Why do automatic erasers fail on see-through objects?
Because a glass pixel is mostly the screen showing through, not an object standing in front of it. This keyer only reacts once something has cut the light behind it to 59 per cent or less, and only keeps it whole below 40 per cent; clear glass passes about 93 per cent, so at the default settings 69.9 per cent of a tumbler is deleted outright and only the hard highlight and the thick rim survive. The AI overlay cannot rescue it either — it merges as a union that can only add opacity, so the best it can do is make whatever is left fully opaque and show the green it transmitted at full strength, which more than doubles the colour error.
A glass pixel is mostly the screen
Put a clear tumbler in front of a green screen and ask what the camera actually receives at a pixel in the middle of it. Most of that light left the screen, crossed the room, went through both walls of the glass, and arrived slightly dimmed and slightly tinted. The rest is the glass surface mirroring whatever stands around it. There is no layer of glass colour in there to find, which is the first reason an eraser has nothing to grab.
Measured on a tumbler with a ten-pixel wall, the thinnest part transmits 93.1 per cent of the screen and reflects 5.0 per cent. Written as a screen multiple, that pixel is the key colour at 0.968 — it is closer to the screen than most of the screen is to itself. Everything downstream follows from that one number.
How dim it has to get before the key reacts
Take the pure case first: a pixel that is the screen, dimmed by a factor c, with nothing else added. Because the encoding curve compresses dark values, dimming the green channel by a third barely moves its code value. At 0.95 the pixel reads (0, 173, 62) against the key's (0, 177, 64), a distance of 0.0171 — the keyer cannot tell it from the screen at all.
It first begins to react at a screen multiple of 0.590, and it only keeps a pixel whole below 0.400. Run that over a real object and the sweep agrees: a tumbler transmitting 0.95 loses 69.9 per cent of its pixels outright, 0.80 loses 67.6 per cent, 0.60 loses 26.0 per cent, and it is not until about 0.45 that nothing disappears. Anything that passes more than roughly half the light behind it is not representable here.
What survives: the highlight and the rim
What is left is not the object, it is the two places where the object stops behaving like a window. The specular highlight off the key light is the first: it reflects the lamp rather than the screen, and 17.8 per cent of the tumbler stays fully opaque there. The second is the rim, where the surface turns away from the camera and reflectance climbs to 20.9 per cent, giving an alpha of 0.749 against 0.000 across the whole body.
So the result is exactly the failure people describe: a bright slash of highlight and a thin bright outline, floating with nothing between them. It reads as a broken matte rather than as a glass, and no amount of feathering puts the middle back.
The tint is not the problem, the transmission is
It is tempting to blame the colour of the glass. The measurement says otherwise. A green bottle glass, whose absorption curve sits right on top of the key colour, loses 70.3 per cent of its pixels; amber loses 69.0 per cent; ordinary clear glass loses 69.9 per cent. The hue changes almost nothing, because a tint shifts the pixel along the same direction the key already measures.
What changes the outcome is how much light gets through at all. Smoke the same tumbler down to 0.70 transmission and nothing disappears — the object holds together at a mean alpha of 0.396. Fill it, frost it, sleeve it, put a label on it: the fix is optical density, not a different colour of glass.
Why both of the obvious controls make it worse
Opening similarity eats what is left. On a screen lit to 0.735 stops of spread, 30 is as low as you can go before the background starts returning — 26 leaves 1.4 per cent of it — and even there 51.8 per cent of the glass is still deleted. At the default 42 it is 69.9 per cent, at 63 it is 83.3 per cent, and at 100 it is 93.4 per cent. There is no setting on this axis that keeps the glass and the screen at the same time.
The AI overlay is no better, for a structural reason. It combines with the chroma alpha as a union, taking whichever is greater, so it can only add opacity. Given an ideal mask covering the whole object, nothing goes missing and the mean alpha goes to 1.000 — and the colour error rises from 28.7 to 63.8 levels, because the glass is now fully opaque and the screen it transmitted is being shown at full strength. You get a solid green-tinted bottle instead of a hole.
Where spill suppression is genuinely the right tool
This is one of the few subjects where the despill control is doing what you want. The green inside a glass is not reflected light baked onto a surface the way it is on skin; it is background that came straight through, and it should be replaced by whatever you composite behind. Pushing spill to 100 takes the composite error from 34.8 levels to 22.6 on the bare screen setup, and from 46.4 to 21.8 once the object has been made visible.
It is still a blunt instrument. It clamps the green channel to the larger of red and blue, so it does not know the difference between screen that came through and colour that belongs to the object, and the pixels it touches move a long way — an average of 102 levels at full strength. Use it here, but expect to pay for it somewhere else in the frame.
What works, in order of effort
Cheapest: stop the object mirroring the screen. The share of its reflected hemisphere filled by green is a setup variable, and it moves the result hard — at 75 per cent, 69.9 per cent of the glass vanishes; at 55 per cent, 61.8 per cent; at 35 per cent, nothing vanishes and the mean alpha is 0.473; at 5 per cent it is 0.884. Flags, cards, or simply moving the object forward off the screen do this.
Next: kill the transmission. A card directly behind the glass makes it fully opaque to the key, and the darker the card the less you have to undo afterwards — measured composite error of 39.8 levels for a black card, 62.3 for mid-grey and 249.2 for white, because the card's colour is what gets baked into the object. Failing all of that, raise the optical density until it passes less than about half the light behind it, or hand-mask it.
Related questions
Can any automatic tool key a glass object out cleanly?
Not from a single colour measurement. This keyer begins to react only when an object has dimmed the screen to 0.590 of its value, and keeps it whole only below 0.400. Clear glass transmits about 0.93 at its thinnest, so 69.9 per cent of a tumbler is deleted outright at the default settings.
Is green-tinted glass worse than clear glass?
Barely. Measured, a bottle-green glass loses 70.3 per cent of its pixels, clear glass 69.9 per cent and amber 69.0 per cent. The tint shifts the pixel along the same direction the key already measures; what decides the outcome is how much light it blocks, and smoking it down to 0.70 transmission is enough to keep all of it.
Should I raise similarity or switch on AI subject mode?
Neither. Similarity 30 is the floor set by the screen — 26 lets 1.4 per cent of the background back — and it still deletes 51.8 per cent of the glass; at 100 it deletes 93.4 per cent. The AI overlay merges as a union that can only add opacity, so it makes the object solid and shows the transmitted green at full strength, raising the colour error from 28.7 to 63.8 levels.
What is the cheapest thing to change on set?
Get the screen out of the object's reflection. Dropping the green share of what it mirrors from 75 per cent to 35 per cent takes the loss from 69.9 per cent of the object to zero. If that is not enough, put a dark card directly behind it, or make it transmit less than about half the light behind it.