Klyply

Turn long videos into short vertical clips on your own PC

Design Tools

Social Media

Artificial Intelligence

Visit WebsiteSee on Product HuntTwitter ⧉

I built Klyply because the tool I was using billed me for what I recorded rather than what I published. A three-hour stream I kept one clip from cost more than a one-hour interview I kept six from, and the fix available inside a metered tool was to trim the source by hand first, which is the job I was paying it to do. The part that took longest was nothing to do with billing. Klyply picks who to frame on, and early on the framing decisions were not a function of the video: adding noise of a single grey level, less than two ffmpeg builds differ by when decoding the same file, changed the layout on most clips. I spent a while blaming the shot detector and the speaker scoring, and both were stable. It was the face detector. Over 216 frame pairs under that noise, the Haar cascades changed their face count 15.7% of the time and their worst horizontal drift was 0.467 of the frame width, which is not a wobble, that is finding a different person. It is how a clip ends up framed on the listener instead of the speaker. Swapping to YuNet took those to 1.9% and 0.006, the layout timeline is now identical across noise on 31 of 33 test clips, and the analysis got 2.6x faster as a side effect. The other thing I changed my mind about: "follow the speaker" used to split the screen whenever it was unsure who was talking. That felt safe and it was wrong, because the option says it follows one speaker. It now widens the crop to hold both instead, and only the Split option ever splits. Honest limits: Windows 10 and 11 only, no Mac or browser version, no free tier, and it renders on your CPU, so a rented GPU will beat a thin laptop on a long file. Happy to answer anything.