Lots of valuable nuggets in the Sora paper, but a few things stood out for me.
— Pietro Schirano (@skirano) 16 février 2024
– Training with different aspect ratios improves output quality for diffusion models (we suspected this for a while but it's nice to have confirmation).
– Better captioning leads to more diverse and… pic.twitter.com/CvORNbUdxJ
Lots of valuable nuggets in the Sora paper, but a few things stood out for me. – Training with different aspect ratios improves output quality for diffusion models (we suspected this for a while but it's nice to have confirmation). – Better captioning leads to more diverse and