THAN A LOT! now it takes 27 sec!
#2
by napalm - opened
THAN A LOT! now it takes 27 sec!
Sadly though the results side-by-side...
Just a completely different model. 8/10 turned to 3/10
looks like nvidia's anyflow was used to train it, which does the same thing and supports n-step distillation
Simple ^^
I wouldn't say it's become a 3/10, at least not in my tests, but the composition has indeed changed, which is expected. Sometimes the composition with the Lora is even better; for a V1, it's not bad at all, but it still needs a bit of refinement to truly show its full potential.
It's promising nonetheless, and I thank you for the work. :)