Why is motion blur on a green screen almost impossible to key out?
Because the green is physically inside the pixel. While the shutter is open a moving subject covers a given pixel for part of the exposure and the screen covers it for the rest, and the sensor adds the two together, so there is no clean colour left to key against. This keyer can still read the mixture — distance from the key tracks time-averaged coverage to within 0.026 — but at the default settings it deletes the outer 27.5 percent of every smear outright, 1,767 pixels at 12.8 pixels of blur, and the part that survives still carries 40 code levels of green above its own red and blue.
What a blurred pixel actually contains
Open the shutter for a fiftieth of a second and a hand crossing the frame occupies a different set of pixels at every instant. One pixel near its edge saw the hand for part of that time and the screen for the rest, and the sensor does not choose between them; it adds them. Written out, the pixel is the time-averaged coverage times the average colour the subject showed while it was there, plus one minus that coverage times the screen. Nothing has been guessed or lost yet. That mixture is the photograph.
A colour key can read the mixture surprisingly well. Distance from the key colour comes out proportional to how much of the exposure the subject covered the pixel: measured across a 12.8-pixel smear, the gap between the distance the keyer sees and coverage times the subject's own distance from the key averages 0.026. The alpha it produces is close to correct. What it cannot do is hand back a colour, because the green sitting in those pixels was really there.
The green is in the pixel, not painted on top of it
Every semi-transparent pixel keeps the red, green and blue it was shot with. This keyer writes alpha and leaves the colour alone, so the green that came from the screen stays in the green channel. After the default spill suppression, pixels in the band still sit 40.2 code levels of green above their own red or blue, and once the result is composited over a dark background the average colour error across the band is 25.1 levels, peaking at 51.4.
That failure is not new to motion blur. A still frame has the same thing on its anti-aliased edge, and worse per pixel — 44.9 levels on average — because every pixel there sits at almost exactly half coverage, which is where the error is largest. It just happens on two columns of pixels instead of on a smear you can see. Motion blur does not invent a new problem; it spreads a familiar one over a band wide enough to notice.
The outer part of the smear is simply deleted
The worse damage is at the faint end. A pixel is removed completely while it stays inside 0.161 of the key colour, and hair measures 0.585 from the key when it is fully present, so every pixel whose time-averaged coverage falls below 0.275 is thrown away without a trace. That is 27.5 percent of the coverage range. Skin, at 0.902 from the key, loses the outer 17.9 percent; a mid-grey jacket, at 0.593, loses 27.1 percent.
Counted along a 360-row edge with 12.8 pixels of blur, that is 1,767 pixels gone and 9.6 percent of the band's coverage destroyed. At 25.6 pixels of blur it is 4,548 pixels. The keyer also invents 1,566 pixels of coverage in the middle of the same band, because opacity climbs to full at 0.443 coverage, so what survives of the smear is both shorter and harder-edged than what the camera actually recorded.
Width is the whole problem
Across the sweep at the default settings the damaged band runs 720 pixels with no motion, 3,600 at 3.2 pixels of blur, 9,360 at 12.8 and 19,440 at 25.6 — twenty-seven times the still-frame area. Total colour error over the band goes 32,357, then 60,638, then 235,310 and 462,685 in code levels. The area grows faster than the error because the average is diluted by the many pixels that are only lightly touched.
Translated into camera settings on a frame one metre wide at 640 pixels per metre, a hand moving one metre per second travels 12.8 pixels during a fiftieth of a second and 3.2 pixels during a two-hundredth. Those two rows are 235,310 and 60,638. Closing the shutter by two stops costs light and takes roughly three quarters of the damage with it, which is the only trade in this whole problem that actually pays.
Direction sets the price
Only motion across the edge does this. Move the same 12.8 pixels along the edge instead and the silhouette never spreads: the band stays at 720 pixels, nothing is deleted, and the total error is 36,457 against 235,310, about a sixth. What is left is the smear of colour inside the two edge columns, which is a much smaller thing than a smear of coverage across thirteen columns.
Camera movement is free for the same reason. A pan moves the screen and the subject across the sensor together, so nothing moves relative to anything else and no new mixture appears. What costs you is the subject moving against the screen, and specifically the part of that motion perpendicular to its own edge. Fast movement that slides along the outline is nearly free; the same speed crossing it is not.
Why no control on the panel fixes it
Opening similarity buys exactly the wrong thing. On a screen lit to 0.735 stops of spread, 26 is where the background starts coming back — 1.6 percent of it — so 30 is as low as you can go, and there 4.5 percent of the band's coverage is still destroyed while the mean error rises to 32.3. At 63 it is 36.9 percent destroyed. At 100 the destruction exceeds the entire band and the body itself starts to go, mean alpha over solid pixels dropping to 0.949.
Spill suppression is the one control that genuinely takes green out, and it is not free either. Pushed to 100 it brings the band error down from 25.1 levels to 8.7 and the leftover green to nothing. Run the same setting on an olive-green jacket and the body loses 20 levels of green it was entitled to, and the error climbs again from 11.2 at spill 70 to 19.8 at 100. You can have the trail or you can have less green, not both.
What to do instead
Shoot it out. A shorter shutter is the only fix that removes the cause rather than trading against it: two stops takes a 12.8-pixel smear down to 3.2 pixels and removes about three quarters of the error, at the price of light you then have to put back — and a green screen already needs a large light budget before you start spending it on shutter speed. Failing that, slow the movement, or choreograph it so that whatever moves fast moves along its own outline.
Do not expect help from the AI overlay. It combines as a union, taking the greater of the two alphas, so it can only add opacity and never remove colour. Modelled as a clean mask at the mid-exposure position it changes nothing at all: the same 447 pixels of coverage still disappear. Modelled as a mask that covers the whole blurred smear, nothing disappears and the error doubles to 51.0 levels, because the green is now being shown at full strength instead of at partial strength.
Related questions
Can any keyer remove motion blur from a green screen shot?
No. The blur is not an artefact the keyer added; it is light that the sensor genuinely summed. A keyer can estimate how much of the exposure the subject covered a pixel — this one tracks it to within 0.026 of the ideal — but it cannot separate the green out of the sum, because the sum is all it has.
Does a faster shutter actually fix it?
Yes, and it is the only fix without a side effect. On a one-metre frame at 640 pixels per metre, a subject at one metre per second smears 12.8 pixels at 1/50 and 3.2 pixels at 1/200; the total colour error over the band drops from 235,310 to 60,638 code levels. You pay for it in light.
Should I raise similarity to catch the blurred edge?
No. Raising it deletes more of the smear: 4.5 percent of the band's coverage is destroyed at similarity 30, 9.6 percent at the default 42, 36.9 percent at 63, and at 100 the destruction exceeds the whole band and solid parts of the subject drop to a mean alpha of 0.949. Lowering it below 30 starts letting the screen back in.
Does AI subject mode rescue a blurred key?
No. It merges as a union with the chroma alpha, so it can only add opacity. A mask at the mid-exposure position changes nothing, and a mask that covers the whole smear removes the holes but doubles the colour error from 25.1 to 51.0 levels, because the green those pixels carry is then composited at full opacity.