Does 10-bit 4:2:2 beat an expensive camera for green screen keying?

No, and the more useful finding is that neither one is what is holding your key back. We ran a scene with a known alpha through a full camera pipeline — Bayer mosaic, shot and read noise, demosaic, quantisation, chroma subsampling — and then through the keyer. Going from 8-bit 4:2:0 on a cheap body to 10-bit 4:2:2 on an expensive one improved overall alpha error from 0.0137 to 0.0133, about three percent. The edge error, 0.266, barely moved at all, because it comes from the keyer's own fixed-slope alpha model and no camera setting touches it. Chroma resolution is the only camera-side thing that measurably helps: on 2.3 px hair strands the error runs 0.1651 at 4:2:0, 0.1618 at 4:2:2 and 0.1461 at 4:4:4. Bit depth does almost nothing — 8-bit already puts 24 quantisation steps across the keyer's 0.098-wide window. And when the screen falls off by 30%, 33.57% of it survives with the cheap body and 33.58% with the expensive one. Buy lights, not a body.

With a flat screen, nothing you can buy matters

The scene is a body with a one-pixel anti-aliased edge, a ring of roughly 2.3 px hair strands, and 6% green bounce. Two camera profiles: a cheap body with a 4000-electron full well and 4 electrons of read noise, and an expensive one with 24000 electrons and 2 electrons of read noise. Both go through the same pipeline and the same keyer window.

cheap + 8-bit 4:2:0 — background left 0.37%, edge error 0.2659, strand error 0.1651, overall 0.0137
cheap + 8-bit 4:2:2 — 0.27% / 0.2663 / 0.1618 / 0.0134
cheap + 10-bit 4:2:2 — 0.27% / 0.2661 / 0.1616 / 0.0134
cheap + 10-bit 4:4:4 — 0.02% / 0.2591 / 0.1461 / 0.0122
expensive + 8-bit 4:2:0 — 0.36% / 0.2648 / 0.1639 / 0.0136
expensive + 10-bit 4:2:2 — 0.25% / 0.2646 / 0.1604 / 0.0133

Reduce that to the two purchases people argue about. Adding 10-bit 4:2:2 to the cheap body improves overall error by 2%. Swapping the cheap body for the expensive one, holding the pipeline at 8-bit 4:2:0, improves it by 1%. Both are rounding errors next to the edge error of 0.266, which does not move.

That edge error is not noise and it is not compression. It is the keyer's own alpha model, the fixed-slope ramp we measured in our piece on hand masking: the alpha saturates long before coverage reaches 1, so the anti-aliased edge gets crushed. Every configuration above sits at 0.258 to 0.266 regardless of what the camera did.

Bit depth: 24 steps is already plenty

8-bit quantisation divides the range into steps of 1/255, which is 0.0039 in normalised RGB. The keyer's window is 0.098 wide. That is about 24 steps across the window at 8-bit and about 100 at 10-bit, and 24 turns out to be enough.

Measured on the strands, holding everything else fixed:

8-bit 4:2:2 — strand error 0.1618
10-bit 4:2:2 — strand error 0.1616

A tenth of a percent. The reason is that quantisation error is bounded and small compared with the window, whereas the keyer's own systematic error is not. This is also why 8-bit footage can look banded in the sky and still key perfectly — banding is a display problem and keying is a threshold problem.

Where bit depth does earn its keep is everything you do after the key: raising exposure, pushing saturation, grading the subject. We tested lifting a frame that was two stops under: the noise on the background's distance measurement went from 0.11 to 0.20 of the window, and surviving background went from 0.37% to 0.41%. Real, but still small.

Chroma resolution is the one that shows up

This is the camera-side setting with a measurable effect, and it shows up exactly where you would predict — on chroma detail.

