How we built our first 117 books, 3,114 pages, and the code that made them

The numbers

Figures are from the live site on 23 October 2026. The making-of numbers cover 93 of the 117 books: we began recording how each book was made partway through, so the other 24 have no such record and every making-of total is a floor, not the whole count. "Pictures on disk" counts the candidates left when each book was published (the model draws several per panel and the best is kept). The 34.6 hours are 19.7 hours from the 66 books that kept a per-stage record plus 14.9 hours from 27 books that kept only a total.

How a book is made

The text is cut into panel-sized pieces. A 7B language model writes the characters and one drawing prompt per panel. An image model draws each panel, three candidates at a time. Two checks look at every picture: one finds words drawn by mistake, one is a vision model that flags what a reviewer would reject. Flagged pictures are redrawn. Then code, not the model, sets the Arabic lettering in a real font, builds the pages, the characters page, the cover and the PDF.

In the pipeline each of those steps is a task, thirteen in all, and a runner puts them in order for a volume. A review task works in rounds: judge, set the flagged pictures aside, redraw, judge again.

The small model that writes the prompts

The language model is the cheapest step and one of the most important. It took 1.4% of the machine time, but it decides what every picture shows. For each panel it gets the passage, the passage before it and the list of characters with their written looks, and it answers with one sentence of 20 to 40 words: who, where, doing what, the shot and the light.

Over the library that was 22,877 calls and about 24 million tokens, roughly 1.7 calls and 1,800 tokens per panel. A 7B model makes mistakes, so most of the work is around it: a character tag is kept only if the passage names that character; words that make the drawing model paint speech bubbles, signs, phones or glasses are removed; and a list of rules written from our reviews goes into every request (how hands and held objects should look, that talk is shown with posture and never with bubbles, that nudity appears only where the passage calls for it and is never explicit).

It still gets things wrong, so the code checks its work: a role written in English is translated by a second call, a character whose Arabic name is not plain Arabic is dropped, and a tag for a character the passage does not name becomes "a figure". For the Nights volumes the cast was wrong enough that we rewrote all of it by hand against the text.

What it takes to run

We run it on one GH200 card with 96 GB: two copies of the drawing model, a 7B text model and a 7B vision model share it (about 86 GB in use). A short story of 64 panels, 11 pages and 192 pictures went from raw text to a PDF in about five minutes (308 seconds).

You can start without a GPU: the tests run anywhere, and a dry run shows the plan for a volume before anything is drawn.

How good are the pages?

We don't have a score for it, so here are the pages themselves.

Character consistency is the hard part. The drawing model has no face reference, so a character stays the same only because the same written description goes into every panel where they appear. In the grid, the long dark hair and the light peasant blouse hold; the hairstyle, the face and her apparent age vary.

What we have not measured is how often a person has to change something. A person looks at the books, in practice after they are on the site, and we have redrawn and republished volumes when a review found a systematic fault (once the cause was a line in the prompt that made the model draw phones and cars). We keep no count of changes per volume, so there is no review rate to quote yet.

What we checked, and what we did not

Drawn pictures are not reproducible across server copies: the same prompt and seed give the same picture on one copy of the drawing model and a different one on another copy. Everything after the pictures is: given the same pictures, the lettering, the layout and the page assembly give the same pages, pixel for pixel.

Not done: the spoken-caption task exists, but the open Arabic voice we tried read the vowels and accent wrongly, so we don't use it. And the new publish task has been tested with a stand-in server, not against sutur.ink.

What is open now

The whole pipeline is in the repository: the tasks, their tests, the documentation for setting up the GPU server, and a diagram of the stages. The tests need no GPU, and GitHub runs them on every push that touches the pipeline.

The reader site is not in the repository yet. Publishing to the site is deliberately guarded in the code: it only runs with an explicit yes, and the production host is never written into it.