Color analysis quiz › Accuracy
How accurate is this color analysis?
There is no single number called “color analysis accuracy”. There are at least three different questions, and this page answers each one separately.
Three meanings of accuracy
| Question | What it would take to answer | What we can say today |
|---|---|---|
| Does the label agree with a professional analyst? | A sample of people with an in-person result, the same photo protocol, and a blinded comparison. | We have not published this yet. A 50-person pilot is planned, not yet recruiting. |
| Does the same person get the same label twice? | Repeated quizzes or photos on different days, with the same inputs. | The Pro report includes three retests and a consensus; we do not yet publish a reliability rate. |
| Does wearing the palette help? | A blinded rating of faces in matched vs mismatched colours. | We found no published experiment that directly tests this for the 12-season labels. |
What the published evidence says
- Ou & Chan (2026) report 55.2–55.6% accuracy on the Deep Armocromia face-image benchmark, with Spring recall below 0.42. That is a model reproducing labels in a dataset, not a validation of the labels themselves.
- Kawase et al. (2026) report 65% accuracy on a four-season task against a 25% chance level, using influencer-labelled lipstick colours, with heavy overlap between seasons in the a*–b* plane.
- We found no peer-reviewed study that validates the 12-season taxonomy or reports how often two trained analysts agree. The science page sets out the detail.
The pilot we plan to run
Planned design (not yet recruiting): 50 adults, in two waves, who had a professional in-person colour analysis in the previous 12 months. Each participant submits three daylight photos and the season they were given. We run the photo check and the quiz without seeing the professional label, then compare.
Stratification: skin tone (Monk scale), lighting quality (good / mixed / poor), device type and age band. The point is not just an overall percentage; it is to see where the method fails. If it performs worse on deep or olive skin, we will say so and show the subgroup numbers.
What we will publish: agreement rate, a confusion matrix, test–retest agreement, subgroup results, the protocol and the limitations. If the result is bad, we will publish that too. The pilot is not a marketing study; it is a reason to trust or distrust the tool.
Status: planned, not yet recruiting. No results have been published yet. This page will carry the date and the sample size when they are.
When you should not trust this tool
- High-stakes decisions: a wedding dress, a uniform, a brand palette, a headshot you will use for years. Use fabric or an in-person analyst.
- Deep or olive skin: the warm/cool axis is the least reliable part of the model on these skin tones. Treat the result as a hypothesis, not a verdict.
- Makeup, filters or dyed hair: the photo check reads the pixels you give it, not the person underneath.
- A close result: if the top two seasons are within a few points, you are on a boundary. Compare both palettes in daylight; do not buy a whole wardrobe on the top label.
What to do instead
Take the quiz and the photo check on different days. If they agree, use the palette as a starting point. If they disagree, read the method page and compare the two neighbouring seasons on the 12-season grid. For a professional second opinion, the online vs in-person guide explains what a human adds.
Try the photo check Read the method
How this page was checked
Sources and dates are stated above where a claim is made. Method pages are checked against the code that ships here; competitor prices are checked against the provider’s public page on the date shown. Corrections: billing@seasoncolortest.com.