Realistic Human Face Generated By Artificial Intelligence
Look at these face. Do you see anything creepy? They all look normal, right? Well, these faces do not belong to any real human beings; these are generated by Artificial Intelligence! Mind blowing, aren’t they?
To generate these realistically unrealistic photos, NVIDIA has developed a system that is based on Generative Adversarial Network (GAN). So, what is a GAN? To just give you a brief idea, let’s look at a simple situation:
You are a director and you are struggling to find an exquisite avatar for your upcoming movie. Can you create an artificial designer who will able create an avatar by looking at past avatars created by designers around the world?
To generate a model that can create a completely new avatar, A GAN system needs a hero (generator) and a villain (discriminator).
The villain, D, is designed to identify all the avatars stored in a database. Now, the hero, G, needs to fool the villain, D, by crafting avatars from a training set (latent space) with a mixture of noise. The training set provides a fine avatar from the database and then the hero tweaks the avatar (using noise) that looks almost right but slightly different than an original avatar. For instance, the hero, G, may generate an avatar by randomly altering its ears or eyes. The result is very similar to a normal avatar stored in the database with the exception of altered eyes or ears.
Now, if the hero cannot fool the villain, it will fine-tune its parameters to beat the villain. And, if the hero wins, the villain will fine-tune its parameters to make sure the hero cannot beat it again. This process continues until a perfect result is achieved and the director is happy!
NVIDIA made a great progress given a very similar case depicted above. Using CelebA-HQ’s database of photos of famous people, it created a GAN system to generate photos of people who do not exist!
First, the hero, G, created an image at a lower resolution. The villain, D, had a chance to analyze the image created by the hero. As the hero progressed beating the villain, higher resolution images were being created by the hero to beat the villain. Here is the result from the GAN system published on YouTube.