How our color analysis works

We use the 12-tone seasonal system. Each of the four classic families splits three ways according to which characteristic dominates: how light or deep the coloring is, how warm or cool, and how bright or muted. That last axis is the one the old four-season system missed, and it is why so many people were told they were a “Winter” in the 1980s and never felt it fitted.

Every season in our data is stored as a position on those three axes. Your photo produces a position too, and the result is simply which season you land nearest, scored against all twelve, not picked from a decision tree.

From your photo: we place 468 facial landmarks and sample only flat skin, both cheeks, the forehead, plus your irises and the hair above your face. Lips, eyes, eyebrows, nostril shadow and background are all excluded, which matters more than it sounds: the naive approach of averaging the middle of a photo quietly mixes dark hair into your skin tone and drags the depth reading down. Everything is converted to CIELAB, a perceptually uniform color space, so that “how far is this from that season” has an honest answer.

Undertone is the hard one, the part other online tools get wrong. A phone camera’s white balance shifts precisely the warm-versus-cool signal you need to read, so before we read it we correct the photo against a neutral reference already in the frame: the whites of your eyes, or the most neutral bright highlights. Add more than one photo, taken in different light, and we keep the reading they agree on, which cancels the color cast of any single lamp or window.

Each axis leans on what measures it best. Depth and contrast come straight off the photo, which reads them reliably. Undertone leans on that white-balance correction and on seeing you in more than one light. The short style profile in the flow tailors the written report to how you actually dress; the season itself is measured from your photo.

Behind the result is a small, deliberately boring measurement. Every season in the 12-tone system is a fixed point in the same three-axis space, warm to cool, light to deep, bright to muted. We place your reading in that space and return the season you land nearest, scored against all twelve at once.

The lines between those seasons are not eyeballed. We calibrated them against Deep Armocromia, a research dataset from a published academic study (ECCV 2024) of several thousand faces, each labelled with its season by the researchers, so the boundaries we draw are checked against real, expert-labelled examples rather than taste.

We also tested the reading on 410 celebrity photos and published the numbers, including the unflattering ones: from a single press photo the exact season is right about one time in nine, which is why the report treats one photo as a first estimate and asks for a second.

It is still a measurement, and measurements have conditions. Colored contact lenses, heavy makeup, a strong tan, dyed hair and unusual lighting all move the reading, and a filter will wreck it entirely. When two seasons come out close we show you the gap rather than rounding it into false confidence, sitting between two neighboring seasons is common and is not a failure of the system.

An in-person analysis with fabric drapes and a trained eye is still the gold standard, and it costs $150–$500. If you want one, take one. This site gives you the same framework, an honest measurement and a usable palette for free, and the full written report for $9.

Ruslan Shchuchkin, generative AI and ML engineer, data scientist, five years in production machine learning. I built the measurement this site runs: the landmark sampling, the white-balance correction against the whites of the eyes, the CIELAB scoring and the multi-photo consensus, and I wrote the tests that pin them. I am not a trained colour analyst, and the site does not pretend to be one. It is an engineer’s attempt to make the measurable part of colour analysis repeatable, and to say plainly where the measurable part ends.

LinkedIn · questions and corrections to [email protected].

It never leaves your device. There is no upload endpoint, no storage bucket and no third-party vision API, we even host the face-detection model and its WebAssembly runtime ourselves rather than loading them from Google, so that running the analysis does not tell anyone else you are doing it. The original photo is discarded once measured; a small cut-out of your face stays in session storage until you close the tab so your report can show it. Details on the privacy page.