What does the percentage beside the picture measure?

Mean absolute change in luminance, taken only over the pixels that actually changed, expressed as a share of those pixels' own original luminance. It is deliberately not averaged over the whole frame: a figure computed that way would read close to zero for any honest portrait retouch and would therefore tell you nothing. Restricting it to the changed area makes it a measure of how hard the edit is working where it is working.

What number should I aim for?

There is no target, and offering one would be the same mistake as a beauty score. What the readout is for is comparison — between two settings of the same control, and between what you thought you had done and what you have done. A figure that climbs while the picture stops improving is the signal. Somebody working on a headshot for print will sit in a different range from somebody cleaning a snapshot, and both can be right.

Who decided the caps?

We did, by reasoning rather than by testing, and that is worth saying plainly. The 80 per cent texture ceiling and the warning above 60 are judgements about where a face stops reading as a photograph, argued from what the bands contain rather than measured against a panel of retouchers. They have not been validated. If they turn out to be wrong they will move, and the reasoning will change on this page with them — but they will not be removed, because the existence of a ceiling is not the part that is uncertain.

Is retouching a photograph dishonest?

Not inherently, and the question is better asked about the specific edit than about the practice. Removing a spot that will be gone next week keeps the photograph truer to the person than leaving it in. Evening a patch of shine corrects the lighting rather than the face. Erasing every line, pore and shadow produces an image of somebody who does not exist and circulates it as a picture of somebody who does. The test this site uses is disclosure: an edit you would be comfortable naming to the person in the frame is one you can make.

Why does the tool warn me rather than stop me?

Because above 60 per cent the result is still a legitimate choice in some contexts and not in others, and a tool that refuses is a tool that gets replaced by one that does not. What it can do is make the consequence visible at the moment of the decision: the note appears, the compare control is one press away, and the figure beside the picture keeps counting. Refusal is reserved for the things that have no defensible setting at all, which is why reshaping is absent rather than capped.

How far is too far, and how this tool measures it

Every control here has a ceiling, and every ceiling is a position rather than a technical limit. This is where each one sits, why it sits there, and what the number beside the picture is counting while you work.

There is a specific failure this whole site is built around avoiding, and it is not over-smoothing. It is a photograph that has been corrected past the point where it is a photograph of that person, published as though it were one. The person in the frame did not agree to it, usually was not asked, and in most cases is the last to see the result.

The reason it happens is structural rather than malicious. A retouch is judged on screen, at a size, in a session where the eye has spent twenty minutes adapting to the version being built. Under those conditions the retouched face becomes the reference and the original starts to look wrong. Every additional step feels like an improvement, because it is an improvement relative to the previous step — and there is no point at which the process announces that it has gone too far.

Three mechanisms in this tool exist specifically to interrupt that. Everything opens at zero. Everything has a ceiling. And there is a number beside the picture at all times that does not care how long you have been looking at it.

Every ceiling, and the argument for it

Texture removal stops at 80 per cent. The high band holds pore structure, fine hair and grain. At 25 per cent — the default — three quarters of it survives and a cheek still reads as skin at full zoom. At 80 a fifth survives and it reads as a surface. Past 80 there is nothing left to reduce that was worth keeping, so the slider does not go there.

A note appears above 60 per cent. That is the setting at which texture loss becomes visible at 100 per cent zoom on a typical portrait. The note says so and points at the compare control, because the decision is legitimate and the information is what is usually missing.

Tone evening is applied at 45 per cent of the slider. The control reaches 80 and what reaches the low band is a good deal less. That gearing is deliberate: flattening the low band is how a face loses the look of having blood underneath it, and unlike plastic texture it is very hard to notice while you are doing it.

Colour cannot move at all. Every tonal operation converts to luminance and chrominance, changes luminance, and writes the original chrominance back. The new luminance is clamped into the window where the inverse conversion cannot clip a channel, because clipping is precisely the mechanism by which a brightening control desaturates a deep skin tone without anyone intending it. A build gate runs every layer at maximum across a reference set and fails if mean hue angle moves more than 1.5 degrees or mean chroma more than 3 per cent.

