ML model and GenAI

Question

When you ask AI to generate a profile picture of a man, on the technical side, I was wondering how much is the variation between the generated image (the man that doesn't exist) and the trained images in the model? To be more clear: would the image generated be very similar to one of the images that have been trained? But with slight change bigger eyes? Longer eyebrows? Etc

Answer 1

OmarSa the ML models are only capable of learning from the trained images and predict them they can't generate any new images. In Case GenAI models are capable of creating realistic images using Generative Adversarial Networks also known as diffusion models which makes a person or AI itself difficult to understand whether it's AI generated or a real image.

Answer 2

So AI generated images must look similar to the trained data? Can we say that? Aysha

Answer 3

See in GAN there are 2 models one is a discriminator and another is a generator. The generator tries to create realistic fake images and the discriminator tries to tell it's fake and gives probability. So it keeps on repeating until the discriminator says you're real I mean the generated image. So it happens like this.

check out this link for detailed info
https://www.techtarget.com/searchenterpriseai/definition/generative-adversarial-network-GAN

Answer 4

Many thanks Aysha