The Just-Noticeable Color Difference: How Small a Gap Can You See?
What a color threshold actually is, why it is not one number, and how a game can measure yours.
Short answer: there is no single just-noticeable difference for color. Under favorable conditions, with the two colors sharing an edge, a difference of about 1 unit of CIEDE2000 is commonly treated as the point where people start to notice; under the older CIE76 formula the average figure is nearer 2.3 units. Both are averages over a space that is not evenly spaced, so your real threshold depends on which color you started from, which direction the difference points, how big the patches are, whether they touch, and what your eyes have been looking at for the last minute.
A threshold is a probability, not a wall
The everyday way of asking the question, "how different do two colors have to be before I can see it," quietly assumes there is a line: below it you see nothing, above it you see the difference. Psychophysics has never worked like that. What actually happens is that your chance of answering correctly climbs gradually as the difference grows, from chance level up towards certainty. A threshold is a point chosen on that curve, usually the difference you get right some agreed proportion of the time.
This matters more than it sounds. It means a single trial tells you almost nothing, because you can be right by luck and wrong by inattention. It means your threshold is a statistic with error bars, so a run that comes out slightly worse than yesterday is probably noise. And it means "I could not see it" and "the difference is not there" are different statements, which is exactly the confusion that makes people argue about whether a gradient is smooth.
MacAdam's ellipses: the finding that will not go away
In 1942 David MacAdam published a study that still shapes how color difference is handled today. He had an observer repeatedly match a test color to a reference, at 25 points spread across the CIE 1931 chromaticity diagram, and recorded how much the matches scattered. Two things came out of it. First, the scatter formed ellipses rather than circles, so sensitivity depended on the direction of the difference, not only its size. Second, the ellipses varied dramatically in size across the diagram: in the green region they were far larger than in the blue region, meaning you can move much further through green before anything looks different.
The uncomfortable conclusion is that the chromaticity diagram, which looks like a map of color, is not a map of perceived difference. Equal distances on it do not mean equal differences to the eye. Every perceptually uniform color space since then, and every color-difference formula, has been an attempt to bend that geometry back into shape. None has fully succeeded, which is why our page on color difference and Delta-E has to keep saying "it depends on the formula".
One detail of MacAdam's method is worth keeping: the two fields he compared shared a border. That is the most sensitive arrangement available to a human observer, and it is not how most real comparisons happen.
Five things that move your threshold
If you take one idea away, make it this: the number you measure is a property of the situation, not just of you.
| Factor | What it does | Where you meet it |
|---|---|---|
| Where you are in color space | Blue-violet regions support far finer discrimination than green regions; the same numeric step is visible in one and invisible in the other. | Some Huekle targets feel much harder than others on the same day. |
| Direction of the difference | Thresholds form ellipses, not circles. A change in lightness, in saturation and in hue are three different jobs for the visual system. | Naming which dimension moved is genuinely part of the skill. |
| Adaptation | Discrimination depends on what your eyes have adapted to; thresholds measured after prolonged exposure to a similar color are not the thresholds you started with. | Long sessions drift. Short ones do not. |
| Shared edge or not | Patches that touch are compared directly. Patches separated in space or time are compared through memory, which adds noise and bias. | Hueseam versus Hueecho: the same difference, two very different tasks. |
| Patch size and duration | Bigger and longer is easier, up to a point. Tiny swatches and quick flashes raise thresholds sharply. | Phone screens are not laptop screens. |
What the formulas are trying to do
A color-difference formula exists so that one number can stand in for "how different do these look". The 1976 CIELAB formula, usually written CIE76, simply measures the straight-line distance in the CIELAB space. It is easy to compute and it is wrong in a predictable way, because CIELAB is not as uniform as it was hoped to be. Work evaluating the uniform color spaces proposed after CIELAB put the average threshold in that space at roughly 2.3 units, which is where the widely repeated figure comes from. Treat it as an average, not a constant.
CIEDE2000 is the current recommendation, and it does something slightly awkward: rather than fixing the space, it patches the distance formula with weighting terms for lightness, chroma and hue, plus an extra rotation term to handle the blue region where CIELAB misbehaves worst. It matches visual data better. It is also fiddly enough that a published paper exists purely to help engineers implement it without introducing errors, which tells you something about how much correction the underlying space needs.
Measuring your own threshold without fooling yourself
You cannot find your threshold by looking at two colors and deciding whether you can see the difference, because you will be guided by what you expect. The standard method is an adaptive staircase: the difficulty follows your answers, getting harder when you are right and easier when you are wrong, so the trials cluster around the region where you are uncertain. Rules of this kind, where a fixed number of correct answers is needed before the task gets harder, have been in use since Harry Levitt formalised them in 1971, and they converge on a specific point of the psychometric curve rather than on your own opinion of yourself.
Huefork works this way. It is a two-alternative task with a staircase behind it, and it reports a threshold rather than a score out of ten. Hueseam attacks the same sensitivity from the other side: instead of asking whether two patches differ, it hides a step inside a bar that looks continuous and asks you to locate it. That version leans on the visual system's edge enhancement, which is a different mechanism and comes with its own illusions, covered in Mach bands and banding.
