Remember "The Dress"? The viral 2015 photo that split the internet into people who saw a blue-black dress and people who saw a white-gold dress? This paper offers a mathematical explanation for why that ambiguity is so sharp and so divisive. The key idea comes from a field called intrinsic image decomposition, which is the problem of taking a photograph and figuring out two separate things that got mixed together to create it: the true colors of the surfaces in the scene, and the color of the light illuminating those surfaces. The Dress is ambiguous precisely because there are two equally plausible solutions to this separation problem, and your visual system commits to one of them.
The authors argue that this ambiguity has a specific geometric character. Imagine the space of all possible images as a landscape. Most images have one clearly preferred explanation, but somewhere in that landscape there is a sharp boundary, and right at that boundary the preferred explanation flips suddenly from one solution to the other, like a light switch. A neural network trained to solve this kind of problem cannot actually implement a sharp flip; instead it smears the transition out into a thin fuzzy zone. The paper predicts two measurable signs of being near such a boundary. For networks that go from image to decomposition, the output should become extremely sensitive to small changes in color temperature near that boundary. For networks that encode images into abstract representations, the geometry of that representation should become highly curved near the boundary.
The authors then test both predictions using real neural networks and real data. Using a large dataset of rendered indoor scenes, they show that a particular measure of color sensitivity predicts where a state-of-the-art decomposition network makes its biggest errors, with a correlation much stronger than simpler measures like brightness or saturation. For The Dress specifically, they feed it through a large vision model called CLIP and find that the internal geometry of its representations spikes dramatically at a color temperature corresponding to standard daylight, exactly the boundary between the two interpretations, while a control image of a non-ambiguous dress shows no such spike. This signature appears consistently across several different network architectures, suggesting it reflects something real about the underlying mathematics rather than a quirk of any one model.