What happens when the green screen emits its own light onto your subject?
A screen that lights itself is not automatically worse than a cloth: at the same on-screen brightness the two deliver the same light to your subject, 0.0848 against 0.0837 in our measurement, 1.3% apart. What hurts is the distance a small panel forces. Our test subject 5 cm in front of a 43 x 24 cm panel sat inside 0.913 of a full hemisphere of green and picked up 40 code levels of it on its edge, while the same subject 80 cm away, which needs an 86 x 48 cm panel, picked up 3. You cannot dial that out either: dimming the panel from 70 IRE to 51 IRE punched holes through 36% of the edge, and pushing similarity from 42 to 60 punched holes through 22% of it.
Emitting is not what makes it green
A surface that fills a cone of half-angle alpha around a point delivers pi times its radiance times sin-squared alpha, and that formula does not care where the radiance came from. A cloth reflecting lamp light and a panel making its own light are the same object to your subject as long as both read the same on the sensor. So the popular explanation — the screen is worse because it is a lamp as well as a backdrop — is wrong on the physics.
We ran both at once. A cloth lit by two fixtures 0.90 m away, reading the same (0, 206, 76) the panel reads, delivered 0.0848 of green-channel irradiance to the subject's rim. The emissive panel at the same distance delivered 0.0837. The ratio is 1.013, which is the cloth's own unevenness showing up, not emission. If the panel is worse in practice, it has to be for another reason, and it is.
The panel you can afford sets the distance
A backdrop has to fill the frame at its own distance, so its metre size is not a free choice: it must be about 1 plus d/D times the frame, where D is the camera distance and d is how far behind the subject it stands. In our 640x360 close-up, a 0.400 x 0.225 m frame shot from 0.70 m, that works out to 0.411 x 0.231 m at 2 cm behind the subject, 0.429 x 0.241 m at 5 cm, 0.514 x 0.289 m at 20 cm, 0.629 x 0.354 m at 40 cm, 0.857 x 0.482 m at 80 cm and 1.314 x 0.739 m at 1.60 m.
How much of the panel the subject sees follows from the same geometry. A flat card facing the panel receives 0.969 of a full hemisphere at 2 cm, 0.913 at 5 cm, 0.756 at 10 cm, 0.511 at 20 cm, 0.292 at 40 cm, 0.165 at 80 cm and 0.105 at 1.60 m. On the silhouette of a round subject, which is what you actually key, the same numbers read 0.335, 0.200, 0.108, 0.048, 0.020, 0.010 and 0.006. Nothing smaller than a 40-inch panel gets you past 0.03.
Where the green actually lands
The panel does not light the side of your subject you can see. A point facing the camera has its normal pointing away from it, so it receives nothing. In our frame 86% of the subject's pixels receive some panel light, but averaged over the whole subject that light is only 2.2% of the key light when the subject is properly lit. Everything you can see happens at the edge.
At the outer 2% of the subject, the panel supplied 20.9% of the green-channel light with a normal key light and 83.6% when the subject was lit only by room ambient. Measured against the same frame with the panel's contribution removed, green excess on the subject peaked at 20 code levels with a proper key light and 31 with ambient only. Holding the panel at 70 IRE and moving the subject instead: 40 code levels at 5 cm, 27 at 10 cm, 14 at 20 cm, 7 at 40 cm, 3 at 80 cm and 2 at 1.60 m.
Why the sliders cannot buy it back
The obvious move is to turn the panel down, and it trades green for holes one for one. At 70 IRE the panel reads (0, 206, 76) and no pixel of the edge is lost; at 51 IRE it reads (0, 151, 53) and 12.31% of the edge is punched out; at 36.9 IRE, (0, 110, 36), 26.50%; at 26.2 IRE, (0, 79, 24), 36.27% plus 0.95% of the subject proper. The green peak falls from 31 to 6 code levels over the same sweep: six code levels for a third of your edge.
Tolerance moves the other way and costs the same thing. From the default 42, going down to 20 leaves the edge intact but the average alpha error rises from 0.246 to 0.437; going up to 60 puts holes through 21.61% of the edge, 80 through 74.48% and 100 through 95.92%. Spill suppression barely engages: it only fires where green is already the strongest channel, 1.3% of pixels on skin. AI subject mode cannot help either, because it is a union that only ever adds opacity and it is a person-segmentation model that returns nothing usable on a hand or a product.
