How to choose a distinguishing mark

What makes a good one

Three properties. It must be concrete — a thing that either is or is not in the frame, not a quality. It must be asymmetric or otherwise off-centre, because symmetry is what a model regresses to. And it must be visible at the distance you actually shoot at: a small mole is useless in a full-length frame.

"Striking green eyes" fails all three. "Heterochromia, the left iris noticeably lighter" passes all three.

The reliable categories

Dentition — a visible gap between the front teeth — is the strongest, because it changes the shape of a smile and a smile appears in most frames. Scars work well for the same reason: a small vertical scar through one eyebrow interrupts a line the model would otherwise draw smoothly.

Pigment marks — a scatter of freckles across the nose, a single dark mole high on one cheekbone, a patch of vitiligo across the jaw — are strong at portrait distance and fade at middle distance. Structural features — a pronounced widow's peak, a cleft chin, deep dimples — survive further out.

Hair details do more than any of them at scroll speed, which is why the studio treats them separately: a single grey streak at the front, an undercut on one side, a saturated dye colour. At a hundred pixels wide nobody is reading a face at all, they are recognising a silhouette.

What to avoid

Anything a beauty filter would remove, unless you also put it in the negative prompt. Models have been trained on retouched photographs, so their default behaviour is to retouch — which means the freckles you asked for are quietly airbrushed away in about one generation in three. The negative prompt AltF Persona generates specifically guards whichever mark you chose, for this reason.

Also avoid anything that reads as a medical condition you would be depicting for effect. Vitiligo is on the list because it is a common, ordinary variation and rendering it respectfully is straightforward; a visible injury is not the same thing.

Choosing none

It is a legitimate choice, and the studio does not block it. It means the face has no textual anchor, which pushes the route recommendation up — usually to reference, sometimes to trained. That is not a punishment, it is the honest cost of the decision.

If you are running the trained route anyway, the mark matters much less: the model has learned the face rather than been described it.