What it does AutoClip is a browser-based video processing system. Paste a YouTube or Bilibili URL (or upload a local file), and it downloads the video, extracts subtitles, runs the text through an LLM to find topic boundaries and “excitement” scores, then cuts the video into clips and assembles collections. The whole thing is packaged as a FastAPI backend with a React frontend, wired together with Celery and Redis for async job handling.

The interesting bit The actual smarts are embarrassingly simple in a good way: it treats the subtitle file as a transcript, feeds it to Tongyi Qianwen (Qwen) for outline extraction and scoring, then maps the LLM’s chosen timestamps back onto the video with FFmpeg. No computer vision, no audio analysis — just a language model reading the dialogue and guessing where the highlights live.

Key highlights

• Downloads from YouTube (via yt-dlp) and Bilibili, plus local file upload
• Six-step pipeline: outline → timeline → scoring → title generation → collection recommendation → video rendering
• WebSocket progress updates so you can watch the LLM think in real time
• Docker Compose and shell scripts for one-command deployment
• SQLite by default, with a migration path to PostgreSQL mentioned

Caveats

• Several headline features are marked 【开发中】 (in development): Bilibili upload, mobile support, multi-account management, and subtitle editing
• Requires a DashScope API key for the Qwen integration; no local model support is documented
• The README badges claim 0 stars/forks/issues, which contradicts the actual 5,548 stars — likely a copy-paste artifact

Verdict Worth a spin if you produce secondary content (clips, compilations, reaction fodder) and want to automate the scissor work. Skip it if you need fine-grained editorial control or were hoping for on-premise LLM inference.