Why does your subject look the wrong size once you drop in the new background?
Because a transparent cutout carries no record of how big anything was. Every frame has a scale, pixels per metre at the subject plane, equal to the frame width times the focal length divided by 36 times the subject distance; your green screen shot has one value of it and the plate has another, and until those match the composite is describing a subject standing somewhere it is not. Shot on a 24 mm lens at 3 m and dropped into a plate whose own reference objects sit 6 m from a 24 mm camera, the subject has to be scaled to 0.50x; into a 50 mm plate at 3 m it has to grow to 2.08x. Measured on a 1280x720 frame keyed with this tool's defaults, applying that scale to the finished cutout is not free either — at 1.5x the soft edge goes from 0.19 px to 1.71 px, while scaling the source before keying holds it at 0.24 px.
A cutout has no memory of the room it was shot in
Every frame carries a scale, and none of it is written into the file you export. Put a camera with focal length f on a 36 mm sensor a distance D from your subject and the frame puts W x f / (36 x D) pixels on every metre at the subject plane — call it ppm. On a 1920-wide frame a 50 mm lens at 3 m gives 888.9 ppm; an 85 mm at the same distance gives 1511.1; a 24 mm gives 426.7. A 1.78 m person is therefore 1582 px tall in the first, 2690 px in the second and 759 px in the third, even though nothing about the person changed.
What comes out of a keyer is an alpha channel and a pixel count. There is no focal length in a PNG, no distance, no sensor width. So the size question is not a matter of taste that you nudge until it looks right — it is a units question with a single correct answer, and the answer lives in the two shoots, not in either file. Your subject is however many cutout pixels tall that ppm_green made it. The plate has its own ppm at the depth where you intend to stand the subject. Until the two are equal, the picture is asserting something false, and the viewer reads that as a person who is the wrong size.
The right scale is a ratio of two f/D values
Since ppm is the frame width times f over 36 times D, and both frames are the same width in the delivery, the width cancels and the scale factor you need is simply (f_plate / D_plate) divided by (f_green / D_green). All four numbers are things you can know. A green screen shot on a 24 mm at 3 m going into a plate whose furniture sits 6 m from a 24 mm camera needs 0.50x. The same cutout into a plate shot at 50 mm and 3 m needs 2.08x. An 85 mm green screen portrait dropped behind that 24 mm plate needs 0.14x — a seventh of its cutout height.
That is the real spread of the problem, and it is why guessing fails. It is not a 10% trim that you can hide with a shadow; between equipment people actually own the honest answer runs from a seventh to a double. Which matters, because a 2x resample of a cutout is not a rounding error on the edge either, and the section below measures exactly what it costs.
This tool resizes your background, and never your subject
Two different rules run here, and they pull in opposite directions. For a still, the working canvas is set to the source image's own naturalWidth, and the export is drawn from that canvas — so a PNG comes out at exactly the pixel dimensions of the file you loaded, and there is no scale control for stills anywhere in the interface. Load a plate as the background image, though, and a second rule fires: the plate is cover-fitted with s = max(w/iw, h/ih) and then centre-cropped to the canvas. Your subject is never touched; the plate is scaled by an aspect ratio you did not choose.
Take a 1920x1080 cutout. Feed it a 1080x1920 phone plate and s becomes 1.778 — the plate is drawn at 1920x3413 and 68.4% of it is cropped away. A 2.0 m door that measured 759 px tall in that plate arrives in the composite at 1348 px. Feed the same cutout a 6000x4000 still and s is 0.320 with 15.6% cropped: a door that was 1580 px in the original lands at 506 px. Same subject, same tool, two incompatible worlds — and the subject's apparent size against the door changes by a factor of 2.7 between them.
The part no scale slider can reach
A lens records more than size. It records how fast things shrink as they move away from it, and that is baked into the pixels permanently. We rendered the same 1.78 m figure at a constant f/D of 12, which holds the subject pixel-identical across rigs: the head measures 9436 px in the 24 mm version and 9418 px in the 85 mm one, 0.2% apart. The figure holds a ball 0.40 m in front of the chest, and that is where the rigs separate.
