Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

iamajmeena/neonclipping

Media & Design

Turn long videos or live streams into captioned vertical clips locally, using audio cues, transcripts and a local LLM.

234
11
Python
MIT
new today

What it does

NeonClipping turns a YouTube, Twitch or Kick link — or a local MP4 — into captioned 9:16 clips with a title and hook for each. Everything runs locally: no account, no upload, and the server only binds to 127.0.0.1. It is aimed at streamers and editors who want short-form clips without a subscription or a cloud service.

Moment selection combines heuristic signals with a small local LLM. Fast mode downloads audio only, uses YouTube's word-timed captions when available (local Whisper otherwise), scores loudness jumps, raised voices and quiet-to-loud transitions, and asks Ollama to score candidates and write titles and hooks. Deep mode also downloads the whole video and adds optical-flow motion and cut density. Star ratings on clips are stored in data/feedback.jsonl, and after 8 ratings the signal weights are re-fitted with a regression pulled toward the defaults. The project describes its moment picking as heuristics plus a small LLM, not guaranteed viral detection.

Quick start

Prepare Python 3.10+ and FFmpeg on PATH. Ollama with llama3.2:3b is optional but recommended for AI titles, hooks and scores; an NVIDIA GPU speeds up NVENC and faster-whisper.

git clone https://github.com/iamajmeena/neonclipping
cd neonclipping
./run.ps1   # Windows
./run.sh    # Linux / macOS

Open http://127.0.0.1:8770, paste a link and press ⚡ Make clips. Clips appear with a preview and MP4 download. If a video has no YouTube captions and you want GPU speech recognition, install the optional CUDA packages inside the virtualenv:

.venv/bin/pip install nvidia-cublas-cu12 "nvidia-cudnn-cu12==9.*"

Set NEONCLIPPING_CPU=1 to force CPU, or NEONCLIPPING_PORT=9000 to change the port. A PDF setup guide in English and Hindi and an agent setup prompt are included in docs/.

Editing and live clipping

After generation, My clips → Edit clip applies any of the 11 speaker layouts, speaker-size slider, editable subtitle words, caption style and caption height. Apply to this video's clips updates the saved clips from the same source without re-downloading or re-transcribing, but it re-encodes the final MP4; source files must stay in data/. Changing a clip's start/end or spoken language needs a separate trim/transcription pass.

The Live clipping panel records a Twitch or YouTube Live stream (via Streamlink), counts anonymous chat messages in 10 s buckets and cuts around chat spikes, starting roughly 24 s before the spike. Clip the last 30 s now cuts on demand. It only works while the stream is live; finished videos go through Make clips. A CLI version lives in live/live_clip.py.

Constraints

  • Caption accuracy follows the source; YouTube auto captions, especially Hinglish, are imperfect.
  • Without YouTube captions, a speech model must transcribe candidates and CPU-only runs are slow.
  • Facecam presets are starting points, not detection; the wrong preset crops the wrong region.
  • Deep mode was verified only on a short video and downloads the whole file first, so it is slow by design on long streams.
  • The LLM reads text only and cannot see the screen.
  • MIT licensed.

HotRepo's take

In HotRepo's 2026-11-19 new-project snapshot, NeonClipping recorded 121 stars and 4 forks, all within a growth window that began that morning, with no all-language daily rank recorded yet. The repository README is unusually concrete about limits: it distinguishes fast and deep modes, reports local timings on a specified RTX 2060 machine, and documents a 796.9-second layout/style regression check. The live-clipping approach — chat-velocity spikes rather than visual scene analysis — is a defensible design choice for stream content, but it depends on chat activity and only applies while a stream is live. The star count comes from a single early snapshot, so it does not establish adoption or quality; the decision should rest on whether local, heuristic-driven clipping fits the workflow.

Project source and documentation

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. haziqahmedqasmi/PeakCutSelf-hosted tool that turns long videos and live streams into captioned vertical clips using your own AI key.Python104
  2. Bodila51/gtavi-claudecodeA Claude Code plugin for finding GTA 6 moments, cutting vertical clips and checking them against clipping campaigns.Python116
  3. SAOS-3624/dsh-video-storyboardPacks a video into one storyboard image plus a timestamped transcript for LLM agents to understand.JavaScript106
  4. kajisho5/ffmpeg-skillGive coding agents a local FFmpeg video editor with 42 structured tools, a machine-readable contract and an MCP server.Python4,443
  5. JayWebtech/autoshortsA local desktop app that transcribes long recordings, ranks viral moments with an LLM, and crops them to 9:16.Rust1,323