Why is lighting a green screen harder than lighting a whiteboard?
A whiteboard only has to look white to a person. A green screen has to land on one code value, because every pixel is tested against a single point in RGB space: removed completely inside 0.161 of our preset #00B140, untouched past 0.259. Converted into light, that means the screen has to be flat to within 0.767 stops across its whole area, exposed inside a window 1.39 stops wide, and lit with about 19 times the total light a whiteboard needs, because it is seven times the size and returns only a third of what falls on it.
One surface is judged by a person, the other by a radius
This keyer runs one test on every pixel: how far is it, in a straight line through RGB space, from the key colour. Our preset is #00B140. At the default settings a pixel is removed completely while it stays inside 0.161 of that value, it starts to soften as it travels out to 0.259, and past that the keyer leaves it alone. There is no second test and no overall impression.
Nothing equivalent is applied to a whiteboard. It stays in the frame, a person decides whether it reads as white, and a corner darker than the middle is shading rather than a fault. In the same lighting model, one light two metres out puts the corner of a 1.2 by 0.9 metre board 0.380 stops below its centre; the same light on a three metre screen puts its corner 1.886 stops down. One of those is a photograph. The other is a key that does not close.
The screen sends back a third of what you put on it
#00B140 is not a paint swatch. It is the code value a correctly exposed chroma-green screen lands on, and decoded to linear light it is (0.00, 0.44, 0.05). Weighted for luminance that is a reflectance of 0.318 against 0.90 for whiteboard paint, so the screen returns 35 per cent of the light and needs 2.83 times the illuminance, one and a half stops, to reach the same reading on a meter.
Only one channel carries much signal either. To bring the screen's green channel up to the code a whiteboard reaches, you need 2.05 times the light. Then there is the size: a three by 2.4 metre screen is 7.2 square metres against 1.08 for an office board, which is 6.67 times the area. Put the reflectance and the area together and lighting the screen takes roughly 19 times the total light that lighting the board does.
Falloff is a threshold on a screen and a gradient on a board
Darkening a pixel walks it toward black, and its distance from the key grows as it goes. Measured on this keyer, a green screen survives 0.767 stops of falloff before the cut stops being perfectly clean and 1.345 stops before holes appear; pushing the similarity slider all the way to 100 buys 3.923 stops. The distance grows at 0.204 per stop.
Green is forgiving here for a structural reason. It is a dark, saturated colour sitting on the steep part of the encoding curve, and one of its three channels is pinned at zero, so only two channels move when the light drops. White is not, at 0.455 per stop, which leaves 0.316 stops before the cut degrades and 0.535 before it fails. That is also why we light green at all: not because it is easy, but because everything else is tighter still.
What it takes to hold that window across three metres
Run the falloff model over the whole surface with the lights on a grid, each aimed at its own patch. One light two metres out gives 1.886 stops of spread and leaves 47.81 per cent of the screen unkeyed at the default settings. Move it back to three metres and it is still 0.991 stops, with 4.91 per cent of the screen surviving.
Two lights at two metres give 1.384 stops and 19.24 per cent left behind. Two lights at three metres reach 0.816 stops and 0.34 per cent. Four lights at three metres give 0.735 stops and a clean 0.00 per cent, which is the first configuration the default settings handle without help; spread them wide at four metres and the spread falls to 0.455 stops.
Too much light breaks it faster than too little
Exposure is a window, not a floor. With the screen lit perfectly evenly, the cut is clean from 0.767 stops under to 0.627 stops over the setting that puts the middle of the screen exactly on #00B140, a window 1.39 stops wide. Beyond that, and out to 1.345 stops under and 0.966 stops over, you get partial transparency first and then holes.
Over-exposure is the worse direction because green clips first. One stop over, the screen reads (0, 241, 90), which is 0.271 from the key and comes back completely opaque, and at one and a half stops the green channel is pinned at 255 while blue keeps climbing. Under-expose by one stop instead and you get (0, 129, 44) at 0.204, sitting in the soft band at alpha 0.407.
Why the similarity slider is not a way out
Each of those configurations can be rescued by opening the slider. One light at two metres needs 77, two lights at two metres need 63, one light at three metres needs 50, four at 2.5 metres need 47 and four at three metres need 41. Every one of those numbers buys a flat screen by spending edge.
Take a strand that covers half a pixel. At the default 42 a half-covered dark strand is already down to alpha 0.452; at 44 it is 0.303, at 47 it is 0.116, at 50 it is 0.010, and by 63 it is gone altogether. A half-covered blonde strand holds to about 97, and a mid-grey shirt holds to 100 and then drops to 0.409. Opening the slider converts a lighting problem into a hair problem.
What this means when you set up
Meter the screen on its own, at several points rather than just the middle, and aim to land every part of it on the same code value inside about three quarters of a stop. Budget the light: two and a half to three times what you would put on a white surface of the same size, and four fixtures rather than two if you can stand them far enough back.
Then keep the subject out of it. Once it is lit, the screen is a 7.2 square metre source, and our earlier measurement had it still supplying 24 per cent of the light falling on a side-facing surface four metres away. Presenting against it is a separate skill as well: there is nothing behind you to react to, and you have to stay inside the lit envelope while you move.
Related questions
How flat does a green screen actually have to be?
At the default settings, 0.767 stops of falloff from the brightest part to the darkest before the cut stops being completely clean, and 1.345 stops before holes start to appear. Pushing the similarity slider to 100 extends the clean range to 3.923 stops. The same test run on a white surface gives 0.316 and 0.535 stops.
Can I light a green screen with one light?
Not a full-size one. Our falloff model gives 1.886 stops of spread with a single light two metres from a three by 2.4 metre screen, leaving 47.81 per cent of it unkeyed at the default settings, and even at three metres it is 0.991 stops with 4.91 per cent left. Two lights at three metres, or four at three metres, is where it starts to work.
Can a green screen be too bright?
Yes, and it fails sooner than going dark. The clean window runs from 0.767 stops under to only 0.627 stops over. One stop over puts the screen at (0, 241, 90), a distance of 0.271 from #00B140, which is past the point where the keyer touches it at all, so the background returns fully opaque.
Why not key a plain white wall instead and skip the green?
Because white is less forgiving, not more. Distance from the key colour grows 0.455 per stop of falloff on white against 0.204 on green, so a white surface survives only 0.316 stops before the cut degrades and 0.535 before it fails, against 0.767 and 1.345 for green. Green is the cheapest surface to light, not an easy one.