Meet The Face Generator AI
AI is a powerful tool. It can be used to do a lot of cool tasks. It can be trained to do anything from recognizing hand-written digits to driving a car. The use cases of artificial intelligence is vast and it is being used almost everywhere in today's world. ThisPersonDoesNotExist.Com is website that demonstrates how an AI can be used to generate insanely realistic human faces. When you go on that website, it shows you a picture of a person that does not exist. The picture it shows is generated by a computer program. You will see how hard it is to tell if the picture it made is real or fake. The AI can generate male, female, and child faces.
The Algorithm
The algorithm that the ThisPersonDoesNotExist website uses is called StyleGAN2. As the name suggests, StyleGAN2 is a GAN (Generative Adversarial Network). What's a GAN? While training an AI, we need a loss function that tells our AI how wrong it was while solving a training problem. A loss function is specific to a problem. A GAN provides a general solution for all use cases. You can train a GAN without having to worry about the loss function. The output of a GAN is also better than a traditional neural network. In some cases, it is also very hard to figure out a good loss function for a problem.
StyleGAN2 is not the only GAN there is. There are also conditional GANs like CycleGAN and pix2pix GAN. These are used to transform an input image into another image. They are called conditional because their output is conditioned on the input image. Common example that is associated with a CycleGAN is that of turning a picture of a horse into zebra. For pix2pix GAN, it is of turning a pencil drawing into a realistic art.
How was ThisPersonDoesNotExist's AI trained?
The AI of ThisPersonDoesNotExist was trained by feeding it images of human faces taken from the internet. Web scrapers can be used to get the images. Sometimes, a little bit of manual effort is also needed for cropping/adjusting the scraped image to make it feasiable for training process. Training process is very time-consuming and computationally expensive. If you don't have a high-end machine, you probably won't go too far. CPU isn't enough for the training process, GPUs are what people use these days to train AIs.