Still trying to get started with video diffusion based on the rentry. I can't get image to image working yet, I am probably going to make a wall of text post about it tomorrow if that ends up still being the case but I am working with t2v now and I have a question about fp16 accumulation aka "--fast".
I have a 3060, so it is slow as shit and despite doing this for hours I haven't experimented comprehensively yet. BUT literally in the first instance of testing --fast I encountered major quality drop. The coherency of the waves in the video, especially the second wave drops massively and very noticeably when I got fp16 optimization enabled. It is only giving circa 10% performance boost (compared to sage attention which is around 30% with little quality loss, at least in this instance)
My question is this, is it usually like this or did I just get unlucky? Because to me this much degradation doesn't seem worth it for 10% faster gens, even in my lower end rig.
Videos in question:
https://streamable.com/radfbh
https://streamable.com/3eeemf