I downloaded the pth files and I used RAFT previously in another project so I was looking forward to upgrade it with a faster version to also keep up with somewhat recent methods. Usually there's less pretrained models to choose from, is there any preference for generalization ability in that list? Meaning should I use a specific one assuming I'll be using this in very different environments?

Also, been reading some papers on this and noticed the memflow one. Is that the current SOTA for practical applications - where low inference times (mainly) and decent quality is a priority? I'm still going to use GMFlow as it's more established for now, but just wanted to ask what you think and if there's a specific flow solution you prefer.

Thanks!