Why do AI children's books look wrong?

AI children's books usually look wrong for four specific reasons: the character changes between pages, the art style drifts, text baked into illustrations is misspelled or badly set, and the print file is well under 300 DPI. Only the last is about raw image quality. The other three are workflow failures, which is why buying a better image model does not fix them.

This is a teardown of each, including where the tools that claim to solve them still fall short.

1. The character is not the same character

Open a poorly-made AI picture book and compare the hero on page 3 and page 11. The face is different. Often the colour is slightly different too, and sometimes the species is arguably different.

This happens because each page was generated from a text prompt, and a prompt describes a category of character rather than an individual. The model samples a new member of that category every time.

Readers are ruthless about this, because human face recognition is far more sensitive than any other visual judgement we make. A reader who cannot articulate what is wrong will still put the book down.

The fix is anchoring each page to one reference image so the model edits a known character rather than inventing one. Covered in detail in how to keep AI characters consistent.

Where the fix is oversold: several tools advertise consistency and deliver it for the main character in front-facing shots, then drift on secondary characters, distant figures and unusual angles. Claiming consistency is not the same as verifying it. Ask whether a tool can show you which pages drifted.

2. The art style wanders

Even with a stable character, books look incoherent when page 1 is soft watercolour, page 6 has hard vector edges, and page 12 has drifted somewhere glossy and three-dimensional.

The cause is usually that the style lives in whatever the author typed for each page — or was invented per page by the model. Small wording differences produce large stylistic differences.

The fix is a single style description applied to every page, character sheet and cover, defined once at the book level. Not per page, and not improvised by the model.

This is also where "AI look" comes from. Left to defaults, image models converge on the same glossy, over-rendered, high-saturation house style, and books made that way are recognisable at a glance. A deliberately chosen style — ink and watercolour, cut paper, gouache — reads as a design decision rather than a default.

3. The text is wrong, and it is wrong inside the picture

Two separate problems get conflated.

Text baked into the image. Image models remain unreliable at spelling, especially at small sizes and in longer strings. A page that renders "The Adventre Begins" in beautiful lettering is unusable, and unfixable without regenerating.

Text laid over the image. Safer, because it is real typography — but it goes wrong when it lands in the gutter, runs off the trim, or sits on a busy area of art at low contrast.

The fix for anything longer than a few words is real typesetting over the illustration, with the layout aware of margins and gutter. Reserve baked-in text for short, deliberate typographic moments — a title, a sign in the scene — and check every one.

4. The print file is soft

This one is pure arithmetic, and it catches almost everyone.

An 8.5 × 8.5 page with bleed is 8.75 in wide. Print wants 300 DPI, so it needs 2,625 pixels. Most image models produce 1,024 or 1,536 pixels on the long edge — about 117 DPI across that page.

On screen it looks perfect. In print it is soft, and KDP's previewer flags it at upload.

The fix is resampling to print size on export. Worth being honest about what this does: it removes the visible pixel grid and satisfies the resolution requirement, but it cannot add detail that was never generated. It makes a book print acceptably rather than making it print beautifully.

What separates the books that work?

Not better prompts. The ones that hold up treat the book as a single object with constraints — one cast, one style, real typography, a print file built to spec — instead of as twelve independent image generations stapled together.

That is a workflow difference rather than a model difference, which is genuinely good news: it is the kind of problem that can be solved once and then stop happening.

Frequently asked questions

Can readers actually tell a book was AI-illustrated? Increasingly, when it exhibits these failures — especially character drift and the default glossy style. A book with a consistent cast and a deliberately chosen art style is judged on whether it is good, which is the right question.

Is it worth ordering a proof copy? Yes. Colour prints darker than it appears on screen, and low resolution is far more obvious on paper. A proof costs a few dollars and catches problems no on-screen check will.

Do these problems apply to ebooks too? Character drift and style drift, yes — those are visible anywhere. Resolution matters much less, and bleed and gutter do not apply at all.

Which of the four is most damaging? Character drift, comfortably. It is the one readers name explicitly in reviews, and the one that makes a book read as careless rather than merely imperfect.