When should you shoot on black and luma key instead of a green screen?
Switch to black when the backdrop is lighting your subject rather than just sitting behind it - glass, water, smoke, chrome, anything see-through or mirror-bright. Measured on this keyer: on a lit green screen 87.5% of a glass tumbler's radiance is the screen shining through it, so the colour key is being asked to delete the light that made the subject visible, and it removes 69.7% of the glass while leaving the survivors carrying 48.6 levels of green. Put the same glass against a matte black backdrop and the screen's share falls to 11.3%: the backdrop clears at 100%, no glass pixel is deleted, and composited additively instead of over, the error drops from 27.3 levels to 10.3.
Your backdrop is a lamp, not a colour
A lit backdrop does two jobs at once. It is the patch of colour you intend to remove, and it is the largest light source behind your subject. For an opaque subject only the first job matters. For a glass tumbler the second one dominates. Traced through the glass in this scene - transmission through the walls, plus Fresnel reflection off the surface - 87.5% of everything the camera receives from the body of the glass is the screen itself: 0.381 of transmitted radiance against 0.055 of reflection. The body reads (30,174,69), which is the screen's green with a little glass in front of it.
Put the same glass in front of a matte black backdrop - 4.5% reflectance, catching only a twelfth of the light that reaches the subject - and the split inverts. The backdrop reads 15.5 code, transmission contributes 0.004 against 0.034 of reflection, and the screen's share falls to 11.3%. The body now reads (37,37,37): dark, neutral, and made almost entirely of the room around it. That is what you are choosing - not a different colour to delete, but the removal of a light source.
Why the colour key cannot win that fight
Run the green plate through this keyer at its defaults - tolerance 42, edge softness 14, spill 35, despill 10 - and the screen itself disappears perfectly: 0.00% of it survives. The glass does not fare so well. 69.7% of it is deleted outright, what remains averages 0.236 alpha, and the pixels that do survive are still carrying 48.6 levels of green above their own red and blue. Dropped over a dark new background the result sits 27.3 levels from the truth, and 58.9 over a white one.
The reason is structural, not a matter of tuning. The keyer writes one number per pixel and leaves the colour alone except to limit the green channel. The green in those survivors is not a rim around the glass, it is the glass: it arrived as light, passed through the object, and was recorded as the object's colour. No tolerance setting separates them, because at that pixel the screen and the subject are the same photons.
What black actually buys you
Point the key at black instead and the picture flips. The backdrop comes out 100.0% removed, not one glass pixel is deleted, and the matte averages 0.881 alpha - mostly partial, which is roughly what a glass tumbler ought to be. The separation is genuine: the backdrop sits at 15.6 code and the body at 37, and this keyer's default threshold throws away everything below 23.7 code.
The catch is that alpha is the wrong tool for the second half of the job. The plate recorded what the glass adds to a dark room, not what it transmits, and source-over compositing treats it as an opaque grey object: 30.7 levels of error over a dark background and 215.6 over a white one, both worse than the green plate you started from. Good matte, wrong colour model.
Composite it additively, because that is what you shot
A plate shot on black is already premultiplied by an absence of light: what the camera recorded is the light the subject contributes. Add it to the new background instead of laying it over, and the same frame lands 10.3 levels from the truth on a dark background instead of 30.7 - better than the green-screen route managed by a factor of 2.7. Screen blending does slightly better again at 9.3.
The trade is that additive can only add. Measured against backgrounds of 24, 64, 128, 200 and 235 code, additive error runs 8.6, 10.9, 17.5, 25.9 and 29.8, while source-over with the same alpha runs 7.6, 40.2, 103.6, 178.6 and 215.6. Below about 25 code the two are equivalent; above it additive wins by more at every step, since its error scales with how bright the plate is. This tool composites only with source-over, so the add has to happen downstream.
The black preset is a radius, not a luma key
This keyer scores a pixel by its normalised RGB distance from the key colour, so pointing it at black leaves you measuring the radius from zero, with all three channels weighted equally. A Rec.709 luma key weights green 0.7152 and blue 0.0722 - a different instrument. On neutral tones the radius is the stricter of the two: it discards everything below 23.7 code, where luma discards below 41.1.
