PNG Sequence vs Transparent WebM

Both exports carry the same alpha channel out of the same keyed frames. They differ in what happens to those frames on the way to disk: one writes every pixel exactly as the keyer produced it, the other compresses them into a video. Which one you want depends on what happens next, not on which is “better”.

The short answer

If the result is going onto a timeline in an editor, take the PNG sequence. Nothing can silently flatten it, every frame stays lossless, and any editor that can import an image sequence can read it. If the result is going onto a web page, into a browser-based compositor, or anywhere a single file is easier to move around, take the transparent WebM — it is one small file that plays where it was made.

What the PNG sequence actually is

The page steps through your clip frame by frame at the rate you set, seeks the video to each timestamp, keys that frame, encodes it as a PNG, and packs the whole set into a ZIP named <your-file>-frames.zip. Inside, the frames are numbered <your-file>_0000.png, _0001.png and so on.

Two properties matter. Each frame is written at the full resolution of your source — no rounding, no scaling. And PNG is lossless, so the alpha the keyer computed is the alpha you get, byte for byte. The ZIP is written stored, not deflated, so it is essentially the sum of the PNG sizes: compressing them again buys nothing worth the CPU time.

What the transparent WebM actually is

The WebM export is a recording, not an encode. The page draws each keyed frame onto a canvas, calls captureStream on that canvas at 30 fps, and feeds the stream to MediaRecorder with VP9 at about 8 Mbps. The file lands as <your-file>-transparent.webm.

So it is a video: one file, small, playable in a browser tab. Two consequences follow from it being a recording. It runs in real time — a sixty-second clip takes about sixty seconds, because there is no faster-than-realtime encoder in the page. And a WebM written this way carries no duration in its header, which is why some editors show an unknown length until the file is played or remuxed. The WebM page covers that artefact.

Where they really differ

• Lossy versus lossless. VP9 drops detail to hit its bitrate. PNG does not. Fine hair and a graded edge are the detail a codec spends its bits on last.
• Resolution. VP9 needs even dimensions, so the recording canvas rounds width and height down. The sequence keeps the original pixel grid.
• Frame rate. The sequence honours your rate, default 12 fps. The WebM records at 30 fps regardless.
• File count. The sequence is many images inside one ZIP; the WebM is one file. For handing something over, that is often the whole argument.
• What can go wrong silently. A sequence has nothing to misread — an editor either imports the images or it does not. A WebM must survive a player that reads alpha, an importer that trusts its duration header, and any re-encode in between.

The ceiling you will hit with long clips

A PNG sequence lives in memory before it is zipped, so it is capped at 1200 frames. At the default 12 fps that is about 100 seconds; at 24 fps, about 50. Go over and the export refuses, and the message names the frame rate that would fit. Your options: lower the rate, trim the clip, or take the WebM — which has no frame cap but costs real time equal to the clip length. The projected frame count sits next to the rate slider before you start.

Which one to pick

1. Going into Premiere, Resolve, Final Cut or After Effects? PNG sequence. Import the folder as an image sequence and drop it on the timeline.
2. Going onto a web page or into a browser demo? Transparent WebM. It is what the browser made it for.
3. Going to a platform that re-encodes uploads? Neither will keep its alpha. Composite onto the background you want first.
4. A single photo? Neither — take the plain transparent PNG, which is the photo export at full resolution.
5. Not sure yet? Sequence. You can always encode it into a WebM later; you cannot recover the pixels a lossy encode already dropped.

When this does not work

• You need one file and lossless. There is no such option here. A sequence is many files; a WebM is compressed.
• Your clip is longer than the frame ceiling allows. Lower the frame rate or trim. Raising the ceiling is not an option — the frames have to sit in memory before the ZIP is written.
• You are in a hurry with a long clip. The WebM records in real time, so a ten-minute clip is a ten-minute wait, and a frame dropped mid-recording is gone.
• Your editor will not import WebM and shows an unknown duration. Use the sequence, or remux the WebM with stream copy — never re-encode to fix the header, that is where the alpha dies.
• Your target flattens uploads. No export format protects you from a service that re-encodes to H.264.
• Your browser cannot record a canvas. There is no WebM button at all in that case; the sequence is the fallback and the page says so rather than offering a broken option.
• The matte itself is fragile. If the key is already eating into the subject, exporting lossless will not repair it — it will just show you the damage more clearly. On our synthetic fixtures, tightening the edge slider to −100 removed 6902 of 21341 subject pixels, dropping IoU to 0.889. The slider page explains what that control does.
• We have no measured size or quality comparison. No bitrate table, no “sequence is N times larger”, no measurement of how much VP9 degrades a soft edge. We have not run it, so no number is quoted.

Where our published numbers come from

The measurements behind the keying claims on this site live in the CSVs of the reference repository on GitHub, and the write-up is archived on Zenodo under DOI 10.5281/zenodo.22916767. Those fixtures are synthetic 320x240 images, not real footage, and they measure how well a key separates a subject from a backdrop — not file size, not codec quality, not player support.

One caveat worth knowing before you read an alpha number off them: in the shadow sweep, every published alpha value is either 0 or 255. Those are binary mattes, so they cannot tell you anything about how a graded, semi-transparent edge behaves. That is precisely the part a lossy encode touches first, and we have no measured data on it.

To make either export, the tool keys the backdrop and writes the result entirely inside your browser — no upload, no account, no watermark. The format comparison covers what each container can carry, and the video page walks through the exports from the shooting side.

Open the tool