Why does an unevenly lit green screen ruin the edge?
Because the key decides by RGB distance rather than by hue, so a fold of the screen that is darker than the lit part reads as a different colour even though it is the same fabric. On this tool's 0-100 Tolerance scale, one stop of falloff across the screen costs about 74 points, and the slider runs out of range entirely at roughly 1.6 stops. Push it that far and you are not cleaning the edge, you are eroding the subject.
The key compares distance, not greenness
A chroma key here works on one number: the distance between a pixel and the key colour in normalised RGB. Each channel runs 0 to 1, the key colour is subtracted, and the differences combine as the square root of the summed squares. Tolerance sets the radius inside which a pixel counts as fully background — 0 to 0.5, from the slider's 0 to 100.
What that formula does not do is separate hue from brightness. The preset key green #00B140 is (0, 0.694, 0.251). Shade a corner to half its brightness and the same fabric reads (0, 0.347, 0.126) — identical hue, plainly green to your eye, but 0.369 away in the only space the keyer measures.
Uneven lighting is not a small quality cost — it is the same error as painting part of the screen a different colour, paid for out of the same budget: the radius.
What one stop of falloff costs you
A patch of screen at brightness k relative to the lit reference sits |1 − k| × 0.738 away from the key colour. On the Tolerance scale, where radius is Tolerance ÷ 100 × 0.5, that works out as follows — from the keyer's own formula, not estimates about footage.
90% of reference brightness (0.15 stops down) — Tolerance 15
80% (0.32 stops) — Tolerance 30
70% (0.51 stops) — Tolerance 44
60% (0.74 stops) — Tolerance 59
50% (1.00 stop) — Tolerance 74
40% (1.32 stops) — Tolerance 89
35% (1.51 stops) — Tolerance 96
Linear and unforgiving: about 74 points per stop of falloff, on a 100-point slider. Past roughly 1.63 stops — a fold at 32% of reference brightness — no setting reaches it, since the radius caps at 0.5.
Shadows cost you slider, hotspots cost you the shot
Under-lit folds and blown-out patches both sit outside the radius, but only one is recoverable. A shadowed fold keeps its hue and is merely scaled down, so it lies on the same line out from the key colour and a larger radius eventually swallows it.
A hotspot does not. Brighten the fabric past the reference and green climbs first, then clips at 255 around 1.44×, after which the patch drifts toward white. Fully white sits 1.286 from #00B140 — about two and a half times the largest radius the slider can produce. No Tolerance reaches it, and spill suppression cannot help, because no green dominance is left to suppress.
Shadows are bought back with Tolerance. Hotspots have to be prevented at the shoot.
Edge softness widens the range as well
Edge softness is not only cosmetic. It adds a soft knee around the radius: pixels closer than (radius − softness) are fully background, pixels beyond (radius + softness) are fully subject, and everything between gets partial transparency on a smoothstep curve. Softness runs 0 to 0.35, so at a high setting the knee is wider than the radius.
At the defaults — Tolerance 42, Edge softness 14 — the fully removed radius is 0.21 and the partial band reaches 0.259, so the screen can sit between roughly 72% and 128% of reference brightness before the background stops being clean.
The trap: raising Edge softness looks like it rescues a mottled screen, because the mottling fades. What happened is that the partial-alpha band moved onto the subject, so shoulders and hair come out semi-transparent — a grey halo on any background.
Pick the middle of the screen, not the bright end
The eyedropper samples one pixel and writes its hex value into the key colour field, so it is a measuring tool as much as a picker — and where you sample decides how much range is left.
Take a screen running from 60% brightness in a dark corner to 100% near a light. Sample the lit area and the dark corner sits 0.295 away, needing Tolerance 59. Sample a mid-tone patch near 80% and both ends land 0.148 away, needing Tolerance 30. Same footage, half the Tolerance, half the damage to the subject.
Load one representative frame and sample a patch near the subject rather than a corner.
Why 4:2:0 makes an uneven screen worse
Nearly every consumer camera records 4:2:0, where colour is stored at a quarter of the resolution of brightness. A chroma key works on colour alone, so it reads the part of the image already discarded — and worst exactly where it matters, at edges, where colour changes fastest.
The two problems compound. Uneven lighting spreads the screen's colour across a range, then subsampling averages that range in 2×2 blocks, so a fold's shadow boundary is smeared across two pixels of colour while the brightness edge stays sharp. The matte follows the smeared map.
If the shoot is ahead of you, 4:2:2 or 4:4:4 beats any amount of post work. If the footage exists, sample the mid-tone, keep Tolerance at the minimum that clears the screen, and let Spill suppression handle the tint.
A wrinkle is three problems in a few pixels
A wrinkle is not merely a darker patch. It is a brightness gradient, a self-shadow where the fold blocks light from its own neighbour, and on anything with a sheen, including most inexpensive backdrops, a specular highlight along the crest. All three happen inside a band a few pixels wide.
That is why a wrinkle ruins a key faster than a soft gradient spanning the same brightness range. A gentle falloff gives a smooth ramp one radius can cover; a fold compresses that range into the pixels against the subject's silhouette, where the matte has to be right.
Fix it on set: stretch or steam the material, use a painted wall or rigid panel, keep the subject far enough forward that their shadow and the folds do not compete for the same pixels, and light the screen from two sources at 45 degrees.
How much unevenness is still acceptable
Measure this before keying anything. Sample the four corners and the centre with the eyedropper, compare the green channel of each hex value, and take the darkest as a fraction of the brightest.
Down to about 72% the default Tolerance 42 still clears it; at 65% you are inside the partial-alpha band and edges go soft and grey; under roughly 32% nothing here will remove it. Keep the darkest patch at 70% or more of the brightest — about half a stop, corner to corner.
A brightest patch of #00B140 has a green channel of 177; a corner reading #007A2D reads 122, or 69% — past the limit, and the first place the edge fails.
Related questions
Can I just raise Tolerance until the dark patches disappear?
Only up to a point: each stop of falloff costs about 74 points on the 0-100 scale, so past roughly 1.6 stops no setting reaches it — and every point added pulls subject pixels inside that radius toward transparency. Beyond that, segment the subject instead.
Does shooting at a higher resolution fix an uneven screen?
Not by itself. More pixels do not change the brightness ratio between a lit patch and a shadowed one, and 4:2:0 keeps colour at a quarter resolution regardless. Recording 4:2:2 or 4:4:4 does help: the colour the key reads has not been averaged with its neighbours.
Should I sample the darkest corner instead of the brightest?
No — that only moves the problem to the lit area. The key measures distance in both directions, so either extreme leaves the opposite end far away. Sample a mid-tone patch and the worst case is halved.
Why is the edge worse on the shadowed side even after the background is gone?
Those pixels were only partly removed: the soft knee gives them partial alpha instead of full transparency, and 4:2:0 has already averaged their colour with the dark fold beside them, leaving a semi-transparent green-tinted band. Spill suppression reduces it; a re-shoot removes it.