4:2:0 — strand error 0.1651
4:2:2 — strand error 0.1618
4:4:4 — strand error 0.1461

The same ordering appears in leftover background once the screen is not perfectly flat. At 20% falloff across the frame, on the cheap body:

8-bit 4:2:0 — 0.71% of the background survives
8-bit 4:2:2 — 0.53%
10-bit 4:2:2 — 0.52%
10-bit 4:4:4 — 0.28%

Note which pair of rows is identical. Going 8-bit to 10-bit moved it by 0.01 points. Going 4:2:2 to 4:4:4 halved it. If your recorder offers you one choice, take chroma resolution over bit depth.

The sensor: six times the photons, one percent better

The expensive body in this model catches six times as many photons and has half the read noise. That is a generous gap — bigger than most body-to-body upgrades — and it buys almost nothing at the key.

8-bit 4:2:0: cheap 0.0137, expensive 0.0136
10-bit 4:2:2: cheap 0.0134, expensive 0.0133

The reason is the threshold. On a flat, cleanly lit screen the keyer's cut-off sits about fourteen standard deviations above the noise floor, so noise cannot flip a background pixel. A camera only starts to matter once pixels are already sitting near that threshold — which is to say, once the screen is uneven. And at that point the failure is systematic, not noise-driven.

The cliff is the screen, not the camera

Here is the test that settles it. Take the screen and let it fall off across the frame, which every real screen does, and measure how much background survives at 10-bit 4:2:2:

0% falloff — cheap 0.27%, expensive 0.25%
10% — cheap 0.31%, expensive 0.30%
20% — cheap 0.52%, expensive 0.52%
30% — cheap 33.57%, expensive 33.58%

Somewhere between 20% and 30% the dark side of the screen crosses the threshold and a third of it stops being removed at all. That cliff is the same one our conditions piece found from a different direction, at 21.8% of brightness variation. And at the cliff, the expensive camera is worth 0.01 percentage points. It is not that the better sensor helps a little; it is that the failure has stopped being about noise entirely.

We also tried tightening the keyer window from 0.098 to 0.042, which is what you have to do to sharpen an edge. Surviving background actually went down, from 0.37% to 0.14%, and the strand error rose from 0.1651 to 0.1662. Noise sensitivity went up and the result still did not depend on the camera.

So what should you spend money on

In the order this data supports:

Light, first. The cliff between 20% and 30% falloff is a bigger effect than every camera setting combined. Flatness of the screen is the purchase that pays.

Chroma resolution, second. 4:2:2 over 4:2:0, and 4:4:4 if your recorder offers it. On fine strands it is worth about a tenth of the error, and on a falloff screen it roughly halves what survives.

Bit depth, third, and mostly for the grade. It does not help the key. It helps the ten things you do to the image afterwards, especially any lifting or saturation push.

The body, last. Six times the photons bought one percent. There are plenty of good reasons to own a better camera. Getting a cleaner key out of an evenly lit screen is not one of them.

And none of this is a substitute for the thing that actually dominates: the keyer's own alpha model. An edge error of 0.266 is present in every single configuration we ran.

Where these numbers stop

This is a simulation, not footage. The camera is one synthetic RGGB pipeline with two arbitrary noise profiles; real sensors differ in quantum efficiency, in chroma noise correlation, in whether they bin or skip lines, and in how their codecs behave before the file ever reaches you.

The expensive body is modelled as more photons and less read noise and nothing else. A real expensive camera also brings better glass, better internal processing and often a cleaner output path, none of which is in these figures.

The screen is a flat colour with a linear brightness falloff. Real screens have wrinkles, seams, sheen and bounce from the room, all of which our wrinkles and distance pieces treat separately and all of which make the falloff numbers worse, not better.

The scene has one subject colour and one strand geometry, rendered at 400 by 400 with 150 angular cycles giving about 2.3 px strands. Finer strands would push the chroma-resolution result further in 4:4:4's favour.

Finally, the keyer is the one running on this page with its default window. A keyer with a narrower window, or with noise-aware filtering, would shift how much noise matters — but its own alpha model would still be the largest term.