AI can help make video archives searchable by what actually happens on screen.
I tested that on 370 hours of public-domain film. The whole index cost about $10 in GPU time to build.
Search “typing on a computer keyboard” and jump straight to the moment it happens.
I processed 1,864 films from the Prelinger Archives on
@internetarchive, cut them into ~60-second chunks, and ran the open 2B video model Marlin-2B on
@huggingface Jobs to generate timestamped descriptions.
The result: 23,148 searchable moments.
Try it:
huggingface.co/spaces/davans…