Is AI colour analysis accurate?

For relative qualities, reasonably. Depth and contrast are comparisons made inside a single image, so they hold up even when the lighting is poor. For absolute skin colour it is not accurate and cannot be, because an ordinary photograph contains no reference of known colour to calibrate against. Treat a season as a shortlist rather than a verdict.

Why did two apps give me different seasons?

Most likely because they measured two different photographs, or because one of them corrects for lighting and the other does not. A warm bulb shifts every pixel in the same direction as a warm undertone, so an uncorrected tool cannot separate the room from the face. Photograph yourself in daylight and in lamplight and run both. If the answer changes, the tool is reading the room.

Is a paid in-person colour analysis better?

It removes the camera and the lighting problem, which are the two largest sources of error, and it adds a trained eye. It also costs money and remains a judgement rather than a measurement. The honest comparison is that in-person analysis is more reliable and less repeatable, and a good tool is less reliable and perfectly repeatable.

Can I do colour analysis without any tool at all?

Yes, and the method is older than any app. Sit by a window in daylight, hold a piece of pure white and a piece of off-white cream under your chin in turn, and look at your skin rather than the fabric. Whichever makes the skin look even rather than blotchy points to your undertone. Repeat with silver and gold. It settles the warm and cool question for most people.

Does undertone change over time?

Undertone itself is stable. Surface tone changes with sun exposure, and that is enough to move a borderline reading from one season to a neighbouring one. If you take a reading in late summer and again in winter and get two answers, the two answers are probably both defensible and you sit between them.

Does AI colour analysis actually work, and why do two apps give you two different seasons?

There are now more than thirty five AI colour analysis platforms and they routinely disagree about the same face. That is not a bug in one of them. It is what happens when a measurement is taken through an uncalibrated camera under an unknown light.

Last reviewed 2 August 20267 min readSravya

Editorial portrait comparing colour placed near the face

The figures

35+
Active AI colour analysis platforms: 12
Seasons most tools sort into: 3
Things a photograph can establish

The short version

It works for the things a photograph can actually carry, and it does not work for the thing most apps claim. A camera records the light in the room as well as the skin, so an absolute skin colour cannot be read from an ordinary selfie. What survives is relative: how dark the skin is compared with the hair, how far apart the lightest and darkest parts of the face are, and which direction the skin sits from neutral once the image has been white balanced. Those three are enough to narrow twelve seasons to two or three. They are not enough to name one and stop, and any tool that names one without showing you the figures behind it is presenting a guess as a result.

01Why the same face gets different answers

A photograph is a record of light reflected off skin, not of the skin. Change the bulb from a warm domestic lamp to daylight and every pixel moves, in a direction that looks exactly like a change of undertone. Nothing about the person has changed.

This is why two apps disagree. They are not reading the same numbers, because the numbers depend on the room. Unless a tool corrects for the light before it measures anything, it is reporting the lamp as well as the face, and it has no way to tell you which part of the answer came from which.

02What a photograph can legitimately establish

Three things survive a change of lighting well enough to be usable, because all three are comparisons within the same image rather than absolute readings.

The three that survive, and the one that does not

Quality	Survives a light change?	How it is read
Depth	Yes, as a comparison	How dark the skin is relative to the hair and eyes in the same frame. Described numerically by the Individual Typology Angle from CIELAB lightness and yellow-blue values.
Contrast	Yes	The distance between the lightest and darkest points in the face. A ratio, so a shift in the light moves both ends together.
Undertone direction	Only after white balancing	The hue angle of the skin in CIELAB, taken after a white balance correction. Without the correction this is the least reliable of the four.
Absolute skin colour	No	Cannot be recovered from an uncalibrated photograph. It needs a reference card of known colour in the same frame, which no selfie has.

The correction that makes the third row possible is a standard piece of colour science, not a proprietary trick. This site corrects against the whites of the eyes, because the sclera is the least strongly coloured surface reliably present in a face photograph. That is a weaker claim than it sounds, and it is the honest one: the sclera is not truly neutral. Russell and colleagues reported in 2014 that sclera grow darker, redder and yellower with age, so treating one as white leaves an error behind, and that error is larger on an older face. A whole-image grey estimate was considered and rejected, because a face filling the frame is exactly the case where it fails: it pushes the dominant region towards grey and removes the signal being measured. So the correction is an approximation with a known bias, not a calibration. It is still a great deal better than measuring raw pixels and calling the result an undertone.

03The three checks that separate a reading from a guess

1Does it show you the numbers? A tool that gives a season and no figures cannot be checked, and cannot be wrong in any way you can catch.
2Does it name the runner up? Twelve seasons sit close together. Soft summer and soft autumn differ by one quality. A tool that never reports a near miss is hiding the most useful thing it knows.
3Does it say what it cannot do? Any tool that claims to read your exact skin colour from a phone camera, with no reference card and no mention of lighting, is claiming something the physics does not allow.

04Twelve seasons, and why four was never enough

The original system had four seasons and sorted almost entirely on warm against cool. It fails on a common case: two people can both be cool, and one of them light and low contrast while the other is deep and high contrast. Put them in the same palette and one of them is wearing colours that flatten them.

Twelve seasons split each of the four along the two qualities that were being ignored, depth and clarity. That is why a modern reading scores four things rather than one, and why a result that only tells you warm or cool has answered a quarter of the question.

Every season, every colour

12 seasons · 72 shades

Light Spring

Warm Spring

Clear Spring

Light Summer

Cool Summer

Soft Summer

Soft Autumn

Warm Autumn

Deep Autumn

Deep Winter

Cool Winter

Clear Winter

These are the actual palettes the colour tool works from, not a sample. Point at one to hold it. Twelve seasons rather than the old four, because four cannot separate someone light and cool from someone deep and cool.

05What to do instead of trusting one answer

1Take the reading in daylight, near a window, with indoor lights off. It removes the largest single source of error at no cost.
2Run it twice on two different days. A season that holds across both is worth acting on. One that moves was never a measurement.
3Test the palette rather than the label. Hold two candidate colours under the chin in daylight and look at the skin, not at the fabric.
4Treat the result as two or three seasons, not one. The neighbouring season's palette usually contains most of what you will actually wear.

Common questions

Sources

Every figure quoted above is listed here with what it was used for and when the source was last checked. Anything not attributed to a source is this site’s own reading and is labelled as such where it appears.

Extent Research
Personal colour analysis market 2026: size, trends and AI growth
Count of active AI colour analysis platforms and the role of social media in driving demand.
Last checked 2 August 2026
Society for Imaging Science and Technology
Shades of Gray and Colour Constancy (Finlayson and Trezzi, 2004)
The grey-world family of white balance methods, described here as the standard approach and the reason an uncalibrated photograph cannot be trusted. This site does not use it, and says why: a face filling the frame defeats the grey-world assumption.
Last checked 2 August 2026
British Journal of Dermatology
Variations in skin colour and the biological consequences of ultraviolet radiation exposure (Del Bino and Bernerd, 2013)
The Individual Typology Angle used to describe skin depth from CIELAB values rather than from a visual guess, and the evidence that the classification tracks real constitutive pigmentation.
Last checked 10 August 2026
Psychology and Aging
Sclera color changes with age and is a cue for perceiving age, health, and beauty (Russell et al., 2014)
The limit on the white balance correction: sclera become darker, redder and yellower with age, so using one as a white reference leaves a bias that grows with the age of the subject.
Last checked 12 August 2026

Check it against your own measurements

Undertone, depth, contrast and chroma, scored separately, with every figure shown and the runner up named. It will tell you when two seasons came close instead of picking one and sounding certain.

Score my four qualities