On colour they disagree outright. A deep blue at (10,10,90) sits 0.357 from black and is kept, while its luma is 0.062 and a luma key deletes it - 5.8x apart. Navy differs by 2.8x, dark red by 3.0x, and a skin shadow at (40,28,24) by 1.8x. So this tool's black key is the more forgiving choice for dark blues and dark reds, and a luma key in an editor will not agree with it. One thing to switch off: spill suppression only fires when green beats both red and blue, which on a black plate is your subject's own colour. Left at 35, it cost a green bottle 1.2 levels of that green and removed nothing.
How dark the backdrop has to stay, and where the window is
Matte black is only black until your subject light reaches it. At 4.5% reflectance the backdrop reads 15.5 code when it catches a twelfth of the subject illumination, and the default tolerance clears it completely - 35 would be enough. Let a quarter through and it reads 26 code, only 34.2% of it clears, and you need tolerance 51. At 40% spill it reads 35 code, nothing clears, and it takes 63; at full subject light it needs 97.
That leaves a usable window - narrower than the green-screen one, but comfortable. The backdrop is gone from tolerance 33, the glass stays fully solid to 37, and the first glass pixel is deleted at 55. The default 42 sits in the middle of it: backdrop clean, nothing deleted, matte averaging 0.881. Push to 60 and you delete 62.8% of the glass. The practical rule is to keep the backdrop under about 20 code: flag it, or set it far enough back that it catches almost none of the subject light.
When black wins, and when it cannot work
Black wins when the subject is made of light rather than of paint: glass, water, smoke, sparks, flames, chrome, anything whose appearance is what it does to the light passing through it. It loses the moment the subject is dark and matte. A matte black plastic sphere under the same key light loses 10.3% of its pixels outright and softens another 12.4% on black, against 0.0% and 0.0% on green - its shadow side reads 25 code, close enough to the 15.6 backdrop for the keyer to treat the two as one thing.
It also loses anything whose colour comes from transmission. The glass body reads (37,37,37) on black and (30,174,69) on green: on black you keep the shape and the highlights and give up the tint, which you then put back in the composite. And it is harder on your camera, because the subject now lives near the bottom of the range where noise bites hardest. At 8 levels of noise, 14.8% of glass pixels swing by more than half an alpha between two frames and 4.93% of the backdrop stops clearing; at 4 levels those fall to 2.86% and 0.08%. Light the subject harder than you think you need to.
Related questions
Can I luma key in this tool, or do I need a different one?
There is a black swatch in the key-colour presets. Clicking it sets the key to #000000 and the keyer then measures each pixel's distance from black, which is a radius with all three channels weighted equally. A Rec.709 luma key weights green far more heavily than blue, so the two are not interchangeable: on neutrals the radius is stricter (it deletes below 23.7 code where luma deletes below 41.1), and on dark blues and reds it is much more forgiving - a deep blue at (10,10,90) is 5.8x further from black by radius than by luma. If your editor's luma key disagrees, that is why.
How do I composite a plate that was shot on black?
Add it, do not lay it over. The camera recorded the light the subject contributes to a dark room, so adding that to the new background reconstructs the shot while source-over with the keyed alpha paints a grey object instead: 10.3 levels of error against 30.7 on a dark background, and 29.8 against 215.6 on a white one. Screen blending measured slightly better again at 9.3 and 24.5. This tool's own background tabs composite with source-over only, so the add has to happen downstream.
How dark does the backdrop really have to be?
Under about 20 code, which is darker than most people expect from something that already looks black. A matte black at 4.5% reflectance reaches 15.5 code when it catches about a twelfth of the subject light, and the default tolerance clears that completely. At a quarter of the subject light it reads 26 code and needs tolerance 51; at 40% it reads 35 code and needs 63. Flag the lights, or move the backdrop back, rather than raising tolerance.
Should I shoot people on black instead of green?
Almost never. A person is opaque and lit from the front, so the backdrop is not inside them and a colour key separates cleanly. The failures are at the dark end: matte black plastic lost 10.3% of its pixels and softened another 12.4% against black while losing nothing against green, and dark clothing or dark hair behaves the same way. Black is for subjects made of light - glass, water, smoke, sparks.