Kanwisher Lab

What is the deal with this goddamned Dress?
Rosa Lafer-Sousa
3/2/2015

This week a photograph of a dress was uploaded to tumblr and it broke the internet. Fights erupted across the globe. The dress received record views in a matter of hours. Why? Because this particular picture (shown below) induces two distinct percepts in people. Some folks see a blue and black dress (or some variation, like blue-brown) others see white and gold. Many explanations have been floated, but it has become increasingly clear that the phenomenon has to do with color constancy, as suggested early in the debate by Bevil Conway here and here.

Most of the time, our visual system does a remarkable job of inferring the ambient lighting conditions at any given time and discounting their contribution to color-computations. The result is that most objects retain their color under varying illumination conditions (bright daylight, shadow, tungsten light, fluorescent light, candle light, etc.). The visual system relies on many visual cues to do this (scene structure, the relative reflectances of objects in the room, their other surface properties, direction of illumination, time of day(?), etc) as well as prior expectations about illuminants and how they behave. But in this image, the cues to the lighting conditions are particularly ambiguous. If you look at the original image (below) you will see that there is a very bright light source in the background. It may be the case that white-gold people interpret this light as backlighting the dress, casting the front of the dress in relative shadow (casting a blueish biased light). The rest of us (I am team blue-black) may be taking that same bright light as a cue that the whole room is very brightly lit, and everything in the space is brightly illuminated (casting a yellowish light on the dress). Is the light illuminating the dress bright and yellowish or is it dim and blueish? Your brain has to make a guess. Some folks infer a blueish biased illuminant and others a yellowish (one end of the daylight spectrum or the other; see explanation of the daylight axis at bottom of page). If your visual system infers blueish ambient light, it will remove some of the blue from the dress, yielding a white-gold percept. Similarly, if your visual system infers yellowish light, it will remove some of the yellow from the dress yielding a blue-black percept (see the Wired article to see what happens when you simulate this in photo editing software).

The original image:

To confirm this hypothesis, I whipped up a demo using Beau Lotto's famous rubik's cube. The illusion puts the dress in two contexts that convey clear cues to either a blue or yellow illuminant. This should allow viewers to see BOTH percepts (I put the image up in an informal poll on Facebook for a night and 94% of my first 130 responders said they were able to see both percepts, many of them for the first time!). This strongly suggests the effect (that people see two different percepts) is a product of differences in illumination estimates rather than something like individual differences in retinal cone-mosaics (as has been suggested by some).

In both panels of the image the dress itself is IDENTICAL (it is the original image!).

The background and the model's skin tone have been set to provide strong cues to the ambient light (either yellow or blue). Recall that the reason people get different percepts in the original image is because people's brains are making different guesses about the illuminant (is it yellowish or blueish? The image doesn't give clear cues, so the brain has to guess). In my demo, the cues are clear and your visual system should pick up on this. Since the brain assumes the light source is adding either extra blue or extra yellow to the dress, it will remove either yellow or blue from the dress... yielding either a blue/black or white/gold percept.

The dress on the left should look blue/black. The dress on the right should look white/gold. (in fact, THEY ARE PHYSICALLY IDENTICAL). If it isn't working for you, zoom in on whichever panel you are having a hard time seeing "correctly" and stare at it for a few minutes without looking at the other panel. eventually it really should work.


The cropped version works best for me (I am normally a blue-black viewer):

And here is the full sized version (it works just as well for most people, but I prefer the cropped version, probably because I tend to see blue black and the white-gold cues are strongest near the top of the dress):

This of course does not resolve the question of why people infer different illuminants in the original image, or why their percepts tend to remain so stable (typically ambiguous illusions are bi-stable). It could have something to do with differences in individuals' priors or perhaps their looking patterns, but it does not seem to be based much on what monitor they are using or those viewing conditions (two people looking at the same screen can have different percepts). In an effort to shed some light on this question, Bevil Conway's lab is conducting some behavioral work on Mechanical Turk and I ask that you participate. You can find the study here: [LINK COMING SOON]

If you would like to share this image please credit me (Rosa Lafer-Sousa) and Beau Lotto (who provided the backgrounds). As well, you agree to not publish the above images in any peer-reviewed journals.

The daylight axis:
Typically, objects directly illuminated by bright daylight will have an orange/yellow bias. Objects illuminated by shadow (diffuse blue skylight) will have a blue bias. As well, the content of daylight varies in the amount of blue and orange over the course of the day. The assumption then is that the visual system, which evolved mostly to see under daylight conditions, is set up to discount blue and orange/yellow biased illumination conditions (and it probably uses cues like scene structure to infer when an object is in shadow vs directly illuminated). And in fact, people are better at consciously detecting illumination changes when they don't happen along the daylight axis (work by Anya Hurlbert and Karl Gegenfurtner), as it seems the visual system is less equipped to discount such illuminants. Our prior work in monkeys (fMRI and ephys) showed that the visual system (V1) is biased for the daylight axis (Conway lab, link to paper ).