From Gemini to Pixar — how we make children’s books with AI
August 13, 2026 · updated August 14, 2026, August 15, 2026 and August 16, 2026
It started with an interview. Connor Boyack, creator of the Tuttle Twins, was a guest of James Corbett in Teaching Children About Conspiracies — #SolutionsWatch. It was about how to tell complex ideas — freedom, economics, power — in a way that children understand and want to read. No textbook. No sermon. A story.
Recently Connor was back, this time staying with Mike Rowe The Way I Heard It #488. The conversation was about the same thing: stories that challenge children, that bring the family to the table, that create a shared language. Mike Rowe called it “spontaneous order” — no one plans it, but it happens when you plant the right seed.
That stuck.
The meeting with Yvonne
And then I met an inspiring lady, Yvonne against on substack who explains in an inspiring way how to use AI tools to optimize your workflow and save time. If you want to make a flipbook yourself, take a look to her explanation.
So I did that. With Gemini, Google’s AI, I made my first children’s book. Three parts about Bram and Emma, based on the Abundomy philosophy. It worked. It came out, it looked like a book, it reads like a book. Read part 1 →
And then the work began.
“AI takes a while” — but that’s not how it works
What you learn: An AI that creates a book “for free” on a big-tech platform is never free. You pay with control over your own story — where it is, who can see it, what rules apply tomorrow. I wanted to add a “for more abundomy.com”. Gemini refused. Not with an error message, but softly, politely, impossible. That’s the first thing you learn: a book that is on a platform is a book that the platform can refuse.
So I started doing it myself. Completely local. No cloud, no account, no external API. Everything on your own hardware.
The evolution — eight books, eight pipelines
What followed was an evolution that you don’t plan in one step. Each book taught me something, and each problem replaced a tool or a step in the pipeline. The links to all flip books are below “The Flip Books”.
Book 1 — Bram and Emma (6-10). Ghibli style SDXL artwork, text with an AI director. It worked. But the illustrations were inconsistent — different clothes, different faces per chapter. And the text had English sentence structures that sounded stiff in Dutch.
Book 2 — Mia and Sophie (6-10). Improved pagination, but still SDXL. The flipbook was better, but the anatomy of the figures was sometimes broken — ghost limbs, missing legs, extra fingers.
Book 3 — Sven and Lotte: The Invisible Power (11-17). Switch to FLUX.1 dev for the illustrations. Urban Graphic Novel Style. New cast for older readers. The images were stronger, but without facial consistency, Sven was a different person every time.
Book 4 — Sven and Lotte: The 9/11 Terror Trade (11-17). Completely new pipeline. And this is where it gets interesting.
Book 5 — Bert and Marieke: The pumpkin, the bet and the question that remained (6-10). The newest step, and in a way the most mature. Here we combined everything we learned and solved the last two persistent problems: the anglicisms and the grandfather’s facial consistency.
Book 6 — The Miraculous Pencil (6-10). A new director flow: deepseek-v4-flash:cloud as director that outsources partial tasks to local LLMs. Sam, Noor and Joep discover how the world works without anyone being in charge — the pencil as a metaphor for spontaneous order. PuLID facial consistency, Ghibli style.
Book 7 — The Invisible Machine (11-17). Same source as book 6, but now for teens. Mila, Jesse and Noor. And here we learned something unexpected: after months of experimenting with PuLID face anchors, the old, simple approach — pure FLUX txt2img with a strong prompt, no face locks — turned out to give by far the best results. Ghibli style, warm gold light, twelve records in one clean run. The text then went through an automatic self-check (perspective consistency + anglicisms sniff) which found and fixed some slips.
Book 8 — Together we can do it! (3-5, preschoolers). The youngest audience — picture book with three young animals (Rabbit, Hedgehog, Mouse). Own LoRA trained on FLUX.1 dev for animal characters. Lesson: three animals in one picture is not possible in FLUX, so one animal per illustration — the animal that has the key role per chapter. Quality over quantity.
You can control every step — but you have to know why
What I learned is that you can control every step in the pipeline, but you have to understand it why send you something. Otherwise you are patching, not improving.
Translation. I translated the first book with a general language model. That works, but it produces anglicisms — English sentence structures that sound unnatural in Dutch. “His eyes glittered” instead of “while his eyes sparkled.” Participatory constructions. Word order that is English, not Dutch. Anglicisms are just as undesirable as Germanisms.
