Why does a clean key still look fake once you drop in a new background?

Because 'clean' only means the matte is correct. Every pixel's alpha is right, and the light is still wrong in four separate ways, none of which alpha can fix. The edge is too hard: with the default smoothness of 14, a light-skin edge against a sharp original gets about 0.25 px of semi-transparent transition, where a real lens would give you two or more, and turning smoothness up to maximum only reaches 1.75 px. The background light that should wrap around the outline is missing entirely — on light skin against a deep blue background, leaving out 15% wrap costs 2.84 dE. Some of the old screen is still in the subject's colour: 1% residual costs 0.94 to 1.23 dE, and 5% costs up to 6.22. And the exposure was never matched, because in a real shot the environment sets both the subject's brightness and the background's, while in a composite nothing does — a white shirt dropped onto deep blue is 2.49 stops out. Alpha is one number per pixel. Lighting is four more.

A clean key proves one thing, and it is not the thing you are looking at

The moment you say a key is clean, you are making a statement about alpha: every pixel that should be background has been removed, and every pixel that should be subject has been kept. That is a real achievement and it is also a small one, because alpha is one number per pixel and it says nothing at all about colour, light, or geometry.

What makes a composite read as fake is almost never alpha. It is that the subject carries the lighting of the room it was shot in, the edge is sharper than any camera would have produced, and the background has not been allowed to touch the subject the way a real background does. None of that is visible while the subject is still on the green screen, which is why it only shows up after the swap, when it is too late to fix at the shoot.

Below are the four that can be measured, with the number for each. Measured on the same keyer this site runs: normalised RGB distance, similarity 42, smoothness 14.

The edge is harder than any lens would have made it

A keyer produces semi-transparency over a fixed distance in colour, not over a fixed number of pixels. With smoothness at 14, that window is 0.098 wide in normalised RGB distance. How many pixels it spans depends on how fast the colour changes across your edge — which is to say, on how sharp your original footage is.

If the original edge transitions across N pixels, and the subject sits at distance d from the key colour, then the width of the soft band comes out at 0.098 x N / d pixels. For a sharp original, where the edge crosses about 2.5 pixels:

white shirt (d = 1.224) — 0.20 px
light skin (d = 1.000) — 0.25 px
brown hair (d = 0.615) — 0.40 px
denim blue (d = 0.473) — 0.52 px

A quarter of a pixel is not a soft edge, it is a hard one with a rounding error. Your original footage had a transition two and a half pixels wide; the keyer has thrown most of it away.

Turning smoothness up helps, and not enough:

smoothness 14 (default) — window 0.098, band 0.25 px
smoothness 30 — window 0.210, band 0.53 px
smoothness 50 — window 0.350, band 0.88 px
smoothness 70 — window 0.490, band 1.22 px
smoothness 100 — window 0.700, band 1.75 px

Even at maximum you do not reach the two pixels your camera would have given you for free. So the softness has to come from somewhere else: either shoot softer to begin with, or feather the matte in your editor afterwards. Blaming the keyer for a hard edge is like blaming the scissors.

The background never got to wrap around the outline

Stand in front of anything and the light coming off it lands on you, most heavily at your outline. That is light wrap, and it is a large part of why a real object looks like it is in a place rather than pasted onto one.

Compositing does not create it. An alpha blend gives you subject x a + background x (1 - a) and stops there, so the outline is missing the background's light entirely. Colour error from that omission, at a coverage of 0.5, for a light-skin subject:

deep blue background — 2.84 dE missing 15% wrap, 5.71 missing 30%
warm orange background — 1.75 dE, 3.51
white background — 1.04 dE, 2.07

The pattern is not about how bright the background is, it is about how different it is from the subject. Deep blue against light skin is the biggest gap and therefore the biggest error; white against light skin is a small gap and the error nearly vanishes. The composite that looks worst is the one where you put a pale subject on a dark, saturated background — which is also the most common thing people want to do.

The fix is to add it back, and the number to start from is 15%: blend that much of the background into the outermost pixels of the matte, weighted by how transparent they already are. It is a colour operation on the edge, not an alpha operation, so no keyer setting will do it for you.

Some of the old screen is still in your subject

A keyer separates pixels. It does not clean them. Subject pixels that picked up light from the screen keep that light, and once the screen is gone there is nothing to explain the green that is still sitting in someone's hair or on the shoulder of a white shirt.

Colour error against what the subject should have been, by how much of the old screen is still in there:

brown hair — 1% residual 1.23 dE, 2% 2.47, 5% 6.22, 10% 12.48
denim blue — 1.14, 2.27, 5.68, 11.31
light skin — 0.97, 1.94, 4.86, 9.74
white shirt — 0.94, 1.89, 4.72, 9.48

One percent is already measurable and five percent is obvious. The reason this survives a 'clean' key is that it lives in the RGB, not in the alpha — you can have a perfect matte and a subject that is still faintly green.

This is what spill suppression is for, and it is worth saying what it costs: it is a colour operation applied to the whole subject, so pushing it hard will start changing colours that were never contaminated. Suppress until the fringe goes, not until the numbers look tidy.

Nobody matched the exposure

In a real frame the room lights both the subject and the background, which means their brightness is related by something more than chance. In a composite, one of those came from a studio and the other came from a stock library, and nothing reconciles them.

Difference in stops between subject and background, as log2 of their Rec.709 luma:

white shirt on deep blue — 2.49 stops
brown hair on white — 2.32 stops
light skin on deep blue — 2.21 stops
light skin on warm orange — 0.42 stops
white shirt on white — 0.06 stops

Two and a half stops is not a subtle mismatch, it is a subject that has been dropped into a different time of day. Brightness is also the first thing the eye reads, well before it reads colour, so this one outs a composite faster than any fringe.

Match the two before touching anything else, and match them by the brightest meaningful part of each, not the average. A background whose highlights stop well below the subject's will never sit right no matter how good the matte is.

What to actually do, in order

Roughly cheapest first, and roughly the order in which each one becomes visible.

1. Match exposure and colour temperature. Two and a half stops is the difference between a composite and a photograph. Nothing else on this list matters until this is close.

2. Put the background's light back on the outline. Around 15% of the background colour blended into the outermost pixels, weighted by transparency. Biggest win on dark, saturated backgrounds.

3. Soften the edge in the editor. The keyer tops out at 1.75 px and you want at least two. A one or two pixel feather on the matte is not a hack, it is restoring what the lens had.

4. Suppress what is left of the old screen. Enough to kill the fringe, not enough to flatten the subject.

5. Add the contact shadow. Not measurable here, and not optional. An object with no shadow and no contact darkening is floating, and no amount of correct colour fixes floating.

Where these numbers stop

Two inputs in this article are parameters rather than measurements, and they are labelled as such so you know what to substitute.

The first is N, the number of pixels your original edge transition spans. We did not measure it — it depends on your lens, your aperture and your shutter. The table gives 1.5, 2.5, 4 and 8; take the one that matches your footage.

The second is the wrap strength, modelled as p. How much light a real background actually wraps onto a real subject depends on how bright it is and how close it is, and we did not measure that either. 15% is a starting point, not a physical constant.

And none of this is a perceptual test. These are colour and geometry measurements on a model; they tell you what changed and by how much, not whether a given viewer will notice. The stop differences and the pixel widths are the ones worth trusting, because those are arithmetic.