Convert MKV to WebP

A library of recordings without preview images is a list of filenames, and a list of filenames is not something anyone browses. The fix is one still per file, which is a small job repeated many times rather than a difficult one — and repeated jobs are where the details that do not matter once start to matter a great deal.

Two of those details drive this page. The image comes out much larger than the card it is going into, and the timestamp that flatters one recording gives you a black frame on another. Neither is a flaw exactly; both are things worth knowing before the twentieth file rather than after it.

There is no resize control on this conversion. A 1080p recording produces a 1920 by 1080 WebP and a 4K rip produces a 3840 by 2160 one, regardless of the fact that it is destined for a 320-pixel card. That is the right default for a tool that is extracting a frame — throwing pixels away by default would be worse — and it is not the file you want to serve.

So plan for a second pass. Scaling the images down afterwards is quick, entirely lossless in the direction that matters, and it is where nearly all of the size saving on an index page comes from: a 3840-pixel WebP sent to a browser that displays it at 320 carries a hundred and forty times the pixels the card can use, and no codec choice compensates for that.

WebP was published by Google in 2010 as a single format to replace both JPG and PNG, and for the ordinary case — a photographic frame at moderate size — it produces a smaller file than JPG at the same visible quality. Every current browser reads it, which is the property that lets it be the only format you serve rather than one you offer conditionally.

On a page showing forty preview images the compounding is the point. A saving that would be trivial on one image is the difference between an index that appears immediately and one that fills in while the visitor is already scrolling. Measure it on your own frames rather than on a published average, because the gap varies with content more than most comparisons admit.

Zero seconds is where the box starts and it is the wrong answer for almost every recording — captures open on black, rips open on a logo or a slate, lecture recordings open on an empty lectern. Something in the first third of the file, past any title sequence, on a settled shot, is what makes a card worth clicking.

When several files are converted together they are all sampled at the same time, which is efficient and imperfect. Pick a number that clears the opening on the longest title sequence in the set — sixty or ninety seconds is often safe — and accept that two or three recordings will need doing again by hand. That is still far less work than choosing individually for all forty.

A frame from a video is not a photograph and it does not look like one. Codecs describe moving parts of the picture approximately and spend their detail where the eye will notice, so a still lifted from the middle of a pan or a crowd is visibly softer than one from a moment where nothing was moving. The format cannot fix this; it can only avoid making it worse.

Choosing a still moment is worth more than any encoder setting available anywhere. At preview size most of it disappears anyway — a 320-pixel card is forgiving of a great deal — which is one more reason to do the scaling step rather than to worry about the codec.

WebP can carry transparency and this output has none. A video frame is opaque everywhere, and the encoder writes no alpha channel when nothing in the frame is transparent — the result is a plain lossy WebP with no extended header and no alpha chunk, rather than a file carrying a channel that is uniformly full. Nothing is wrong with it and nothing about it needs a background colour chosen.

It also carries no metadata at all. There is no EXIF block, no source filename, no timestamp, nothing tying the image back to the recording it came from. For a catalogue that has to be handled outside the file — the filename or the database row is where the association lives — and the habit of naming each image after its recording as it is downloaded saves an hour of matching later.

No quality control is offered on this conversion. The frame is encoded at a fixed setting chosen to be clean on ordinary content, which for an image that will be displayed at card size is comfortably past the point of visible difference.

Where a page has an unusually tight budget or an unusually demanding image, the route around it is to take the frame as a PNG — which stores exactly what was decoded — and to encode the WebP yourself at whatever setting the budget calls for. One extra step, complete control, and no risk of a slider on this page quietly making a hundred images worse.

Only enough of each file is read to reach the timestamp, and the decoding and encoding both happen in the browser tab. Nothing carrying the recordings is sent to a server, which for a library of internal training material or unpublished conference footage is not a preference but a condition of doing the work at all.

The free tier accepts an MKV up to 100 MB per file. Longer recordings will be past it, and for those the practical route is to have the media server produce its own poster — Plex, Jellyfin and Emby all generate thumbnails — or to take a snapshot in VLC and convert that image here instead. The result is the same picture by a different road.