Short answer: sample four to six points across the backdrop, including the darkest corner, before touching the tolerance slider. A key sampled in the middle of the shade range needs only about half the spread — 9.4% on an 18.7% falloff. If that number reaches the closest material you must keep, no single slider can work and you split the frame into zones.
Measure the spread before you touch the slider
Drop a still frame, then click four to six clean patches of backdrop: the two corners the light hits hardest, the two in shadow, and anything uneven in between. Switch to the material mode and click the one colour you cannot lose — a green shirt, a product, hair. The meter returns the spread, the best single key colour, the tolerance that key needs, and whether that tolerance leaves your material opaque. Nothing is uploaded.
| Sample | Colour | Distance from key |
|---|
Waiting for an image. Click the backdrop in at least two places.
Distance is sRGB straight-line distance divided by 441.67 — the same scale as the keyer's Tolerance slider. The best single key is the centre of the smallest sphere around your samples; its radius is the tolerance that clears every sampled point. The verdict assumes the shade changes smoothly between the points you clicked.
One backdrop, three strategies, measured
A synthetic 800×600 frame with a stand light on the left: the same fabric runs from #00B140 at the lit edge to #00612B in the shadowed corner, a spread of 18.73%. Five solid swatches sit on it, including a green shirt at #2F7D3F that must survive. Every row below was run through this site's keying formula at edge softness 3% on 2026-10-28.
| Strategy | Key colour | Tolerance to clear the backdrop | Backdrop pixels left | Green shirt α | Navy denim α |
|---|---|---|---|---|---|
| One key, lit corner | #00B140 | 19% | 0 | 0 | 255 |
| One key, middle tone | #008936 | 10% | 0 | 99 | 255 |
| 12 zones (4×3), own key per zone | per zone | 3–4% | 0 | 255 | 255 |
Read the green-shirt column. The lit-corner key needs 19% to reach the dark corner, and the shirt is already gone at 16%. The middle tone is the smarter single key — it halves the reach to 10% — but 10% still clips the shirt to alpha 99, which is 61% transparent and visibly wrong next to skin at 255. Only the zone pass keeps the backdrop at zero and the shirt at 255.
Left: one key from the lit corner at 19%. Middle: one key from the middle tone at 10%; the green shirt turns into a see-through patch. Right: 4×3 zones, each with its own sampled key and a 3–4% tolerance. Checkerboard means transparent.
Why the middle sample halves the number
A keyer removes every pixel within the tolerance radius of the sampled colour. Your backdrop is a line of colours in RGB, from the lit green to the shadowed green. The worst case for one sample is the distance to the far end of that line, so the best sample is the middle: each end is then half the spread away.
spread = distance(litCorner, darkCorner) best key = midpoint of that line tolerance = spread / 2 (round up) 18.73% spread -> 9.36% exact radius -> 9.39% once the key rounds to a hex -> set 10% 31.03% spread -> 15.52% radius -> set 16%
That is the whole arithmetic behind “sample the middle”, and it is also why sampling a corner doubles the number you need. The exact midpoint is not always a whole hex colour, so the tool rounds the key and re-measures: #008936 sits 9.39% from the darker end, still a 10% setting, and 11.17% from the green shirt. A material survives only while the tolerance stays below its own distance minus edge softness, so the shirt starts fading at 8.17% — and the backdrop needs 10%. The window is negative, which is why alpha lands at 99 instead of 255.
Push the falloff further and the same failure gets louder. On a 31.03% version of the frame, the middle key needs 16%, the green shirt is at 0, and navy denim drops to 225. The zone pass still clears everything at 5–6%, with every swatch at 255.
| Backdrop spread | One key at the middle: tolerance | Green shirt α | Navy denim α | 12 zones: worst tolerance | Shirt α |
|---|---|---|---|---|---|
| 18.73% | 10% | 99 | 255 | 4% | 255 |
| 31.03% | 16% | 0 | 225 | 6% | 255 |
Two numbers decide the case: the tolerance the backdrop needs, and the fade start of the closest material. If the first is larger than the second, the frame is a tiling job. Nothing in the slider can fix it, because the same pixel range has to cover both the dark corner and the wardrobe.
What the zone pass changed, pixel for pixel
The 12-zone run split the frame into 4 columns by 3 rows. Each zone was sampled at its own backdrop pixel nearest the zone centre, and each zone's tolerance was set to that zone's own worst backdrop distance, rounded up. The resulting matte cleared all 411,200 backdrop pixels in both test frames and left every swatch at 255.
Compared with the middle-tone run that also cleared the backdrop, the zone matte differs on exactly 12,800 pixels in the 18.73% frame — the 80×160 green shirt, and nothing else. In the 31.03% frame it differs on the same 12,800 pixels, but by the full 255 levels, because the middle-tone run deleted the shirt completely and the zone run brought it back. That is the honest measure of what tiling buys: the backdrop result is identical, the difference is entirely the materials you were about to lose.
It also tells you what tiling costs. You need one pass per zone, a way to put the results back together, and a check at every join.
Seams are the one thing tiling can break
Two neighbouring zones use two slightly different key colours. A pixel that sits near the key — a spill edge, a soft hair strand — can fall inside one zone's tolerance and outside the other's. Measured on the test rig, two zone keys 2.36 distance points apart turned the same spill pixel into alpha 92 on one side and alpha 0 on the other. That is a visible step in the matte, and it is exactly the kind of defect a checkerboard preview hides.
