---
title: "We measured character drift instead of describing it"
description: "One character, 13 appearances, scored detail by detail. Five fixes tested three against three. What worked, what did not, and what is still open."
canonical: https://kiddobooks.app/blog/character-drift-data
published: 2026-11-11
last-updated: 2026-11-11
---

# We measured character drift instead of describing it

We scored one character, Bumble, on all 13 of his appearances in a test book. The model kept his age, face and trousers, but added stockings in 9 of 12 leg shots and a collar in 6 of 13. Of five one-change fixes, only one clearly worked. Drift got smaller across our tests, but it is not solved.

## Why we stopped describing drift

"The character looks different on every page" is the most common complaint about AI picture books. It was ours too. For months we described it in words, and words like "he looks off again" don't tell you what to fix.

So in November 2026 we picked one character and scored him. Bumble is a character in one of our test books. He appears 13 times, often next to Santa. We cropped every appearance at 4K and checked each detail against his character card: hair, clothes, size, face. All numbers below are from small tests on our own books. They show what happened to us; they are not general rates for every tool.

## Bumble's scorecard

| Detail | What the card says | What we saw |
|---|---|---|
| Stockings | None | Added in 9 of 12 shots where his legs show |
| Collar | None | Added in 6 of 13 |
| Striped tunic | Yes | Lost in 5 of 13 |
| Hair | Auburn | Blond in 4 of 13 |
| Eyes | — | Extra detail added in 5 of 13 |
| Size | Knee-high | About 0.8–0.9 times Santa's height |
| Age, face shape, green trousers | — | Held on every appearance |

Two things stood out. First, the drift was not random. The same inventions came back again and again: stockings, a collar. Second, the big features held: age, face shape and trousers. The small things, the ones a careful author notices, slipped.

## Five fixes, tested three against three

We then tried five one-change fixes. Each one was painted 3 times next to 3 control paints with no change. The round cost about $3.50 in images and checks. The four that taught us most are below; only one clearly worked.

| Fix | Result |
|---|---|
| "Plain green legs, no stockings" in the description | No effect: 1 of 3, same as the control |
| A sentence saying his hat reaches only Santa's knee | A small effect at most |
| A side-by-side size picture as an extra reference | No clear effect (height target met 1 of 3 times) |
| One moment per spread, so he is not painted twice | Worked |

The one that worked was about structure, not description. Some scene notes put Bumble in both halves of a spread ("Left half: Bumble flat on the ground… Right half: Bumble sits up"). The model then painted him twice. Our first duplicate check only looked at names at the start of a sentence. A better check found 27 duplicate spreads instead of 23 across 21 recent books.

We also tried a redraw with a "match the sheet" note. It fixed the stockings 3 of 3 times, but lost the striped tunic 2 of 3 times. Fixes trade one detail for another.

## What other tests told us

Bumble was not our first test. Earlier ones changed how we build the product:

| Test | What we measured | Result |
|---|---|---|
| Sheet attached on every spread (one book, Aug 2026, Nano Banana Pro) | Pages that followed the sheet | About 82%; 18% invented something new |
| Same crowded spreads, two models (Aug 2026) | Kept the committed outfit | Nano Banana 2: 6 of 6. Pro: 2 of 6 (a small test: 12 paints) |
| Full book on three image models (Sept 2026) | Named character missing | 28% (16 of 58 spreads) on all three |
| A "start from a cast lineup" method (Sept 2026) | Drift rate | Worse: 31% became 50% |
| Long descriptions cut to fit (Oct 2026) | Outfit sentence kept | Lost on 10 of 10 spreads the lead was in |
| Same spread at two sizes (Nov 2026) | Invented stockings | 3 of 3 at 4K, 0 of 3 at 2K |
| Redraw notes (Nov 2026) | Outfit fixed | "Fix the outfit" 1 of 3; naming the outfit words 6 of 9 |

A few lessons from this table:

- **The model often ignores a reference it already has.** Pages where the outfit text was in the prompt drifted as often as pages where it was cut.
- **Missing is not the same as drifting.** About a quarter of what looked like drift was a character not painted at all, and switching models did not change that.
- **Levers don't stack.** Adding more references and more description made drift worse, not better.
- **Words help in a redraw, not in the first paint.** A longer description did not move the first paint. A redraw note that names the card's exact words did.

## What we changed

From these tests, Kiddo now works like this. Each character has one short description, at most 300 characters, so it fits without being cut. The painter is told to take body traits only from that description. Each character has one clean reference sheet, labelled for that character only. The author checks every character in the **CastCall** step before anything is painted. Since 2026-11-05 every page and cover is painted with Nano Banana 2.1, which held the character cards better than Pro in our comparison. And single pages can be redrawn with a note; every book includes 5 free redraws.

## What is still open

Drift is smaller, not gone. We found no fix for Bumble's height or his stockings at first paint. Our automatic outfit check still misses small details; it caught the stockings 3 of 9 times. So the author's own eye stays the final check, and redraw stays part of the process.

If you want the practical version of this post, read [Why your AI character looks different on every page](/blog/ai-character-consistency). And since these pictures are AI-generated, see [how to answer KDP's AI disclosure question](/blog/kdp-ai-disclosure-picture-book).

## Questions

### How often do AI picture-book characters drift?

In one of our test books, with the character sheet attached, the model followed it on about 82% of pages. That is one book and one model, not a general rate.

### Does writing "no stockings" in the prompt stop the AI adding stockings?

Not in our test. "Plain green legs, no stockings" in the description gave the same result as the control: 1 of 3 paints right in both.

### Which AI image model keeps characters most consistent?

In one test on crowded spreads, Nano Banana 2 kept the committed outfit 6 of 6 times and Nano Banana Pro 2 of 6. Nano Banana 2 had other problems on crowded spreads. Kiddo paints with Nano Banana 2.1 since 2026-11-05.

### Can an AI redraw fix one wrong detail without breaking another?

Not always. A redraw that asked to match the sheet fixed the stockings 3 of 3 times but lost the striped tunic 2 of 3 times. Check the whole character after every fix.

## Sources

1. [Kiddo: How AI picture-book illustration works, step by step](https://kiddobooks.app/blog/how-ai-picture-book-illustration-works)
2. [KDP: Content guidelines — AI-generated content](https://kdp.amazon.com/en_US/help/topic/G200672390)

Kiddo illustrates your own picture-book manuscript and exports KDP print-ready PDFs.
