Hacker News Reader

I wonder what served as the dataset for the model. Videos on YouTube presumably, since messing around with the film industry would be too expensive?

View on HN

How would they access youtube tough?

yt-dlp

Almost certainly troves of stock footage. The type of exaggerated motion seen in these examples is very reminiscent of stock footage. And it is heavily textually annotated for search.

Reply on news.ycombinator.com