How accurate is color analysis from one photo?
We tested our own photo analysis on 410 celebrity photos, each of a person whose season every published source agrees on. From one press photo, the exact season was right 11.2% of the time, and the best model we could train on the same measurements reached 21.1%. Here are the numbers, what went wrong, and why one photo can only ever be a first estimate.
By Ruslan ShchuchkinML engineer and data scientistPublished
The short answer
From a single photo, a color season is a guess with a measurable error, for us and for everyone. Our measurement gets the exact 12-season answer right 11.2% of the time on these photos and the family (spring, summer, autumn, winter) right 30.7% of the time. A model trained on the same measurements gets to 21.1% and 38.8%. A random guess would score 8.3% and 25%. The gap between a photo tool and a draped session is real, and the reason is not the maths. It is the light.
Every photo colour-analysis tool shows you a season with the same confident screen. Ask any of them for their accuracy number. This page is ours.
What we tested
- 1
An agreed label for each person
We compiled 5,422 published celebrity season classifications from 62 sites and kept the 473 people every source agrees on. Sources disagree about the rest half the time; that study is its own page.
- 2
One public photo each
The lead photo from each person's Wikipedia article, 422 in all. Press photos: flash, stage light, makeup. Not the daylight selfie the analysis asks for. 12 had no readable face and were dropped.
- 3
The exact production reading
Each photo went through the same code a visitor's photo does: 468 face landmarks, six skin patches, hair, irises, white balance from the eye whites, then the nearest of 12 seasons. Then we trained statistical models on those measurements to see how much better the best possible reading of one photo could be.
The results
| Reading | Exact season | In the top 3 | Right family |
|---|---|---|---|
| chance | 8.3% | 25% | 25% |
| majority class | 12.9% | — | 26.8% |
| rule-based engine (production) | 11.2% | 34.1% | 30.7% |
| random forest · tabular | 20.2% | 50.8% | 42.4% |
| MLP neural net 64-32 · tabular | 15.9% | 34.5% | 36.4% |
| MLP neural net 128-64 · tabular | 13.6% | 36.8% | 41.5% |
| PCA-64 + MLP neural net · image (ResNet-50 embedding) | 8.6% | 26.7% | 31.1% |
| random forest · image (ResNet-50 embedding) | 20.2% | 47% | 36.3% |
| PCA-64 + MLP neural net · both | 11.5% | 26.6% | 30.7% |
| random forest · both | 21.1% | 50.6% | 38.8% |
Model figures are 5-fold cross-validation repeated three times, so they are out-of-sample. Neural networks were tried on the measurements and on a pretrained image embedding of each face; with about 34 examples per season they did worse than the simpler models, so nothing learned is in the product.
Where one photo fails
The production reading, per agreed season: how often the exact season and the family were right.
- Light Summer
- 38 photos · exact season 39% · right family 82%
- True Summer
- 24 photos · exact season 29% · right family 63%
- Soft Summer
- 31 photos · exact season 13% · right family 52%
- True Winter
- 31 photos · exact season 13% · right family 42%
- Deep Winter
- 39 photos · exact season 13% · right family 23%
- True Autumn
- 16 photos · exact season 13% · right family 31%
- Soft Autumn
- 41 photos · exact season 12% · right family 15%
- Deep Autumn
- 53 photos · exact season 6% · right family 28%
- Bright Winter
- 31 photos · exact season 3% · right family 42%
- Light Spring
- 35 photos · exact season 0% · right family 3%
- True Spring
- 41 photos · exact season 0% · right family 2%
- Bright Spring
- 30 photos · exact season 0% · right family 3%
The pattern is one failure, not twelve. On average every family read cool: the mean temperature reading was -0.45 for spring and -0.31 for autumn, where warm is positive. Springs and autumns, who are warm, landed in True Summer. The depth reading held up: the two deep seasons were the only ones the model learned well, because how light or dark you are survives a flash. Undertone does not.
We checked whether photo quality explained it. Keeping only photos with readable eye whites, readable hair and a readable iris changed nothing. The problem is not photos we can filter out. It is that a red-carpet photo does not contain the undertone information at all.
What we do about it
The reading treats one photo as a first estimate, and says so. The report shows the runner-up season and the confidence gap between the two, never a lone verdict. You can add up to 2 photos free (more in the full report); every photo votes on each axis and the reading keeps the median, so the cast of any single lamp or window cancels against the others. When a new photo changes the season, the report announces it rather than switching silently.
We have not yet measured how much the second and third photo improve the number above, and we will not claim a figure until we have. What we can say is what the test shows: the light is the error, and the only fix for one light is another.
What this test does not prove
- Press photos, not selfies. Flash, stage light, retouching, makeup. The analysis asks for a bare face in daylight. The numbers are a floor, not the product’s accuracy.
- Published labels, not drape sessions. “Agreed” means every blog and tool that named a season named the same one. Sites copy each other.
- One photo only. The multi-photo reading the product uses was not scored here.
The method is in how our analysis works, and the disagreement study in celebrity color analysis: do the experts agree?.
Frequently asked
Is color analysis accurate?
In-person draping by a trained analyst is the standard, and even there analysts disagree with each other. From a photo, accuracy depends on the photo far more than on the method. In our test on 410 celebrity press photos, the exact 12-season answer was right 11.2% of the time from our measurement and 21.1% from the best statistical model we could train, against 8.3% for a random guess. On the broader question, warm or cool and light or deep, the best model was right 38.8% of the time. One lit, filtered, made-up photo does not show undertone, so no photo tool should claim a confident season from one.
How accurate is AI color analysis?
No AI colour-analysis tool we know of publishes an accuracy number, so the honest answer is that nobody outside those companies knows. Ours, measured on press photos of celebrities whose season every published source agrees on: 11.2% exact season, 30.7% correct family, from a single photo. We publish it because a number you can check is worth more than a confidence you cannot.
Why is one photo not enough for color analysis?
Because the axis that decides your season, warm versus cool undertone, is the one a camera's white balance shifts most. In our test every family read cool on average: springs and autumns, who are warm, came out as True Summer. The measurement is fine; the light lied. A second photo taken in different light gives a second vote, and the reading keeps the answer the photos agree on.
What does the test not prove?
It does not measure the product as used. These are red-carpet photos, lit by flash and stage light, with professional makeup, not the bare-faced daylight selfie the analysis asks for. The labels are what bloggers and tools published, not drape sessions. And the test scores one photo; it does not yet measure how much a second and third photo improve the reading. Treat the numbers as a floor.