Identity-Safe AI Photo Editing: What to Preserve
A checklist for editing portraits without drifting into a different person, from skin texture and age to pose and expression.
Wan 2.6 Editorial
Tool aus diesem Guide testen
Most AI portrait edits fail in the same quiet way: the result looks good, but it isn't the person you started with. A jaw gets slightly narrower, eyes shift a few millimeters apart, the skin loses its pores, and the subject ages down five years. Nothing looks broken in isolation, yet the image no longer passes the "is that really them?" test. Identity-safe editing is the discipline of deciding, before you touch a slider, which pixels are load-bearing and which are fair game.
This guide is a working checklist for that decision. It applies whether you're cleaning up a headshot, swapping a backdrop, or generating a short clip from a still. The embedded editor handles the mechanics; the judgment about what to change and what to protect is yours, and that judgment is what keeps a portrait recognizably human.
Dieses Rezept ausprobieren
KI-Bildeditor
Identity is more than a face
Recognition is holistic. People identify each other through a bundle of cues: the ratio between face length and width, the spacing of the eyes, the shape of the brow ridge, the way skin sits over the cheekbones, hairline, posture, the resting set of the mouth. Change one and the others compensate; change several at once and the brain registers a stranger even if each individual edit seems minor.
That means identity preservation isn't a single feature you enable. It's a constraint you impose across an entire edit. The practical framing: treat the face as a reference, not a canvas. Edits that respect the geometry, texture and age of the subject read as retouching. Edits that redesign those properties read as a different person wearing similar clothes.
A second point that matters more than most tutorials admit: no workflow guarantees identity preservation. Diffusion-based editing is probabilistic. You can strongly bias the output toward the original, and you should, but you still have to verify every result against the source. The review step isn't optional politeness — it's the actual safety mechanism.
The preservation list
Before opening any editor, write down the attributes you are not willing to lose. This takes two minutes and saves an hour of re-rolling. A reasonable default list:
| Attribute | Preserve or change | Why it matters |
|---|---|---|
| Face geometry (width-to-length ratio, jaw, chin) | Preserve | The strongest single recognition cue |
| Interpupillary distance and eye shape | Preserve | Small shifts read as a different person |
| Nose bridge height and tip shape | Preserve | Highly individual, easy to "improve" by accident |
| Skin texture, pores, freckles, scars | Preserve | Texture is identity; smoothing is the most common failure |
| Apparent age | Preserve unless asked | Retouching trends younger by default |
| Hairline and hair volume | Preserve | Frames the whole face |
| Expression at rest | Preserve unless asked | Changes the muscle pattern around eyes and mouth |
| Pose and camera angle | Preserve unless asked | Alters foreshortening of every feature |
| Subject count | Preserve | Extra or merged people are the worst failure mode |
| Wardrobe and accessories | Your call | Often the point of the edit |
Everything not on the list is where you do your work: background, lighting balance, color grade, dust and blemish removal, crop, and — if the brief calls for it — deliberate changes to wardrobe or setting. If you're replacing the environment, the background replacement guide covers the geometry and lighting matching in detail; keep the face locked while you do it.
A short prompt example for an instruction-based editor, written to encode the list rather than fight it:
Retouch the portrait of the woman in the green jacket. Keep her face geometry, eye spacing, nose shape and jawline exactly as in the source. Keep skin texture: visible pores and the small scar above the left eyebrow must remain. Do not change her apparent age, expression, head angle or pose. Remove the three blemishes on the right cheek and the dust spot on the collar. Replace the cluttered office background with a plain warm gray studio backdrop, matching the existing light direction. Do not add or remove any people.
Notice what the prompt does: it names the protected attributes explicitly, scopes the retouch to specific defects, and closes the door on the two most common failure modes — age drift and extra subjects. Vague prompts like "make it look professional" invite the model to redesign the person.
Skin, age and expression
Skin is where identity-safe editing most often collapses. Models are trained on flattering imagery, so "improve skin" tends to mean "remove texture." Pores, fine lines, freckles, moles and scars are not defects; they are the fingerprint of a face. If the brief is a natural retouch, keep texture and remove only transient blemishes — a healing spot, a dust fleck, a stray hair. If the brief explicitly asks for editorial smoothness, do it in a layer you can dial back, and check the result at 100% zoom rather than in a fitted preview.
