The scene changed. An 8x8 grid of brightness caught it.

A cut in a video is a moment where almost every pixel changes at once. The frame before and the frame after share nothing: different room, different light, different faces. That single property is enough to find every scene boundary in a clip without understanding a thing about what's on screen, and it's cheap enough to run on 64 numbers per frame.

That is roughly what Framepick does. You drop a video on the page and it hands back one keyframe per scene, a contact sheet of the shots that actually represent the clip, each stamped with its exact timecode. The file never leaves your machine.

Why this was annoying before

If you want the representative frames from a clip today, you have two options and both are friction. Upload it to a web tool, which ingests your footage to a server, runs ffmpeg, and hands back thumbnails, meaning your private or unreleased video now sits on someone else's disk. Or run ffmpeg yourself with -vf "select='gt(scene,0.4)'", which works if you have ffmpeg installed, remember the incantation, and enjoy tuning the threshold by hand. Neither is something you send to a non-technical collaborator.

Framepick does the whole thing in the browser tab, on-device. It demuxes the container with mp4box.js, feeds the encoded chunks to a hardware-accelerated WebCodecs VideoDecoder, and walks the actual decoded frames. The decode runs in a Web Worker, off the UI thread, so the page stays responsive while a 200 MB clip gets analyzed with zero bytes crossing the network. The "server" is your own GPU.

Signatures, not pixels

Here's the part worth knowing even if you never open the app. Comparing full frames to find cuts would be slow and, worse, jumpy: film grain and compression noise wiggle individual pixels constantly, so a pixel-diff spikes on things that aren't cuts at all.

So Framepick doesn't compare pixels. Each sampled frame (about four per second) gets reduced to an 8x8 grid of average brightness, sixty-four numbers, a fingerprint of the frame's large-scale light and dark structure. Averaging over each cell throws away exactly the high-frequency noise that fools a pixel-diff, while keeping the layout that actually changes at a cut. Then it's just a distance between two 64-value vectors. Where that distance spikes, a scene boundary gets marked; Framepick picks one establishing frame per scene and drops the near-duplicate shots so the contact sheet is shots, not noise. The math lives in framework-free, unit-tested modules (src/lib/signature.js, src/lib/scenes.js), so the interesting logic is verifiable in Node without a browser anywhere in the loop.

One knob, and it doesn't re-decode

There's a single sensitivity slider: more keyframes or fewer. Users think in that, not in distance thresholds. The nice property is what happens when you move it. The expensive work is the decode, pulling frames off the GPU, and that already produced the cached signatures. Re-segmenting is just re-running the cut detection over sixty-four-number vectors that are already in memory. So the slider responds live, re-cutting the whole video the instant you drag it, with no second decode. Cheap comparisons are what make that feel instant.

The honest limitation is the browser. WebCodecs is required, so you need Chrome or Edge 94+ or a recent Safari; Framepick detects an unsupported browser on load and says so plainly rather than failing silently. Containers and codecs are bounded by what the browser's decoder plus mp4box.js can handle, which is MP4 and MOV with H.264, HEVC, or AV1 as the platform allows. A DRM-protected or otherwise undecodable file surfaces a designed error state instead of a spinner that never stops. And Framepick extracts, it doesn't edit: no trimming, no audio analysis, no transcription. It pulls the frames out and gets out of the way.

Try it

Open Framepick, drop in any clip with a few scene changes, and watch the contact sheet fill in as it decodes. Drag the sensitivity slider and notice the sheet re-cut instantly, no reload, no re-decode. Click a keyframe to save it as a PNG, or grab the whole set as a zip for a shot list. Do it with your network panel open if you want proof: nothing uploads. The code is on GitHub if you'd rather read the signature math than take my word for it.