Blog · Behind the scenes
How Sutur turns a book into a manga volume, step by step
The whole pipeline in plain words: from a Hindawi text to a published volume, with the models, the hardware and the checks.
· 3 min read
In short
- A language model plans, an image model draws, a vision model checks, and code does the lettering, the layout and the publishing.
- One volume takes about thirty pages and a few hundred candidate pictures, and a person reviews it before it goes live.
- Everything here is open: the code is MIT, the books are CC BY 4.0, and you can rebuild the pipeline yourself.
Start from a text we are allowed to reuse
Every book starts as Arabic text from the Hindawi Foundation, published under CC BY 4.0. That licence is the reason the whole project can be open: we can adapt the text, show it beside the pictures and let anyone reuse the result, as long as the source is credited. Each book page names its source.
Cut the text into panels
The text is split into short pieces of about thirty words, cut at the end of a paragraph, and six pieces make a page. Each piece keeps its position in the book, so a picture always knows which words it illustrates, and the original diacritized Arabic is never rewritten.
A small language model plans the volume
A seven-billion-parameter language model reads the volume and writes two things: a short cast of characters with a look for each, and a drawing prompt for every panel. It is fast and cheap, and it is also the weakest link, which is why a person reads its cast before anything is drawn.
An image model draws, many times
Pictures come from Z-Image Turbo, an open image model that makes a picture in nine steps. Three copies run side by side on a single NVIDIA GH200 graphics processor. For every panel we draw three candidates with different random seeds and keep the cleanest. A test volume of 210 pictures took under three minutes.
Other models check the work
A vision model of the same size looks at each picture for nudity, text drawn inside the image, animals drawn as people, bad hands and cages, and flagged pictures are redrawn. A text-recognition pass catches stray English words. These checks are filters, not judges: a person still looks at the flagged pages and at a sample of the rest.
Code does the lettering and the layout
The model never draws the Arabic letters. Code sets the real text in a font with proper Arabic shaping, places it in a caption box above each picture and assembles the page. This is why the words are exact, with every letter joined correctly, even when the picture is not.
Publish, then watch
A finished volume becomes a PDF and web-sized page images, and is uploaded through the site's own admin interface.
Tech notes
- Text sourceHindawi Foundation (safahat.org), CC BY 4.0
- PlanningA 7B-parameter language model (Qwen 2.5) writes the cast and the panel prompts
- DrawingZ-Image Turbo, 9 steps, three copies on one NVIDIA GH200
- CheckingA 7B vision model (Qwen 2.5 VL) plus a text-recognition pass, then a human review
- LetteringReal Arabic text shaped and placed by code, never drawn by the model
- OutputWebP page images at 1,400 px wide, plus PDF and EPUB
- LicencesCode MIT, books CC BY 4.0
Try it yourself
- Pick a book under a licence that lets you adapt it, and keep its credit line.
- Split it into short panels at paragraph ends and keep each piece's position.
- Let a language model draft the cast and prompts, then correct them by hand against the text.
- Draw three candidates per panel and keep the cleanest.
- Set the text in code, not in the picture.
- Check with a vision model, then look at every flagged page yourself.