· 4 min read · AI, Fun
Four books with AI in one year

Between mid-2023 and mid-2024 I made four small illustrated booklets with AI. The first one took weeks. The last one took an hour or two. The tools got better, but I also learned what to ask of them.
The four booklets
Adam (mid-2023) is a poem about a boy. He stands for the birth of my son, and also for AI as a new kind of mind. It needed rhymes and the same boy on every page. Both were hard. MidJourney v5 had no reliable way to keep one character. ChatGPT struggled with structured poems, so I rewrote a lot of the text by hand. It took weeks. (PDF)
Musical instruments (mid-2023) pairs each instrument with the style of music it plays. I used AI for the pictures and for font suggestions. This one took two days. Nobody needed to stay the same from page to page, and ChatGPT is good at descriptive text. (PDF)
The Magical Chest (mid-2024) is about a boy whose chest takes him to different worlds, depending on the clothes he wears. I wanted to test character consistency again. MidJourney v6 had a way to keep a recurring character, still imperfect. The whole book took a few hours. (PDF)
Pasta (mid-2024) is a simple children's book about making pasta. I designed the story to avoid what AI is bad at, like keeping a character the same. It took one or two hours. (PDF)
You can read the original write-up on the project page.
The same problems, seen from outside
I was not the first. In December 2022 Ammaar Reshi made Alice and Sparkle with ChatGPT and MidJourney in about a weekend. He said getting consistent images took hours. So my Adam problem was already known six months before I started.
The rhyme problem has a likely cause. These models read text as chunks, not letters or sounds. Researchers at Bielefeld University built ByGPT5, a model that works on bytes, to get better at rhyme and meter. They report it beats ChatGPT on their poetry tests. The paper does not explain exactly why ChatGPT struggled, so treat the chunk idea as a guess.
Consistency was a research topic too. A November 2023 paper, The Chosen One, tries to make the same character appear from a single text prompt. In May 2024, StoryDiffusion took on consistent images for stories. Both are research, not products I used.
The product fix came first from MidJourney. Version 6 started in alpha in December 2023. In March 2024 it got character reference: add --cref and a link to an earlier image, and MidJourney tries to carry the face, hair and clothes into a new scene. That matches what I saw in The Magical Chest. It also matches my "still imperfect". Reviews say it misses small details like freckles and logos. It is meant for characters MidJourney made, not photos of real people.
Who owns the pictures?
This is a side question, but it matters. In February 2023 the US Copyright Office looked at the comic Zarya of the Dawn. It kept protection for the text and for how the author arranged everything. It refused protection for the MidJourney images. It said prompts work more like suggestions than orders. A month later its guidance said people must disclose AI material in applications, and that protection can still cover the human part.
This is US law only, and I am not a lawyer. But the split is worth knowing: the human part, such as the text and the arrangement, can be protected, and the raw AI pictures may not be.
My opinion
The lesson I trust most is the Pasta one. Do not fight the tool. Pick the job it can already do. A simple story, one or two pictures per idea, no hero who must look the same on page twelve.
The lesson I trust least is "AI is getting better fast". It is true for these four booklets. But a gain in speed is not a gain in taste. Human oversight was still needed for quality control, and as the speed goes up, that checking is the part that matters most.
Sources
- Creating four books with AI (my project page) - the original write-up of all four booklets
- Adam, PDF - the first booklet (link loads; contents not described here)
- Musical instruments (Glasbila), PDF - the instruments booklet
- The Magical Chest, PDF - the third booklet
- Pasta, PDF - the fourth booklet
- Children's book generated with ChatGPT and MidJourney (80.lv, 13 Dec 2022) - Alice and Sparkle: weekend project, hours spent on consistent images
- ByGPT5 (arXiv, Dec 2022) - token-free poetry model, reported to beat ChatGPT on rhyme and meter tests
- The Chosen One: Consistent Characters in Text-to-Image Diffusion Models (arXiv, Nov 2023) - research on keeping a character the same
- StoryDiffusion (arXiv, May 2024) - research on consistent images for stories
- Midjourney V6 is here (VentureBeat, 21 Dec 2023) - V6 alpha, and the warning that things will change
- Midjourney debuts character consistency (VentureBeat, 11 Mar 2024) - the
--creffeature and its limits - Midjourney's consistent characters (The Decoder, 12 Mar 2024) - second report on
--cref, with weak spots - Copyright Office: no protection for certain AI-generated works (Volokh Conspiracy, 23 Feb 2023) - the Zarya of the Dawn decision
- Copyright Registration Guidance: Works Containing AI-Generated Material (Federal Register, 16 Mar 2023) - the official guidance
Comments