Why does an evenly lit green screen still bounce green onto skin?
Because "evenly lit" describes the front of the screen, not the light leaving the side of it that faces your subject. A screen that measures perfectly flat still radiates green in every direction, and every point of the subject that can see it receives some. On a head modelled at five distances, the point facing the camera picked up no screen light at any distance, while the rim at 0.4 m got 57% of its illumination from the screen and swung 86 levels greener. Standing further back barely helps: at 4 m the screen was still supplying 23.7% of the rim's light.
Even describes the cloth, not the spill
When you meter a screen and get the same reading top, middle and bottom, you have measured how evenly the screen is lit. That tells you every part of it will key at the same tolerance, which is genuinely useful. It says nothing at all about the light leaving it.
A lit screen is a large surface radiating in every direction, and it is usually bigger than the person standing in front of it. Somewhere between a tenth and a half of everything illuminating your subject can be coming off the cloth rather than off your lights. Lighting the screen more evenly does not reduce that by a single photon — if anything it makes it more uniform, and therefore harder to spot until it is on the edge of your subject.
Who receives it, and who does not
This is decided entirely by geometry, and we worked it out by integrating over the screen rectangle cell by cell. Take a point on the head facing straight at the camera: the screen is behind it, so every part of the screen is on the wrong side of that point's surface, and it receives exactly nothing. Measured as green excess over the larger of red and blue, that point read -39.1 with the screen 0.4 m away and -39.3 with it 4 m away — the same number, because distance is irrelevant to a point that cannot see the source.
Now move to the rim, where the surface turns side-on and does see the screen. At 0.4 m the rim read +47.5: an 86-level swing from the point beside it on the same head. At 0.8 m it was +16.8, and at 1.5 m it had fallen back to -5.5, still well above the untouched value of about -39 but no longer obviously green to the eye.
So the bounce never lands where you are looking when you check the shot. It lands on the silhouette, on the sides of the head, on the tops of the shoulders — exactly the pixels the keyer is also trying to decide about.
Why moving back barely helps
The instinct is to put more distance between the subject and the screen. It works far less well than you would expect, because a 3 m screen at a couple of metres is not a point source — it is an extended one, and the fraction of the hemisphere it fills shrinks slowly as you step away.
Measured as the share of the rim's total illumination coming from the screen: 57.0% at 0.4 m, 50.1% at 0.8 m, 43.0% at 1.5 m, 37.4% at 2.5 m and still 23.7% at 4 m. You have to multiply the distance by ten to cut the contribution by a bit more than half.
That is the practical trap. Distance is the fix for the subject's shadow on the screen, and it is a poor fix for light coming back the other way.
When it stops being a fringe and becomes a hole
Green on the rim is ugly, but there is a worse version. The keyer measures distance from the key colour, and a pixel that has taken enough bounce can travel far enough to arrive inside the key's own radius — at which point the tool starts removing your subject.
At 0.4 m, the closest any pixel on the head came to the key colour was 0.142, comfortably inside the 0.259 where softening begins at the default settings, and 0.70% of the head was softened. At 0.8 m the closest approach was 0.267, just outside, and nothing was touched. So the failure mode flips at a distance you would not think of as unusually close.
If you see the edge of a shoulder or the side of a head going translucent as well as green, that is this. Raising tolerance will make it dramatically worse, because it widens the same radius the bounce is already pushing pixels into.
What a shiny scalp changes
Gloss matters, but less than people assume, and it changes the distribution more than the amount. We ran the same geometry twice: once with a broad lobe, like skin and fabric, and once with a tight one, like a shaved head or a forehead with some oil on it, each scaled by the Fresnel reflectance, which climbs steeply as a surface turns away from the camera.
At the rim the Fresnel factor averaged 0.247, nearly a quarter, which is why edges are where this shows up. Going from the broad lobe to the tight one moved the rim from +16.8 to +18.6 at 0.8 m and from -5.5 to -3.2 at 1.5 m. So a glossy head is a little greener at the edge, but the dominant term is still the ordinary diffuse bounce, not the highlight.
What gloss really changes is the shape of the damage: a tight lobe concentrates the bounce into a small, bright, clearly green patch that is easy to miss in a thumbnail and obvious in a still, rather than spreading it evenly around the whole silhouette.
What spill suppression actually takes out
The control for this in the keyer is a green-channel limiter: wherever green is stronger than both red and blue, it is pulled back toward that larger value. It is aimed precisely at this artefact, so it is worth knowing what it costs.
On the 0.8 m glossy case, the worst green excess on the head was 61.1 levels with the limiter off. At the default strength of 35 it came down to 39.7, at 70 to 18.3, and at 100 to zero. The rim's green channel went from 0.39 to 0.36, 0.33 and 0.31 over the same steps.
So the default takes out about a third of it. That is a deliberate choice: pushing to 100 also removes green that belongs to the subject, and warm skin with its green pulled down starts to look grey. Take it up until the edge looks neutral, then stop — and remember it is treating the symptom on the pixels that survived, not reducing the bounce that caused it.
What actually reduces the bounce
Stop the light leaving the cloth in the first place. A screen made of a less reflective material, or flagged so the subject stands in its shadow, or simply turned down relative to the subject's own key, all reduce the source term that every number above is proportional to.
Then make the subject brighter. The bounce is a fixed quantity of light; if you raise the key light illuminating the person, the bounce becomes a smaller share of the total, and the share is what decides how green the pixels look. This is the one lever that helps without any cost to the matte.
Matte the scalp if it is a bald head, and keep the subject's own light coming from a direction the screen is not. And when you check the result, look at the silhouette on a mid grey background rather than at the front of the face — the front is the one part of the subject the screen cannot touch.
Related questions
If my screen meters flat, why is there green on my subject?
Because metering the screen measures how evenly it is lit, not how much light it throws forward. A perfectly even screen still radiates in all directions, and any point on the subject that can see it receives some. Measured on a head, the rim at 0.4 m got 57% of its illumination from the screen.
Does standing further from the screen fix it?
Much less than you would expect. The share of the rim's illumination coming from the screen went 57.0% at 0.4 m, 43.0% at 1.5 m and still 23.7% at 4 m. A large screen fills a big part of the hemisphere, so its contribution falls off slowly with distance.
Can the bounce make the keyer eat my subject?
Yes. At 0.4 m the closest a pixel on the head came to the key colour was 0.142, inside the 0.259 where softening starts at default settings, and 0.70% of the head was softened. At 0.8 m the closest approach was 0.267 and nothing was touched.
Why does spill suppression not remove all of it?
At its default strength of 35 it took the worst green excess on the head from 61.1 levels down to 39.7, about a third. At 100 it removes all of it, but it also removes green belonging to the subject, and the rim's green channel drops from 0.39 to 0.31, which starts to look grey.