Iris Portrait Studio

The project at a glance.

What it does.

How an eye photo turns into a portrait

1. 01

   Photograph

   Take a photo with the camera or upload one, following three short capture tips.

2. 02

   Frame the eye

   Pinch and drag so the iris fills the square, or use the whole image.

3. 03

   Check the boundaries

   Confirm the detected iris and pupil rings, or nudge them into place.

4. 04

   Create the portrait

   The app corrects white balance, repairs lashes and reflections, lifts detail on lightness only and composes a perfect circle.

5. 05

   Save or print

   Pick a print size the photo can support and a background, then save the PNG, JPEG and print-ready PDF.

How it's built.

Technology

The hard parts, solved.

1. 01

   Colour the enhancement stages cannot touch

   All tone, contrast and sharpening work runs on the CIELAB lightness channel, and the colour channels are separate buffers the detail stage is never handed. The only steps allowed to touch colour are white balance from the sclera and two bounded controls that scale intensity or separation with the hue held fixed. Tests assert that hue moves by less than 2 degrees under the strongest preset, that chroma stays within 10%, and that the intensity control stays within its 1.25x ceiling.

2. 02

   Telling lashes and eyelids apart from real iris structure

   Dark crypts are real tissue that must survive, while eyelid skin over a brown iris is only about 20 degrees away in hue. The occlusion analyser scores each sample against the statistics of its own radius, uses chroma per unit of lightness, finds lashes by morphological thinness and thresholds at two scales. Tightening the mask cut the share of the iris being rebuilt on a real development photograph from 30.2% to 10.8%, as recorded in the commit history, so far more of the artwork is measured tissue rather than interpolation.

3. 03

   Doing the hard maths in polar space

   Resampling the iris into an angle by radius buffer turns several difficult problems into simple ones. Sharpening the radial fibres becomes a choice of axis, repairing a lash means interpolating along the angle at a fixed radius, and an off-axis, elliptical iris is straightened into a circle as part of the resampling.

4. 04

   Fast enough in pure Dart on a phone

   Keeping the engine in plain Dart made it portable and testable, but the project's own analysis expects pure Dart to run roughly 3 to 8 times slower than native OpenCV for this kind of filtering. VNE replaced a quadratic morphology filter with the van Herk/Gil-Werman method (checked identical to the naive version across 152 configurations), turned the mask growth into a breadth-first fill and estimated each ring's statistics from a sample instead of sorting all of it; the commit history records a high-resolution render falling from 37.9 s to 14.6 s. On Android, detection and rendering also run on a background isolate, so the screen stays responsive while they work.

5. 05

   Letting AI in without letting it invent

   The AI layer is designed so a model can help without being trusted. Provider keys stay on the studio's own gateway, a learned segmenter's answer is only used when it is more confident than the built-in detector, and insight narration receives measured numbers rather than the photograph, with its reply length-limited and stripped of any colour swatch that was not measured. For future pixel-changing steps such as restoration or upscaling, a fidelity guard already compares a result with its input for hue shift, colour difference and structure drift, and tests confirm it rejects both a hue shift and reinvented structure.