of people that are completely imagined.
The general architecture of the model, a Generative Adversarial Network or GAN, is shown below. There is a generator network that tries to create fake images. These fake images along with real images are fed into a discriminator that then tries to learn the difference between fake and real. It is called adversarial because the two networks are working at cross purposes: the generator learns to create more and more realistic images based on feedback from the discriminator, and the discriminator also improves over time, trying to prevent from being fooled by fake images. Ideally, the networks converge when the discriminator can’t tell the difference between real and fake, and the fake images approach what humans would think of as real images.
What I found particularly fascinating by this process is that the generator takes as input random noise. It then learns how to convert that noise into an amazing photo of a person that doesn’t exist. I find that to be a metaphor for my life. Over and over again, I see things that don’t exist. Life is going well, so I think God is looking out for me. Life takes a turn for the worst, God has something for me to learn. People that mean something to me, and I mean something to them. But I think it is me just imagining up things from the noise of life. I see what I want to see, but there’s no substance there. And of course I have a poem to go along with the musing:
Theotherpersondoesnotexist
I am a God
I create
Nothing into Something
I see value when it isn’t there
Visions of happiness but it’s just thin air
Loving people that don’t really care
A meaningful life but it’s just a prayer.
Structure that came from noise
Is that all life is?
Here are a couple of the fake images at full resolution.
