Quick Answer: NeonClipping is a free video clipping tool that runs entirely on your own computer. Paste a YouTube, Twitch, or Kick link (or use your own MP4), and it returns captioned 9:16 clips with titles and hooks in about two minutes for a short source. It scores moments from audio, speech, and motion, then downloads only the sections it keeps.
Most clip generators want an account, a monthly fee, and a copy of your footage on their servers. NeonClipping skips all three. It's an open-source project, and you run it on your own machine, so the video never leaves your computer. The part worth understanding is how little of the video it needs to look at, because that's where the speed comes from.
Why the fastest method never watches the whole video
The common assumption is that a clipping tool has to process every frame to find the good parts. NeonClipping does the opposite in its default Fast mode, which kicks in for videos over 10 minutes. The approach works in five steps:
- Download audio only. For an 85-minute video that's about 79 MB, and it takes roughly 10 seconds.
- Read the transcript from YouTube's own word-timed captions, so no speech model is needed. For other sources, a local Whisper model (through faster-whisper) is the fallback.
- Score the whole audio track for sudden loudness jumps, measured against a two-minute rolling baseline, plus raised voices in the 2–4 kHz band.
- Send the best candidates' transcripts to a local LLM through Ollama. The model scores each moment and writes a title and hook. Start and end points snap to sentence boundaries inside your Min and Max length.
- Download only the chosen sections, trim them precisely, and render 1080x1920 clips with burned-in captions.
Here's the counter-intuitive part. Audio is a far cheaper signal than video, and for talk-heavy or shout-heavy content it's often the stronger one. Decoding every frame costs time and adds little when a streamer's loudest moments are already audible. The video step only happens after the audio has narrowed the field.
How it decides which moments are worth keeping
Each candidate moment gets scored from several signals, and the best ones survive. The signals break into two groups.
Signals from the sound and the words:
- Loudness jumps that stand out against the last two minutes
- Raised voices, measured as energy in the 2–4 kHz band
- Keyword triggers such as "oh my god," "run," and "police," including Hinglish and Hindi phrases
- AI score from the local LLM, covering emotion, punchline, and payoff
Signals from the picture:
- Optical-flow motion using OpenCV's Farneback method, which catches fast objects, impacts, and camera whips
- Explosion-like flashes, red and blue light flicker, and scene cuts
Here's where most guides go wrong: they treat the LLM as the whole system. In NeonClipping, Ollama is optional. The README says that without it, clips are chosen by the audio and motion score alone. A pull of llama3.2:3b adds AI titles, hooks, and scores, but the pipeline still works without it. That matters if you're on modest hardware and want to test the tool before installing another model.
Fast mode vs Deep mode: which one to use
Fast mode is the default. Deep mode trades time for thoroughness. The table below sets out the difference based on the project's documentation.
| Setting | Fast mode (default, videos over 10 min) | Deep mode |
|---|---|---|
| Source download | Audio only (79 MB for an 85-minute video) | Whole video |
| Visual analysis | None; scoring relies on audio and transcript | Adds optical-flow motion and cut density to the audio signals |
| Candidates shortlisted | Baseline shortlist | 4× more candidates |
| Context given to the LLM | The moment's transcript | 15 seconds before and after each candidate |
Use Deep mode when the visual action is the story, such as a game where the impacts matter, or when Fast mode misses moments you can see were good. Stick with Fast mode for long streams and talking-head content, where the audio carries most of the signal.
What the benchmarks show (and what they don't)
The author measured Fast mode on one machine (RTX 2060, 16 threads) and produced three clips each time:
- 85-minute YouTube stream with YouTube captions: 97 seconds
- 3-hour Twitch VOD with no captions, using a local speech model: 136 seconds
- 90-second YouTube trailer: 21 seconds
- 90-second local MP4: 14 seconds
The Twitch result is the one to notice. A three-hour recording without captions still finished in just over two minutes, but only because the speech model ran locally on a GPU. Without that GPU, expect the speech step to take much longer. The author notes that your numbers will depend on your network and CPU, and that's the honest reading. Treat these as a sense of scale, not a promise.
