Generating meaningful text in an image is very complex. Most of these models like Dall-E and simple diffusion are essentially guided denoising algorithms. They get images of pure noise, and are being told that it’s actually just a very noisy image of whatever the description is. So all they do is remove some noise for many steps in a row until a clear image emerges. You can kinda imagine it as the “AI” staring into the noise to see the image that you described.

Most real-world objects are of course quite complex. If it sees a tree branch in the noise, it also need to make sure that the rest of the tree fits. And a car headlight only makes sense if the rest of the car is also there. But for text these kind of correlations are even way way harder. In order to generate meaningful text it not only needs to understand how text is usually spaced, and that letters usually are written in a consistent font, it also needs to learn the entire English language. All that just to generate something that is probably overall of less influence to it’s “score” on images form the dataaset than learning how to draw a realistic car.

So in order to generate meaningful text, the model requires a lot of capacity. Otherwise, since it’s not specifically motivated to learn to write meaningful text, it’ll do whatever it’s doing now. Honestly I’m sometimes quite impressed with how well these models do generate text, given all these considerations.