NeonClipping: open-source tool auto-generates vertical clips

A desktop computer fused with a film reel and sound waves, glowing from within, surrounded by floating vertical video frames that form a halo—symbolising local, automated video clipping.

NeonClipping, a free and open-source tool by Ajay K Meena, enables creators to turn long-form YouTube, Twitch, Kick, or local MP4 videos into polished 9:16 vertical clips in under two minutes. The entire process runs locally—no cloud processing, user accounts, or data uploads. It’s gaining traction among content creators focused on Shorts, Reels, and TikTok, especially those prioritising data control and cost efficiency.

How NeonClipping works

The tool operates in two modes: Fast and Deep. Fast mode, the default for videos over 10 minutes, downloads only audio (e.g., 79 MB for an 85-minute video) and uses YouTube’s built-in word-timed captions when available. If no captions exist, it falls back to a local Whisper model. Audio signals—like sudden loudness jumps, raised voices (2–4 kHz energy), or trigger phrases such as “oh my god”, “bhaago”, or “arre yaar”—are scored. A local LLM (via Ollama, e.g., llama3.2:3b) then reviews transcript snippets to generate titles, hooks, and emotional scores. Only the selected video segments are downloaded, trimmed frame-accurately by matching audio, and rendered with burned-in captions.

Deep mode downloads the full video and adds visual analysis: optical flow (using OpenCV Farneback) detects fast motion, impacts, camera whips, or scene cuts, combining them with audio cues to shortlist four times more candidate moments. This mode is more resource-intensive but yields richer context.

Performance varies by hardware and source. On an RTX 2060 with 16 threads, Fast mode processed an 85-minute YouTube stream in 97 seconds, a 3-hour Twitch VOD (using local speech recognition) in 136 seconds, and a 90-second local MP4 in 14 seconds.

Local processing and user control

All data stays on the user’s machine. The app runs a local server bound to 127.0.0.1 only. Star ratings (under each clip) are stored in a local JSON file (data/feedback.jsonl) and used to retrain the moment-scoring algorithm after eight ratings via ridge regression, gradually adjusting signal weights toward user preferences—favouring shout-heavy moments for gamers or calm insights for educators—without sending data off-device.

Users can customise clip length (e.g., 30–80 seconds), choose from 11 native layouts (e.g., facecam overlays, gameplay-only, circle facecam, side-by-side), and pick from five caption styles including MrBeast, Hormozi, and GTA Neon. Captions are word-timed, editable, and exportable as SRT or WebVTT. Hinglish output (Hindi in Latin script) is supported by default; full Hindi transcription requires enabling the local large-v3 Whisper model, which may be slower on low-end GPUs.

Facecam positioning presets (top-left, bottom-centre, etc.) are remembered between sessions. Users can also manually draw face and gameplay regions in the editor. Opening headings (2–60 seconds, default 8) with black backing are supported, and AI-generated curiosity questions can be applied using the “Improve opening heading” feature.

Live clipping and automation

NeonClipping supports live stream clipping for Twitch and YouTube Live. By monitoring anonymous chat activity, it detects spikes exceeding 2.5 sigma and 2.5x the 5-minute baseline, then generates a clip starting ~24 seconds before the peak—accounting for typical reaction delay. Users can add creators to a watchlist; the app records and clips live streams unattended, using the layout, facecam boxes, and caption style selected at the time of adding.

Only one stream is processed at a time. The app must remain open; unattended recordings stop after 3 hours (a new session starts if the stream continues), or when less than 8 GB of disk space remains. The recording file is deleted after clipping. Pressing Stop pauses auto-clipping for 30 minutes; toggling Auto off and on resumes immediately.

Automated live clipping has been tested on Twitch end-to-end. YouTube channel detection works via yt-dlp, but full auto-clipping of YouTube streams has not been tested end to end.

The system requires Python 3.10+, FFmpeg on PATH, and optionally Ollama (with llama3.2:3b recommended) and an NVIDIA GPU for NVENC encoding and faster-whisper. Windows users can install via a one-click batch script (NeonClipping-Setup.bat); Mac and Linux support is experimental (install.sh, not yet tested on real Mac hardware).

What it means

For creators and small studios in the UK, UAE, US, and Kazakhstan, NeonClipping offers a low-cost, private alternative to cloud-based clipping services. Unlike tools from Runway, Opus, or HeyGen, it avoids subscription fees and data exposure. Its reliance on local LLMs and speech models reduces dependency on external APIs—critical for businesses managing compliance, bandwidth, or vendor lock-in.

However, performance depends heavily on local hardware. Users without a GPU may face delays, especially in Deep mode or when processing uncaptioned videos. The lack of end-to-end testing for automated YouTube live clipping suggests caution in relying on it for mission-critical workflows.

The tool’s open architecture allows inspection and modification, appealing to technically skilled teams. Its regression-tested codebase (187 passing end-to-end and unit tests, documented in tests/REPORT.md) and bilingual documentation (English and Hindi) lower the barrier to adoption.

Businesses should consider NeonClipping as a scalable, one-time-setup solution for content repurposing—especially where data sovereignty, cost control, or independence from cloud providers is paramount. Pairing it with local infrastructure (e.g., a dedicated clipping workstation) could streamline social media pipelines without recurring fees.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *