rectified-flow-pytorch
rectified-flow-pytorch copied to clipboard
What's better, ReFlow or Consistency FM?
I've been reading a lot of Flow matching papers recently and I'm a big fan of anything that gets close to optimal transport, like we had with Wasserstein GANs in the past. Since ReFlow and Consistency FM have the same goal of getting perfectly straight paths, I was wondering which of them is better. The Consistency FM paper claims to outperform a 2-Rectified Flow, but you know how it is, sometimes papers exaggerate their results. Since you implemented both techniques in you repo, can you share your experiences with them? Do you have a clear favorite?