Methodology
How AI Color Analysis Works
Personal color analysis used to mean a $300 appointment and a consultant draping fabric against your face. Our AI color analysis reaches the same three judgments — undertone, depth, and chroma — from a single selfie in seconds. Here is exactly how it does that, how accurate it is, and where it still falls short.
By Aiyi · Color Analysis Editor
What AI Color Analysis Actually Measures
Every seasonal color system, from the classic four seasons to the modern 12-season model, rests on three measurable qualities of your natural coloring. Good AI color analysis does not guess your “vibe” — it estimates these three axes and nothing more:
- Undertone — whether your skin reads warm (golden), cool (rosy/blue), or neutral. This is the foundation of the entire analysis.
- Depth — how light or dark your overall coloring is, and how much contrast sits between your hair, skin, and eyes.
- Chroma — whether soft, muted tones or clear, saturated ones suit you best.
Plot those three values together and you land somewhere specific on the seasonal map. That coordinate is your season. If you want the full breakdown of the system itself, our guide to the season system walks through every sub-type.
Where Those Three Axes Come From
Undertone, depth, and chroma are not something we invented to make AI color analysis sound rigorous. They are inherited, and it is worth knowing from whom — because it is the difference between a personality quiz and a measurement.
- Munsell (1905) — the original three dimensions
- Albert H. Munsell built the first color system to order color along three independent axes: hue, value, and chroma. Depth is his value; chroma keeps his name outright. It remains the vocabulary colorists actually work in.
- Sci\ART — Munsell applied to human coloring
- Kathryn Kalisz adapted that framework to faces, and the twelve seasons are what came out of it. That is why the seasons subdivide the way they do — Light, True, and Soft or Bright variants are positions on Munsell's axes, not moods.
- CIELAB — how we compare two skin tones
- Raw RGB is a poor judge of whether two tones look different to a person. CIELAB was designed so that equal numeric distance approximates equal perceived difference, so the sampled tones from your cheeks, forehead, and chin are compared there rather than in the pixel values a camera happens to record.
None of this makes the result infallible — a century-old system applied through a camera still inherits every problem a camera has. It does mean the axes being estimated are real, defined quantities rather than something a model made up on the spot.
The AI Color Analysis Pipeline, Step by Step
Image intake & quality read
The pipeline first checks the selfie itself: is there enough natural light, is the face unobstructed, are filters or heavy makeup distorting the skin? AI color analysis is only as good as the photo it reads, so this step quietly flags conditions — warm indoor bulbs, strong shadows, a beauty filter — that would otherwise push the result toward a wrong season.
Undertone estimation (warm / cool / neutral)
The vision model evaluates how your skin, cheeks, and the area around your eyes behave under the available light, estimating whether your coloring reads warm (golden, peachy), cool (rosy, blue-based), or neutral. Undertone is the single most decisive axis in any color analysis, and it is also the one self-report quizzes get wrong most often.
Depth & contrast reading
Next the model gauges depth — how light or dark your overall coloring is — and the contrast between hair, skin, and eyes. A high-contrast face (dark hair, light skin) routes very differently from a low-contrast one, even when the undertone is identical. This is where AI color analysis pulls ahead of a flat questionnaire.
Chroma reading (muted vs. clear)
The third axis is chroma: does your coloring suit soft, dusty, greyed tones, or clear, saturated ones? Together, undertone + depth + chroma form the three-dimensional coordinate that the 12-season system is built on.
Season classification & palette
Finally the three readings are combined and mapped to one of twelve seasonal sub-types — Light Spring through Dark Winter. You receive a personalized palette of your best colors for clothing, makeup, and accessories, plus the shades to avoid. The whole AI color analysis runs in seconds. Signed out, your photo is not saved; signed in, a reduced-size copy (about 600 px) is kept privately in your Style Profile until you remove it.
Why We Run This Engine (and Threw Out Two)
The accuracy of AI color analysis depends almost entirely on the vision model behind it. We learned that the hard way — four engines in nine months: two thrown out, one outgrown, and the one running now.
Open-source vision model
AbandonedThe first public version wired up an open-source vision model — lightweight and fast, but it could not separate the 12 sub-seasons cleanly enough to trust. Shelved after about two weeks live.
A self-trained classifier
AbandonedNext came a small classifier trained on 5,000 hand-labeled selfies and a rented GPU. It taught us what real data work costs — and that 5,000 images were nowhere near enough to generalize across twelve seasons. Shelved again.
