benchrWhen it breaks
The eyes change color and the kid looks older on every page
Also asked as my character's eyes keep changing color · the kid looks older on every page · can't keep six characters consistent · character ages between pages · hair color changes between images · multiple characters get mixed up
Widely reported; no vendor documents the cause Not tested by benchr First logged Last checked
What is actually happening
Not the whole face drifting — specific attributes drifting, which is worse, because the image looks right until somebody reads it.
Why
- A text description under-specifies. 'Brown hair' and 'blue eyes' each name a range, and every generation samples somewhere different in that range.
- With more than one character in frame, models redistribute attributes between them: the tall one gets the curls, then next time the short one does.
- Style drift compounds it. Even when the features are close, a shift in rendering reads to a person as a different character.
- benchr has found no vendor documentation promising attribute-level consistency across separate generations, so this is a reported pattern rather than a documented behavior.
The quick fix
Name the attributes that must not move — eye color, hairline, one distinctive mark — and check those three specifically rather than judging the face as a whole.
The real fix
One reference image per character, passed in every time, and one character per generation wherever the composition allows. Composite two single-character generations rather than asking for both at once.
Step by step
- Fix one hero image per character and treat it as the character's definition.
- Write down the three attributes that carry recognition for this character.
- Generate one character at a time; composite in an editor when a scene needs two.
- Change one variable per generation — pose, or light, or outfit.
- Check the named three attributes against the hero image before accepting anything.
Grounded in
- arxiv-2604.21279One of several 2026 papers whose stated motivation is that diffusion models struggle to preserve attributes they were not asked to change.