AI Portrait Retouching Without the Plastic Look

11:00

Everyone can spot an over-retouched portrait, even people who have never opened a photo editor. The face looks fine, the lighting looks fine, but something reads as false.

What they are noticing is the absence of texture. Real skin has pores, fine lines, slight colour variation and specular highlights that break up unevenly. Take those away and the face stops looking like a person and starts looking like a rendering of one.

AI retouching made this failure mode much easier to fall into. The tools are fast enough now that you can wreck a hundred portraits in the time it used to take to carefully retouch one. And they lean plastic by default, because the aesthetic they learned from is a decade of heavily smoothed beauty imagery.

The good news: this is largely a prompting and process problem, not a capability ceiling. Current models can remove a blemish while leaving the surrounding skin structure untouched. They just need to be told to.

Why does AI retouching make skin look plastic?

Two causes. Prompt wording — "smooth", "flawless" and "perfect" all push the model toward the airbrushed aesthetic that dominates its training data. And stacking — running a retouch, then an enhancer, then an upscaler, each reducing micro-contrast until the pores are gone.

Cause one: your words

Certain words are effectively instructions to destroy texture. Smooth, flawless, perfect skin, beautiful, glowing, airbrushed — each moves the model toward the region of its training distribution where skin is a matte gradient. "Retouch her skin and make it smooth and flawless" is, from the model's point of view, a request for exactly the thing you do not want.

This catches people out because the words sound like the goal. You do want the skin to look good. But "good" in retouching means clean, not smooth, and the model does not infer the difference.

Cause two: stacking

Someone runs a retouch, is nearly happy, then runs a skin enhancer to polish it, then upscales for delivery. Each stage independently reduces micro-contrast a little. None is wrong on its own; three in a row strip the pores out entirely.

How much texture does the wrong prompt actually cost?

27% of it, measured. One portrait was retouched twice from the same source: once asking for "smooth, flawless, perfect" skin, once naming pores, freckles and fine lines as things that must survive. The first kept 72.7% of the original's pore-scale detail. The second kept 93.9%.

The magnified strip is where it becomes obvious. In the middle image the pores and freckles are simply gone, replaced by an even wash of colour — and that image was produced by asking politely for exactly what most people ask for.

Two things are worth taking from the numbers. The damage is real but not total, which is why over-retouched skin often passes at thumbnail size and fails when someone looks properly. And the texture-preserving prompt does not score 100%, because blemish removal legitimately removes some detail — the goal is to lose the spots, not the surface.

What does real skin actually look like?

Under raking light, real skin shows four things: pores that vary in density across the face, fine vellus hair catching light at the edges, uneven specular highlights where skin is oilier or drier, and subtle subsurface colour variation. None of these are flaws — they are what makes a face read as real.

• Pores. Denser and larger across the nose and cheeks, finer at the temples and jaw. They are not a uniform pattern; the variation is what reads as real.
• Vellus hair. The very fine hair covering most of the face. It catches light at the edges and is the first thing to disappear in an over-smoothed image.
• Uneven specular highlights. Real skin is slightly oily in some places and matte in others. Plastic skin has one continuous sheen.
• Subsurface colour variation. Faint redness across the cheeks and nose, cooler tones near the eyes. Averaging this into a single tone is what makes skin look painted.

Notice that none of these are flaws. Acne, scars, razor burn and stray hairs are what a client wants gone. Pores, vellus hair and colour variation are what make the result believable. Good retouching is precisely the ability to distinguish between the two — and that is what your instruction has to encode.

What prompt keeps skin texture intact?

Separate temporary blemishes from permanent features, name pores and fine lines as things that must remain visible, block the smoothing verbs explicitly, and lock pose, lighting and identity. Deliberately unglamorous language produces conservative edits — which is what retouching actually is.

Retouch this portrait. Remove temporary blemishes only:
acne, spots, razor burn, stray hairs on the face, and any
distracting marks that would not be there next month.

Keep permanent features exactly as they are: moles,
freckles, scars, birthmarks, the shape of every feature.

Preserve all skin texture. Pores, fine lines and vellus
hair must remain clearly visible at full resolution. Do
not smooth, blur or even out the skin surface. Keep the
natural variation in skin tone, including redness across
the cheeks and nose.

Do not change the lighting, the pose, the expression, the
background, or the person's identity. This is a
conservative clean-up, not a beautification.

Four things are doing the work here.

"Temporary versus permanent" is the key distinction. It gives the model a principle rather than a list, and it maps almost exactly onto what clients actually want. A spot that will be gone next month should go; a mole someone has had their whole life should stay. Far more reliable than enumerating every blemish type.

Naming texture as a requirement. "Pores, fine lines and vellus hair must remain clearly visible" is an explicit constraint. Without it, texture is simply not something the model optimises for.

Negative instructions for the smoothing verbs. "Do not smooth, blur or even out" blocks the default behaviour directly.

An identity lock. Editing models will happily drift a face toward a more conventionally attractive average if not told otherwise. Naming pose, expression, lighting and identity as fixed prevents that drift.

What is the right order to retouch in?

Blemishes first, then verify texture survived, then lighting as a separate instruction, then upscaling last if at all. Order matters more than people expect — upscaling before retouching wastes money on pixels you are about to re-render, and stacking passes is what sands the texture off.

Step 1: Remove blemishes

One pass with the prompt above, at the image's native resolution. Nothing else in this step — no lighting changes, no background work. Compare against the original at 100% before moving on, specifically checking that pores survived.

