How to stop AI photo edits from changing your face

Ask an AI to clean up a portrait and it will often hand back a stranger: the same haircut and shirt, but a face that has been quietly idealised. Skin smoothed, jaw narrowed, eyes enlarged, a few years removed. Each change is small; together they cross the line from your photo to a photo of someone who resembles you.

This is the single most common reason people abandon AI portrait edits, and it is fixable. The face does not drift because the model is malicious — it drifts because most prompts never tell it that the face is the one thing it must not touch.

Why faces are the first thing to drift

Faces sit at the centre of a portrait, carry the most detail, and are the region a model has the most training pressure to "improve." Many image models lean toward a learned average of attractiveness — symmetrical, smooth, youthful. Left unconstrained, an edit slides toward that average, and the average is not you.

So the goal is not to ask for a better face. It is to ask the model to leave your face exactly where it found it while it changes everything around it.

Write a lock-the-face clause

Add an explicit instruction that protects the identity-bearing features, and be specific about the ones that usually get "helped":

• "Keep the exact same face, identity, and bone structure — do not slim the face or jaw."
• "Preserve skin texture, freckles, lines, and pores; do not smooth or airbrush the skin."
• "Keep the original eye size, eye colour, and gaze direction; do not enlarge the eyes."
• "Do not change apparent age, do not whiten teeth, do not reshape the nose or lips."

Then describe the edit you actually want in the same breath — new background, better lighting, a suit instead of a t-shirt — so the model knows the face is the fixed point and the rest is fair game.

Change the surroundings, not the surface of the skin

Most portrait goals — a professional headshot, a cleaner background, warmer light — can be reached without touching the face at all. Direct the model to relight the scene and replace the backdrop while treating the face as a protected region. If you need genuine cleanup (a stray hair, harsh shadow, a blemish), name only that one thing rather than asking for "flawless skin," which is an open invitation to airbrush.

Verify identity before you use the result

Open the edit next to the original at full size and check the features that carry identity: the distance between the eyes, the shape of the nose, the line of the jaw, the texture of the skin. If a feature reads as "better" rather than "same," the model changed it — discard that generation and re-run with a firmer lock-the-face clause. A headshot that does not survive this check is not your headshot.

Key takeaways

• Faces drift because models pull toward a smoothed, idealised average — not because you asked for it.
• Add an explicit lock-the-face clause naming bone structure, skin texture, eyes, age, and teeth.
• Reach your goal by changing the background and lighting, not the surface of the skin.
• Replace "flawless skin" with the single specific fix you actually want.
• Compare identity-bearing features to the original before trusting any portrait edit.