AI and Translation — How These Books Are Made
These books are made with a fully local AI pipeline on a MacBook Pro M5 Max. No cloud APIs, no external services — everything runs on our own hardware.
How we got here
We started by translating English source material into Dutch — a two-step pipeline (a dedicated translation model, then a polish pass) that produced Dutch children’s books. It worked, but we kept hitting problems: anglicisms that slipped through, stiff translated phrasing, and a constant fight to keep the Dutch natural rather than a thin wrapper around English syntax. The more we looked at it, the clearer it became: the source is already in English, so writing native English directly is the logical path — and it lets us reach an international audience without a translation layer.
So we switched to an English-to-English pipeline: the director distils the English source directly into an English essence, and the local AI writes native English. No translation step, no anglicisms to chase — just a story told in the language it was born in.
The pipeline
Each book goes through several steps, from English source to finished flipbook:
- Step 1 — Source essence: the director (pi) reads the English source and distils each chapter into a core scene, a metaphor and the source logic. The real names and facts are kept — children and teens must be able to look them up.
- Step 2 — Story generation with
gemma4:26b-mlx, a local model that writes native English. For the 11–17 books we use a creative brief (narrative non-fiction thriller: 40–50% dialogue, show-don’t-tell, cliffhangers, no moralizing, a mentor who doubts, a protagonist who grows). - Step 3 — Illustrations with FLUX.1-dev (FP8) via ComfyUI, with PuLID for face consistency (the mentor is the face anchor) and FaceDetailer for face detail. Per chapter, a small vision model (
qwen3-vl:4b-instruct) picks the scene from the chapter text. - Step 4 — Human editing: every chapter is read and corrected by hand.
How we keep the quality up (autoresearch)
Making books with AI is a process of trial and error — different models, prompts and settings give very different results. To keep our internal flow honest, we use automated self-checks and an autoresearch loop:
- Text self-check: every chapter is checked for required real terms (did the real names appear?), element-timing (are names revealed too early?), 8-gram repetition across chapters, and Dutch words leaking into English. A chapter only passes when the check is clean.
- Image self-check: every illustration is checked by a vision model — are all three characters present? Is there a ghost figure? Is the mentor’s face consistent? Is the canon setting recognisable?
- Autoresearch loop: we tune the pipeline parameters (model choice, prompt rules, PuLID weight, scene-description style) in an experiment loop that measures the self-check results and keeps the winning settings. This is how we found, for example, that a “description-variety” rule (don’t describe the mentor’s glasses in every chapter) cut repeated phrases from 13 to zero.
The checks are a proxy, not a guarantee — the human editor has the final word. But they catch the obvious failures before a chapter or illustration is accepted.
Why human editing is still needed
AI models have habits: favourite phrases, favourite gestures, favourite ways to open a chapter. The self-check catches repetition, but the editor catches tone, pacing and the moments where the story needs a human’s ear. Every book you see here has been read and corrected by hand.
Working together
We are happy to work with others. If you have difficult content — a text, a PDF, or a video — that contains facts, sources and grown-up issues, and you want it turned into a children’s book that explains the core more clearly and gently, we would like to help. Think of it as the approach Connor Boyack took with The Tuttle Twins: taking complex ideas and making them accessible for young readers. Get in touch if you have a source you would like us to work with.
This page is updated as the pipeline evolves.