Why do wrinkles ruin a chroma key faster than almost anything else?
Because a crease turns a smooth, gradual change across your screen into a local extreme: a deep shadow sitting a few pixels away from cloth that is lit normally. On the flat, well-lit screen we modelled, the whole spread was 0.29 stops and the default key removed every pixel of it; adding a single 45-degree crease pushed the darkest point to 1.11 stops and forced the required tolerance from 33 to 79, which is past what the subject survives. A fold is also asymmetric and superlinear — the shaded flank fell to 0.702 of reference brightness while the lit flank only reached 0.891 — and a 60-degree crease landed at 1.96 stops, further than the tolerance slider can reach at all.
What a wrinkle does that a soft falloff does not
We modelled a 3.0 m screen lit by four lights 4.0 m back, and shaded it from surface normals, so a patch turned away from a light loses exactly what the geometry says it should. On a perfectly flat screen the brightness ran from 1.000 at the point you would eyedropper down to 0.817 in the corners — a spread of 0.29 stops. At the default tolerance of 42 the key removed all of it: 0.00% of the screen left behind, and every part of the subject intact.
A wrinkle changes the kind of problem it is. Falloff is smooth: neighbouring pixels differ by hundredths of a stop, so one radius comfortably covers a wide band of them. A crease is a discontinuity — cloth at the reference exposure, and a few pixels later cloth that is half a stop down. A single radius cannot cover a region and its neighbour when the gap between them is larger than the radius itself, which is why the advice on production Q&A sites has been consistent for years: watch out for large wrinkles, because they ruin the key quickly.
The cost grows faster than the fold gets deeper
We added one gaussian crease and swept its angle. The tolerance the whole frame then needed was 33 flat, 46 at 25 degrees, 79 at 45 degrees, and at 60 degrees no slider position could clean it at all. The darkest point on the screen went 0.29 stops, to 0.52, to 1.11, to 1.96.
That is not a straight line, and it is not an artifact of the measurement. The tilt of the cloth enters through the cosine of the incidence angle plus the tilt, so as the fold deepens, each extra degree costs more than the one before it. Going from 25 to 45 degrees — under twice the angle — moved the requirement from 46 to 79, and the residual left behind at the default settings roughly doubled, from 1.52% to 3.00%.
Why one flank is worse than the other
A fold has two sides and they do not behave alike. Sweeping the same crease from 10 to 60 degrees, the flank that turned toward the lights moved only from 0.924 to 0.832 of reference brightness. The flank that turned away went from 0.873 all the way down to 0.575 — a distance from the key colour of 0.314 against the bright side's 0.124.
The physics is that the cosine curve is flat near zero and steep further along. Turning a surface toward a light that is already roughly frontal buys you almost nothing; turning it away costs you a great deal. So a crease does not average out to the same brightness it replaced. Its mean drops, and the average screen colour you eyedroppered is no longer the average colour of the cloth.
What you need against what your subject survives
Every point of tolerance you spend on the cloth is spent on the subject too, because the radius grows in every direction at once. On this scene the subject's hair, carrying green spill, held full opacity up to similarity 53 and then fell apart. The 45-degree crease needed 79. There is no overlap between those two numbers.
Running the frame at 79 showed exactly what that costs: the screen came back clean at 0.93% residual, but spill-contaminated hair went from alpha 1.000 to 0.000 — all 5,588 pixels of it, 8.70% of the subject punched into holes — and the semi-transparent rim lost half its mass, dropping from 3,738 to 1,906. A single crease had consumed the entire budget the subject needed.
One crease is a local problem, and that is the good news
The remedy follows from where the damage sits. Splitting the frame into 32-pixel blocks and giving each its own tolerance, the flat screen needed between 18 and 35. With the 45-degree crease in place, the worst block anywhere needed 86 — but the blocks covering the subject still needed only 22, unchanged. The crease was somewhere else, so the aggressive setting never touched the person.
The result was a full recovery: 0.11% residual, hair back at 1.000, no holes, and rim mass at 4,285 against 1,906 for the single global setting. A crease you can draw a box around is a crease you can fix, because the fix stays where the crease is.
A crumple is not local, and that is the bad news
Now crumple the whole cloth instead. With a corrugation running at a 20-pixel period, the default key left 14.21% of the screen behind against the crease's 1.52%, and the global requirement rose to 57, which already cost the hair over a third of its opacity. Unlike the crease, the damage was everywhere at once: 1,427 separate broken runs across the frame against 31 for a single fold.
Zoning no longer rescues it cleanly, because every region now contains both a trough and a crest. The per-block tolerance over the subject rose from 22 to 44, and although the hair came back to full opacity, the rim still measured 3,305 against the crease's untouched 4,285 — a 22.9% loss that no arrangement of regions recovers. At a 60-pixel period it was worse, at 3,190.
What to do about it
Flatten the cloth before you light it. The flat screen in this comparison needed 33 and lost nothing at all, so every stop you take out of the fabric is budget handed straight back to the subject. Pulling the lights back and lighting from both sides is what made that flat case 0.29 stops rather than one, and it also softens creases, since a flank turning away from one source is still facing another.
Match the fix to the defect. One crease gets a bounded patch, and the measurement says it will cost you nothing. Crumple gets steam, tension or a reshoot, because no regional setting fixes it and every region pays for it. And a deep fold is simply gone: at 60 degrees the shaded flank sat 1.96 stops down, while the slider's maximum radius of 0.500 covers 1.632 stops. No setting on this page reaches it, and the only real repair happened before the camera started rolling.
Related questions
Can't I just raise the tolerance until the wrinkles disappear?
Not without paying for it elsewhere. A 45-degree crease needed similarity 79, while the subject's spill-contaminated hair on the same frame held full opacity only up to 53. At 79 the hair went to alpha 0.000 and the semi-transparent rim lost half its mass, so the creases went away and the edge went with them.
Why does only one side of a crease go dark?
Turning a surface toward a light that is already frontal gains very little, while turning it away loses a lot, because the cosine curve is flat near zero and steep further along. Measured across a 10 to 60 degree sweep, the flank facing the lights moved only from 0.924 to 0.832, while the shaded flank fell from 0.873 to 0.575.
Is a fine crumple worse than one big crease?
Yes, measured two ways. A 20-pixel corrugation left 14.21% of the screen behind at the default settings against 1.52% for a single 25-degree fold, and it broke the frame in 1,427 separate places against 31. Even after per-region keying it cost 22.9% of the rim, where the single crease cost nothing.
Is there any setting that rescues a really deep fold?
No. At 60 degrees the shaded flank sits 1.96 stops below the reference exposure, and the tolerance slider's maximum radius of 0.500 corresponds to 1.632 stops. The distance involved is beyond anything the control can produce, so a fold that deep has to be fixed physically — steamed, stretched or shot again.