Vision · a stratum in the mind seam
The Grey the Edge Invented
Two patches of the exact same grey can look nothing alike, and a smooth ramp of light can grow bright and dark stripes that are not in the light at all. Your eyes are not photometers. They report where light changes, and rebuild everything between the changes, so the edge gets the last word over the middle. Operate it below. Every pixel value is recomputed in front of you.
Point a light meter at a wall lit unevenly by a window and it reads a hundred different numbers across one flat coat of paint. You see one wall, one colour. That is not sloppiness, it is the whole trick of seeing: a visual system that measured absolute light would be helpless, because the same surface throws wildly different amounts of light at your eye depending on the sun, the lamp, the shadow. So vision throws the absolute number away almost immediately and keeps only the thing that carries meaning, the edges, the places where light changes. Then it reconstructs the surfaces between the edges by, in effect, believing them.
The illusions on this page are not tricks bolted onto normal seeing. They are normal seeing, caught in the act. Each one is the same machine doing the same job on a stimulus built to make the job visible. And they line up as a ladder: the first needs almost nothing to explain, the last is still argued over in the journals. That climb is the real subject here.
Rung one · the edge decides the whole
Two greys that are one grey
Below is a single rectangle. The left half looks lighter than the right half. They are not: away from the seam down the middle, the two halves are the identical grey, pixel for pixel. The only real difference is a thin cusp right at the border, a bright lip on the left, a dark lip on the right. Your visual system finds that cusp, decides it is a real step between two surfaces, and paints the decision across both whole halves.
Do not take my word for it. Drag the slider to slide a neutral strip over the seam. The moment the cusp is covered, the two halves snap together into one flat grey, because that is what they always were.
The check
The left plateau and the right plateau are written to the canvas with the identical grey value (128 of 255); read the flat line on the profile plot, and eyedrop the pixels if you like. The cusp reaches 128±48 and no further. Nothing away from the seam differs. research/lightness-edges/verify.mjs re-derives all of it (27/27).
This is the Craik–O'Brien–Cornsweet effect, noticed by Kenneth Craik in the 1940s, demonstrated by Vivian O'Brien in 1958, and made famous by Tom Cornsweet in 1970. It is the cleanest proof that you do not see the light that arrives. You see a reconstruction, and the reconstruction can be talked into a difference the world never sent.
The machine · a cell that is blind to how bright things are
What "reporting change" looks like in arithmetic
The first stage of seeing, in the retina and the thalamus, is built from cells with a centre-surround shape: a little excitatory centre ringed by an inhibitory surround. Model one as a difference of two Gaussians, a narrow bump minus a broad one, both balanced so they sum to the same weight. Slide it across an image and it answers a single question at every point: is the middle here brighter than its ring?
Because the two Gaussians are balanced, the cell has a startling property, and it is the whole thesis in one line: feed it a field of any uniform brightness, black, grey, white, and its answer is exactly zero. It cannot tell you how bright anything is. It only ever speaks up where light changes. Watch it below: set the field to any level, the response stays flat on zero, until you break the flatness.
Said honestly
That 1-to-1.6 width ratio between centre and surround is the Marr–Hildreth (1980) figure that makes this difference-of-Gaussians best imitate an ideal edge detector (a Laplacian of a Gaussian). It is not a measured retinal anatomy number; real surround-to-centre sizes run larger (roughly 6-to-1 in diameter, Enroth-Cugell & Robson 1966). And this is a first-order model: it fits the retina's linear "X" cells and breaks on the nonlinear "Y" cells. "The eye computes a derivative" is a useful picture, not a claim that neurons do calculus.
Rung two · a straight ramp, bent by the eye
Bands that are not in the light
Now feed the machine a plain ramp: a dark plateau, a straight run up to a light plateau. The light itself is a perfect straight line between the two levels, no bumps. Yet at the foot of the ramp you will see a band darker than the dark plateau, and at the top a band brighter than the bright plateau. These are Mach bands, described by the physicist Ernst Mach in 1865, and the centre-surround response shows exactly where they come from: it overshoots at the knee and undershoots at the foot, because the change in the slope is what the cell reacts to.