On the 12-zone chart itself, zero boundary pixels ended up with partial alpha, so the seam test had nothing to compare: no subject edge crossed a zone boundary. That is not luck to rely on, it is the rule to follow. Put boundaries through empty backdrop, never through hair, a shoulder, or a product edge.
The geometry is plain arithmetic, so write it down before you start. The measured chart was 800×600 split 4×3: 200×200 tiles with vertical boundaries at x = 200, 400 and 600, and horizontal ones at y = 200 and 400. A 4000×3000 photo at the same split gives 1000×1000 tiles and boundaries at 1000, 2000, 3000 and 1000, 2000. Keep the numbers on hand while you crop — dragging boundaries by eye is how a seam lands in someone's hair.
- Boundaries in the background. A seam that runs through flat green disappears once both sides reach alpha 0.
- No overlap on the subject. Stacking two keyed PNGs with normal alpha gives
a = a_top + a_bottom × (1 - a_top). Two 50% edge pixels become 75%, so a 1 px overlap on a hair edge is visible. - Same canvas size. Export every zone at the original frame size and paste at the same coordinates; cropping each zone to its own rectangle and stacking at 0,0 shifts the seams.
Do this in order
- Measure the spread. Four corners plus the worst patch in between, then the material you cannot lose. If the meter says the best single key needs less than the material's fade start, use that key and one slider — the middle tone is a real setting, not a compromise.
- If the window is negative, fix the light first. A falloff under about 10% usually leaves room for one key. Moving the subject away from the backdrop, adding a fill on the dark side, or flagging the hot spot is cheaper than tiling, and it helps every frame in the shoot.
- Tile only the frames that need it. Four columns by three rows was enough for a 31% falloff; a smaller spread needs fewer zones. Keep boundaries in empty backdrop and give each zone its own sampled key.
- Set each zone's tolerance to its own worst point. Not the frame's worst point — that is the single-slider mistake with extra steps. The measured zone values were 3–4% on the 18.73% frame and 5–6% on the 31.03% frame.
- Approve the matte, not the composite. Look at the alpha channel at 100%: no grey sheet over the backdrop, no step at a seam, no half-transparent shirt. The edge checker counts all three.
This is a still-frame workflow
The keyer on this site reads one image and exports a PNG. It does not decode a video, and this page will not pretend otherwise. For a clip, the same measurement applies to each frame you sample, but the work becomes a per-frame key: export the frames, key them, rebuild the clip. The video workflow covers the container and alpha checks, and the batch queue covers running more than one clip.
If a shot has both a green and a blue backdrop, the zone idea and the two-colour idea stack: separate keys per colour first, then per zone. The measured recipe for two mattes in one pass is on two colours at once.
Checklist before you export
- The darkest backdrop sample and the best single key are less than the closest material's fade start apart.
- Tolerance is set from that measurement, not from a preset copied out of a tutorial.
- Every zone's tolerance was measured inside that zone, not inherited from the frame.
- No zone boundary crosses hair, a shoulder or a product edge.
- The alpha channel shows zero residue on the backdrop and a solid subject, with no step at a join.
Uneven green screen FAQ
Why is half my green screen still there after I key it?
Because one sample cannot be equidistant from the lit part and the shadowed part. In the measured 18.73% frame, keying from the lit corner at 14% left 27.6% of the backdrop pixels with a mean alpha of 175.6 — the far corner stayed almost fully opaque. Raising tolerance to 19% cleared the backdrop completely, but the green shirt had already gone to alpha 0 at 16%. The fix is not a bigger number: sample the middle tone, which needs 10% instead of 19%, and if that still reaches your material, split the frame.
Should I raise the tolerance or fix the lighting?
Fix the lighting when you still can. The measured difference is large: an 18.73% falloff needed 19% from a corner sample and left the green shirt at 0, while a 31.03% falloff needed 16% even from the best middle sample and still deleted the shirt and dented navy denim to 225. Adding a fill on the dark side costs one light and lowers the spread for every take. Tiling is the fallback for footage you already shot, or for a still where the backdrop cannot be relit.
Do I really need to light the green screen evenly?
Yes, and the measurement above is the reason. Someone asking this on Video Production Stack Exchange put it plainly: “As long as there are no shadows, would it be ok to not go through complicated steps to evenly light it and just use it?” The catch is that “no shadows” is not the same as “one green”. One panel on one side leaves a smooth gradient even when nothing casts a shadow, and the measured 18.73% falloff needed 19% from the lit end. You do not need a full kit; you need a spread small enough that one tolerance reaches both ends without touching the wardrobe.
Do the seams between zones show in the final PNG?
Only where a boundary crosses something that is not pure backdrop. On the 12-zone test chart every boundary sat in flat green, so there were zero partial-alpha pixels to compare and no visible join. In a controlled stress case, two zone keys 2.36 points apart gave the same near-key spill pixel alpha 92 on one side and 0 on the other — a hard step. Put boundaries through empty backdrop, do not overlap tiles on the subject, and inspect the alpha channel at 100% along each join before you call it done.
Can I use this on a video clip?
Not in this browser tool. It keys one still frame and exports a PNG; there is no video decoder in the page. For a clip, export frames, measure the spread on a frame from each lighting setup, then key per frame and reassemble. A shot whose backdrop drifts during the take is worse than a still: the spread changes over time, so a single setting that works at the start can fail a minute later. The frame-by-frame workflow and the alpha checks are on green screen remover for video.
Next step
Measure your frame above, then key it with the returned colour on the green screen remover. If the backdrop is clear but the edges glow, go to spill removal; if you need to approve the matte, use the edge checker. To set a tolerance by material instead of by backdrop, start with the measured tolerance windows.