Skip to main content
Back

NeonClipping Turns 3-Hour Streams Into Vertical Clips in 2 Minutes

A free, local open-source clipping engine turns long YouTube, Twitch and Kick videos into captioned 9:16 shorts in about two minutes.

Pro

Repo, Full Access
  • NeonClipping is a free, MIT-licensed local clipping engine that turns long videos into captioned 9:16 shorts.
  • Fast mode downloads audio only, reuses YouTube captions, and never processes the full video file.
  • Produces 3 clips from an 85-minute YouTube stream in 97 seconds on an RTX 2060.
  • Uses a local Ollama LLM (llama3.2:3b) to score moments and write titles and hooks.
  • Learns per-user taste via 5-star ratings and ridge regression on signal weights after 8 ratings.
  • Supports Hinglish captions, facecam layout presets, NVENC export, and runs entirely on 127.0.0.1.

NeonClipping is a new MIT-licensed project that turns YouTube, Twitch, Kick, and local MP4 videos into captioned 9:16 clips. Processing runs on the user’s machine, local footage stays off third-party clipping services, and the software requires no account or subscription.

The Python repository reached roughly 1,800 GitHub stars within days of release. Its two-minute claim comes from Fast mode benchmarks on an RTX 2060, with runtime varying by video source, caption availability, hardware, and the number of requested clips.

NeonClipping interface showing generated vertical clips
NeonClipping generates vertical clips with editable captions, titles, and hooks.

Audio first, pixels last

Fast mode limits video work by treating audio and text as its primary selection signals. For remote sources, it downloads the audio track and any available captions, identifies promising moments, and fetches video data only for the sections chosen for export.

The smart.py pipeline reduces a long recording through five stages:

  1. Fetch audio and captions. YouTube captions provide word-level timestamps when available. In the published 85-minute test, the audio download was 79 MB and took about 10 seconds.
  2. Score the audio timeline. The system measures loudness jumps against a rolling two-minute baseline, energy in the 2 to 4 kHz voice band, and quiet-to-loud transitions.
  3. Build candidate transcripts. Existing captions supply the text for YouTube sources. When captions are unavailable, faster-whisper transcribes shortlisted sections locally.
  4. Rank and shape each clip. An optional local Ollama model scores the transcript, writes a title and hook, and adjusts clip boundaries to complete sentences within the configured minimum and maximum duration.
  5. Fetch and render the selection. NeonClipping downloads the chosen video ranges, aligns them against the audio, and renders 1080-by-1920 clips with burned-in captions.