Step 2: Verify texture before continuing

Zoom to 100% on a cheek. If the pores are gone, do not try to fix it with another pass — go back to the original and make the instruction more conservative. Texture cannot be restored by adding more processing; it can only be preserved.

Step 3: Fix lighting, if it needs it

Only now, and as a separate instruction: lift shadows under the eyes, soften a harsh key, warm the overall balance. Keeping this separate means that if the lighting edit goes wrong you still have a good retouch to fall back on.

Step 4: Upscale last, if at all

Upscaling is a delivery step, not a quality step. Do it after everything else, once, and only if the client needs the resolution. Upscaling before retouching wastes money because you pay for pixels you are about to re-render, and upscaling repeatedly is one of the fastest ways to sand the texture off.

Which AI model is best for portrait retouching?

Nano Banana Pro for anything where identity must survive exactly — it follows conservative instructions most reliably. Nano Banana 2 for general work at a third less cost. Freepik's Skin Enhancer is a different tool entirely: it rebuilds texture on images whose detail has already been destroyed.

The distinction worth internalising is between instruction-led retouching and texture rebuilding. Nano Banana Pro and its siblings do the former: you describe a change and they apply it. Skin Enhancer does the latter: it takes an image whose texture has already been destroyed — by a previous retoucher, a beauty filter, or an over-eager AI pass — and reconstructs plausible skin detail.

They solve different problems, and the order matters. If a client sends you an image that has already been through a phone beauty filter, texture rebuilding comes first; there is no point asking a model to preserve pores that are not there any more.

For a fuller comparison of the two Nano Banana models, see Nano Banana 2 vs Nano Banana Pro.

How do you retouch a whole shoot?

Cull first, then retouch only the selects with one consistent instruction across the set. Check the first three results carefully before committing the rest. Consistency across a gallery matters more than any single frame being perfect — a set handled slightly differently reads as sloppier than one handled uniformly.

One portrait is a demo. A hundred is the job, and it is where the economics get interesting.

Retouching everything is the most common way people overspend, because roughly eighty percent of any shoot never reaches the client. Cull first. Then run the whole batch through one consistent instruction rather than tuning per image.

One caveat that saves real money: check the first three results carefully before committing the other ninety-seven. If your instruction is producing plastic skin, you want to discover that at three images, not at a hundred.

How much does AI portrait retouching cost?

Three to nine cents per image on your own API key, depending on model and resolution. A hundred-portrait shoot costs roughly $6 to $9. A human retoucher charges $2 to $15 per image; subscription retouching software runs $20 to $40 a month regardless of volume.

Provider rates as of August 2026, checked against Rangy's live pricing tables. Rates change — verify before budgeting.

The number that changes behaviour is not the absolute cost but the cost of a retry. At six cents, trying a more conservative instruction three times on the same portrait is pocket change — which means there is no reason to accept a plastic result.

When should you not use AI retouching?

Skip it for high-end beauty and advertising where every pore is art-directed, for anything where the image is evidence rather than decoration, when a client contract forbids AI processing, and when the subject has features a model may "correct" without being asked — scars, vitiligo, asymmetry, visible disability.

I build one of these tools, so weigh this accordingly — but there are four cases where I would not reach for it.

• High-end beauty and advertising. When every pore is art-directed and the retouch is the craft, AI is a first pass at best. Use it to clear the tedious eighty percent, then hand the file to a retoucher.
• The photograph is evidence. Documentary, journalism, medical or legal work, and any before-and-after claim about a real result. Once a model has reinterpreted the image, it no longer documents anything.
• Client contracts that exclude AI processing. Increasingly common in editorial and some commercial work. Read the terms before you upload; some also restrict sending images to third-party services at all.
• Subjects with features a model may "fix" unasked. Scars, vitiligo, facial asymmetry, visible disability. Models trend toward a conventional average, and silently erasing part of how someone looks is worse than not retouching at all. If you retouch these, check the result against the original at 100% specifically for changes you did not request.

That last one deserves emphasis. The identity lock in the prompt above helps, but it is not a guarantee. Where the subject's appearance is part of who they are, verify rather than trust.

What mistakes ruin a retouch?

The five that recur. Using the word "smooth", stacking multiple passes, removing permanent features like moles and freckles, retouching before culling, upscaling first, and judging the result at fit-to-screen instead of 100%. Five of the six are process errors rather than prompting errors.

1. Using the word "smooth". Or flawless, or perfect. These are instructions to remove exactly what you want to keep.
2. Stacking passes. Retouch, then enhance, then upscale — each strips a little texture. One pass, then stop.
3. Removing permanent features. Moles, freckles and scars are identity, not blemishes. Removing them is how a portrait stops looking like the person.
4. Retouching before culling. Paying to retouch images that will never be delivered.
5. Upscaling first. More pixels to re-render, at higher cost, for no quality gain.
6. Judging at fit-to-screen. Plastic skin looks fine zoomed out. Check at 100% on a cheek, every time.

Frequently asked questions

The bottom line

The plastic-skin problem is a default, not a limitation. Separate temporary blemishes from permanent features, name pores and fine lines as protected, block the smoothing verbs, run one pass, and check at 100% before doing anything else. That is most of the fix.

Models were trained on a decade of imagery where smooth skin was the goal, so smooth skin is what they produce when you do not say otherwise. Saying otherwise is the whole technique.

Do that and AI retouching stops being the thing that makes portraits look fake and becomes what it should be: the tool that handles the tedious eighty percent, so the time you spend is on the frames that actually deserve it.