You drop a YouTube or Twitch link into NeonClipping, wait a couple of minutes, and get back vertical 9:16 clips with captions, titles, and hooks—no cloud, no subscription, and nothing leaves your machine. NeonClipping is a Python-based, open-source clipping engine designed for anyone who wants to turn long-form video (streams, lectures, or your own MP4s) into short, shareable highlights. Everything runs locally: it grabs transcripts, finds the moments with the most energy or dialogue, and lets a local LLM write the title and hook for each clip.
The process is fast. On a test machine (RTX 2060, 16 threads), NeonClipping processed an 85-minute stream in 97 seconds, all offline. The trick is to avoid downloading or analysing the whole video unless you want to: it defaults to audio-first processing, using YouTube's word-timed captions where available, and falls back to Whisper locally for other sources. It scores the audio for sudden loudness, raised voices, and quiet-to-loud transitions, shortlists the best moments, and then uses a local LLM (Ollama by default) to score, title, and hook each candidate. Only the chosen video sections are downloaded and rendered, so bandwidth and disk usage stay low.
Fast and Deep Modes: How It Works
- Fast mode (the default for videos over 10 minutes) never looks at the whole video. It downloads only the audio (about 79 MB for 85 minutes), pulls free YouTube captions, and analyses audio spikes to find candidate moments. The LLM reads the transcript for each candidate and generates a title and a hook. Only the relevant video segments are downloaded and trimmed, with captions burned in.
- Deep mode downloads the full video and adds motion analysis (OpenCV Farneback for object movement, impacts, and camera whips) and cut density to the audio signals, shortlisting more candidates and giving the LLM a wider context. This is slower but can find more nuanced highlights.
Local Learning and Feedback
NeonClipping keeps all feedback and preferences on your computer. You can rate generated moments with a five-star system; after at least eight different ratings, the local ranking weights adapt to your preferences. Nothing is uploaded—your feedback lives in . The ranking works as a small preference ranker, not a full model retrain, so it’s fast and private.




