What has to be true before a background can be removed automatically?

Three conditions get quoted: the subject takes no light from the background, the background is a single colour, and the background colour does not appear anywhere in the subject. Measured against the same keyer this site uses, only two of those are independent — spill is not a separate failure, it is the third condition arriving late, because spill injects background colour into the subject. The thresholds you can actually check are these. The subject must contain nothing within 0.259 of the key colour in normalised RGB distance, or it gets keyed out rather than kept — forest green sits at 0.232 and sea green at 0.251, so a green shirt is not a subject to this keyer at all. The background may vary by no more than 21.8% in brightness before residue survives the default settings, and 30% falloff toward a corner is ordinary, which means most setups are already out of spec before anyone touches a slider. Each condition fails differently: broken spill costs 5.83 dE of colour error, broken evenness costs 37.22, and a broken colour condition deletes the pixels outright.

The three conditions, and the one that is not real

The version usually given runs like this: the target gets all its light from white lamps, nothing from the single-colour background, and the background colour is absent from the target. Taken literally that is three requirements. It is really two.

The reason is that the first and the third are the same statement at different times. Spill — light bouncing off the screen onto your subject — does not just tint the subject, it adds the key colour to it. A subject that has taken 6% of its light from a green screen now literally contains green. So the condition about lighting is not a separate test; it is one of the ways the condition about colour gets violated.

That matters practically, because it means one of your checks is redundant and you can spend the time on the two that are not.

The other thing worth knowing up front is that the three do not fail equally, and the differences are not small. Measured as CIE76 colour error against what the pixel should have been:

spill condition broken, 6% — 5.83 dE
evenness condition broken, 30% falloff — 37.22 dE
colour condition broken — the pixels are gone, and there is no error to measure

One of those is a colour correction you can do in a minute. One is a background that will not come out. One is a hole where your subject used to be. They are not the same problem and they do not get found by the same check.

Condition one: the background colour must not appear in the subject

This is the one that destroys footage, because when it fails the result is not a bad edge. It is a missing object.

Normalised RGB distance from key green #00B140, with the default key window running from 0.161 to 0.259. Anything under 0.259 is at best semi-transparent; anything under 0.161 is keyed out completely:

key green itself — 0.000, gone entirely
forest green #228B22 — 0.232, semi-transparent
sea green #2E8B57 — 0.251, semi-transparent
grass green #4CAF50 — 0.305
teal #008080 — 0.316
olive #6B8E23 — 0.456
denim blue — 0.473
black shirt — 0.619
light skin — 1.000
white shirt — 1.224

Read the top of that list again. A forest green or sea green shirt is not a difficult subject for this keyer, it is simply not a subject. It will be keyed out along with the screen, and what you get is a hole with a person's head and hands floating in it. Nothing you do with the sliders brings it back, because raising the tolerance to save the shirt also stops the screen from being removed at all.

The practical version is a wardrobe check, and it is the cheapest of the three: anything in the green-to-teal range goes back in the wardrobe before you shoot, not after.

Condition two: how single-coloured does 'single' have to be?

No screen is lit perfectly evenly. The question is how much variation you can get away with, and that has a number.

A screen that is dimmer in one corner is not just darker — it is a different colour to a colour-distance keyer, and its distance from the key colour grows in proportion to how much dimmer it is. Key green #00B140 sits 0.738 from black, so a 30% falloff puts that corner 0.221 away from what the keyer is looking for. Residue, as a percentage of the pixel left behind where 0% means cleanly removed:

5% falloff — 0% residue at every setting tested
10% falloff — 0%
15% falloff — 0% at default, 30.3% at similarity 26
20% falloff — 0% at default, 27.2% at similarity 34, 68.0% at 26
25% falloff — 24.0% at default, 64.8% at 34, 100% at 26
30% falloff — 61.7% at default, 100% at 34 and below
40% falloff — 100%, nothing is removed

The clean threshold, below which the default settings still take everything: 21.8% brightness variation. Tighten similarity to 34 and your budget drops to 16.4%; at 26 it is 11.0%. Every step you take to rescue an edge costs you evenness you did not have to spare.

And here is the uncomfortable part. A 30% falloff toward the corners is not a broken setup — it is what an ordinary two-light screen does. Which means the common case is already outside the budget, and the reason most people reach for the sliders is that their lighting broke this condition before they started.

Condition three: spill, and why it is the first condition arriving late

Spill pushes the subject's colour toward the key colour, which shrinks its distance from the key colour by exactly the proportion of light it took: d_eff = (1 - s) x d. Once that distance falls below 0.259, the keyer starts treating your subject as background.

How much spill each item tolerates before that happens:

grass green #4CAF50 — 15.0%
teal #008080 — 18.1%
olive #6B8E23 — 43.2%
denim blue — 45.2%
navy — 52.2%
black shirt — 58.2%
dark skin — 59.5%
light skin — 74.1%
white shirt — 78.8%

For most of a normal wardrobe the answer is 'more spill than you will ever have'. Sixty or seventy percent of someone's illumination coming off the screen is not a spill problem, it is a lighting plan. So on its own, spill rarely kills a key.

It matters for the opposite reason. Look at where the numbers get small: the same colours that were already close to the key colour. Grass green was 0.305 away and needs only 15% spill to fall inside the window. Spill does not create a new class of failure, it drags the items that were already marginal over the line. That is why it is not an independent condition — it is the colour condition with a delay on it.

What each failure actually looks like

Worth separating, because they get diagnosed as the same thing and they are not.

Spill is a colour error, 5.83 dE at 6%, and it is the only one of the three you can fix after the fact. It lives in the RGB, not in the matte, which is exactly what spill suppression is for.

Unevenness leaves background behind: 37.22 dE at a 30% falloff, because a chunk of screen is still sitting in the frame at 61.7% opacity. No slider fixes this cleanly, since the control is global and the problem is a corner.

Colour present in the subject is not an error at all. Those pixels are deleted. You cannot measure the difference because there is nothing left to measure.

The order of severity is worth memorising in reverse, because it is the opposite of the order people check in. Everyone checks spill first, because it is visible. It is the least damaging of the three.

A three-step check, run before you shoot

All three conditions are cheaper to satisfy on set than to repair afterwards, and all three can be checked in about two minutes.

1. Wardrobe against the key colour. Anything in the green-to-teal range, or anything that reads close to the screen to your eye, is out. This is the condition that deletes pixels, so it goes first.

2. Measure the screen's falloff. Point the camera at the empty lit screen and compare the brightest area to the dimmest. If the difference is more than about 20% in brightness, you are already outside what the default settings can absorb, and the fix is a light or a flag, not a slider.

3. Look for screen light landing on the subject. Pull the subject away from the screen and light them separately. This is the least urgent of the three, but it is also the one that quietly pushes marginal colours over the edge.

If all three pass, automatic removal is genuinely automatic. If one fails, the failure tells you which tool to reach for — and if it is the colour condition, the tool is a different shirt.

Where these numbers stop

The item colours are representative sRGB values we chose, not samples measured from real clothing. Your sea green shirt is a different sea green. Use the ordering, not the third decimal place.

The falloff model treats the screen as getting uniformly darker, which is what a lighting falloff does. A screen with a colour cast, or one catching light from a different source, moves in other directions as well and this table does not cover it.

Everything here is computed on a single-pixel keyer with one key colour and a fixed window. Software that builds a spatial model of the background, or that lets you sample several key colours, will tolerate more than these thresholds suggest — but the ordering of severity does not change, because a hole is still a hole.