If LLMs are text models, how do they generate images?

If LLMs generate images by learning to turn words into patterns instead of just strings of letters, much like how you can feel a shape even if you cannot see it.

You know how when someone says apple, you picture that red, crunchy fruit? Your brain is doing a kind of translation. An Image LLM does something very similar but with pixels instead of text words. Imagine your brain has a giant closet full of pictures. When the AI hears "blue dog," it doesn't just see two separate boxes labeled color and animal. It pulls out all those visual memories of blue things and dogs, then mixes them together until the right picture appears.

Think of it like cooking soup. A traditional text model is like a chef who only reads recipes written on paper. It knows what "salt" means because it has read about salt thousands of times. But an image model is like a chef who has tasted the soup before. It knows that salt changes the flavor even if it cannot see it. The AI looks at your sentence and asks, "What does this look like?" It then builds the image piece by piece, starting with rough shapes and adding details until the image matches your words perfectly.

The secret is that text and images are both just patterns. Text is a pattern of letters. Images are a pattern of colored dots called pixels. The AI learns to connect these two worlds. It sees that the word "fluffy" often appears near pictures of clouds or cotton candy, so it learns to make those parts look soft in the image.

So, the AI is not guessing randomly. It is using what it has learned to paint a picture that matches your words, just like you might draw a dragon if someone described one to you. You use the details (scales, wings) and add your own style. The AI does the same with data.