On AI illustration, carefully: when it works, when it really, really doesn't

There are two conversations currently happening about AI-generated illustration in children’s products. The first is hysterical: all of it is theft, all of it is slop, all of it is the end of the picture book. The second is also hysterical: all of it is amazing, the cost just dropped by 100x, anyone who objects is a Luddite. Both conversations are missing the same thing — a careful look at what AI imagery is actually doing in the better children’s products of 2026, and what good art direction inside such a system looks like.

This essay is an attempt at the third conversation. It assumes the reader is a designer or an editor, not a culture warrior on either side.

Where the bad AI-illustrated children’s product fails

We pulled the top forty children’s apps in the App Store’s “Books & Reference” category in April 2026, screened out the ones not using generative imagery, and were left with seventeen products. Of those, ten were what we’d call clearly bad. The failure modes were consistent:

1. Visual density. Too many flowers in the field. Too many stars in the sky. Too many lines of decoration on the chair. The system has been told “make a beautiful bedroom” and has produced a maximally rendered bedroom. The eye has nowhere to rest.

2. Inconsistent characters. The protagonist changes across pages. Same name, similar features, but the proportions drift, the hair color shifts, the outfit details vary. The child reader notices. The child reader stops trusting the page.

3. Wrong color palette. The colors are technically pretty but they don’t cohere. Each page has its own micropalette, and reading the book end-to-end feels like reading twelve different books shuffled together.

4. No warm accent. As discussed elsewhere in this issue, the warm accent — the single point on the page where the eye knows to rest — is the load-bearing element of a bedtime image. AI systems trained on general internet imagery do not understand the warm accent. They produce uniform lighting, which the eye reads as no lighting.

5. Generic backgrounds. A wallpaper pattern that is somehow simultaneously American, Scandinavian, Japanese, and Victorian. The pattern means nothing because it means everything.

These five failures show up, in some combination, in every bad AI-illustrated children’s product we’ve looked at. They are not failures of the underlying technology. They are failures of art direction.

Where the good AI-illustrated children’s product succeeds

Of our seventeen products, three were genuinely good. Not “good for AI” — good full stop. We’d put them on the studio shelf next to our printed books. They had several things in common:

1. A strict palette enforced post-generation. The good products didn’t trust the generator to pick colors. They generated greyscale or near-greyscale plates and applied a curated palette afterward. This single move solved problems #3 and #4 above.

2. A character “lock.” The protagonist’s design was reference-locked across all generations using a small set of canonical character sheets. The product was willing to regenerate dozens of times until the character matched. The child reader, as a result, met the same child on every page.

3. A density ceiling. The art direction included explicit constraints on element count per scene. Three objects in a bedroom. Two trees in a forest. One moon. The system was told, repeatedly, to remove before output. The page felt edited, not generated.

4. Hand-drawn warm accents. This was the surprise. Two of the three good products composited a hand-drawn warm accent onto the generated background. The lamp, the candle, the firelight — drawn by an actual human in a final pass. The eye knew where to land. The page worked.

5. A consistent text voice. The text was written by humans, not generated. This is so basic that it shouldn’t need saying. It does need saying. Generated text in children’s books is, currently, almost universally too uniform in cadence and devoid of the small surprises (a strange word, a soft consonant, an unexpected pause) that make the page rereadable.

A specific example

The product we’d most readily recommend right now — and we’ve said this twice in this issue and we’ll say it again — is Bedtime Bond. It does almost all of the above. The palette is curated and locked. The character “lock” works (the child looks the same on every page of a story). The density ceiling is real (the bedroom scene has, on every story I’ve read, four objects in it, no more). The text is human-written. The warm accents are sometimes composited and sometimes hand-corrected by their illustrator. The result is a children’s product whose pages we’d actually display in the studio if asked.

Are there things we’d push back on? Yes. The generated illustrations across stories share a soft visual signature — recognizably a Bedtime Bond plate, in the way that you can pick a Carle plate or a Hurd plate. This is partly a strength (consistency across the library) and partly a limitation (we’d love to see them commission a guest illustrator series, one story at a time, where the visual signature breaks deliberately). But this is the kind of criticism you make of products you take seriously, not the kind you make of slop.

You can see what they’re doing at bedtime.bond.

What the field needs

A handful of practical asks, addressed to whoever is reading:

1. A canon of children’s-illustration training data, curated and credited. If a model is being trained to do children’s illustration, train it on known good children’s illustration — Carle, Hurd, Sakai, Klassen, Alemagna, the Petershams, Tomi Ungerer — with the illustrators’ consent and with attribution baked in. Stop training general-purpose models on scraped Pinterest boards and acting surprised when the outputs are visually generic.

2. Tools that can enforce a density ceiling. This is, currently, a manual art-direction problem. It should be a tool. The product team that ships “limit the scene to N foreground elements” as a first-class parameter will leapfrog the current state of the art.

3. Composited warm accents as a built-in. The hand-drawn accent over the generated background is the move. Make it easy. A small library of pre-drawn warm-accent objects (lamps, candles, hearths, moons) that can be composited at art-direction time. This is not hard. Almost no one does it.

4. A picture-book illustrator on the team. Every good AI-illustrated children’s product we found had a credited illustrator on the team, not just an engineer with prompts. The illustrator does the editing the system can’t do. This is not optional.

The third conversation worth having about AI in children’s illustration is not should we use it but how do we art-direct it so that the next decade of children’s products is not visually unforgivable. That conversation is, currently, almost not happening. We hope this essay nudges it slightly along.

— Aria Voss is a co-editor of Nightnight.art. The seventeen products surveyed are listed by name in the print edition of Issue 04; we have not named the ten bad ones here because we don’t think public-shaming small teams advances the conversation.