How generated portraits actually work, without the marketing

Strip the language and the mechanism is short. You upload a handful of photographs of your face from different angles, a model learns what is constant about that face, and it produces new images of it in circumstances you did not photograph.

The part that decides your result

The system separates what is stable about you from what is circumstantial. If every source photograph came from one afternoon in one shirt, the shirt is stable — so the shirt becomes part of you.

Different days, different clothes, different light give it enough variation to tell you apart from your circumstances. That single input decision moves the outcome more than choosing between vendors does, and it appears on no product page because it is not a feature.

Where it fails

Consistently in the same places: strong prescription glasses, tightly curled hair, darker skin tones, pronounced asymmetry. And the failure is not obvious — the image looks real, it just looks like a professionally photographed person who is almost you.

Plausible and identical are different targets. That gap is the actual limitation, and it never appears in a sample gallery.