The First Ouch

When I had the grand idea to generate illustrations for my books with an AI image generator, it put a smile on my face.

"I'll tell the AI what I want and it will give it to me," was at the top of my thoughts.

My children's book, The Hat Who Lost Its Head, was the focus of my first foray into discovering what I could do.

There I was, all smiles, looking to create the first image for the book.

One of the scenes in the book is the hat sailing in the air, having been thrown. In my head, it was a simple scene — a hat floating in the air along a street with a house or several in the background.

I typed "floating hat" and specified 4 images, then sat back to await gorgeous results.

Here are two of them. The other two got deleted as not even hat-ish.

The AI insisted that a floating hat had to have a head.
[floating hat]

Seems the AI can't imagine a hat floating along without a head.
[floating hat]

The AI is not intuitive, I concluded. It can't read my mind. The unfortunate AI had no inkling what I wanted.

My approach needed to be reconsidered. The decision became to approach this learning period of image generation one aspect at a time. Let's first tell it that the hat is on a street, I decided.

This time, I requsted just one image instead of a confident four. I told it "hat on street".

Well, the hat is interesting. But not quite right for the story. Further, the hat has a head.
[hat on street]

Something about the hat images hit me:

All the hats were different!

That would be unsuitable. For a book, the protagonist must be recognized as the same from one illustration to the next.

What I had been thinking of doing simply would not do. As I found out with a bit of additional practice, I had less control over the exact type of hat or how it looked than I felt I needed.

Of course. It made sense. Now.

I had known, intellectually, that the AI was creative in its image generation. Now, I also knew it experientially.

What to do? I was unwilling to let the idea go. My children's book would be illustrated with AI-generated images.

Poking around the NightCafe dashboards, I found something that might work — generating an image as if it was a painting, created from an image I provided.

"Maybe paintings would do," I thought. "If it worked, both the hat and me would be ahead."

My experience with that is in the next blog post.

(Back to blog index page)