Green screen for Zoom and meetings — what a real key adds

Every meeting app now ships its own background replacement, and it runs while you are on the call. Producing a transparent asset you can reuse is a different job. This page is about the seam between the two. It does not benchmark any meeting app: we have not measured one, and nothing below is a claim about how any of them behave.

Two jobs that look alike

A meeting app's background feature reads the frame from your camera, decides which pixels are you, and sends the composited result into the call. Nothing is written to disk. The only artefact is the call itself.

A keyer like this one takes a file and writes a new one: a transparent PNG, a ZIP of PNG frames, or a transparent WebM. Nothing is live, and what you get is a reusable asset.

So: if all you need is a tidy background on this afternoon's call, the built-in feature is the right tool. If you need the subject with the backdrop actually gone, to put over something else more than once, you need the second job — and footage recorded separately.

The recording you get back from a call is already baked

This is the one that catches people. A recording from a meeting app is a video of the composited call: the background was flattened into every frame before the file was written, and MP4 has no alpha channel to carry a matte anyway. No keyer can separate a subject from a backdrop merged on someone else's machine.

To get a keyable version of yourself, record locally from the camera, before the meeting app touches it. That file is what you drop into the tool. Which formats keep transparency at all is on the transparent background video page.

Shooting at a desk is harder than shooting in a studio

Desk setups fail predictably: the screen is lit by the same lamp as the subject, so it is never even; it is narrower than the frame, so the border of the shot is room; and a laptop webcam sits close and low, so the subject casts onto it.

All three map onto measurements we do have. These come from synthetic 320×240 fixtures scored against a binary ground truth — read them as what a clean setup can do, not a prediction for your webcam.

A shadow on the screen is the common failure

In the shadow sweep, a backdrop at 65% brightness sits 29.3412 units from the key colour in the Cb/Cr plane and keys clean at the default tolerance of 25. The same screen at 60% sits 33.5337 away and comes out at alpha 255 — fully opaque, not keyed at all. That step is a cliff, not a gradient.

On the shaded fixture the default tolerance leaves 21,201 of 55,459 backdrop pixels behind: IoU 0.617718. At tolerance 35 it leaves 12,720; at 40 every backdrop pixel goes, IoU 1.000000. At 90 it takes 21,341 subject pixels with it, IoU 0.722122. The fix is real and it has a ceiling — which is what the tolerance page is about.

A screen that does not fill the frame breaks the auto pick

The key colour is guessed, not asked for: it is the median of the outer 5% band around the frame. That median shrugs off a subject touching one edge, but assumes most of the border is backdrop. Hang a screen that stops short of the frame and the median is your wall.

Shade the screen and the pick drifts: on the shaded fixture the detected colour is #00a23b where the real screen is #00b140, a maximum channel delta of 15. You can always set the colour by hand.

If the backdrop has almost no colour in it — a white wall, which is what most home offices actually have — the keyer falls back to a luminance key. The band widens to 37.5 units at the default tolerance instead of 29.5, and so does what gets caught. The honest margin on a white wall is on the white background page.

What you get out of the tool

Three exports, all built in your browser with no upload: a transparent PNG at the source resolution, a ZIP of PNG frames, and a transparent WebM. The sequence is capped at 1,200 frames; the WebM records in real time — a 30-second clip takes 30 seconds. Picking between them is on the PNG sequence vs WebM page; resolution and frame rate are on export settings.

No watermark, because there is no server: the file never leaves the tab. That matters more than usual here, since footage shot for work is often internal. More on the no-watermark and no-upload pages.

When this does not work

• A recording that came out of a meeting app. The background was baked in. There is nothing to key.
• A live call you want to look better right now. This tool works on files. It is not a camera effect and it does not create a virtual camera device.
• A screen that does not reach the frame edge. The auto pick reads the border: either fill the frame or set the colour by hand.
• A shadow across the screen, at default settings. Below about 60% brightness the default tolerance leaves solid pixels. Raise tolerance — and stop before it eats the subject.
• Hair, smoke, glass — anything genuinely semi-transparent. At the default edge of 0 the two thresholds coincide, so alpha is only ever 0 or 255: no partially transparent pixels describe a soft edge unless you dial one in. We have no measurement of how that holds up on real hair.
• Noise that changes frame to frame. Every frame is keyed on its own, with no temporal or spatial filtering, so noise can make an edge boil between frames. Not measured: the fixtures are single frames.
• A long recording. The sequence is capped at 1,200 frames, the WebM records in real time, and everything happens in one tab's memory. Nothing checks the file size on the way in, so nothing warns you before the browser runs out of room.
• Anything compressed. Every number here comes from uncompressed synthetic fixtures. A webcam file or a re-encoded MP4 discards colour detail before the keyer sees it. Not measured, not guessed.

Where the numbers come from

The IoU figures come from the synthetic-fixture benchmark behind this site: 320×240 frames, a binary ground truth, Cb/Cr distance keying against two naive baselines. On flat green the shipped default scores 1.000000; on the shaded fixture 0.617718 at tolerance 25, 1.000000 at 40. Data and code are in the reference repository; the write-up is on Zenodo under DOI 10.5281/zenodo.22916767. Fixture, not footage.

For the shooting side — what to light, how big a screen to hang, why wrinkles ruin the matte — start with the lighting guide and the DIY screen page.