Reshaping is absent rather than capped. No jaw, no eyes, no nose, no waist, no height, at any setting. That one has a page of its own.

Texture removal: 80% maximum, 25% default
Visible-loss note: above 60%
Tone evening: 80% slider, geared to 45%
Hue movement allowed: 1.5°
Chroma movement allowed: 3%
Heal history: 30 strokes
Red-eye strength: none — the artefact is binary
Reshaping: not present at any setting

The readout, and what it is not

The figure beside the picture is mean absolute luminance change over the pixels that actually changed, as a share of their own original luminance. Two properties make it worth having. It does not drift as your eye adapts. And it counts where the work is, rather than diluting the result across a frame that is mostly background — an average over the whole picture would read near zero for almost any real portrait edit and would be a decoration.

It is not a score and there is no good value. It is a second opinion that has not been in the room with you.

Two questions worth asking about an edit

Would this be there next month? A spot, a stray hair, a patch of shine from a warm room, a red pupil from a flash — none of those are properties of the person, and removing them arguably makes the photograph more accurate rather than less. A line at the corner of an eye, the texture of skin, the shape of a jaw are properties of the person. That distinction does more work than any percentage.

Would you say what you did? If you would be comfortable telling the subject, and the viewer, exactly which controls you moved and how far, the edit is fine. If the answer depends on nobody asking, that is the answer. This is the standard the whole tool is built to: it is why the caps exist, why the defaults are conservative, and why the compare control is a hold rather than a toggle — a held comparison is the only look that reliably tells you what you have done.

Where the occasion changes the answer

A headshot for a company page is expected to be cleaned, and cleaning it well is a professional service rather than a deception. A wedding or matrimonial portrait carries a strong convention of retouching that everyone involved understands. A press photograph is governed by a standard most outlets publish, and the permitted set is usually close to nothing. An identity document is checked against a living face, and a materially altered photograph can invalidate it — background colour, pixel dimensions and file weight are what those forms actually specify, and none of them is skin work.

The tool cannot know which of those you are doing. That is precisely why its ceilings do not move: a control that would be wrong in one of those contexts is a control that is simply not offered, because nothing here can tell the contexts apart.

What people ask

What does the percentage beside the picture measure?: Mean absolute change in luminance, taken only over the pixels that actually changed, expressed as a share of those pixels' own original luminance. It is deliberately not averaged over the whole frame: a figure computed that way would read close to zero for any honest portrait retouch and would therefore tell you nothing. Restricting it to the changed area makes it a measure of how hard the edit is working where it is working.
What number should I aim for?: There is no target, and offering one would be the same mistake as a beauty score. What the readout is for is comparison — between two settings of the same control, and between what you thought you had done and what you have done. A figure that climbs while the picture stops improving is the signal. Somebody working on a headshot for print will sit in a different range from somebody cleaning a snapshot, and both can be right.
Who decided the caps?: We did, by reasoning rather than by testing, and that is worth saying plainly. The 80 per cent texture ceiling and the warning above 60 are judgements about where a face stops reading as a photograph, argued from what the bands contain rather than measured against a panel of retouchers. They have not been validated. If they turn out to be wrong they will move, and the reasoning will change on this page with them — but they will not be removed, because the existence of a ceiling is not the part that is uncertain.
Is retouching a photograph dishonest?: Not inherently, and the question is better asked about the specific edit than about the practice. Removing a spot that will be gone next week keeps the photograph truer to the person than leaving it in. Evening a patch of shine corrects the lighting rather than the face. Erasing every line, pore and shadow produces an image of somebody who does not exist and circulates it as a picture of somebody who does. The test this site uses is disclosure: an edit you would be comfortable naming to the person in the frame is one you can make.
Why does the tool warn me rather than stop me?: Because above 60 per cent the result is still a legitimate choice in some contexts and not in others, and a tool that refuses is a tool that gets replaced by one that does not. What it can do is make the consequence visible at the moment of the decision: the note appears, the compare control is one press away, and the figure beside the picture keeps counting. Refusal is reserved for the things that have no defensible setting at all, which is why reshaping is absent rather than capped.