The glass reflects a room you forgot about
A panel is glossy, and gloss reflects. At the angles this framing produces, 0 to 18.1 degrees of incidence, Fresnel reflectance is flat at 0.0400 to 0.0402, so the panel carries a 4% mirror image of whatever sits behind the camera on top of its own emission. The angle term is not what matters; how bright that room is does.
With the room at a quarter of the panel's luminance the panel reads (15, 207, 78) and the worst panel pixel sits 0.0659 from the key, needing similarity 23. At equal luminance it reads (37, 209, 86), 0.1262, needing 35. At four times the panel's luminance — a dim screen in a bright room, which is the usual failure — it reads (77, 218, 108), the worst pixel is 0.2582 away and 9.75% of the panel survives the default setting. That reflection is a mirror image, not a flat wash, so it also destroys the panel's uniformity, which is otherwise its best property.
What an emissive panel genuinely fixes
Two things, and they are worth having. The first is uniformity. A panel has no falloff and no wrinkles, so in our frame the worst backdrop pixel sat 0.0078 from the key — that is 8-bit rounding, not unevenness — against 0.0474 for the cloth lit to the same level, which needs similarity 19.3 where the panel needs 11.4. The second is shadows: an emitter is not occluded by the subject, so the contact shadow that costs so much on a cloth simply does not exist here.
It is also the only backdrop you can switch off. Displaying near-black, our panel read (6, 6, 6); the worst backdrop pixel sat 0.0088 from that colour and the nearest subject pixel 0.0839, a gap of 0.0752, just under the 0.098 the default smoothness demands. Drop smoothness to 0 and a window opens at similarity 2 to 17. That is narrow, and a properly lit subject would open it much wider, but it means you can shoot both versions without moving anything and pick the one that keys.
What to do with it
Buy distance, not brightness. Every step back is worth more than any slider: the silhouette's share of the panel drops from 0.335 at 2 cm to 0.048 at 20 cm and 0.010 at 80 cm, and the green on the edge follows it from 40 code levels to 3. Getting to 40 cm needs a 0.63 x 0.35 m panel and getting to 80 cm needs 0.86 x 0.48 m, so this is a purchase, not a setting.
Then light the subject on its own. With a real key light the panel supplied 37.7% of the green-channel light on the rim; with nothing but room ambient it supplied 150.8%, and the green peak rose from 20 to 31 code levels. Keep the panel at about 70 IRE, re-pick the key colour from the captured panel rather than the hex you typed into the display, and darken the room behind the camera first.
Related questions
Does a screen that makes its own light throw more green than a lit cloth?
No. Irradiance from a surface filling a cone is pi times radiance times sin-squared alpha however that radiance was produced. Measured side by side at the same on-screen reading of (0, 206, 76), the cloth delivered 0.0848 of green-channel irradiance to the rim and the emissive panel 0.0837, a ratio of 1.013 that comes from the cloth's unevenness. Distance is the variable that matters, not whether the light is reflected or emitted.
How far behind my subject should an emissive panel sit?
As far as the panel you own allows. In our close-up the silhouette received 0.335 of a hemisphere at 2 cm, 0.108 at 10 cm, 0.048 at 20 cm, 0.020 at 40 cm and 0.010 at 80 cm, and the worst green contamination fell from 40 code levels to 3 across the same range. Because the panel must fill the frame, those distances require panels of roughly 19, 21, 23, 28 and 39 inches diagonal.
Should I turn the screen's brightness down to stop the green?
Only if you want holes instead. Dropping our panel from 70 IRE to 51 IRE cut the green peak from 31 code levels to 19 but punched out 12.31% of the edge; at 36.9 IRE it was 11 code levels and 26.50% of the edge; at 26.2 IRE, 6 code levels and 36.27%. Raising similarity is no better: 60 puts holes through 21.61% of the edge and 100 through 95.92%.
Can AI subject mode rescue a shot like this?
Not here. In this tool the AI mask is combined with the chroma matte as a union, taking the larger of the two alphas, so it can only ever add opacity and can never remove the green the panel threw onto the subject. It is also a person-segmentation model: on a hand or a product it may find nothing, and the code discards any mask covering under 1% of the frame, leaving you with the chroma result unchanged.