Ball diameter over chest height comes out 0.3162 at 24 mm from 2.00 m, 0.2899 at 35 mm from 2.92 m, 0.2645 at 50 mm from 4.17 m and 0.2339 at 85 mm from 7.08 m. Four frames in which the subject is exactly the same size, and a 26% swing in how large the near hand reads against the body. A uniform scale multiplies the ball and the chest by the same number, so it cannot move that ratio by anything. If the hands look wrong, the fix is a different lens on the day, not a different number in the transform.
What the rescale does to the edge you just keyed
Say the f/D ratio comes out at 1.5x. We keyed a 1280x720 green screen frame with this tool's defaults — similarity 42, smoothness 14, spill 35, shrink 10 — and then applied the scale to the finished cutout. Counting pixels with 0.05 < alpha < 0.95 per unit of silhouette perimeter, the soft edge is 0.19 px at 1:1 and 1.71 px after the enlargement, with 5.2% of the subject left semi-transparent. Push to 2x and it is 2.43 px. Shrinking is not the safe direction people assume: at 0.25x the edge measures 1.27 px and 13.5% of the subject comes out as partial alpha.
Against that, the same shot keyed at the size it was always going to be used at holds the edge between 0.12 and 0.20 px across the entire range. The difference is not that the resample invents detail — it cannot. It is that a chroma key is a hard decision made about each pixel, and resampling an alpha channel averages a decision that was already made on a pixel grid that no longer exists. The threshold has to be evaluated on the pixels you are going to ship.
Scale the source, not the cutout
So reverse the order: work out the scale from the f/D ratio, resample the green screen file to the target size, then load the resized file and key it. Same 1280x720 frame, same destination size. At 1.5x the soft edge measures 0.24 px instead of 1.71; at 2x it is 0.25 px instead of 2.43; at 0.25x it is 0.15 px instead of 1.27. Mean alpha error against a native key of the same shot drops 4.6x at a quarter scale, 5.9x at a half and 2.5x at 1.5x.
Resampling an image changes the evidence and lets the keyer decide again; resampling an alpha channel just smears the verdict. In practice that means one extra step outside the browser, because this tool has no scale control for stills and will not do it for you — resize the file, then key it. Video has a Long edge control that will scale on the way out, and it is worth knowing that it scales in both directions: a 720p clip with 1080 selected is being enlarged, with the same cost as above.
What to check before you call it finished
Three numbers settle the size question. First, a reference object in the plate whose real dimension you know — a 2.03 m door, a 0.74 m desk — against the subject's real height; the pixel ratio between them has to match the metre ratio, and that is checkable with a ruler tool in seconds. Second, the eye line: if the plate was shot with a level camera at 1.6 m, the subject's eyes belong at that height in the plate, not wherever the cutout's own framing puts them.
Third, check the depth cue last, because it is the one item on this list you cannot repair afterwards. If hands or anything held forward read too large for the body, that is the lens. Shoot the green screen at the focal length the plate implies, or accept the compromise. The clean version of this job is boring: same lens, same distance, one key, no resampling at all. See why a clean key still looks fake for the other half of the same problem.
Related questions
Does exporting at a higher resolution make the subject bigger?
No. A still export is drawn from a canvas sized to the source file's own natural width and height, so a 1280x720 green screen shot produces a 1280x720 PNG. There is no scale control for stills here at all; the Long edge setting applies to video output only.
How do I work out the scale factor without guessing?
Divide (f_plate / D_plate) by (f_green / D_green), where each D is the distance from that camera to the thing whose real size you trust. A 24 mm green screen shot at 3 m against a plate whose reference objects sit 6 m from a 24 mm camera gives 0.50x; against a 50 mm plate at 3 m it gives 2.08x.
Why does my subject still look wrong after I got the size right?
Perspective. At a constant subject size, a ball held 0.40 m in front of the chest measures 0.3162 of chest height at 24 mm from 2 m but 0.2339 at 85 mm from 7.08 m — 26% difference. A uniform scale multiplies both by the same number, so it cannot change that ratio. Only re-shooting at the plate's focal length can.
Is it safer to shrink a cutout than to enlarge it?
No, and shrinking is the direction people wrongly trust. Keying first and then scaling to 0.25x left 13.5% of the subject as partial alpha and a 1.27 px soft edge. Resizing the source first and then keying kept the edge at 0.15 px. Enlarging had the same shape of problem: 2.43 px against 0.25 px at 2x.