How Neural Networks enable AI’s to become artists
Published in · 8 min read · Sep 5, 2022
--
Usually, throughout the month I like to look at a little bit of news in Data Science just to see what’s changing or to find anything interesting I haven’t already seen. This past week, however, I did not have to do any searching. This time, my girlfriend has been talking about a few of her favorite YouTubers using popular AIs (Dall-E and Midjourney) to create images, or they are asking their fans to create and send the images to them. Either way, she has been talking about how these AI’s are creating images based on text prompts. Some work out, others come out with strange adaptations. Whichever way, she told me how fascinating this technology is. Neither of us knew exactly how AIs created images based on text, so I thought that would be something interesting to learn.
In this week’s article, I plan to dive into how an AI can create images based on text prompts. Because the technology is still growing, this will be a high level, but I thought it would still be interesting to get an idea of how it works. First, I want to start with how an AI can create images and the methods it can use. Next, we’ll look at these methods in a little more detail. Finally, I’ll just list a few answers to some questions I had about the AIs(Dall-E in particular). So without any further delay, let’s talk about AI-generated art.
For decades now, computer scientists have been using Machine Learning to train computers to understand the world. Computers have filtered through images to learn the differences between particular objects and items. The next progression is to learn to create images. But how can an AI create images? As you probably guessed, it is a complicated process. To generate images, the machine uses two neural networks. The first neural network is used to create the image based on the text input by the user. The second neural network analyzes the generated image with reference images. By comparing the photos, it creates a score to determine the accuracy of the generated image. The score is sent back to the original AI system, which can return an altered image for another score. The systems will continue to score likeness until the generated image…
