Amount, radius and threshold: the three controls, in full

Start with the operation itself, because all three controls are modifications of it. A copy of the picture is blurred. That blurred copy is subtracted from the original, leaving a residual that contains only what the blur removed: the rapid changes. Edges are rapid changes. So is texture. So is sensor noise, and at this stage nothing in the arithmetic can tell the three apart. The residual is then multiplied by some factor and added back to the original, so wherever the picture was rising it rises more sharply and wherever it was falling it falls more sharply.

Two implementation choices here change what the controls mean, and both are worth knowing. The mask is computed on luminance alone, with the colour channels left as they were, because sharpening the three channels independently produces coloured fringes along any saturated edge. And the addition happens in linear light rather than in the gamma-encoded values the file stores, because in the encoded space the same residual brightens a pixel more than it darkens it — so the white side of every halo arrives before the black side and is always the stronger. Working linear makes the two sides symmetric, which is what allows the overshoot and undershoot clamps to be two separate controls that mean the same thing as each other.

Radius is the standard deviation of the blur used to make the residual. It is called a radius by thirty years of convention rather than by description. What it sets is the size of detail the pass responds to, and — the part that matters when a result looks wrong — the width of the rim it draws, which runs to roughly three times the number.

The radius that is correct depends on the size of the frame and on where it is going, not on how blurred it looks. A 24 megapixel file viewed on a screen at fit size hides a radius of 3; the same file printed at A3 shows it as an outline. This is why the output presets exist separately from the creative pass, and why a preset meant for matte paper looks wrong on a monitor and is nevertheless the right number.

Amount is a plain multiplier on the residual, expressed as a percentage: at 100% the full residual is added back, at 200% twice as much. It is the control most people reach for and the least interesting of the three, because it scales a decision the other two have already made. Doubling the amount at a radius of 0.6 px produces a crisper picture; doubling it at 3 px produces an outlined one.

The useful way to work is to set the radius against the size and destination of the file, then raise the amount until the rim on the highest-contrast edge in the frame becomes findable at 100%, then come back about a fifth. The stage rings that edge as soon as the rim crosses the visibility boundary, so the judgement is not being left to memory.

Threshold is the reason a portrait can be sharpened without the skin turning to gravel. Every residual value smaller than the threshold, measured in 8-bit code values, is discarded and the pixel is left exactly as it arrived. Since noise is by definition small and edges are by definition not, a threshold set just above the noise floor sharpens the edges and leaves the noise where it was.

The catch, and the reason a naive implementation of it is worse than none, is the cut itself. A hard boundary at, say, 4 levels means that a residual of 3.9 is thrown away entirely and one of 4.1 is multiplied in full — so across a smooth gradient where the residual drifts past the boundary, a contour appears exactly along the line where it crossed. The pass here ramps the residual in smoothly over the four levels above the threshold, which costs nothing and removes the contour.

• Set threshold from the measured σ, not from habit. Step one prints it. A threshold of roughly twice σ is a sound starting point and the tool opens there.
• Threshold acts on amplitude; the edge mask acts on location. They are not alternatives — a noisy sky has small residuals everywhere, which threshold handles, while a busy foreground has large residuals in places that should not be touched, which the mask handles.
• A threshold above about 12 on a clean file will visibly skip real detail. If a high number seems necessary, the balance control is usually the thing that is actually wanted.

These are starting points for a file of roughly 12 to 24 megapixels that has not been sharpened before, destined for a screen. They are not answers. Every one of them should be moved after looking at the result at 100%, which is the only reason the inspection view is compulsory.

Threshold first, from the measured noise figure, because it decides what the other two are allowed to touch. Radius second, from the size of the file and where it is going, because it decides what scale they act at. Amount last, because it is the only one that is a simple matter of how much. Setting amount first and then hunting with the other two is the common approach and it is why so many sharpened photographs have rims: the amount ends up compensating for a radius that was never right.