Guides
8 guides that build one step of a clipping pipeline with open source code on public domain NASA videos, then send the same video to one API call that costs $0.03 per source minute.
Clipping guides: 8 guides, each run for real on a NASA video
Word timing for captions, the person talking, a 9:16 face crop and cuts between shots all have open source answers, and each one here comes with code we ran and the output it printed.
Updated
All guides
Each guide ends with the same job sent to POST /v1/clips and the clip it returned.
| Guide | What you build |
|---|---|
| Active speaker detection in Python | Find who is talking in a two-person shot on a CPU with YuNet landmarks and mouth motion, then check it against the transcript. |
| AutoFlip alternative | Google ended AutoFlip support in 2023. Rebuild its shot-by-shot vertical crop in Python with MediaPipe Tasks and PySceneDetect, or send one request. |
| Build an OpusClip clone | The whole clipping pipeline in 5 Python files: transcribe, pick moments with an LLM, reframe to 9:16, caption and render, with real costs. |
| Face tracking crop from 16:9 to 9:16 | Track a face with OpenCV YuNet and move an ffmpeg crop with sendcmd to turn a 1920x1080 lesson into a 1080x1920 vertical clip. |
| Find viral moments with an LLM | A tested prompt that makes an LLM pick clip-worthy moments from a transcript by hook, complete thought and payoff, plus the checks its answers need. |
| Scene cut detection for clips | Find hard cuts with PySceneDetect and ffmpeg scdet, split a clip into shots, and see why a vertical reframe must restart at each one. |
| TikTok safe zones | TikTok, Reels and YouTube safe zones for captions in pixels at 1080x1920, measured from TikTok's own template, with a script that checks your clips. |
| Whisper word timestamps | Word-level times from faster-whisper, whisper.cpp, WhisperX and openai-whisper, cut to one clip and paged into karaoke captions. |
How each guide works
Every guide starts from one of the 4 public domain NASA videos the API demo uses, so you can download the same file and get the same starting point. We run the open source tools on a CPU, pin the library versions in the install steps, and paste the real output under each code block.
Then we send the same video to the API and embed the clip it returned. The code is short on purpose, so each idea is easy to see, and every guide says where its code falls short on real footage: shots with no face, two people in frame, a word that lands on a cut.
What a clipping pipeline needs
A long video becomes a short clip in a few steps. You need a transcript with a time for every word, a way to pick the moments worth posting, a crop that keeps the speaker in a vertical frame, cut detection so that crop never slides across an edit, and captions that light up in time with the words.
Each step is a few dozen lines on its own. The work is in the joins: word times have to move to each clip's own zero, the crop has to restart at every cut, and the active speaker has to win by a clear margin before the frame moves.
When to build it and when to call the API
Build a step yourself when you need control over it, like your own caption design or a detector tuned to your footage. Call the API when you want the whole chain from one request. It takes a direct file URL or an upload, not a YouTube or other page link, and returns ranked clips as MP4 files, 1080x1920 for 9:16, with captions in 3 presets plus SRT and VTT.
It costs $0.03 per source minute, billed per second, and $0.02 past 2,000 minutes in a month. New accounts get $1 of free credit, enough for 33 minutes of source video. See pricing for the details.
Questions, answered
Are these guides free to follow?
Yes. The tools are open source and run on a CPU, and the videos are public domain NASA footage. Only the API comparison at the end of each guide costs money, at $0.03 per source minute.
Do I need a GPU?
No. Every code block in these guides ran on a CPU. Some models in the comparisons, like audio-visual speaker detectors, ship demo code that expects a GPU, and the guides say so.
Why NASA videos?
They are public domain, so anyone can download and test the same files. They are also the demo videos in the dashboard, so a Try it link opens New clips with the same video loaded.
How is the output in the guides checked?
Every code block was run on a real video, and the output under it is pasted from that run. The install steps pin the library versions we used, so you can rebuild the same setup.
Start free
Your back catalog is full of shorts
Your first $1 is free: 33 minutes of video, every clip included up to 1080p60 and 20 Mb/s.
