Why does color range refuse to select your green screen past 50%?
Because a colour range never says 'this is green'. It scores every pixel by how far it sits from the colour you sampled, and that score is a ramp: full inside the inner radius, zero past the outer one, and exactly 50% halfway between. If the closest pixel on your screen sits at that midpoint - 0.210 in normalised RGB at the default tolerance, which is roughly 0.8 stops of extra light or 10% of room light washing onto the screen - then every other pixel is farther out and nothing in the frame can score above 50%. The message is not a glitch; it is telling you the colour you asked for is not on your screen any more.
What 'percent selected' is actually measuring
A colour range does not sort pixels into two piles. It gives each one a number. In this keyer the number comes from the normalised RGB distance to the key colour: pixels inside the inner radius score 1 and are gone, pixels past the outer radius score 0 and are kept, and in between the score is a smoothstep across the gap. Alpha, the preview and the exported cutout are all derived from that one number - alpha is simply 1 minus the score.
So '50% selected' means alpha 0.5: half the screen is still there. And 'no pixels are more than 50% selected' means the most screen-like pixel in the whole frame is only halfway into the ramp, with everything else behind it. At the defaults on this page - similarity 42, smoothness 14 - the inner radius is 0.161 and the outer one is 0.259. The entire decision lives inside a band 0.098 wide, which is about a third of a stop of exposure.
The 50% line is a distance you can name
The band is placed symmetrically around a radius the tool calls tolerance, and that radius is just similarity divided by 200. So the 50% mark is a number you can read straight off the control: similarity 20 puts it at 0.100, 42 at 0.210, 60 at 0.300, 80 at 0.400 and 100 at 0.500. Smoothness widens the band on both sides and never moves the midpoint, so the 50% line does not shift when you soften the edge.
Measured on a flat screen at the default settings, with the closest pixel at a known distance: 0.180 gives a best score of 90.2%, 0.210 gives 50.0%, 0.240 gives 9.8% and 0.259 gives 0.0%. The rule is exact rather than approximate - no pixel in a frame can score above 50% precisely when the closest pixel on the screen is at least similarity/200 away from the key.
What 0.210 looks like in a room
Leave the key at the nominal studio green and light the screen hotter, and you walk into it quickly. At 0.68 stops over, the screen reads (0,218,81), a distance of 0.175, and the best score is still 94.4%. At 0.80 stops it reads (0,227,84), distance 0.210, score 50.0%. At 0.96 stops it reads (0,238,89) and the selection is empty. From fully selected to nothing at all there are 0.33 stops of exposure - a third of a stop.
A light that simply drifts during a take cannot get you there: a 20% drift moves the distance by 0.064 and a 60% drift by 0.175, still a 94.4% score. What does get you there is a different lighting setup, or light from the room landing on the screen. Ten percent of ambient white on the screen reads (50,183,81), distance 0.209, score 52.3% - and at 15% the distance is 0.263 and the selection is gone entirely.
Why it says 'no pixels' rather than 'some pixels'
Because the ramp only runs one way. The score depends on distance alone, so the nearest pixel sets the ceiling and every other pixel sits below it. Put the nearest one at the midpoint and everything else is past it, which is why the complaint is about the whole frame and not about a corner. A screen whose near side is fully selected and whose far side is not would have to straddle the band - and the band is only a third of a stop wide, so a screen uneven enough to straddle it is a separate problem.
That is also why the message is more useful than it sounds. It is not a verdict on your screen; it is a distance reading. It tells you the key colour is at least one tolerance radius away from every pixel in front of you, which narrows the cause to a handful of things: the key came from a preset or a paint chip, from another take, or from a screen that is no longer lit the way it was when you sampled it.
Fix one: take the key from the frame you are keying
The distance that matters is the one between the key and the pixels in front of you, so the cheapest fix is to make it zero. This page's eyedropper reads a single pixel - one getImageData call on the spot you click - so clicking anywhere on the screen guarantees that pixel scores 100%, and the question becomes how much of the rest follows.
Measured over screen areas with real shading: at 0.14, 0.43 and 1.04 stops of spread across the screen, 100.0% of the screen reaches full selection from a single click. At 1.97 stops it is 88.2%, and at 7.3 stops - a screen lit by one lamp from the side - it falls to 62.0%. Under about a stop of uneven lighting a sampled key removes the failure outright, and it is the only fix that costs nothing else.
Fix two: raise the tolerance, and pay for it in hair
When the key cannot be re-sampled - a preset baked into a batch, say - the tolerance that brings the whole screen back to fully selected is 200 times the distance plus 9.8. For a screen sitting at 0.210 that is 52. Getting it above the 50% line alone is cheaper: any tolerance at or beyond 200 times the distance will do.
The cost is measured on a 640x360 frame with hair, skin and a shirt against a flat screen lit 0.80 stops hot, key left at the nominal green. At similarity 42 the screen averages 48.7% selected and 35.7% of hair pixels are already more than half gone. At 52 the screen is 96.9% selected and half the hair - 50.2% - is gone. At 60 the screen is fully selected and 59.6% of hair is gone; at 100 it is 93.4%, and the shirt's average alpha has dropped to 0.454. Hair strands go first because a half-covered hair pixel is already most of the way to the screen's own colour.
Why a green-channel matte does not have this failure
The standard way out is to stop using a colour range at all: duplicate the green channel and set levels on it. Levels clamps at both ends, so 100% and 0% are both reachable by construction - which is exactly what a distance ramp cannot promise. On a frame whose screen reads 177 on green, levels set to [120,149] puts 100.0% of the screen at full selection, with no ceiling anywhere in sight.
The price is that one axis is not three. On the green channel the screen sits at 177 and skin at 155 - 22 levels apart, and only 8 at the brightest skin pixel - while the same two colours are 0.849 apart in the 3-D distance, roughly ten times more room. Drag the white point down to 149 to catch the dim part of the screen and 100% of skin crosses 50% selected; hold it at 176 to protect the skin and only 55.9% of the screen reaches full selection. Anything bright in green simply disappears: a yellow prop reading 205 on green is removed 100% by the channel matte, and left untouched at alpha 1.000 by the colour distance. That trade is why the workaround always ends with brushes.
Related questions
Is 50% selected the same as half transparent?
Yes. The score the keyer computes is the fraction of that pixel treated as background, so a score of 0.5 leaves alpha at 0.5 and the screen showing through at half strength. It is not a partial selection of the area - it is every affected pixel sitting at the same half-and-half value, which is why the result looks like a flat green wash rather than a ragged edge.
Can I just push tolerance to 100 and be done with it?
It removes the screen, and it takes the subject with it. On the frame measured here, similarity 100 puts the screen at 100% selected but leaves 93.4% of hair pixels more than half gone and drops the shirt to an average alpha of 0.454. Tolerance is a radius around one colour; past a certain point the radius reaches into the subject faster than it reaches into the rest of the screen.
Can this happen with a colour I sampled from the screen myself?
Not in the same frame - the pixel you clicked is at distance zero and scores 100% by definition. It happens when the key outlives the frame it came from: a preset, a paint chip, another take, or a screen whose lighting changed by around 0.8 stops after you sampled. A light that merely drifts will not do it; 60% of drift still leaves the best score at 94.4%.
Why does the green channel work when the colour range does not?
Because levels clamps. A colour range maps distance to a score and can only saturate if the key colour is inside the cluster of pixels it is scoring. A channel with levels maps one axis and clamps at both ends, so full black and full white always exist. The cost is separation: on the green axis the screen and skin are 22 levels apart, against 0.849 in the full colour distance.