Take the red out of a flash photograph

The scan works on a reduced copy — around 480 pixels on the long edge, which is plenty for a blob a few pixels across and keeps the pass instant on a large file. Every pixel is tested for red dominance: a red channel comfortably above both others, and a red share of the total above 42 per cent. The flagged pixels are then grouped into connected regions and each region is judged on three things — area within a band, an aspect ratio between roughly 0.45 and 2.2, and at least 42 per cent of its bounding box filled. What survives all of that is small, round and solid, which is what a pupil is and what a stripe of red fabric is not.

The correction desaturates and darkens. Each pixel inside an accepted disc takes a neutral built from its own green and blue, scaled down, faded in from the edge of the disc so there is no hard rim. Two things are deliberately left alone: any pixel not actually red-dominant, which is how the iris around the pupil survives, and any pixel that is bright and neutral in all three channels, which is the catchlight.

Landmarks would narrow this search to the eye sockets and cut the false-positive rate substantially. This release has no landmark model, so the size, roundness and fill tests do that work instead — and the confirmation step exists because they do it less well.

Let the scan finish, then look at the rings before you press anything. Zoom in with 2 and check each one is on a pupil. Accepting is all-or-nothing in this release: if the scan has offered you a genuine pupil and a red bead together, you are choosing between correcting both and correcting neither, and correcting both is usually still the better picture.

Then hold the compare control. Red-eye is the one layer where the corrected version almost always looks right and the original looks wrong, which makes it the easiest place to stop paying attention. Look specifically at the catchlight and at the boundary of the iris; if either has gone soft, the disc that was found is larger than the pupil.

What the scan runs on

Scan resolution

≈480 px long edge

Red dominance threshold

44 levels over max(G, B)

Red share of total

above 42%

Region aspect ratio

0.45 to 2.2

Bounding-box fill

at least 42%

Candidates offered

up to 6, largest first

Correction

neutral from G and B, darkened to 74%

Catchlight

kept — bright neutral pixels are skipped

• There is no manual pupil tool. If the scan misses an eye — too small in frame, partly behind a lens, lit orange rather than red — there is no way to point at it and say “there”. That is the honest gap in this layer and the one most worth closing.
• It cannot separate a pupil from something else that is small, round and red. Without landmarks the search covers the whole frame, so a bead, an out-of-focus tail-light or a bright lip highlight can be offered. The rings exist precisely because the test cannot be trusted on its own.
• It does not correct animal eyeshine. Green and yellow reflections from a tapetum are a different phenomenon with different colour, and neither the dominance test nor the shape tests apply to them.

Why is there no strength slider here when every other layer has one?

Because there is no partial right answer. Flash red-eye is a physical artefact — the retina, lit straight on through a dilated pupil, reflected back into the lens — and a pupil is either showing the colour of the back of an eye or it is showing the colour of a pupil. Half-correcting it produces a dull maroon that is not a real state of either. Every other control on this site is a judgement with a defensible range; this one is a repair.

Why does it ask before correcting, if there is only one right answer?

Because the question is whether the thing it found is a pupil, not how much to change it. A colour test that looks for small, round, red-dominant blobs will sometimes offer you a bead on a necklace, a tail-light out of focus behind a shoulder, or a bright lip highlight. Each candidate is ringed on the picture and you accept them together, so the thing being confirmed is the identification.

It found the pupils but the correction looks flat and dead.

That is usually the catchlight, and this layer is built to keep it. The catchlight is the small specular reflection of the flash itself, the one bright neutral thing inside a red pupil, and painting over it is the single clearest tell of a careless fix. Pixels that are bright in all three channels are left exactly as found. If the eye still looks flat, the catchlight was lost when the photograph was taken and nothing here will invent one — a painted highlight in the wrong place is worse than none.

It found nothing and the pupils are clearly red.

The test asks for three things at once: a red channel at least 44 levels above the higher of green and blue, red taking more than 42 per cent of the total, and a blob that is small, roughly round and solidly filled. An eye lit orange rather than red, a pupil under an arm of sunglasses, or a face small enough that the pupil is three pixels across will all miss. There is no manual pupil tool in this release, which is a real gap.

Does this work on the green or yellow eyes in a photograph of a cat?

No. Animal eyeshine comes from the tapetum lucidum, a reflective layer behind the retina that people do not have, and it returns green, yellow or blue rather than the red of human retinal blood. The colour test here is built for the human case and the shape tests assume a human pupil. Nothing stops you loading the picture; the scan will simply find nothing.

Should I run this before or after the other layers?

It does not matter what order you press things in — the stack always runs heal, then red-eye, then skin. Red-eye sits in the middle deliberately: it is colour work and should be evaluated against luminance that no other layer has moved yet, and it needs the discrete marks already gone so a healed spot near an eye is not mistaken for one.

• Spots and blemishesOpens with the heal brush in hand and the region already over the face.
• Skin smoothingOpens with the skin layer at 25 per cent, which leaves three-quarters of the texture.
• Why nothing here reshapes a faceThe one exclusion that is a decision rather than a boundary, and the reasoning behind it.
• The retoucherEvery layer at zero, the working region waiting to be placed.