Explaining AI-Generated Portraits to My Wife: A Tale of Art from Algorithms

Yesterday night I faced an intriguing challenge: explaining to my wife how the AI-generated portraits I created were of people who don’t exist and never will. Despite her background in the IT field and a some knowledge of models and AI, the concept of these hyper-realistic images still provided a delightful twist.

The Challenge

My wife was both amazed and a bit puzzled when I showed her a collection of portraits created using a stable diffusion model. Her first question was quite natural: "Who are these people?" She thought these were actual photographs of real individuals, perhaps from some hidden part of a social media site or a photographer’s portfolio. The concept that they were born not from a camera but from an algorithm was a bit mind-bending.

How I Explained It

To make it simple, I compared the AI model to a very imaginative artist. Here's how I usually broke it down:

Imagine you give a super creative artist a box of crayons and a task to draw a completely new face, one that's never been seen before. Instead of drawing from sight, they close their eyes and think of different features like eyes, noses, and mouths they've seen before. Then, they mix these ideas up to create a face that doesn’t exist in real life.

That's similar to what the AI does. It has 'seen' thousands of faces through the data it was trained on, but when it makes a new face, it's just combining parts in new ways—ways that don’t replicate any real individual.

How Stable Diffusion Models Work

For a kid-friendly explanation of stable diffusion models: Imagine you have a big, blurry picture. This picture is so blurry that you can’t tell what it’s supposed to be. Now, imagine you have a magic wand that you can wave over the picture. Each time you wave the wand, the picture gets a little less blurry. After waving the wand enough times, the picture turns into a beautiful, detailed drawing. The AI works like that wand, starting with something random and blurry (noise) and gradually making it clearer and more detailed until it looks like a face.

Key Parameters to Tune

When making these images, I can adjust a few important settings to change how the final picture looks:

• GFPGAN Steps: This helps in restoring faces in the portrait to make them look more natural.
• Sampler: This is a tool that decides how each wave of the wand changes the picture. Different samplers can make the picture clearer in different ways.
• Steps: This is like deciding how many times to wave the wand. More steps usually mean a clearer picture.

Creating these portraits has been an enjoyable and creative outlet, and explaining it all to my wife helped demystify the process and share the fun.