Identity retention is the degree to which a generated portrait still looks like you — not merely like a plausible person with roughly your features. It is the single most useful quality measure in AI portraiture, and the one most tools avoid discussing.
The failure mode
Image models are trained on enormous numbers of faces and carry a statistical pull toward the average attractive face. Generate a portrait and the output often drifts: the nose straightens, the jaw squares, the skin smooths, asymmetries disappear.
The result looks good. It looks like a sibling.
Why it matters more than polish
A headshot's job is to be recognised. If a hiring manager compares your photo to you in an interview, or a client meets you after seeing your listing, a drifted photo creates a small credibility problem that a slightly less polished but accurate photo never would.
Polish is easy for any modern model. Recognisability under generation pressure is the hard part.
How to evaluate it
- Show it to someone who knows you and watch their first reaction. Hesitation is the signal.
- Check the asymmetries. Real faces are asymmetric. If yours has become symmetrical, the model has drifted.
- Look at the eyes and the space between features. Proportion drift shows there first.
- Compare at full resolution, not thumbnail size. Small images hide everything.
The tradeoff
Aggressive beautification produces images people like more in isolation and trust less in context. Reference-based approaches that keep a source photo in view during generation generally hold identity better than heavy stylisation.
Related: what is an AI headshot and best AI headshot generators.