benchrWhen it breaks
Same character, second image, different face
Also asked as AI keeps changing the person's face · my character isn't consistent · the model changed her nose · identity drifts between generations
Widely reported; no vendor documents the cause Not tested by benchr First logged Last checked
What is actually happening
Each generation is a fresh sample. Unless something carries identity from one image to the next, there is nothing forcing the second face to be the first face.
Why
- Identity preservation while changing pose, expression or lighting is an open research problem, not a setting. The 2026 literature is full of methods proposed specifically because current models do not hold identity by default.
- A text description of a face is not an identifier. Two samples that both satisfy 'woman, 30s, dark curly hair' can be two different people.
- benchr has found no vendor documentation that promises identity consistency across separate generations, which is why this record is marked as a reported pattern rather than a documented behavior.
The quick fix
Stop re-describing the person and start re-supplying them: pass the first image back in as a reference for every subsequent generation.
The real fix
Fix one canonical reference image and generate every variation from it, changing one variable at a time. Where the tool offers a character or identity reference feature, that feature — not the prompt — is the mechanism that carries the face.
Step by step
- Produce one hero image and accept it as the definition of the character.
- Feed that image as a reference into every later generation instead of describing the face again.
- Change one thing per generation — pose, or lighting, or outfit — never three.
- Keep the seed and settings if the tool exposes them, and record them next to the image.
- Check the details that carry identity — the eye spacing, the hairline, a mole — rather than the overall impression.
Grounded in
- arxiv-2604.21279LatRef-Diff (April 2026) is one of several 2026 papers whose stated motivation is that diffusion models struggle to preserve unchanged attributes, identity in particular, while editing.