The method CALLUME · HOW IT WORKS

How a color reading is actually made.

Put a known reference in the frame and correct the light against it, sample traced regions instead of pixels, measure in CIELAB with CIEDE2000, fuse the measurement with what your own eyes pick in the drape rounds, then publish the uncertainty alongside the answer. Five steps, in that order — the first one is what most tools skip.

The pipeline

Most color-analysis tools read your season straight off the pixels, which means they are reading the room as much as the face. Here is the whole pipeline instead, in the order it runs — including the parts that make the answer less certain rather than more.

  1. 01 the light

    Correct the light, before anything else

    A reference in the frame is something whose true color the math already knows: a sheet of plain white printer paper, your phone’s own screen stepped through four known colors — black, white, yellow, cyan — or an 18% gray card if you own one. The screen method works by subtraction: what the room was doing is whatever is left when you take the dark frame away from the lit ones, so the room cancels instead of being guessed at. Every photo is adapted to standard daylight (D65) with a Bradford transform in a sharpened cone space, then normalized against whichever reference was there. Skip this step and every number after it inherits whatever the room was doing.

  2. 02 the sample

    Sample regions, not pixels

    Skin, hair, and eye color come from traced regions rather than single pixels — face landmarks carve the zones, with lips, eyes, and brows excluded from the skin mask. A robust median with an inlier test throws out stray hair, glare, shadow, and blemish pixels before any color is computed.

  3. 03 the read

    Read it in CIELAB, compare with CIEDE2000

    Color is measured in CIELAB and compared with CIEDE2000 — a perceptual distance, the metric a color scientist would use, rather than raw RGB arithmetic. Four continuous axes come out of it: undertone, depth, clarity, and contrast. Those four coordinates place you against the twelve seasons.

  4. 04 the check

    Let your eyes overrule the math

    Real colors go against your face, two at a time, and your picks are fused with the measurement in a Bayesian tournament. That ordering matters: a confident-but-wrong measurement can be overturned by what you actually see on the day. Each pass narrows the answer toward the season the drapes and the math agree on.

  5. 05 the honesty

    Publish the uncertainty with the answer

    Every axis carries an error bar and every structural measurement carries a reliability tier, because a forehead width read from a photo is trustworthy and a jaw width is not. Confidence is capped at 95% on purpose: we expect a photo to land somewhere around 75 to 85 percent agreement with an in-person analyst — our estimate, not a published figure, because none exists — and claiming certainty on top of that would be a lie.

What step one actually does

One skin sample, as three rooms rendered it The same sample, each corrected to D65

Three rooms, one face. Read straight off the pixels, that face is three different people, and a tool scoring it would name three different seasons. Correct each frame against a known-neutral gray and they converge — which is why the same session shot on two different days should land on the same season, and why that repeatability is a better sign a read is right than any confidence number.

The terms, in plain words

Gray card
A card of known-neutral gray, held in the frame. Because its true color is known, the math can see exactly what the room's light did to it — and undo that on every other pixel. About eight dollars, once.
D65
The standard daylight white point every reading is corrected to. Once two photos are both adapted to D65, their colors are comparable even if one was shot at a window and one under a bulb.
Bradford transform
The chromatic-adaptation method used to move a photo from the room's light to D65, applied in a sharpened cone space. The crude alternative — scaling the red, green, and blue channels independently — distorts saturated color.
CIELAB
A color space built so that distance roughly matches how different two colors look to a human eye. It describes a color with three readings: how light it is, how far it sits between green and red, and how far between blue and yellow. Your undertone, depth, and clarity are read here.
CIEDE2000
The formula used to measure how far apart two colors are in CIELAB, built so its answers line up with what an eye actually notices — a distance of about one is the smallest difference a person can see. It is how every palette color is scored against your measured coloring.
The four axes
Undertone (warm to cool), depth (light to deep), clarity (muted to clear), and contrast (the gap between your hair, skin, and eyes). Twelve seasons are twelve regions of that four-axis space.

What it still cannot do

A trained eye in the room

An in-person analyst with drapes and daylight is still the gold standard for the drape step, and this page says so. What a calibrated photo adds is measurement and repeatability, not a replacement for a professional.

Equal precision for every skin tone

Like all photo-based color analysis, accuracy may not be identical across every skin tone. A melanin index cross-checks depth so it is not confused with undertone, and every reading states its own uncertainty rather than hiding it.

Certainty

Confidence is capped at 95% deliberately. A tool reporting 99% is telling you something about itself, not about your face.

Questions

How does color analysis actually work?

A calibrated reading runs in five steps. The light is corrected first, using a gray card as a known-neutral reference so the photo can be adapted to standard daylight. Skin, hair, and eye color are then sampled from traced regions rather than single pixels. Those samples are measured in CIELAB and compared with CIEDE2000, placing you on four axes: undertone, depth, clarity, and contrast. Drape rounds let your own eyes overrule the math where they disagree. Finally the uncertainty is published alongside the answer, rather than hidden.

Why do I need a gray card?

Because a photograph records the light as much as the face. A gray card is the one object in the frame whose true color is already known, so the math can see exactly what the room did to it and undo that everywhere else. Without one, a warm bulb or a cool window is baked into every number that follows. A card costs about eight dollars and never wears out. The free instant read deliberately skips it, which is why it reports a lower, honest confidence.

What is CIELAB, and why not just use RGB?

CIELAB is a color space built so that distance roughly matches how different two colors look to a human eye. It describes a color with three readings: how light it is, how far it sits between green and red, and how far it sits between blue and yellow. RGB distance does not work that way — two pairs of colors can be equally far apart in RGB while looking obviously different to a person. Undertone, depth, and clarity are read in CIELAB for exactly that reason.

What does it mean for two colors to be 'close'?

Closeness here is a measured quantity, not an impression. Two colors are compared with the CIEDE2000 formula, which is built so that its distances line up with what a human eye actually notices — a distance of roughly one is the smallest difference a person can see at all. That is why a palette can say a color is a very close match rather than merely a similar sort of shade, and it is what turns a palette from an opinion into a measurement.

If the analysis is measured, why are there drape rounds?

Because a measurement can be confidently wrong. Photo conditions vary, and a single number should not get the last word on your face. Real colors go against you two at a time, and what you pick is fused with the measurement in a Bayesian tournament, so your eyes can overturn the math. Each pass narrows the answer toward the season where the drapes and the measurement agree.

How accurate is photo-based color analysis?

Photo-based analysis tops out around 75 to 85 percent agreement with an in-person analyst, so stated confidence is capped at 95 percent on purpose — a tool claiming certainty is telling you something about itself. Every axis carries an error bar and every facial measurement carries a reliability tier. The strongest sign a read is right is not a big confidence number but repeatability: the same face, shot on two different days, landing on the same season.

Every number in a calibrated Callume session traces back through those five steps, and the bound document prints the working — every formula, every correction, with your actual values in it. The free read runs the same math with no reference at all, so you can watch exactly how much your camera was guessing.