What color is the dress: the answer, and the measurement behind it

The dress is blue and black. It is a real garment, it was sold by a British retailer called Roman Originals, and the manufacturer confirmed the colors while the argument was still going on in 2015. If that is all you came for, you can stop here.

The part worth staying for is why so many people looked at that photograph and saw white and gold instead, and why they were not being stubborn or dishonest. Most explanations stop at “your brain assumes different lighting”, which is true and which explains nothing, because it does not tell you how common that kind of ambiguity is or why the argument landed on blue against gold rather than on any other pair of colors.

So I measured it. I took every color an sRGB screen can display, sampled on a grid of 636,056 of them, and asked a simple question of each one: if you assume this pixel is lit by a warm indoor bulb, what surface color does that imply, and if you assume it is lit by cool blue shade instead, what surface color does that imply? Four things fell out, and the last one is the actual answer to why it was blue and gold.

How the measurement works

When you look at an object, the light reaching your eye is the color of the object multiplied by the color of the light falling on it. Your visual system wants the first of those and has to guess the second. That guessing is color constancy, and it is the reason a banana stays yellow under a sunset and under a fluorescent tube.

The standard way to model it is a chromatic adaptation transform. I used CAT02, the one inside CIECAM02, which converts a color into a cone-response space, divides out the assumed illuminant, and multiplies back in a reference illuminant. Run a pixel through it with a warm assumed light and you get the surface color a warm-light observer would report. Run the same pixel through with a cool assumed light and you get a different surface color. Neither output is wrong. They are answers to different questions.

For the two assumptions I used CIE illuminant A at 2856 K, which is an ordinary tungsten bulb, and an 18000 K daylight illuminant, which is the blue light of open shade under a clear sky. Both are ordinary. Neither is exotic. The lighting figures come from the Planckian locus and the CIE daylight locus, the same machinery behind our color temperature chart. I counted a color as decisively flipped only when it read as clearly blue under one assumption and clearly gold under the other, with a dead band in the middle so that near-neutral colors drifting slightly either way do not get counted.

Finding one: the dress is not a freak pixel

37.20 percent of all sRGB colors flip between decisively blue and decisively gold depending only on which of those two lights you assume. 236,636 of the 636,056 colors tested. More than a third of everything your monitor can display is dress-ambiguous.

That number reframes the whole phenomenon. The usual telling treats the photo as a freak, a one in a million capture that happened to land on a knife edge. It did not. Ambiguity on this scale is the normal condition of color, and what is unusual about ordinary photographs is not that they avoid the trap but that they hand you enough context to escape it. A patch of sky, a white shirt cuff, a shadow with a visible edge: any of those pins down the illuminant and collapses the ambiguity before you are aware there was one. The dress photograph was cropped so tightly, and blown out so badly, that it offered almost none of that.

Here is one color from the middle of the ambiguous set, sitting almost exactly on the knife edge.

Nothing about that pixel is unusual. It is a drab pale green of the sort that turns up in any photograph. The disagreement is not in the pixel. It is in the question you did not know you were answering.

Finding two: pale washed-out colors are the worst offenders

Splitting the results by lightness produced the cleanest gradient in the whole exercise. The darker a color is, the more likely both observers agree about it. The paler it is, the more likely they split.

From 10.6 percent in the near-blacks to 61.7 percent in the near-whites. The colors that flip are pale and washed out as a group: mean lightness 64.6 against 53.4 for everything else, and mean chroma 39.7 against 68.6. Ambiguous colors are light and weakly saturated. Confident colors are dark and strong.

Which is a precise description of an overexposed phone photo. Overexposure does two things at once. It pushes everything up in lightness and it crushes chroma, because channels clipping at the top of their range converge toward each other. Both of those moves march the image directly into the region of color space where the flip rate is highest. The dress photo was taken in poor light on a phone camera, and the result was not merely a bad photo of a dress. It was a bad photo in exactly the direction that maximises disagreement.

The same logic runs the other way, and it is useful. If you ever need someone to see a color the way you see it, do not send them a bright washed-out crop. Send them something dark and saturated with a reference white in the frame. That is not an aesthetic preference. It is a four-to-one difference in the odds of being understood.

Finding three: color space itself is lopsided

Now the question that most write-ups skip. Why blue against gold? Why did nobody claim the dress was green and pink?

My first guess was that the blue-yellow direction is special in the structure of color space, so I tested it. I built a second pair of illuminants that differ along the green-magenta axis instead, sitting either side of the daylight locus at 5500 K, and I tuned their separation until the two white points were exactly as far apart perceptually as the warm and cool pair, 46.98 CIEDE2000 in both cases. Same perceptual gap, different direction. Then I ran the same sweep.

• Blue against gold: 37.20 percent of colors flip.
• Green against magenta: 14.91 percent of colors flip.

So there is a real asymmetry, and it is a factor of 2.50. Colors are two and a half times more likely to be ambiguous along the blue-yellow axis than along the green-magenta one, for lighting changes of identical perceptual size. Some of that is the shape of the sRGB gamut, which is long in the blue-yellow direction and comparatively short across it, and some of it is the way lightness and the blue-yellow axis interact in CIELAB.