Age is the second trap. Skin smoothing, eye brightening and jaw definition all push a face younger in combination. A person can lose a decade without any single edit looking extreme. If the subject's age is part of the image's purpose — a founder headshot, a family photo, an ID-adjacent portrait — treat apparent age as a hard constraint and compare side by side.
Expression deserves its own caution. Changing a mouth from neutral to smiling also changes the cheeks, the eye corners and the shadow under the nose. The result may be a pleasant image of someone who is almost the subject. If you need a different expression, the honest approach is to say so in the prompt and then check that the face still reads as the same person in the new expression — not to assume a small tweak stays small.
For video, these risks compound frame to frame. Flicker in skin texture or eye shape reads as uncanny even when any single frame looks fine. Preserve more aggressively on video than on stills, and keep clips short so review stays feasible.
Pose and subject count
Pose changes are legitimate edits — turning a three-quarter view to frontal, or relaxing stiff shoulders — but they are not identity-neutral. Rotating a head changes the apparent width of the face and the position of every feature. Small adjustments are usually safe; large ones invite the model to invent a new face to fit the new angle. If the pose must change substantially, supply a second reference of the same person at the target angle and verify the result against both.
Subject count is the least glamorous and most damaging failure. Models occasionally add a background person, merge two people into one, or duplicate a limb. On a still this is obvious on inspection. On a clip it can appear mid-sequence. Always compare the number of people and limbs against the source. If an edit adds someone who wasn't there, discard it — no amount of cleanup makes that acceptable in most contexts.
Consent and disclosure
Editing a portrait is editing a person. Before you start, be clear on whether you have the subject's permission to alter their likeness, and whether the output will be presented as a photograph of them. Retouching with consent and disclosure is ordinary professional practice. Producing a realistic image of someone in a pose, outfit or setting they never agreed to, and presenting it as real, is not — regardless of how good the tooling is.
Practical rules that keep you out of trouble: get explicit consent for any change to age, body, expression or attire; don't generate a person into contexts that imply endorsement; keep provenance notes for client work; and disclose AI retouching where the outlet or client requires it. For clips derived from a still, treat the source portrait's consent as covering the clip only within the scope it was given.
A side-by-side review method
Review is where identity-safe editing actually happens. Use a consistent method so you don't talk yourself into a bad result:
- Open source and output at the same zoom, side by side. Not stacked, not sequential — simultaneous.
- Check the preservation list in order: geometry, eyes, nose, skin texture, age, expression, pose, subject count.
- Flip between them rapidly. Differences in identity pop out in alternation even when they're invisible in isolation.
- Zoom to 100% on the face for texture. Zoom out for silhouette and proportion.
- For video, watch once at normal speed for flicker and once paused on every fifth frame for drift.
- Score each attribute pass or fail. If anything on the preservation list fails, re-roll or tighten the prompt — don't patch the output with more generation unless the patch is scoped and re-reviewed.
Keep a short review note per deliverable: what you protected, what you changed, and which frame or region you're least confident about. It takes a minute and makes client revision cycles far calmer.
When to use a different tool
Identity-safe editing is the right frame when the subject must stay the same person. It is the wrong frame when they must not. If the brief requires a new face, a different body, or a person who never existed, don't try to bend a portrait edit into that — use a generation workflow and be explicit that the result is synthetic. If the main task is environmental, the background replacement guide is the better starting point and keeps the face out of the model's decision space. And if you're planning a batch of edits, check what the run will cost first: the exact credit cost is shown in the studio before submission, and pricing tiers are on the pricing page.
The editor itself is at /tools/ai-image-editor. Load a portrait, write your preservation list into the prompt, make one scoped change at a time, and review side by side before you ship. That loop — protect, change, verify — is the whole method. It won't guarantee a perfect match every time, because nothing does, but it turns identity drift from a surprise into something you catch before anyone else sees it.