Aseiel/VideoHighlighter 是一个托管在 GitHub 上的开源项目，主要使用 Python 编写，采用 GNU Affero General Public License v3.0 许可证，目前获得 385 个 Star、46 次 Fork。
仓库全名
Aseiel/VideoHighlighter
README 摘要
VideoHighlighter English | 中文 | 日本語 | Español Find and explain the moments that matter in footage you won't upload — then export a cut, on your machine. A local desktop app. Drop in raw video; it scores strong moments from scene, motion, audio, objects, actions and the transcript, shows you why each one scored on a signal timeline and in a report, then exports a highlight reel and the individual clips. Nothing is…
Open-source local AI video analyzer powered by Ollama. Visual search, automatic highlights, scene/action/object detection, audio analysis, and subtitle generation. Free, offline alternative to Twelve Labs, Runway, and Descript.
Find and explain the moments that matter in footage you won’t upload — then export a cut, on your machine.
A local desktop app. Drop in raw video; it scores strong moments from scene,
motion, audio, objects, actions and the transcript, shows you why each one
scored on a signal timeline and in a report, then exports a highlight reel and
the individual clips. Nothing is uploaded for analysis, and the basic pipeline
needs no API key.
Free and open source (AGPL-3.0). Windows, macOS, Linux. Intel, AMD and NVIDIA
GPUs.
App interface languages (About → Interface language): English, 简体中文,
Español, Português (Brasil), Français, 日本語, Deutsch, 한국어, Bahasa
Indonesia, Türkçe, Italiano, Polski, Tiếng Việt, Nederlands. This README is
still English-first; translated copies above cover 中文 / 日本語 / Español.
It’s free. To make sure you see new releases in future, please click the
motivation button: the ⭐ at the top of the page. It’s the cheapest payment we
accept.
What people use it for
A 4-hour Twitch or YouTube VOD → a few minutes of highlights. Crowd noise,
your own reactions and scene changes all score, so the loud parts surface
without scrubbing.
A night of CCTV or trail-camera footage → the minutes something moved.
Object detection plus motion, on hours of nothing, locally — which matters
when the footage is of your own property.
A GoPro or drone card → a finished film. The Auto tab finds the card,
copies it off, cuts the highlights, builds the reel and lays music on the
beat, as one resumable job.
A match → the goals. Write a rule for “ball inside net” and it scores that
event by where things are, instead of hoping an action model can tell it
from how the frame looks.
A long interview, lecture or podcast → chapters and subtitles. Local
Whisper transcript, chaptered video, optional local translation.
Dashcam or bodycam review → an account you can hand to someone. The report
states what was measured, what was only said, and what it could not determine.
Not sure which detector fits your footage?
docs/DETECTION-GUIDE.md covers what each one is good
at and where it falls down.
What makes it different
Every run explains itself. The report is the arithmetic
behind each kept moment — the per-signal point breakdown, what fired, what
scored well and still missed the cut. It names the claims that came from the
transcript and were never measured, instead of scoring them anyway.
Open a real one →
Composition rules. You
say what a combination of detections means for your footage — one class inside
another, counted, held over a window — and that becomes an event under a name
you choose. Because a rule re-reads detections that already exist, editing one
and re-running costs milliseconds.
VideoHighlighter — this repository — is free software under AGPL-3.0, and
stays that way. It includes offline analysis, live face detection, VR
side-by-side playback, CLIP search, the composition engine, model training —
including teaching a model from one box drawn in the player, which then keeps
improving while the app is idle — and the model hub.
VideoHighlighter Pro is a
separate paid edition that adds real-time work on top: live object and action
overlays during playback, teaching a live category mid-playback,
find-more-like-this search, open-vocabulary detection and counter/scoreboard
detection.
Explanation is not among them. The report, the findings and the advisor are
identical in both editions — a cloud tool gives you a button and a result you
cannot interrogate; answering “why”, locally, is what this is instead.
Community
VideoHighlighter occasionally has feelings about your footage. When it does:
join the Discord and yell in #support, I’m
usually around. Bugs and ideas are welcome in
Issues.
Frequently asked questions.Read the FAQ for help with:
Released under the GNU Affero General Public License v3.0 — use, modify and
distribute it freely, provided modified versions, including ones offered over a
network, make their complete source available under the same license. Full text
in LICENSE; notice in COPYRIGHT.
Contributors keep copyright in their own work — see
CONTRIBUTING.md and CLA.md. VideoHighlighter is
also offered under a separate commercial license by the copyright holders.
Contributing
See CONTRIBUTING.md for guidelines on code quality, testing,
and what makes a good contribution. The project welcomes bug reports, small fixes,
and feature proposals — please check the CLA before your first PR.
评论