A factor of 2.5 is worth something. It is not enough to explain a clean split of the entire internet into two camps. So I looked at the other half of the problem.

Finding four: real light barely moves on the other axis

A chromatic adaptation model will happily assume any illuminant you hand it. Your visual system will not. It has a prior, learned from a lifetime of daylight and firelight and bulbs, about which lights are plausible. So I measured how far real light actually travels on each of the two axes.

Along the blue-yellow direction, natural and domestic lighting runs from roughly 1900 K, a candle flame, up to about 20000 K, the blue of a clear northern sky. Walking that whole range along the Planckian and daylight loci covers an arc of 0.1699 in CIE 1960 uv coordinates.

Along the green-magenta direction, real lights hardly move at all. That direction is the Duv axis, the distance a light sits off the locus, and the ANSI C78.377 binning standard that lamp manufacturers work to allows a tolerance of about plus or minus 0.006. Total spread 0.0120.

Real light varies 14.2 times further on the blue-yellow axis than on the green-magenta one. Multiply that by the 2.5 times greater ambiguity of the blue-yellow axis and the puzzle dissolves. Blue against gold is not the argument people happened to have. It is the only argument the physics of ordinary lighting leaves available. A green-magenta dress could exist in principle, and 14.91 percent of colors would support it, but no one has ever stood under a light far enough off the locus to develop the habit of guessing that way.

There is a corollary I like. Sodium street lamps sit a long way off the locus, and the one place people reliably report bizarre unresolvable color arguments in real life is a car park at night under orange sodium light. The prior fails there because the light is somewhere the prior was never trained.

Where the tipping point sits

One more measurement, because it explains why the split was roughly even rather than ninety-ten. For every pale, low-chroma color that flips at all, I solved for the exact assumed color temperature at which its reading crosses from gold to blue. Across 116,526 such colors:

• 10th percentile tipping point: 4798 K
• Median tipping point: 6570 K
• 90th percentile tipping point: 10098 K

The median sits at 6570 K, which is average daylight, give or take. The tipping point is not tucked away at the end of the range where few people would land. It is parked in the exact middle of the lighting people spend their lives under. And the middle 80 percent of tipping points span 5300 K, which is wide enough that the answer depends on a judgement rather than a fact.

Watch a single pixel walk across it. This is #abb090 again, read under a series of assumed lights:

The crossing happens between 4500 K and 5500 K, and around there a thousand kelvin of error in the estimate moves the reading by 7.2 CIEDE2000. For scale, a just noticeable difference is around 2.3. So a misjudgement most people would never notice making, of the order of the difference between imagining a room lit by a window and imagining it lit by a ceiling fitting, is worth three visible steps of color. That is the whole illusion in one number.

Why you cannot flip it at will

Almost nobody can consciously switch their reading of the dress, and a fair number of people never see the other version at all. That is worth a word, because it is the part that makes the argument feel personal.

The illuminant estimate is not a belief you hold. It is settled early in visual processing, before anything you would recognise as thought, and it is committed rather than provisional. Being told the correct answer updates your knowledge and leaves your perception exactly where it was. This is the same stubbornness you meet in afterimages, which keep drifting across the wall no matter how firmly you know the wall is blank, and it is a cousin of the effect behind metamerism, where two surfaces match under one light and separate under another. In all three cases the visual system has made a call on your behalf and does not offer an appeal.

Researchers who studied the split found it correlated loosely with chronotype, with people who wake early leaning toward white and gold and night owls leaning toward blue and black. The proposed reason is exposure: a life spent mostly under daylight builds a different prior from a life spent mostly under artificial light, and the prior is what does the work here. That result has been debated since, and I would not lean on it hard, but the mechanism it points at is the right one. What you see is downstream of what you have spent years looking at.

What the numbers add up to

• The dress is blue and black. The photograph is genuinely ambiguous anyway, and both of those statements are true at once.
• 37.20 percent of all sRGB colors flip between decisively blue and decisively gold on nothing more than a change of assumed lighting.
• The flip rate climbs from 10.6 percent in the darkest colors to 61.7 percent in the palest. Overexposure moves an image into the worst zone on both axes at once.
• For lighting changes of equal perceptual size, blue-gold ambiguity is 2.50 times more common than green-magenta ambiguity.
• Real light varies 14.2 times further along blue-yellow than along green-magenta, so the second axis never gets exercised.
• The median tipping point sits at 6570 K, in the middle of everyday light, and a 1000 K error near the crossing is worth 7.2 CIEDE2000, about three just noticeable differences.

The lesson I take from it is that color is a conclusion rather than a measurement, and conclusions can differ on identical evidence. That is also, incidentally, why matching a color from memory is so much harder than it looks: you are not recalling a number, you are re-running an inference with the context stripped out. If you want to find out how much of it you can actually do, the color memory game is the five minute version, our tour of color illusions covers the neighbours of this one, and training your eye for color is the long game. None of them will let you see the dress both ways. Not much will.