A filter radius fixed in pixels is calibrated to one subject size, not to the image

Fact

Any spatial filter tuned in pixels — a smoothing blur, a high-pass, an unsharp mask, a morphological open — is implicitly tuned to one size of subject. When the subject's size in frame varies, the same radius means a different thing at every distance.

A face fills 40% of the frame width at arm's length and 12% across a room: a radius that softens pores at the near scale erases the whole face at the far one. The fix is to key the radius to a feature the detector already measures, not to the frame:

radius = clamp(faceWidth * imageHeight * k, lo, hi) * userSlider;

The slider then multiplies a derived value instead of being the value.

Why it matters

The symptom is "this setting is wrong", not "this setting is scale-dependent", so it gets retuned against whatever the subject happened to be doing during testing — which bakes that one distance into the defaults. Every later report disagrees with every other, because each reporter sat at a different distance.

It is the same failure whenever scale is free: a document scanned at another DPI, a map at another zoom level, a render at another output size.

How to apply

• Derive the radius from a measured feature; the detector that finds the subject almost always hands you one for free.
• Clamp the derived value at both ends, so a bogus measurement — a detection on a reflection, a zero-size box — cannot explode or collapse the filter.
• Keep the user's control as a multiplier on the derived radius.
• Verify by asserting the derived radius scales: two subject sizes in, two proportional radii out, both inside the clamp.

Related: [[radial-kernel-on-a-non-square-grid]], [[palette-is-calibrated-to-its-shading-model]].