Why do you end up polishing the part of the frame you're about to throw away?
Because the cloth is the biggest thing in the frame, so it attracts the work — but a chroma key only needs the cloth inside a fixed radius, and once it is inside, extra evenness changes nothing at all. On a 1280x720 talking-head frame the cloth was 81.55% of the pixels and the edge was 1.26%; once every cloth pixel sat inside that radius — roughly 1.1 stops of falloff corner to corner — every further improvement changed 0 of those 751,533 cloth pixels in the delivered file, while the edge was still off by 17.96 code levels per pixel. The cloth has a finish line you can measure, and the edge does not, which is the whole reason the effort ends up in the wrong place.
How much of the frame you are actually working on
I built a 1280x720 talking-head frame — head, hair with flyaways, shirt, trousers — in front of a 3.0 x 1.69 m cloth, and counted pixels by what the keyer does to them. The cloth came to 751,533 pixels, 81.55% of the frame. The subject was 158,437 pixels, 17.19%. The edge, the only place where a pixel is part cloth and part subject at once, was 11,630 pixels: 1.26%.
That imbalance is the whole habit in one line. The cloth is 65 times the size of the edge, it is the part you can see is uneven, and it fills the preview while you work. A green background remover deletes it entirely once the key is right, so time spent making it smoother is time spent on 81.55% of the frame that will not exist in the output.
The cloth has a finish line, and it is a number
This keyer measures each pixel's distance to the key colour in normalised RGB. A pixel turns fully transparent once that distance is at or below lo, which is similarity/200 minus smoothness/280; at the default 42 and 14 that radius is 0.161. Turned around, that gives the number worth knowing: the worst cloth pixel in the frame decides the similarity you need, and it is 200 times its distance plus 9.8.
I swept one rig — two fixtures 1.05 m high at x = -1.05 m and +1.05 m, pulled further back each time — and read that figure off the worst cloth pixel. At 1.5 m the cloth fell off 2.268 stops corner to corner and needed similarity 70.7. At 2.0 m, 1.691 stops and 59.0. At 2.5 m, 1.357 stops and 50.6. At 3.0 m, 1.111 stops and 43.6, a hair over the default. At 8.0 m it needed only 19.0.
The finish line is where that figure drops to 42 or below, because at 42 the tool hands you the cloth for free. In this frame it sits between 1.357 and 1.111 stops of falloff. Cloth residue runs 14.19% at 1.5 m, 3.48% at 2.0 m, 0.23% at 2.5 m and 0.00% at 3.0 m, where the number of cloth pixels still carrying more than half their opacity is exactly zero.
Past the finish line, more light buys exactly nothing
I composited every version over mid grey and compared each to a mathematically flat cloth, pixel by pixel. At 1.5 m, 241,629 pixels of the delivered image changed by more than three code levels, and 238,098 of them were cloth pixels. At 2.0 m it was 86,502 and 83,319. At 2.5 m, 20,256 and 17,332. At 3.0 m the total drops to 2,806 and the cloth count reaches 0.
From there it stays 0. At 4.0 m, 6.0 m and 8.0 m the cloth region of the output is identical to the flat-cloth version, while the flux needed to hold the same exposure climbs from 3.42x to 5.27x, 10.34x and 17.32x. Going from 3.0 m to 8.0 m costs five times the light and moves zero cloth pixels. The 1,339 to 2,806 pixels that still move are edge pixels, because the local cloth colour mixed into them is what changed.
Where the error lives once the cloth is clean
Measured against a physically correct composite, the split flips the moment the cloth crosses the line. At 1.5 m the cloth carries 90.90% of the total error. At 3.0 m it carries 0.03%, and per pixel its error is 0.00 code levels against the edge's 17.96. The cloth is 81.55% of the frame and contributes nothing; the edge is 1.26% of the frame and contributes 25.03%.
The largest slice left over, 74.94%, sits on the subject at 3.95 code levels per pixel. That is the green the cloth bounces onto skin and shirt, and it is not an evenness problem either — it is set by how much cloth the subject can see, which is a distance question. So of everything still wrong after the cloth is clean, none of it is fixed by making the cloth cleaner.