Measured skin tones, then a vision model
Replaced, October 2026From May to September the engine split the job in two. Google's MediaPipe face landmarks run in your browser and sample your actual skin from four patches — both cheeks, forehead, chin — so the coloring is measured rather than guessed, and Meta's Llama 3.2 Vision on Cloudflare Workers AI read the face. Accuracy jumped roughly 20 points to around 80%. In August the model's job was narrowed to perception only — four 0–100 readings, no season names — with the 12-season decision moved into tested code; in September every photo began to be read three times, keeping the median of each reading, after the same photo came back as two different seasons.
A larger model for the reading
Live, still tuningSince October 1, 2026 the four readings come from Google's Gemma 4 (26B) on Cloudflare Workers AI, a larger model that costs more per photo than the one it replaced. Your skin is still measured in the browser first, and the measured tone still weighs into depth and undertone; the model describes, and the code decides the season. This is the AI color analysis you use today.
Keeping the Same Photo on the Same Season
An AI color analysis you can't repeat is not worth much. In September we ran one photo twice and got True Winter, then Light Spring. These four changes are how we closed that gap.
A larger model, even though it costs more
In October we moved the reading from an 11B model to Google's Gemma 4 (26B), which costs more for every photo. On our 20-photo test set it gave the same numbers on 17 of 18 repeat reads, and its warm-versus-cool score tracked the measured skin color at a correlation of 0.83, against 0.50 for the old model.
Finer steps on every scale
The model scores warmth, depth, clarity and contrast from 0 to 100, with randomness switched off. The old model only answered in steps of ten, so the smallest wobble was a full step — enough to push one face across a season boundary. On the same photos Gemma 4 used 13 distinct score combinations where the old model used 9, so neighbouring faces stop collapsing into one reading.
Three reads and the median on borderline photos
From September every photo was read three times and the median of each score was kept. Since October a photo is read once, and when that read sits within a few points of the line between two seasons (confidence under 66%), it gets two more reads and the median decides. In testing, 96.8% of analyses landed on the photo's majority season, with 1.96 reads per analysis instead of 3.
Watching what the model returns
No change to the model or the prompt goes live until a set of licensed sample photos has been run through the real model and passed its hard checks — a deep-skinned face, for instance, may never come back as a light season. A repeat test sends the same face through the production call again and again to catch a photo that flips. And every live analysis logs how many reads it took and how far they disagreed, so drift shows up in the numbers before it shows up in a reader's inbox.
The trade-off we accepted: a clear-cut photo is read once, so a confident first read that happens to be wrong is no longer outvoted by two more. In testing that cost 0.3 points of consistency and saved a third of the reads.
Accuracy & the Honest Limits
On a clear, well-lit, filter-free selfie, our AI color analysis agrees with a careful human read about 80% of the time. We would rather state that number plainly than pretend the tool is infallible. The most common things that throw it off are the same things that challenge a human eye: warm indoor lighting, strong shadows, beauty filters, and dyed or unusually colored hair.
Because of those limits, we treat AI color analysis as a fast, free starting point rather than the final word. If your result feels off, retake the photo in daylight, or cross-check with our color analysis quiz, which arrives at your season from a different direction. When two independent methods agree, you can trust the answer; when they disagree, your lighting is usually the culprit. Our practical accuracy guide explains how to repeat the test for dark, Asian, and olive skin without treating skin depth as undertone.
AI Color Analysis FAQ
How does AI color analysis work?
AI color analysis works by passing your selfie to a vision model that estimates three things — undertone (warm vs. cool), depth (light vs. dark), and chroma (muted vs. clear) — while accounting for the photo's lighting. Those three readings are combined and mapped to one of the 12 color seasons, which then defines your personalized palette.
Is AI color analysis accurate?
Our current AI color analysis lands around 80% accuracy on clear, well-lit selfies — a number we reached only after replacing two earlier engines. Accuracy drops with poor lighting, heavy filters, or unusual hair color, which is why we flag those conditions and suggest pairing the scan with our quiz for a second read.
Is AI color analysis as good as an in-person consultant?
A skilled human consultant draping fabric in controlled light is still the gold standard. AI color analysis trades a little of that precision for being free, instant, and available anywhere — a strong starting point rather than a total replacement.
Do you store my photo?
Only if you are signed in. Signed out, the photo is analyzed and not saved; only the season and palette are kept so you can return to them. Signed in, a reduced-size copy (about 600 px) is kept in private storage under your account, so it can sit beside your result and let you compare seasons on your own face. You can remove it from the result page at any time, and it is erased within 30 days. For every analysis we keep a one-way digest of the photo, which cannot be turned back into it and only makes the same file return the same result. Photos are never used to train any model.
See your own AI color analysis
Upload a selfie and the pipeline above runs in about 30 seconds — free, no sign-up, and your photo is not saved unless you sign in.
Want the story behind the tool? Read about Aiyi.