Simo Ryu

That's interesting. I'm not really sure...

Wow @CookiePPP thank you so much for the comparisons! I would be so happy if you let me use your benchmarks on my paint-with-words repo as well, do you mind...

> @cloneofsimo Feel free to use anything I posted on this thread. Thank you so much! I will certainly credit you when I add some of these materials! And another...

Awesome! Thank you so much for sharing that information as well! I agree 100% that certainly much better configuration could be possible as the model structure differs with eDiffi.

@CookiePPP Based on your findings, repo I've added user-defined weight scaling function as well as some findings in my repo. I hope this feature gets added to A1111's repo as...

Well, eDiffi does use literally three different conditionings and something like 5 times more parameters, so... Just in case anyone is interested, there are LOT of room for improvements in...

> One idea for improvement: What if we'd generate unique noise for each of the word zones ? Then when re-rolling we could choose which words we want to re-roll...

Something to worth noting, during conversation with andreas * Microbatching is `not guranteed` to always benefit *effective throughput* . It ultimately depends on memeory-boundedness of the model operations To actually...

Huh, it happens in the middle of the steps.....

Thanks for the issue! Ive merged @timh 's branch so have a look if its still the case!