Why the edge never finishes
The cloth stops at a threshold; the edge does not, because the edge is a coverage problem. At 3.0 m, with the cloth already perfect, keyed alpha differs from true coverage by 0.2418 on average, and 83.53% of edge pixels are off by more than 0.10. A mathematically flat cloth only improves that to 0.2276. The soft band between transparent and opaque simply does not line up with how much of the pixel the subject actually covers.
You can trade for it, which is the difference. Smoothness 60 cuts the average error to 0.1225, but it brings back 9.12% cloth residue — you are buying the edge with the cloth again. Spill suppression is the honest lever: on the subject side of the edge, green excess runs 66.24 code levels at 0, 42.87 at the default 35 and 19.51 at 70, and it costs only 1.56 levels of the subject's own green.
What a minute of edge work buys next to an hour of cloth work
On the badly lit cloth at 1.5 m, one click beats the whole afternoon. Take the key colour from the mean of the cloth and residue is 14.19%. Take it from the brightest pixel and it is 43.92%; from the darkest, 67.39%. Those two need similarity 110.4, past the end of the slider, so no setting rescues them. Pushing similarity from 42 to 50 drops residue from 14.19% to 4.95%, to 60 gives 0.36%, to 80 gives 0.00%, and the subject survives untouched until 100.
On the already-clean cloth at 3.0 m the same slider buys nothing: residue is 0.00% at every step, and the only thing that moves is the edge getting worse — 0.2418, 0.2895, 0.3214, 0.3366 — before similarity 100 deletes 81.12% of the subject. That is the signature of work spent past the finish line. The number you are chasing cannot improve, and the number you are not watching quietly does.
A stopping rule you can check in thirty seconds
Sample the mean of the cloth, not a corner, and work out what the worst cloth pixel needs. If that figure sits at 42 or below with smoothness at 14, stop lighting the cloth. If it is above 42 you will pay twice: once in light, and once in the similarity you have to add, which comes out of the subject and the edge. On this frame the whole decision turned on about 1.1 stops of corner-to-corner falloff.
Then inspect the cloth over a flat background instead of the checkerboard, because the default preview is near black and hides exactly the low-alpha residue you are hunting. If no cloth pixel survives there, the cloth is finished and the rest of your time belongs to the two numbers that are still wrong: how far the subject stands from the cloth, and how much green the edge is carrying.
Related questions
How even does a green screen actually need to be?
In this frame, about 1.1 stops of falloff from the brightest part of the cloth to the darkest. The practical test is the worst cloth pixel: 200 times its distance from your key colour plus 9.8 tells you the similarity it needs, and if that is 42 or below the default already removes the cloth completely. At 1.357 stops the worst pixel needed 50.6 and left 0.23% behind; at 1.111 stops it needed 43.6 and left nothing.
Does a perfectly even screen give a perfect edge?
No. With a mathematically flat cloth the average edge alpha error was 0.2276, against 0.2418 with the cloth falling off 1.111 stops, so the edge barely notices the difference. The floor comes from the width of the soft band measured against each pixel's true coverage, which is a coverage problem rather than an evenness one. With the cloth already perfect, 83.53% of edge pixels still sat more than 0.10 away from their true coverage.
Should I buy more lights, or move the subject away from the cloth?
Move the subject. Once the cloth is inside its radius, more light on it changes zero output pixels, while the green it bounces onto the subject is untouched by evenness and accounts for 74.94% of what is left. Distance reduces that bounce at the source, and the spill slider reduces what remains on the edge, from 66.24 code levels of green excess at 0 down to 19.51 at 70.
What is the quickest way to see whether my cloth is already good enough?
Switch Background from Transparent to a flat colour and look at the cloth area rather than the subject; any patch that survives is cloth still outside the radius. If nothing survives, stop. In this frame the three biggest further improvements, at 4.0 m, 6.0 m and 8.0 m, changed the cloth region of the output by zero pixels while costing 5.27x, 10.34x and 17.32x the light.