What I learned making an illustrated book with AI-generated images

I have a son, and I wanted to make him a book.

I thought it would be easy and fast. I had played with image generation before, so I knew the images would be good, and the story was ready. I only needed to generate some images: one afternoon, a couple of hours, and I'd be done.

I was wrong. It took a really long time, even for a short story. Here is what I learned doing it three times. If you want to tell stories with images but don't know how to draw, this is for you.

First book: by hand

The characters were similar from page to page, but not the same: different clothes, different faces. The places didn't look like themselves from one frame to the next. And every time I tried to fix something, the image got worse.

It took about two weeks, across many sessions.

I finally got an image I liked, and I used it as the reference for everything else. But every time I had to type in the changes and keep track of them in my head. If the character got a drawing in his hand in one scene, I needed to keep that drawing in the next scene, so I had to write it into the prompt again. The prompt kept growing, and so did the number of changes from the original.

I tried the opposite, basing each image on the previous one. Sooner or later I ended up with an image that didn't look anything like the first ones.

I used Gemini Nano Banana and OpenAI's gpt-image, because they were the ones recommended for preserving characters and locations best.

We printed this book in a store. It was expensive: the images were colorful and high contrast, so it used a lot of ink.

The trick of the craft

Illustrators and animators don't redraw the character from memory on every page. They have a design sheet, and every page starts from it.

This was the unlock: first create the characters, clear and in high quality, from many perspectives. Then make every new image from those designs, instead of correcting the last image.

I took the same idea to places and to objects.

The app

I decided to create an app for this. It was rudimentary and not great visually, but it helped me create what I needed.

• A world has one visual style, so everything in it looks like it belongs together.
• A world has characters, locations and objects. Each one starts as a name and a description, and gets its own set of reference images.
• A shot is a mix of characters, locations and objects, plus a prompt.

Second book: with the app

The second book was much faster because of the app. It took three days, and the characters were the same from the first page to the last.

This book was printed too, and we used it many times. I think it helped all of us. This time I prompted the style differently, to get line art like a coloring book, and I printed it at home. It was cheaper.

Third book: thirty minutes

After that I improved the interface and added scenes to organize the projects better, along with a few other utilities.

Then I asked Claude to make The Nightingale and the Rose, Oscar Wilde's short story. I chose it because it's well known and probably in the model's training set. Claude picked the style on its own and created the characters, locations and objects. Then it wrote a script of scenes and generated all the images automatically using the app.

It looked amazing. I was completely shocked at how well it worked. The best part: it took only thirty minutes, from prompt to finished book.

Two weeks, three days, thirty minutes

The app took me longer to build than any of the books. If you only make one book, it probably isn't worth it.

But now I want to make even more books and stories, maybe even go to animation and voice generation. I never thought I could do this on my own, or that I'd have to spend so much time learning to draw before starting the book itself. Now I can turn a well-defined story into a book in less than thirty minutes.

If you want to look inside the app and make your own illustrated story with it, the code is here: Creative AI Tools.