Three practical rules if you want your numbers to mean anything across weeks. Use the same device and roughly the same room light, because the display's transfer function and your adaptation state both shift the measurement. Run enough trials that the staircase settles, rather than stopping at the first good result. And read the trend across a fortnight rather than any single day, as our guide to recording and evaluating your scores sets out.
The screen is part of the experiment
An ordinary 8-bit channel offers 256 steps, and those steps are not evenly spaced in perceived lightness. In dark tones neighboring code values can be close to, or even above, a visible difference, so the smallest change your display can render may be larger than the smallest change you could detect on better hardware. In bright tones the opposite happens and several code values may be perceptually identical. That is why a threshold measured in a game is a threshold for you on that screen, and why a friend's number on a different laptop is not comparable to yours.
It is also why we do not publish leaderboards for threshold scores. The number is real and useful to you over time; it is not a rank.
When a low threshold is not the point
Fine discrimination is one skill among several. Being able to see a small difference does not mean you can name it, reproduce it from memory, or judge it consistently under changing light, and those are separate abilities handled on other pages here: color working memory for reproduction, color constancy for stability under changing illumination. If your threshold is persistently very high in one particular direction, especially along the red-green axis, that is worth an eye examination rather than more practice. Our page on color vision differences explains why.
Find your own number. Huefork runs an adaptive staircase on lightness differences and hands you a personal threshold, not a score out of ten. It takes about two minutes, and it is the same task every day so the trend means something.
Frequently asked questions
What is the smallest color difference a person can see?
There is no single number. Under good conditions, with the two colors sharing an edge, a difference of roughly 1 unit of CIEDE2000 is often treated as the point where a difference becomes noticeable, while the older CIE76 formula puts the average threshold nearer 2.3 units. The real threshold changes with the color, the direction of the change, the size of the patches and the observer.
Why do people quote 2.3 as the just-noticeable difference?
That figure comes from work evaluating how uniform CIELAB actually is, where the average threshold in the 1976 formula worked out at about 2.3 units. It is an average across color space, not a constant, and CIELAB is uneven enough that the true threshold in one region can be several times the threshold in another.
What are MacAdam ellipses?
In 1942 David MacAdam measured how much an observer's repeated color matches scattered around 25 reference points on the chromaticity diagram. The scatter formed ellipses rather than circles, and the ellipses differed enormously in size and orientation across the diagram. That is the classic demonstration that the 1931 chromaticity diagram is not perceptually even.
Does the direction of the difference matter, not just the size?
Yes. Two colors the same distance apart in a color space can be easy or hard to tell apart depending on which way the difference points, and on what you have been looking at beforehand. Discrimination thresholds shift with the observer's state of adaptation, so the same pair can be easier after a rest than after staring at a similar color.
Why is a difference easier to see when the two colors touch?
Because the visual system is built to compare across edges. When the patches share a border, the comparison happens locally; separate them by a gap or a few seconds and you are now comparing against a memory, which is both noisier and biased. This is why Hueseam, where the step sits inside one continuous bar, can be harder or easier than it looks depending on where the seam falls.
How does Huefork find my threshold?
By moving the difficulty up and down in response to your answers rather than running a fixed list. Adaptive staircase procedures of that kind have been standard in psychophysics since the 1970s, and they converge on the difference you get right a fixed proportion of the time instead of asking you to judge your own limit.
Does my monitor change my threshold?
It changes the measurement. An 8-bit channel gives 256 steps, so the smallest difference your screen can show may already be close to the smallest difference you can see, especially in dark tones. Automatic warm-tint modes and ambient light shift things further, so keep the device and the lighting the same if you want your numbers to be comparable.
Is a threshold a hard wall?
No. It is a probability. Just below your threshold you are not blind to the difference, you are just right less often than chance plus a margin. That is why a single trial tells you almost nothing and a run of thirty tells you something.
Keep reading
- Color difference and Delta-E: the formulas that try to turn a threshold into one number.
- Mach bands and banding: why your eye finds edges that the signal does not contain.
- Gamma and perceptual uniformity: why equal steps in code are not equal steps in appearance.
Sources: MacAdam, D. L. (1942), Visual sensitivities to color differences in daylight, Journal of the Optical Society of America 32(5), 247; Mahy, M., Van Eycken, L. and Oosterlinck, A. (1994), Evaluation of uniform color spaces developed after the adoption of CIELAB and CIELUV, Color Research & Application 19(2), 105; Sharma, G., Wu, W. and Dalal, E. N. (2005), The CIEDE2000 color-difference formula: implementation notes, supplementary test data, and mathematical observations, Color Research & Application 30(1), 21; Krauskopf, J. and Gegenfurtner, K. (1992), Color discrimination and adaptation, Vision Research 32(11), 2165; Levitt, H. (1971), Transformed up-down methods in psychoacoustics, Journal of the Acoustical Society of America 49(2B), 467; International Commission on Illumination, CIE 015:2018 Colorimetry, 4th edition.
Understand it, now train it. The daily games stay free forever. A Personal Pass adds progress tracking (trends + export); Teacher / School tools let you assign it and follow a whole class. For teachers & self-learners →