Slide the ramp from wide to a sharp step and watch the honest crack open.
Where the simple story cracks
Slide all the way to a sharp step. The response peaks even harder there, yet nobody sees a Mach band at a clean edge (Ross, Holt & Johnstone 1984). So a single centre-surround cell already over-predicts: it shouts at the step where no band appears. Lateral inhibition is the classic textbook account of Mach bands and it gets the direction right, but it is known to be incomplete and is still argued (response-normalization and feature-detector models compete). This is rung two, and the model is already only an approximation.
Rung three · the surround tips the scale
The same grey, weighed against its neighbours
Two small squares, the identical grey, one on a dark field and one on a light field. The one on dark looks lighter; the one on light looks darker. This is simultaneous contrast, the oldest illusion in the book, and here the centre-surround model earns its keep: read at the coarse scale where a cell's surround reaches out into the field, it answers positive for the patch on dark (brighter than its ring) and negative for the patch on light. It orders the two patches exactly the way you see them.
Drag the surrounds toward a common grey and the illusion dissolves, because now both patches share one ring and there is nothing to weigh them against.
The check
Both squares are written with the identical grey (128). The coarse centre-surround response is + on the dark field and − on the light field, so the model ranks them the way people report. Pull the surrounds to a common grey and the response difference falls to zero along with the illusion. Recomputed in verify.mjs.
Rung four · where the simple model gets it backwards
The illusion that breaks the rule
Here is the same trick one more time, and this time the tidy story fails. A grating of black and white bars, with two identical grey patches: patch A lies on the black bars, patch B on the white ones. Patch A looks lighter. But look at what borders each patch: patch A's long edges run against white bars, so it has the whiter surround. Simple contrast, the rule that just worked, says the patch with the whiter surround should look darker. It looks lighter. The rule is not just weak here, it points the wrong way.
The check
Patch A and patch B are the identical grey (128). Patch A's border is measurably whiter (mean 168) than patch B's (mean 87), so a contrast/lateral-inhibition account predicts A darker, the reverse of the reported percept, where A looks lighter (White, 1979). The pixel facts here are computed; the perceptual direction is cited, not measured on this page. Lift both patches onto a plain field and they are plainly one grey.
What White's illusion does and does not overturn
It defeats the simple, single-scale centre-surround model, not spatial filtering as an idea. An oriented, multi-scale version (the ODOG model, Blakeslee & McCourt 1999) reproduces both White's illusion and plain contrast. Other leading accounts are mid-level: the patch is grouped with the bar it lies along (belongingness, T-junctions), or assigned a lightness by the framework it sits in (anchoring, Gilchrist). Which is right is not settled. That is the honest frontier this ladder climbs toward.
The ladder, and why it matters
Read the four together and a single shape appears. The same operation, report change and rebuild the rest, runs under all of them, but each needs a little more of the machine to explain:
Mach bands need only the front-end filter (and even it over-shouts at a step). The Cornsweet effect needs the filter plus a second move, integrating the kept edges back into surfaces while discounting slow gradients as mere lighting, which is how a local cusp gets painted across a whole plateau. Simultaneous contrast the filter handles at a coarse scale. White's illusion defeats the simple filter outright and needs oriented, multi-scale, or mid-level grouping that is still being worked out.
So perception is not one trick but a stack, and the top of the stack is a live research question. The reason any of this is worth an hour is not that your eyes can be fooled, everyone knows that. It is why they can be fooled: because seeing absolute light was never the goal. A visual system that reported the meter reading would be lost the moment a cloud crossed the sun. Yours reports the edges and reconstructs the world, and the price of that competence, the tiny, revealing tax on an otherwise brilliant strategy, is that a grey with the right edge can be any grey it likes.