The Farnsworth-Munsell 100 hue test is the color vision test that looks like a puzzle. You get four trays of small colored caps, the caps at each end of a tray are fixed in place, and everything between them arrives out of order. Your job is to put them back into a smooth sequence. No plates to read, no letters, no naming anything. Just the question of whether you can see which of two nearly identical caps belongs closer to the orange end.
Dean Farnsworth published it in 1943 and it has barely changed since. It is still the reference test for graded color discrimination, used in ophthalmology clinics, in paint and textile quality control, and as a screen for jobs where getting a shade slightly wrong is expensive.
Browser versions of it are everywhere, including a hue sorting mode on this site. Most of them are honest about being an approximation. What none of them tell you is the specific way the approximation breaks, which is the thing I measured for this piece. The short version: if you generate the caps by stepping evenly around the sRGB hue wheel, the difficulty of the test stops being about your eyes and starts being about which part of the wheel you happened to land on.
What is actually in the box
The physical kit has 85 movable caps and 8 fixed anchor caps, 93 in total, split across four trays. Each tray holds 21 removable caps between its two anchors, except the first, which holds 22. Laid end to end the trays walk once around the full hue circle, through the reds, yellows, greens, blues and the purples that close the loop.
The name is a leftover. Farnsworth designed the thing around a hundred hue steps and the pigments available could not hold a hundred steps apart, so the shipped test has 85 movable caps and a name that describes the intention rather than the contents. People still call it the 100 hue test, or the hue test, or the Munsell hue test, and they all mean this.
The caps are Munsell chips. That matters more than it sounds. The Munsell system was built by measuring people, not by dividing a wheel into equal arithmetic slices. Albert Munsell asked observers which pairs looked equally far apart and spaced his hues accordingly, which is why a step from one Munsell hue to the next is meant to feel like the same size step wherever you take it. Farnsworth chose those chips precisely because the difficulty is even all the way around.
How the score works, and why a low score is not zero
Scoring is mechanical. For each cap you take the difference between its color number and the number of the cap on its left, add the difference against the cap on its right, and subtract 2. A cap sitting in its correct position scores 0. Swap one pair of neighbors and four caps have their score bumped by 1 each, so a single transposition costs 4. Add up every cap and you have the total error score.
The published norms are wider than most people expect. Roughly 16 percent of people score 0 to 16, about 68 percent land between 16 and 100, and about 16 percent score above 100. A perfect zero is unusual rather than impressive, and near-perfect is common. The adjacent caps are sitting close to the threshold of perception by design, so a normal observer with normal vision transposes a few caps and that is the expected result, not a warning sign.
Age moves the number more than most people realize. Kinnear and Sahraie collected fresh norms across every year from 5 to 22 and by decade from 30 to 70, and found performance follows a U shape with the low point, meaning the best scores, at around 19 years of age. Children do worse. So does everyone over about 40, steadily. If you take a browser hue test at 55 and score worse than your colleague at 25, the test is working correctly.
Where the errors fall matters more than how many there are. Cluster them in one part of the circle and you have an axis of confusion, which is what the test is really for. Scatter them evenly and you have a normal observer having an ordinary day. For normal observers, the mistakes tend to concentrate in the greens.
The measurement: what an even hue step actually costs
Here is the part nobody publishes. A browser version has no Munsell chips. It has sRGB, and the obvious way to build 85 caps in sRGB is to take the HSV hue wheel, fix saturation and value, and step hue by 360 divided by 85, which is 4.235 degrees per cap.
So I asked what that step is worth perceptually. I walked the full saturation hue circle in quarter-degree increments, converted each pair of colors 4.235 degrees apart to CIELAB under D65, and ran them through CIEDE2000, the current industry standard for how different two colors look to a person. It is the same metric this site uses to score every guess. Under CIEDE2000, a difference of about 2.3 is the classic just noticeable difference, the point where an ordinary observer starts to see two colors as two colors.
The results are not close to even:
- Mean difference for one cap step: 3.26. That part is reasonable, and it is almost exactly the 3.27 you would get if the circle were spaced evenly.
- Largest step anywhere on the circle: 7.33, at hue 182, in the cyans.
- Smallest: 0.15, at hue 116, in the yellow-greens.
- Ratio between them: 50 to 1.
- Share of the circle where one cap step falls below the 2.3 threshold: 34 percent.
A third of the wheel produces caps that are, by the standard measure, not distinguishable from their own neighbors. Two concrete pairs, both exactly one cap step apart:
- #15ff00 and #03ff00, at hue 115. Difference 0.16. These are the same green.
- #00f7ff and #00e4ff, at hue 182. Difference 7.33. Nobody would put these in the wrong order.
Put the same question the other way around and ask how far you have to travel from a starting hue before the difference becomes visible at all:
- From hue 115, yellow-green: 19.0 degrees, which is 4.5 caps.
- From hue 240, blue: 14.8 degrees, 3.5 caps.
- From hue 0, red: 9.3 degrees, 2.2 caps.
- From hue 210, azure: 1.45 degrees, 0.3 of a cap.
- From hue 180, cyan: 1.55 degrees, 0.4 of a cap.
Thirteen times the distance in one region compared to another, for the same nominal step. In the yellow-greens you would have to slide a cap four and a half places out of position before the error was even visible to the person being tested. In the cyans, a third of one cap width is enough.
