AI-powered video clipper with a visual dashboard. Paste a video URL, get short-form vertical clips with animated captions — ready to post anywhere.
No cloud account required. Everything runs locally. Works out of the box — no API key needed.
Paste a URL, hit Clip, and watch the pipeline run in real time. Adjust clip count, duration, AI model, caption style, fonts, and more — all from the UI.
The AI reads the full transcript and picks the best moments worth clipping — things with a strong hook, a clear story arc, and a satisfying payoff. It avoids intros, dead air, and mid-sentence cuts. Each clip is scored on how well it would stop someone from scrolling.
For landscape videos (standard YouTube, podcasts, screencasts), vid-clipper automatically detects the speaker using computer vision (YOLO person detection with MediaPipe and Haar cascade fallbacks) and reframes the shot into vertical 9:16.
The output is a split-panel layout:
- Top third — the full horizontal video, letterboxed, showing context (slides, screen, B-roll)
- Bottom two thirds — tight face crop, automatically centered on the speaker, scaled to fill the frame
No manual cropping. The speaker stays in frame even if they're off-center in the original.
Word-by-word captions are burned into each clip with highlight sync — the current word lights up as it's spoken. Font, size, color, and position are all configurable.
Before each clip plays, a 6-second teaser pulls the most compelling spoken moment from inside the clip and prepends it as a hook — the kind of thing that makes someone stop scrolling before the full clip even starts.
- Downloads any video via yt-dlp (YouTube, Twitter/X, and 1000+ sites)
- Local transcription with faster-whisper — no API key, runs on your machine
- AI moment selection via OpenRouter — free by default, supports any model
- YOLO + MediaPipe face detection for automatic speaker framing
- 9:16 split-panel output (top context + bottom face crop)
- Animated word-by-word captions with highlight
- Hook teaser auto-generated from the best spoken moment
- Visual dashboard — job queue, live progress, clip preview, history
- Everything saved locally — no uploads, no cloud, clips are yours
Requirements: Python 3.10+, Node.js 18+, ffmpeg
# 1. Install ffmpeg
brew install ffmpeg # macOS
apt install ffmpeg # Linux
# 2. Clone and set up backend
git clone https://github.com/anl331/vid-clipper.git
cd vid-clipper/backend
pip install -r requirements.txt
cp settings.example.json settings.jsonOpen settings.json. No API key needed — it uses a free model by default. Optionally add an OpenRouter key for faster/better paid models:
{
"openrouter_api_key": ""
}Free by default. Leave
openrouter_api_keyblank and vid-clipper usesgoogle/gemini-2.0-flash-exp:freeautomatically. Add a key to unlock paid models like Gemini Flash or Claude.
# 3. Start the dashboard
cd ../dashboard
npm install
npm run devOpen http://localhost:5180 — that's it.
Everything is configurable from the Settings tab — no code edits needed.
| Setting | What it does |
|---|---|
| Max clips | How many clips to generate per video (default: 5) |
| Min duration | Shortest a clip can be, in seconds (default: 20s) |
| Max duration | Longest a clip can be, in seconds (default: 90s) |
| AI model | Which LLM picks the moments — swap anytime |
| Setting | What it does |
|---|---|
| Clip format | split (top context + bottom face), center (tight face crop only), fullscreen (original framing scaled) |
| Crop anchor | auto (face detection), left, center, or right — controls where the speaker is framed |
| Setting | What it does |
|---|---|
| Font | Caption font (Squada One, Impact, Arial, etc.) |
| Font size | Caption text size |
| Vertical margin | How high/low captions sit on the frame |
| Words per line | How many words appear at a time (caption chunk size) |
| Highlight | Toggle word-by-word highlight on/off |
| Highlight color | Color of the active word highlight |
| Setting | What it does |
|---|---|
| Title enabled | Show/hide the clip title overlay |
| Title font | Font for the title (can differ from captions) |
| Title font size | Title text size |
| Title position | intro (shows at start then fades) or persistent (stays on screen) |
| Intro duration | How many seconds the title shows before fading (default: 3.5s) |
| Title margin | Vertical position of the title on the frame |
| Setting | What it does |
|---|---|
| Provider | local (faster-whisper, free, private) or groq (cloud, ~10s for 1hr video) |
| Groq API key | Optional — only needed for cloud transcription |
Any model on OpenRouter works. Pick based on speed vs quality:
| Model | Speed | Quality | Notes |
|---|---|---|---|
google/gemini-2.0-flash-exp:free |
Fast | Good | Default — free, no key needed |
google/gemini-2.0-flash-001 |
Fast | Great | Paid, best all-around |
google/gemini-2.5-flash |
Fast | Excellent | Newer, slightly better |
anthropic/claude-3.5-sonnet |
Medium | Excellent | Best for nuanced content |
openai/gpt-4o |
Medium | Excellent | Solid alternative |
Switch models anytime from the Settings tab — or per-job from the clip form.
On a modern machine (M-series Mac or equivalent), a 1-hour video goes from URL to 5 finished clips in under 2 minutes:
| Step | Time |
|---|---|
| Download | ~15s |
| Transcription (Groq) | ~10s |
| AI clip selection | ~15s |
| Rendering 5 clips | ~60s |
| Total | ~2 min |
Running transcription locally (no Groq key) adds ~3 min for a 1-hour video. Everything else stays the same.
By default, transcription runs locally using faster-whisper — free, private, no API key needed. It takes ~3 min for a 1-hour video on a modern machine.
Want faster transcription? Add a Groq API key in Settings. Groq's Whisper API is free and cuts that to ~10 seconds.
Clips are saved to backend/clips/<video_id>/:
backend/clips/
dQw4w9WgXcQ/
clip_01_The_hook_moment.mp4
clip_02_Key_insight.mp4
- Backend: Python, yt-dlp, ffmpeg, faster-whisper, OpenRouter, YOLO, MediaPipe
- Dashboard: React, Vite, Tailwind CSS
- Storage: Local JSON files — no database, no cloud required
PRs welcome. Open an issue first for big changes.
MIT