AI picture book page art: text-free art, story text set by you
Make each picture book page as text-free art in one style, then set the story text in your layout. Reference the first page so the style holds.
To make AI picture book pages, generate each spread as text-free art in one fixed style, then set the story text in your layout tool. Send the first approved page as a reference for every later page so the characters and palette hold. Words in a layout file can be proofread and translated; words baked into pixels cannot.
Fix the style first
Write a short style block once and paste it into every prompt: the medium (gouache, pencil, paper cut), the palette, the line weight and the lighting. Generate three takes of page one with n: 3, choose one and keep its URL as the style anchor.
Send the anchor with each page
Put the anchor page in input_references and describe only what changes: the scene, the action, the time of day. State that the character design, clothing and palette must match the reference. Nano Banana Pro takes up to 10 references and GPT Image 2.5 up to 16, so you can add a character sheet as well.
curl -X POST https://api.sume.com/v1/images \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"google/nano-banana-pro","prompt":"Same gouache style, characters and palette as the reference. Page 4: the fox and the girl cross a wooden bridge at dusk. Leave the top third quiet for text. No text in the picture.","aspect_ratio":"4:3","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/page-1.png"}}]}'Leave room for type
Ask for a calm area at the top or bottom of each page, then set the type there. Plan the same zone on every page so the book reads evenly. Keep the safe margin in mind if the book will be printed and trimmed.
Watch for drift across pages
Check each page against page one: hair, clothing, proportions and the main color. Drift builds up if you use the latest page as the only reference, so keep sending the original anchor. If one detail is wrong, fix it with a masked edit instead of re-rolling the page, as in the mask_url guide.
Cost planning
Check the live pricing lines on the endpoints call and multiply by your page count and takes per page. A 32-page book with three takes per page is 96 generations before fixes, so draft on a lower-priced model or quality and finish on the one you keep.
Related posts
More in Use cases
- AI price tag image: make the blank tag, set the price in code
For a price tag or shelf label image, let the model draw the tag and write the number yourself. Prices must be exact, and generated digits can drift.
- AI product photo with a hand holding it: packshot plus pose
To put a hand in a product shot, send the packshot as the first reference and a pose photo as the second, and say to keep the product identical.
- AI thumbnail readability test: check the text at 168 by 94 pixels
Before you publish an AI thumbnail, shrink it to a small display size and read the text. If it fails there, regenerate with fewer words and larger type.
- All-hands recap with an AI presenter: a 60-second video and its cost
A weekly company recap read by a Sume avatar: script structure, the 60-second cap, a cost table for one week and a year of 52, and the disclosure to add.
Written by Sume