Prompting for Symmetry (and Where It Falls Apart)

Symmetry is one of the few things you can ask an image model for that has a right answer. Either the left half matches the right half or it does not, and you can check by folding the image in your head. That makes it unusually easy to prompt for, and unusually easy to catch when the model gets it wrong.

Say Which Symmetry You Mean

The word “symmetrical” on its own is weak, because it covers several different arrangements. Bilateral symmetry — a mirror line down the middle — is what most people want, and naming it that way, “bilaterally symmetrical”, “mirrored along a vertical axis”, gets you closer than the bare adjective.

Radial symmetry is a different request entirely: petals around a centre, a mandala, a rose window, spokes from a hub. If that is what you want, say radial or say the object that has it. Translational symmetry — the same element repeated at even intervals, a colonnade, a row of identical windows — is a third thing again, and asking for “symmetry” will rarely produce it.

There is also symmetry of the frame versus symmetry of the object. A perfectly symmetrical vase photographed from the side is one thing; a symmetrical composition where two unrelated figures balance each other on either side of the frame is another. The second is closer to a balanced layout than to a mirror, and asking for it in mirroring language will get you two suspiciously similar figures.

What Holds and What Slips

Architecture holds symmetry well. Facades, arches, corridors, altars, gateways — these are symmetrical in the training data, and asking for a head-on view of one gives you something close to exact. Objects held straight on hold reasonably well. Ornament, tiling and kaleidoscope patterns hold at a glance and fall apart under inspection: the overall structure mirrors, the small details do not.

Faces are the classic failure. Real faces are not symmetrical and neither are generated ones, so a request for a perfectly symmetrical portrait produces something almost-but-not-quite, which reads as unsettling rather than clean. If you need genuine mirror symmetry in a face, it is easier to generate a frontal portrait and mirror one half yourself afterwards.

Text and hands break symmetry every time. So does anything with an inherent handedness — a person holding a tool, a car, a clock face. Asking for a mirrored version of these tends to produce a mirrored scene with the handed elements quietly corrected back, which is usually what you wanted anyway but not what you asked for.

Centring Is a Separate Request

A centred subject and a symmetrical one are not the same, and prompts often confuse them. “Centred composition”, “subject dead centre”, “shot straight on” place the thing in the middle of the frame without demanding the thing itself mirror. That is frequently the real goal — the calm, formal, deliberate look — and it is a much easier ask.

If you want both, ask for both explicitly: a symmetrical object, centred in the frame, viewed straight on. Leave the viewing angle out and the model will happily hand you a symmetrical building shot from three-quarters, where none of the symmetry survives the perspective.