How do you check a cutout before you ship it?

Never check it against the colour you just keyed out. We measured four real defects on six backdrops: a residual green haze at alpha 0.08 scores dE 7.71 against mid grey, 7.65 against magenta, 7.64 against white — and exactly 0.00 against key green #00B140. One defect, one pixel, and the most obvious backdrop erases it completely. Then check the alpha channel directly, because the defect your eyes miss is not the dramatic one: green spill averages dE 4.68 across all six backdrops while a hardened edge averages 38.26, so the thing you are most likely to ship is the thing you are least likely to notice.

The four things that go wrong

Almost every bad cutout that reaches a client has one of four defects, and they are not equally easy to see.

Green spill. Light bouncing off the screen tints the subject itself. The edge is in the right place, the alpha is right, and the colour is wrong. Nothing about the silhouette looks broken.

Residual haze. Pixels that should have gone to alpha 0 keep a thin layer of the key colour. At alpha 0.08 it reads as nothing at all on most screens, until the shot is composited over something bright.

A hardened edge. Semi-transparent pixels — hair, fur, motion blur, fabric fuzz — get flattened to fully opaque or fully transparent. The outline looks tidy and the fine detail is gone.

A dark seam. The colour was already multiplied by alpha once and gets multiplied again on export, so every partial pixel comes out too dark. It looks like a thin dirty outline.

Three of those four change the shape or the brightness. One of them only changes the hue, and it is the one that survives review most often.

How we measured it

Opinions about what looks bad are useless here, so we built a pixel with a known answer and broke it on purpose.

Take an edge pixel: a foreground colour F at alpha 0.5 over a backdrop B. The correct result is F x 0.5 + B x 0.5. Now inject one defect and composite the same way. The distance between the broken result and the correct one, in CIE76 dE, is how visible that defect is on that backdrop. Higher dE means easier to catch.

We ran that for five foreground colours — light skin #E8C4A8, dark skin #8A5A3C, blonde hair #D8C078, denim blue #3B5C7E and a white shirt #F0F0F0 — over six backdrops: white, black, mid grey, the key colour #00B140, magenta and sky blue. Every blend is done in 8-bit sRGB, which is what a browser canvas and most editors do by default. The figures below are the average across the five foreground colours.

The backdrop that hides the defect completely

Here is the number that decides how you should check your work.

Residual haze at alpha 0.08, by backdrop:

mid grey — dE 7.71
magenta — dE 7.65
white — dE 7.64
sky blue — dE 7.18
black — dE 6.37
key green #00B140 — dE 0.00

Zero is not a rounding error. It is exact, and it has to be: the haze is made of the key colour, so compositing it over the key colour returns the key colour. The defect and the backdrop are the same pixels. Any inspection surface made from the colour you removed will always read clean, no matter how dirty the matte is.

This is the trap, because the key colour is the easiest backdrop to reach for. It is sitting right there in the source clip, it is what the subject was shot against, and putting the cutout back on it looks like a fair test. It is not a test at all — it is a way of guaranteeing you see nothing.

The rule that follows from the number: inspect on a backdrop that is not the key colour, and preferably not even close to it.

Which backdrop catches which defect

No single backdrop wins everything. Here is each defect ranked by the backdrop that exposes it best.

Green spill (10 percent of the key colour mixed into the subject):

black — 5.72
mid grey — 5.06
white — 4.69
sky blue — 4.67
key green — 4.25
magenta — 3.70

Residual haze (alpha 0.08):

mid grey — 7.71
magenta — 7.65
white — 7.64
sky blue — 7.18
black — 6.37
key green — 0.00

Hardened edge (alpha 0.5 forced to 1.0):

magenta — 79.64
key green — 48.58
black — 34.54
sky blue — 27.01
white — 21.68
mid grey — 18.08

Dark seam (premultiplied colour multiplied by alpha a second time):

black — 19.58
mid grey — 17.37
sky blue — 16.49
white — 16.17
key green — 16.11
magenta — 15.81

Two things stand out. The hardened edge is by far the loudest defect — it is the one everybody spots, and it is also the only one where magenta is dramatically better than anything else, at 79.64 against 18.08 for mid grey, a factor of 4.4. And the dark seam barely cares what you put behind it: the whole spread from best to worst is 19.58 down to 15.81, a factor of 1.2. For that defect the backdrop hardly matters.

Green spill is the defect nobody catches

Average the six backdrops together and the four defects sort like this:

green spill — dE 4.68
residual haze — dE 6.09
dark seam — dE 16.92
hardened edge — dE 38.26

The most common defect in real chroma key work is also the least visible one, by a factor of 8.2 against the least subtle. That is the whole reason cutouts ship with green in the hair.