In book 4 I replaced the translation step: first translategemma:12b, a dedicated translation model that does EN→NL without chat behavior, after that gemma4:26b-mlx with strict style rules for the polish. Two steps instead of one.
With book 5 I went one step further and eliminated the separate translation step altogether. Instead of first translating and then rewriting, I leave the English source to the director directly distill it into a Dutch essence — a core scene, a metaphor and the source logic per chapter. The local AI then writes native Dutch, straight from that essence. No intermediate step, no anglicisms. It’s the same lesson as with PuLID: sometimes the best direction is the tools you don’t use.
Story generation. On the first books, I dictated the style tightly — “write like this, use this tone, this length, this structure.” That worked, but it felt programmed. I chose book 4 minimal direction: the AI gets the characters, theme and context, and determines the scene, dialogue and pace. The AI is aya-expanse:32b, and the result is more vivid, less constructed.
I added a tight one to book 5 story arc to it. Instead of twelve chapters explaining the same thing, I built a beginning-middle-end arc with one concrete child metaphor: a school swap meet with click cards, a giant pumpkin being destroyed, a secret bet, a missing notebook. Each chapter has its own beat, and a self-check tool checks that all beats are hit and that no sentence is repeated too often.
Illustrations. This was the biggest learning curve. SDXL gave inconsistent figures. FLUX.1-dev gave nicer images but without facial consistency. I added to book 4 PuLID — one reference photo per character, and PuLID enforces the same facial identity through all 12 chapters. Plus FaceDetailerfor facial detail and ControlNet Union Pro for pose control.
But PuLID had a problem: when I used Sven’s reference, it also made Grandpa’s face young. The solution was no more steering, but less — Reduce PuLID weight, or turn it off completely for chapters where Grandpa is prominent.
In book 5 we solved that by using PuLID with the grandpa himself as a facial anchor. The grandpa appears in all twelve chapters, so I took his face from Chapter 1 and used that as a reference for all the plates. The result: the same grandfather throughout the book, with his silver-gray hair, gold-colored glasses and dark green cable cardigan. And we added a negative prompt against the “ghost” artifact — the semi-transparent extra figure that FLUX sometimes places next to a character.
The scene by chapter. I didn’t want every illustration to have the same composition. So a vision model reads every Dutch chapter and generates per scene: who is standing where, what is Bert doing, what is Marieke doing, what is Grandpa doing, what lighting suits the atmosphere. That means steering at the level of the scene, not the style.
I used that for book 4 qwen3-vl:4b. I upgraded to book 5 qwen2.5vl:32b — a larger vision model that describes the scenes in much richer and more detail. And I now use the same model for the quality control: an automatic check that confirms per illustration that the grandfather is in it, that there is no ghost figure, and that the grandfather is consistent with chapter 1. That catches problems before a human has to see them.
The human editorial team — why it remains
Despite the two-step translation, despite the anti-anglicism rules, despite the polish — things keep coming through that a person needs to correct.
“Counter bet” instead of “against bet.” “After the attack” instead of “after the attacks.” “Adjustable tires” instead of “identifiable tires.” They’re little things, but they’re the difference between a book that’s right and a book that’s almost right.
I read and corrected each chapter of book 4 manually. Not because the AI couldn’t do it, but because an AI has no feeling for the word “counter-bet.” I do. That’s not a limitation of the AI — it’s the limit of the tool. And that boundary is exactly where the human work begins.
In book 5 we tackled this at the source: by eliminating the separate translation step and having the AI write native Dutch, far fewer Anglicisms came through. But the human editing remains – it cannot be automated away, and that is not the intention.
The flip books
All eight books are available online as an interactive flip book Brainfusion.nl/flipboeken. From the first experiment with Gemini to the current Ghibli style with PuLID facial consistency. Look at them side by side and you will see the evolution.
The fifth book — The pumpkin, the bet and the question that remained — is for the youngest readers (6-10), with Bert and Marieke who, together with their grandfather, discover how secrets, bets and power work. It’s the first time we’ve fully controlled the grandpa consistency and the ghost artifact.
And now there is a sixth book: The Wonderful Pencil — Sam, Noor and Joep discover how the world works without anyone being in charge. It is the first book we made with a new director flow: deepseek-v4-flash:cloud as a director who outsources partial tasks to the best local LLMs.
And now there is a seventh book: The Invisible Machine — Mila, Jesse and Noor discover spontaneous order, for teenagers (11-17). The same source as book 6, but in its own teen version. The illustrations were made with pure FLUX txt2img (no PuLID, no IP Adapter) — back to the simple approach that gave the best results.