Your ratings teach the scorer what you like
Every clip has a five-star rating. Ratings stay in data/feedback.jsonl on your machine. After eight ratings, the tool refits the signal weights using ridge regression, which is pulled toward the defaults and blended in gradually.
The practical effect is that the scorer adapts to your channel. A gamer whose audience loves shouting will see loudness and raised voices weighted higher. A calm explainer's ratings push those same signals down. Most tools lock in one idea of a "viral" moment, and that's the lazy default here too. Adapting to your own taste is the better design, though eight ratings is a small sample, so the early adjustments should be treated as a nudge rather than a verdict.
Editing clips after they're rendered
Once clips are exported, you can change how they look without re-processing the source. Open My clips and choose Edit clip. The editor offers 11 speaker layouts, including camera above or below gameplay, blurred fill, circle facecam, and side-by-side. It also has a speaker-size slider, editable subtitle words, and five caption styles: GTA Neon, Impact, Clean, Hormozi, and MrBeast.
You can apply a change to one clip or to every saved clip from the same source URL. Each clip keeps its own timestamps, transcript, and title. The author's validation on a single 796.9-second source exported all 11 layouts and five caption styles, and a seven-clip layout update took 119.58 seconds while reusing each clip's source range and word list.
There's a real trade-off here. Burned-in captions and framing mean every change re-encodes a new MP4. Previous exports are kept, but the library shows the latest version, so check disk space if you iterate often. Also, editing layout and captions does not change a clip's start or end. Moving those boundaries needs a separate trim workflow.
Limits and risks to check before you rely on it
A responsible guide has to name the rough edges. The README is candid about several of them:
- Mac install is untested. The
install.shscript exists but hasn't run on a real Mac yet. Windows has a one-click setup, and Linux is the tested path. - Auto-clipped YouTube streams were not run end to end. Twitch creator clipping was tested.
- Best Hindi captions use the large-v3 model locally and can be slow on small GPUs.
- Unattended creator recordings stop after three hours and when free disk falls below 8 GB, and the app must stay open.
- NVENC is optional. Without an NVIDIA GPU, export falls back to x264, which is slower.
- Rights still matter. The tool doesn't decide whether you're allowed to clip and post a source. Check the creator's terms and copyright before you publish anything you didn't make.
Frequently Asked Questions
what is a free video clipping tool like neonclipping?
A free video clipping tool like NeonClipping takes a long video and cuts out short, shareable clips. NeonClipping runs on your own computer with no account or upload. It returns 9:16 vertical clips with burned-in captions, titles, and hooks, and it works from YouTube, Twitch, Kick, or your own MP4 files.
how do i make captioned shorts from a long youtube video for free?
Install NeonClipping and run it locally, then open http://127.0.0.1:8770. Paste the YouTube link, set your minimum and maximum clip length, and press Make clips. Fast mode reads YouTube's own captions, so the setup costs nothing beyond your own hardware. Ollama is optional for AI titles and scores.
why does the fast mode skip the whole video?
Fast mode downloads only the audio, then scores loudness, raised voices, and transcript content to find candidates. Only the chosen sections get downloaded and rendered. This keeps the initial download small, around 79 MB for an 85-minute video, and it's much faster than decoding every frame.
do i need ollama to use neonclipping?
No. Ollama adds AI titles, hooks, and scores using a local model such as llama3.2:3b. Without it, clips are chosen by the audio and motion score alone. The LLM improves the wording and the ranking, but the core pipeline runs without it.
Pass this to someone who's stuck scrubbing through a three-hour stream for ten usable seconds. Try a free video clipping tool on one video you already know well this week: set Min/Max to 30–80 seconds, run Fast mode, and count how many of the clips you'd actually post. The result will tell you more than any feature list.