claude-real-video
Let Claude — or any LLM — actually watch a video
Start new thread

claude-real-video - Let Claude — or any LLM — actually watch a video

by•
Open-source CLI that turns any video into scene-aware keyframes and a timestamped transcript, processed locally. New since the first launch: exact per-frame timestamps, faster-whisper transcription, a trilingual interactive viewer, and honest dedup for slow-changing footage.

Add a comment

Replies

Best
Maker
📌
Hey PH! I'm Leo, a non-engineer from Taiwan. I kept pasting YouTube links into AI chats and realizing they only read the transcript. The picture - the cuts, the pacing, the visual hooks - was invisible. So I built crv with Claude Code. It grabs a frame at every scene change (not a fixed 1 fps), throws away near-duplicates, and transcribes the audio. Everything runs on your machine. On a 58-second test clip that's 26 frames instead of 58, with nothing missed. It hit the Hacker News front page last week and just passed 700 stars, which still feels unreal. There's also a bundled Claude Code skill: paste a video link into the chat and Claude watches it by itself. It's MIT and `pip install claude-real-video`. Ask me anything - especially where it breaks. :)

How does the scene-change detection decide which frames actually matter, and can I tweak the threshold if it's pulling too much or too little from longer videos?

Maker

  Yes - the --scene flag is exactly that dial (default 0.30, lower = more frames). Long videos usually want a slightly higher threshold plus --grid to pack survivors into contact sheets so token cost stays flat.

how does the scene-change detection actually figure out which frames matter across different video styles like fast cuts versus long talking heads?

Maker

  Different failure modes: fast cuts give you lots of frames (dedup then collapses the repeats), talking heads barely change so the density floor guarantees you still get periodic frames, and the transcript carries most of the meaning there anyway.