What this does to a test
A hue test built this way is not measuring one thing. In the cyans and the oranges it is an easy sorting task that almost everyone passes. In the yellow-greens and the deep blues it is close to a coin flip, regardless of how good your color vision is, because the information needed to answer correctly is not on the screen.
That produces a specific and misleading failure mode. Errors pile up in the greens, which is exactly where clinical FM100 errors pile up for normal observers, so the output pattern looks plausible. It looks like a real result. It is a property of the sRGB coordinate system.
Saturation makes it worse, and browser tests love a soft palette. Holding value at maximum and dropping HSV saturation:
- S = 1.00: mean 3.26, and 34 percent of the circle below threshold.
- S = 0.60: mean 2.62, 50 percent below.
- S = 0.30: mean 1.71, 78 percent below.
- S = 0.20: mean 1.37, 89 percent below.
A pastel hue test with evenly spaced HSV caps is close to unanswerable across nine tenths of its range. If you have ever taken one of these, scored badly, and felt the caps genuinely looked identical, you were probably right.
Which browser tests are affected
I want to be careful here, because I cannot see how any particular site generates its caps. Some online hue tests are built from measured approximations of the real Munsell chips and will be spaced roughly evenly. X-Rite, who own Munsell, run the best known one and have every reason to use the real values. The critique lands on the generic case: any test that generates caps by stepping HSV hue at a fixed interval, which is the simplest way to build one and the way most weekend projects do it.
You can check a test yourself in about a minute:
- Screenshot the tray, pick two adjacent caps and read their hex codes with any color picker.
- If the two hex codes differ only in one channel by a small constant, and the same constant shows up between the next pair along, it is an even HSV walk.
- Compare a pair from the green tray against a pair from the blue tray. If the green pair looks identical and the blue pair looks obviously different, the test is measuring the palette, not you.
There is also a display problem underneath all of this. The clinical test is run under a specified illuminant with the caps at a controlled angle. Your monitor is not calibrated, your brightness is set to whatever, and a warm-tinted night mode will quietly compress exactly the hues the test cares about. Any browser hue test is a rough indicator at best. If you have a real concern about your color vision, an optometrist can run the physical version, and an Ishihara style screen is a better first filter for red-green deficiency anyway, since it is designed to survive imperfect viewing conditions.
What our own hue sort does instead
Since I was pointing at other people's tests, I ran the same measurement on ours. The Hue Sort mode generates each round by picking two anchor colors, interpolating between them, and shuffling the middle. I generated 100,000 rounds at each of the five difficulty levels and measured every adjacent chip gap:
- Round 1, 4 chips: mean gap 22.4, none below threshold.
- Round 3, 6 chips: mean gap 13.9, 0.1 percent below.
- Round 5, 8 chips: mean gap 10.1, 0.8 percent below.
Every gap is three to seven times larger than a real cap step. That is a deliberate design choice and it means something specific: our hue sort is not a color vision test. The chips are comfortably distinguishable, so it does not measure discrimination threshold. What it measures is whether you can hold a sequence in order and reason about where a color belongs, which is a different and more game-shaped skill. Calling it a Farnsworth-Munsell test would be a stretch, and I would rather say so than let the comparison do marketing work it has not earned.
The reason the gaps stay wide is that the generator interpolates between two anchors that are 60 to 200 degrees apart and then splits that span across at most 8 chips. Spread a wide arc over few chips and every step is large, wherever on the wheel it sits. The non-uniformity of the sRGB hue coordinate is still there, it just gets swamped.
If you want to actually train this
Hue ordering is a narrow skill and it does respond to practice, though not in the way most people assume. What improves is your strategy, not your retina. Better sorters stop comparing caps one at a time and start working from the anchors inward, placing the obvious extremes first and binary searching the ambiguous middle. That is worth several points of error score and it is learnable in an afternoon.
Your actual discrimination threshold barely moves. This is the same conclusion I reached looking at whether brain games work: you get better at the task, and the improvement mostly stays inside the task. It is still worth doing if you enjoy it, and it is genuinely useful if your job involves judging shades, because the strategy part transfers even when the perception part does not.
If you want the harder version of the same idea, color memory is a separate skill from color discrimination and a much weaker one in most people. Holding a hue in your head for ten seconds costs you far more accuracy than telling two hues apart side by side.
Where to go from here
- Hue Sort, the arrangement game described above. Five rounds, chips get more numerous as you go.
- The color blind test, an Ishihara style screen, which is the better first check for a red-green deficiency.
- The eye test for acuity and contrast, which are separate from hue discrimination and often the real culprit.
- The tetrachromacy question, since a very low hue test score is the claim most online tetrachromacy tests are built on, and it does not support the weight they put on it.
- How many colors can you see, which is where the ten million figure comes from and why it is softer than it sounds.
Method
sRGB to CIELAB under D65 using the proper sRGB transfer function rather than a gamma 2.2 approximation, then CIEDE2000 with all three weighting factors set to 1, implemented from Sharma, Wu and Dalal. The implementation was checked against known values before use: #ffa500 returns lightness 74.9, chroma 82.5, hue angle 73.1; #ff0000 returns hue angle 40.0; #ffff00 returns 102.9. The circle sweep steps hue in 0.25 degree increments at full saturation and value, comparing each hue against the hue 4.235 degrees ahead of it. The hue sort figures come from replaying the shipped round generator 100,000 times per difficulty level and measuring every adjacent pair. Cap and tray counts describe the current physical kit. Norm figures are the published ranges for the clinical test and do not transfer to browser versions.