DIY green screen at home
You do not need a studio. You need the right fabric, enough of it, and not a single wrinkle. This page covers what to buy, how big to make it, and why folds wreck your key — with the measurements behind each claim.
The material matters less than the wrinkles
The keyer does not care whether you bought a branded pop-up or taped green card to the wall. It cares about one thing: is every pixel of the backdrop the same colour. A cheap bedsheet pulled tight and lit evenly will key cleaner than professional muslin that is bunched up and lit from one side.
The chroma key measures distance from the key colour in the Cb/Cr plane. On a flat green fixture the key colour is #00b140, chroma magnitude 83.66, luma 111.20. At full brightness every pixel sits at distance 0 and is removed. As the screen darkens — a shadow, a fold — the distance grows. At 65% of full brightness the distance is 29.34, still inside the default tolerance radius of roughly 29.5, so the pixel still keys out. At 60% the distance reaches 33.53, past the threshold, and the pixel turns solid. That is a 5% brightness drop. One fold can be more than that.
What to buy
- Fabric (muslin or polyester). The cheapest option. Look for matte green with no sheen. Polyester holds its shape; muslin is heavier and drapes flatter. Either keys fine flat. A green double bedsheet covers a head-and-shoulders shot; 3 metres covers a seated person.
- Pop-up collapsible screen. A wire-framed disc that folds into a bag. Flat by design, which solves wrinkles for you. Most are 1.5 to 2 metres — enough for one person, not a wide shot.
- Green paint on a wall. The flattest surface you can get and never needs folding. Use pure green with no blue or yellow tint. The preset on this tool is
#00b140. - Green paper or card. Poster board works for small product shots; butcher paper on a roll lays flatter. Avoid glossy paper: the sheen reflects light as a white hotspot the key cannot remove.
How big to make it
Big enough that the subject and their shadow do not spill past the edge. Measure the widest the subject will be, add 50 cm per side, and the same above their head. For someone at a desk, 2 m wide by 1.5 tall is workable. For standing shots, 3 m wide and 2.5 tall. If the subject gestures past the edge, that part of the arm is gone — there is no backdrop to key, and the tool has no brush to paint it back.
Do not confuse screen size with subject distance. Farther from the screen means less green bounce on the subject, but the screen stays the same size. The trade is distance against framing: the screen must cover every pixel behind the subject, including where their shadow falls.
Wrinkles: the one thing that wrecks it
A wrinkle is a shadow you cannot light away. Each fold catches light on one face and casts a shadow on the next, so the fabric has two or three brightness levels side by side. The keyer sees them as different colours, and the dimmer ones fall past the threshold.
On a synthetic shadow fixture — a smooth ramp with no fabric texture or camera noise — the default tolerance of 25 scores a mean IoU of 0.618: roughly two thirds of the backdrop left in the frame. Raising tolerance to 40 takes it to 1.000. Same image, no code change: the shadow was a lighting artefact the keyer was compensating for. The cost is fine detail — hair, translucent fabric and motion-blurred edges start disappearing.
Kill wrinkles before you shoot; raise tolerance only as a last resort. Three ways:
- Stretch it. Clamp the fabric top and bottom, or use a frame. Tension removes most folds.
- Steam it. A garment steamer relaxes muslin or polyester in minutes.
- Weight the bottom. Sew a channel at the bottom edge and thread a dowel through it. Gravity pulls the fabric flat.
If none of those work, switch to a pop-up or paint the wall. A wrinkled screen is a material problem, not a slider problem.
Check it before you record
Take one still with the subject in place and drop it into the tool. Toggle the checkerboard and compare corners against the centre: if the screen keys in one place and not another, you have a wrinkle or lighting problem, not a tolerance problem. Check the auto-detected key colour: on a shadowed screen it reads darker — in our fixture run it lands on #00a23b where the screen is #00b140, a channel difference of 15. Small enough to key fine; for the exact colour, pick it by hand off a lit part of the screen.
The lighting guide goes deeper on even lighting: two lights from each side at 45 degrees, aimed only at the backdrop, subject lit separately.
When this does not work
A DIY screen solves material, size and wrinkles. It does not solve these:
- A room too small for distance. Subject close to the screen: their shadow lands on the backdrop, green bounce lands on them. Distance fixes both. A blue backdrop reflects less light — see the blue vs green page.
- Semi-transparent detail. Flyaway hair, fur, veils, smoke. A colour key produces one matte; these need a hand-drawn mask. There is no brush on this tool.
- A subject wearing green. Green clothing keys out with the screen. More tolerance makes it worse. Shoot on blue — see the blue screen page.
- Wrinkles you cannot get out. If the fabric will not steam flat, a pop-up or painted wall is the only fix. No tolerance setting saves a deeply creased screen.
- Bad light, not bad fabric. A flat screen lit from one side has the same gradient as a wrinkle. Fix the light — see the lighting page.
- Heavy compression. A phone recording smears colour across the boundary before the keyer sees it. See the video page.
- Glossy fabric. Satin reflects light as white hotspots the key cannot separate from any other white. Choose matte.
- A green that is not green enough. Teal, lime and olive sit at different points in the Cb/Cr plane. Auto-detect picks the wrong colour; set the key by hand.
Then key it locally
Once the screen is up and flat, drop the clip or still into the tool and the backdrop is keyed out in the browser — no upload, no account, no watermark. The no-upload page covers what keeping the file local avoids. The fixture data is in the reference repository, and the write-up is archived at doi.org/10.5281/zenodo.22916767 — synthetic test fixtures, not real footage, so read them as the shape of the problem.