Stream clipping with computer vision and a local judge model
Claude Fable 5 wrote a local script using computer vision and ffmpeg that scans a downloaded stream for events like death screens or killfeed names, with a local Qwen model judging the clips.
Credits
- Reddit poster TangeloOk9486 Source ↗
Earliest dated source:
The full story
A streamer's clipping setup, described second-hand: Claude Fable 5 wrote a script using computer vision and ffmpeg that runs on a Mac Studio and scans a whole downloaded stream for on-screen events, like the Dark Souls death screen or the streamer's name in a Call of Duty killfeed, without spending tokens watching video. A local Qwen model then checks each clip's video and transcript for events, swearing or louder audio, and survivors go to a phone for approval and on to TikTok drafts. The big model is paid once to build the tool, and the local one does the daily work.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: Stream clipping with zero tokens spent finding clips and a local qwen as the judgeReddit poster: TangeloOk9486Posted
A post describing the setup, which was first shared on X.