And now there is an eighth book: Together we can do it! — Rabbit, Hedgehog and Mouse discover for the little ones (3-5) that together everything works. Sweet short sentences, and illustrations with our own trained LoRA on FLUX.1-dev. One animal per plate — the quality we were looking for. The first toddler picture book in the series.
The vision LLM that read the source notes
One of the best moments of this project was with book 6. The original source notes of the book Report — the bibliography with 21 sources — existed only as two photos in a notebook. No text file, no copyable list. Photos only.
We gave those two photos to the vision LLM qwen2.5vl:32b. The model read the photos, recognized the handwritten and typed notes, and converted them into neat, structured text. That text now appears as “The original source notes” at the back of the flipbook, with archive links so you can look up each source yourself.
That’s the advancement that new LLMs bring: not just better text and better illustrations, but also the ability to turn a physical notebook into a readable, searchable part of a book. A vision model that reads photos and turns them into usable text — amazing stuff. Progress.
And now again: in the teen version of the same book (book 7, 11-17), the vision LLM has again read two photos from the source notes and placed them as text in the flip book. Same technology, own production.
And because the story is based on the book Report, we also neatly mention that: the book was written by James Corbett, and his work can be found at corbettreport.com.
Not every AI just joins in
An honest story should also mention the moments when an AI refused. For book 5 — based on the chapter on the 9/11 Terror Trade from Report — we asked a local model, muse-glimmer:30b-mlx, to turn it into a children’s book for 6-10 year olds. The model refused. Twice, even after we explained that we believe the source only contains verifiable facts.
The rejection wasn’t from a platform or a company — it was the model itself guarding its own border. It called the material “highly historically sensitive”, “constructed from conspiracy claims presented as fact” and “not age-appropriate for 6-10 year olds”, and said that making such a children’s book would “transmit misinformation about a major historical trauma to children”. It did offer to write a safe, pedagogical story about media literacy and critical thinking, without copying the specific claims.
We thought this was exemplary, and in line with the Main Stream Media narrative: a model that knows and monitors its own boundaries is not necessarily more reliable than a model that does everything well. It’s exactly the difference between the tools you can use and the tools that Big-Tech narrative has and is on your own hardware. Should you trust this as your tool? The assignment was then taken over by another local flow, who did produce the book — but muse-glimmer’s refusal is a story we’d like to tell What do you think…
Update (Aug 2026) — book 7a: The Invisible Machine, improved version
After the initial publication of book 7 (The Invisible Machine, 11-17) followed a thorough second round of editing. The result is book 7a: the same twelve Ghibli illustrations, but the text of all twelve chapters has been critically revised. An AI hallucination in Chapter 2 (the AI came up with ‘school turkey bags’ instead of the pencil investigation) has been corrected; false promises in cliffhangers have been removed; the connection between successive chapters is more logical; language is more appropriate to the age group; and the term ‘spontaneous order’ now appears consistently once per chapter, with the explanation. ‘For the curious reader’ now has four concepts (Spontaneous Order, Shared Space, Restorative Justice, Self-governance). The original version remains available in addition to the improved version — both on the collection page. A nice illustration of the section above: the human editors that remain.
What now
We contacted the author of the source book. We want his approval, and we want to know under what conditions we can use his work. That’s not a formality — it’s respect for the work that goes into it. Without his book, no Sven, no Lotte, no Grandfather, no Bert, no Marieke.
And we investigate how we can make these books more international. The current edition is deliberately Dutch — idiomatic, not a thin wrapper of English syntax. But the stories are about themes that transcend boundaries. An English edition, perhaps other languages, a parallel project that must remain in sync with the Dutch original. We accept that challenge.
We also look at a character LoRA: a trained model that captures the faces of Bert, Marieke and Grandpa Gijs even more firmly, so that consistency no longer depends on a reference image per generation. That is the next step in evolution.
Support the work
Everything you see here — the translations, the illustrations, the flip books, the local AI pipeline — runs on proprietary hardware and is made possible by donations. Unfortunately, we are asking for sponsors to make all AI projects possible. The hardware, the power, the time — it costs what it costs, and without support it can’t go any further.
Support the fundraiser for local AI at Brainfusion →
Subscribe
Would you like to follow how this project develops? Subscribe to the Brainfusion Substack:
Patrick — Brainfusion.nl