Keeping a child looking like themselves on page four

Our first internal build made lovely pictures and terrible books. Page one had a girl with two brown braids in a yellow raincoat. By page three the braids were loose, the coat was red, and the face belonged to somebody else. Adults call this character drift. A six year old calls it that is not me and closes the book.

What did not work

• Re-reading the prose for each page. The prose does not describe the child, on purpose, so the illustrator invented a new one every time.
• Repeating the parent's words verbatim in every prompt. Better, but the words drift in meaning depending on the scene around them, and short lists lose to long scene descriptions.
• Generating page one first and passing it as a reference to page two, then page two to page three. Errors compound. By page five you are painting a copy of a copy.

The character sheet

What worked is boring and it is the whole product. When a book is created we write one paragraph, once, that describes the child in terms of shape and colour and nothing else. No mood, no setting, no action. That paragraph is stored with the book and sent unchanged with every single page request, in the same position in the prompt, after the scene and before the house style.

The scene is the only thing that varies. The house style is fixed too: soft gouache, warm light, cream paper, no text in the image, not photorealistic. Three fixed parts and one moving part. Drift dropped by an order of magnitude the week we did it, and the failure mode that remains is nearly always a fight inside the parent's own words, like a hat and a hairstyle competing for the same head.

Why not just fine-tune

We looked at it. Training a small model per child would mean holding images of children, which is exactly what we have promised not to do. There is no photo upload in Quillow and there is nowhere to put one. A text character sheet holds far less about a real child than a photograph does, and it turns out to be enough.

The pipeline retries a failed page once on its own, then leaves a soft placeholder with a re-paint button rather than shipping a page that is wrong. A missing picture is a problem you can fix in ten seconds. A stranger on page four is a book that gets put down.