It gets worse when you scan the spill strength instead of fixing it at 10 percent. These are dE values on white, black, mid grey and key green:

2 percent spill — 0.94 / 1.14 / 1.01 / 0.87
5 percent spill — 2.34 / 2.86 / 2.53 / 2.15
10 percent spill — 4.69 / 5.72 / 5.06 / 4.25
15 percent spill — 7.03 / 8.57 / 7.59 / 6.31
20 percent spill — 9.37 / 11.41 / 10.11 / 8.32
30 percent spill — 14.03 / 17.01 / 15.12 / 12.18

At 2 percent, every backdrop sits under dE 1.2. You will not see that on a laptop at 100 percent zoom, and you will not see it in a thumbnail. It only becomes obvious somewhere past 20 percent, by which point the subject has been tinted for the entire delivery.

Note also that black is consistently the best backdrop for spill and magenta is consistently the worst — the same magenta that is unbeatable for catching a hardened edge. That is why the answer cannot be one magic inspection colour.

A haze you cannot see until you scan it

Residual haze behaves differently. It does not depend on the subject at all, only on how much alpha was left behind. Scanning alpha against white, black, mid grey and magenta:

alpha 0.01 — 0.95 / 0.76 / 0.96 / 0.94
alpha 0.02 — 1.90 / 1.53 / 1.93 / 1.88
alpha 0.05 — 4.76 / 3.82 / 4.82 / 4.74
alpha 0.08 — 7.64 / 6.37 / 7.71 / 7.65
alpha 0.12 — 11.48 / 11.18 / 11.55 / 11.73
alpha 0.20 — 19.21 / 24.49 / 19.15 / 20.84

The growth is almost perfectly linear in the low range — every 0.01 of leftover alpha is worth roughly one dE. That makes it predictable and, more usefully, makes it checkable: you do not have to trust your eyes, you can read the number. Anything still sitting at alpha 0.20 is not a subtle problem any more, it is a 19 to 24 dE problem, and on black it is the loudest number in the whole scan.

The practical version: if your tool shows an alpha view or lets you sample pixels, sample the corners and the empty areas first. Those are supposed to be exactly zero, and any reading above zero there is haze by definition.

Checkers are not better than flat grey

The checkerboard is the default transparency view in most editors, so it is worth asking whether it actually outperforms a plain 50 percent grey. A checker gives you two inspection surfaces at once, but the defect only has to hide on one of them, so the honest figure is the worse of the two.

green spill — white square 4.69, black square 5.72, checker floor 4.69, flat grey 5.06
residual haze — white square 7.64, black square 6.37, checker floor 6.37, flat grey 7.71
hardened edge — white square 21.68, black square 34.54, checker floor 21.68, flat grey 18.08
dark seam — white square 16.17, black square 19.58, checker floor 16.17, flat grey 17.37

Flat grey wins on haze, 7.71 against 6.37. The checker wins on the hardened edge, 21.68 against 18.08. They tie for practical purposes on spill and on the dark seam. There is no overall winner, which is the point: the checkerboard is convenient, not superior, and treating it as a real inspection is how a dE 6 defect gets approved.

What the checker genuinely gives you is two looks in one pass, and that is worth something when you are scrubbing through two hundred frames. It just is not worth pretending it is a test.

Read the alpha channel instead of trusting your eyes

The three defects that hide best — spill, haze and hardening — are all measurable without compositing anything. The most useful single number is the share of pixels that are genuinely semi-transparent.

We generated a test matte of 20,000 pixels with a realistic profile: 79.92 percent fully opaque, 7.94 percent fully transparent, and 12.12 percent somewhere in between, standing in for hair and soft edges. Then we binarised it the way an over-aggressive choker or a badly exported PNG will, rounding every pixel to 0 or 1.

After binarising, the semi-transparent share is 0.00 percent. Fully opaque goes to 85.92 percent and fully transparent to 14.09 percent. The semi-transparent band does not shrink, it vanishes — 12.12 percent down to nothing.

That is a check you can run in seconds on any tool that shows a histogram or lets you scan pixel values: if a shot with hair in it has almost no pixels between alpha 0.02 and 0.98, the edge has been cut hard, no matter how clean it looks on the checkerboard. Be careful with the honest caveat, though — the 12.12 percent figure is a synthetic distribution we built for this test, not a measurement of a real photograph. What transfers is the shape of the signal, not the number: on real footage, expect the semi-transparent share to be small, and treat a reading of zero as a failure rather than a success.

A two-minute checklist

In the order that catches the most damage per second:

Never review on the key colour. The haze you are looking for scores exactly dE 0.00 there. Use mid grey or white as the default.
Sample the emptiest corner. It should read alpha 0.000. Every 0.01 of leftover alpha is worth about one dE once it is composited.
Count the semi-transparent pixels. Hair and motion blur should produce a real population between alpha 0.02 and 0.98. Zero means the edge was cut hard.
Then look at the edge on magenta. It is the best single surface for a hardened edge at dE 79.64, more than four times better than mid grey.
Then look at the subject on black. Black is the best surface for spill at dE 5.72, and it is also where a dark seam shows up first at 19.58.
Finally, composite on the real background. Not as a test — as a confirmation. By then everything that could hide has already been caught.

The ordering matters more than the individual steps. Eyes first, numbers second is the wrong way round for this job, because the defect that survives eyes is the defect that is most common.