How do you keep AI characters consistent in a children's book?

Keep an AI character consistent by anchoring every illustration to a single reference image rather than re-prompting for each page. When the model is given the character as an image and asked to place that character in a new scene, it edits something known. When it is only given words, it invents someone new every time — which is why the hero's face changes between page 4 and page 5.

That distinction is the whole problem, and almost every complaint about AI-illustrated children's books traces back to it.

Why does the character change between pages?

Because a text prompt is a description, not an identity.

"A young unicorn with a golden horn and a cream coat" describes an enormous family of possible unicorns. Each time you send that prompt, the model samples a different member of that family. It is not making a mistake — it is doing exactly what you asked, twelve separate times.

Adding detail helps less than people expect. You can specify the horn, the mane, the eye colour, the exact shade of cream, and still get a different face, because faces are where small sampling differences are most visible to us. Humans detect facial identity at a resolution far finer than any prompt can specify.

So the more detail you add, the more consistent the description becomes and the more obvious the remaining inconsistency looks.

What actually fixes it?

Three approaches exist, and they are not equally good.

Reference-image anchoring (works). You supply the character as an image, and each page is generated as an edit of that image into a new scene. The model has the actual character in front of it. This is the only approach that reliably holds facial identity across a whole book.

Seed locking (partial). Reusing the same random seed makes generations reproducible, but only if everything else stays identical. Change the scene — and every page changes the scene — and the seed no longer protects the character. It helps with style continuity, not identity.

Prompt engineering alone (does not work). Character sheets described in text, long consistency preambles, "the same character as before" phrasing. The model has no memory of before. This is the approach most people try first and abandon.

Does anchoring have limits?

Yes, and it is worth knowing them before you plan a book.

• Extreme angles drift. A reference showing a character from the front gives the model less to work with for a sharp rear three-quarter view. Faces held from the front stay tightest.
• Distant figures lose detail. A character occupying 5% of the frame has few pixels to be identifiable in.
• More characters, more risk. Two anchored characters in one scene can blend features unless each reference is explicitly bound to a named character and the model is told not to merge them.
• Deliberate change needs handling. A character who is supposed to transform — a colour-changing creature, a costume change — fights the anchor unless the change is specified as scene-level, not identity-level.

None of these are fatal. They are reasons to plan compositions rather than accept whatever comes back.

How do you check consistency rather than assume it?

This is the part most workflows skip. Anchoring makes drift unlikely, not impossible, and drift is easiest to miss on the pages you have looked at most.

The check that works: put the reference image and the finished page side by side and score identity only — explicitly ignoring pose, angle, lighting and scene, because those are supposed to differ. A vision model can do this at scale and flag the pages worth a second look. What matters is that the check can fail. A check that always passes tells you nothing.

Pageling runs exactly this pass over a finished book and reports which pages drifted and on which feature, before you send anything to print.

Does this matter for a book you intend to sell?

More than for one you print for a single child.

A personalised gift book survives a slightly different face on one spread. A book listed on Amazon does not — reviewers name it specifically, and "the character looks different on every page" is among the most common one-star complaints on AI-illustrated children's titles. It reads as carelessness even when everything else is good.

If you are publishing to sell, treat character consistency as a pass/fail production requirement, not a nice-to-have.

Frequently asked questions

Can ChatGPT or Midjourney keep a character consistent across a whole book? Partially. Both support reference images now, which helps considerably. What neither gives you is the surrounding book workflow — a persistent cast reused across pages and books, a single art style applied to every page, and a verification pass — so consistency depends on you re-supplying the right reference every time and catching drift by eye.

How many reference images do I need per character? One good front-facing reference is enough for most picture books. A second reference at a different angle helps if your story needs many side or rear views.

Does anchoring make every page look the same? No. The character stays the same; pose, expression, scene and lighting all change freely. Anchoring constrains identity, not composition.

What about a character who changes colour or clothing on purpose? Specify it as a scene-level change against a fixed identity — same character, different colour — rather than editing the reference. Pageling has an explicit setting for characters whose colour varies by page.