Bring Out the Eyes Without Faking the Light

How it works

Your portrait is scaled to about one and a fifth megapixels and repainted at a low denoise with a fresh seed each run. The instruction asks for detail in the iris and the lashes and a slight lift of the shadow around the eye, and then states the two constraints that make the page work: keep the existing catchlight where it is, and keep the whites their natural color.

The negative prompt names the standard artifacts of this category directly — pure white sclera, an added catchlight, a moved reflection, enlarged eyes, changed eye color, and a glassy stare.

The reflection is a measurement

The eye is a wet sphere, so it behaves like a mirror, and what it reflects is the scene in front of the person. In practice the brightest thing there is the light source, and it appears as a small bright shape on the cornea.

That shape carries real information. High on the iris means the light was above. Off to one side means it was to that side. A large soft blob means a window or a softbox; a hard small dot means a bare flash; a rectangle means a screen. Two separate reflections mean two sources, and both then have to be consistent with every shadow on the face.

People do not analyze this consciously and they use it constantly. It is a large part of why some portraits feel real and others feel assembled, and it is the first thing that goes wrong when a tool decides to add sparkle to an eye.

So the constraint here is conservative on purpose. Whatever reflection your photograph recorded is the correct one for that photograph, and it stays.

What is actually recoverable

The iris is a textured structure — radial fibers, a visible collarette, color variation between the pupil edge and the rim — and in most portraits it is recorded but underexposed, sitting in the shadow of the brow and the lashes.

That is a good situation. The detail is present at low values, and lifting it brings out something genuinely photographed. The same is true of individual lashes, which are often merged into a dark mass by softness rather than absent from the file.

The bad situation is an eye in real shadow, under a cap or a heavy brow, or a small face in a wide frame. There the iris occupies very few pixels and the structure that appears in the output was decided by the model. It will look convincing. It is not a recording of that person's iris, and the color in particular should not be trusted.

The tells of an overdone eye

Three, in the order people notice them.

Pure white sclera, which reads as startled and slightly ill, because no healthy eye is uniformly bright in ordinary light. A second catchlight that the lighting does not justify, which registers as fake without being identified. And eyes enlarged relative to the rest of the face, which drifts the portrait towards a doll and away from the person.

None of these is subtle once you know to look, which is the useful part: checking takes one zoom and one comparison against the original.

Publication gate

This page ships once the workflow has been run against a portrait with a single soft catchlight, one with a hard flash reflection, eyes in deep shadow, and a subject wearing glasses, confirming in every case that the reflection has not moved and the sclera has not been flattened to white.