bad 4nfe generation performance
#4
by gaga-lee97 - opened
I tried the fl2va lora and used it for 4/8step T2VA generation. I find the 8 step results look good, although I met some blur hands/ layouts in high-dynamic secnes.
But when I trid 4 step inference with the same lora weight, the videos are not clear and seem not good.
I wonder why would this issue happened? I thought the lora should support 4step generation as it said.
Agree with your assessment, 4 is not very good. 8 is the way to go. I believe it is a limitation of the model unfortunately