AI children's book illustrations with the same character on every page

To keep one character recognisable across a children's book, generate a single anchor illustration first, then send it as an input reference with every page prompt and repeat the same description words. On Sume that means one POST /v1/images per page, with the anchor URL in input_references and a page-specific scene in the prompt. Models that take references hold a character far better than prompt-only runs, but they still drift, so plan to check each page.

Build the anchor

Spend your budget on the anchor. Generate four options with n: 4 on a model that supports it, pick the best, and keep that image URL. Describe the character once in a fixed block: age, hair, clothes, colours, art style. Paste that block into every page prompt unchanged.

Pick a model that takes references. GPT Image 2.5 takes up to 16, and Nano Banana Pro and Nano Banana 2 are in the same catalog (Image API). Read each model's input_references range from GET /v1/images/models before you depend on it.

• One anchor image, one fixed character block.
• Style words identical on every page: medium, line weight, palette.
• Scene prompt changes only: place, action, time of day.
• Ratio fixed for the whole book, for example 4:3 landscape spreads or 3:4 portrait pages.

Page loop

Run pages one by one, each call independent. Sume has no seed field, so you cannot rerun to get a matching face; the reference is your only anchor. If page six drifts, regenerate that page from the anchor, not from page five, so errors do not accumulate.

Text on the page

Keep story text out of the image. Typeset it in your layout tool so you can fix typos and translate. If a title must be in the picture, quote it and check the spelling on every render (text in images).