ArcFlow is a new way to make text-to-image models draw great pictures in only 2 steps instead of 50, giving about a 40× speed boost.
This paper teaches a video-understanding AI to think in 3D plus time (4D) so it can answer questions about specific objects moving in videos.