“B-roll” is the term for video footage that you cut to during talking head footage to make your videos more visually interesting—it gets boring to just watch someone talk.
I’ve been trying to get AI to make b-roll every six months for the last few years, checking to see if the models would improve enough to be usable.
This is the first time it’s been good enough to share publicly.
My goal was to find a style with animations that was visually interesting, didn’t burn up a ton of tokens, and wasn’t too cheesy (which is hard with AI).
It’s not amazing, but it finally feels like a very good b-roll is possible in the near term. Below are gifs of some of the b-roll I was able to create:
The workflow:
- find an image that fits the theme
- drop it into a Claude Code script
- Gemini restyles the whole photo as a paper collage
- Claude Code then has Gemini cut each figure out of that collage separately
- Those cut-outs get animated locally, stop-motion style, and assembled into an mp4
It’s cost me ~40 cents a clip.
If you want to see finished videos with the clips in them, here’s are two:
- The story of my brazilian friend with a “fake podcast” that gets him clients
- AI models are getting scary good, so back up your stuff
If anyone wants to experiment with it, the github repo is here. If I were you, I would wait for me to spend another 10 hours refining this to get to even better results, which I’ll share when I have it.