How to Understand AI Image Generation From a Kindergartener’s Perspective
A brief introduction to Artificial Intelligence (AI) image generation that anyone can understand
Artificial Intelligence (AI) Gives Artists Headaches.
On the one hand, AI makes image generation far easier than scanning through thousands of stock images, and it also makes it cheaper than commissioning an artist. On the other hand, it feels like AI is doing this by stealing from the artists it is putting out of a job.
This article addresses the second point. Is AI stealing artists’ works, cobbling those works together and publishing them as its own?
The answer is no, but to help readers understand why, I’ve written a short story to illustrate, at a basic level, how a Generative Adversarial Network (GAN) generates a new image.
Here is a Simple Story Using Kindergarteners to Help Explain AI.
Tim and Archibald are in Kindergarten. The teacher, Ms. Wembley, teaches Tim about shapes and Archibald about cats.
First, Ms. Wembley puts Archibald in the closet. Then she teaches Tim all about shapes and colors and shows Tim how to make them out of colored sand.
Next, she brings Archibald out of the closet and sends Tim out to recess.
While Tim is gone, Ms. Wembley shows Archibald a picture of a sturgeon. “What’s this?” she asks.
“No, that is a fish.”
Then she shows Archibald a picture of a walleye. “Is this a cat?”
“Good,” she encourages.
Next, with a wry smile, she holds up another image and raises her eyebrows.
Archibald squints at the picture for a moment with his chin resting on his knuckles. “Cat?”
“No, this is a raccoon.”
Eventually, Archibald is able to correctly identify a cat from any picture Ms. Wembley shows him.
On the following day, the janitor, Lord Sebastian Harrington, comes into class and asks Tim to make him a cat out of the pile of colored sand. Tim fiddles with the sand and asks Archibald if it looks like a cat. Archibald looks at the picture and tells Tim where he may have gone wrong and Tim makes adjustments.
Other children in the room throw Tim pudding if they like the cat.
This keeps going until Tim has made a decent-looking cat and gives Sebastian the results.
Eventually, after many tries, Sebastian has a nice sand painting of a gray tabby.
How Machines Learn and Create Images Using GAN.
In a Generative Adversarial Network (GAN), we find two essential roles: the generator and the discriminator. Imagine the generator as our friend Tim, an artist who doesn’t know what a cat looks like, yet is adept at moving sand (or pixels) to craft shapes. The discriminator (Archibald) recognizes a cat but doesn’t know how to make one. Through a dynamic interplay, the generator forms an image, guided by the discriminator’s feedback on whether it resembles a cat.
Take a look at the image above, divided into four sections, showcasing Midjourney Bot’s progression in image generation. It starts with an unformed blob and evolves step-by-step into a recognizable image. This metamorphosis is achieved as the generator refines the image, utilizing the feedback from the discriminator, until an entirely unique creation emerges.
Conclusion: Understand How the Distinct Roles of the Generator and Discriminator Ensure the Authenticity of AI-Created Art.
What sets this process apart from theft is the discriminator’s role. It doesn’t save or replicate existing art but offers general feedback, steering the generator towards originality. Thus, the GAN’s collaboration doesn’t steal or reproduce but invents and innovates.
Challenge: Navigate Ethical Considerations in AI’s Intersection with Artistic Fields.
What ethical considerations should we bear in mind as AI continues to evolve and intersect with various artistic fields? How do we balance innovation with artistic integrity? Feel free to comment below.
See also: “Get Your Stories out Faster with ChatGPT.“