/vt/ - Virtual Youtubers

LHC-anon here.
I have disappointing news, worrying news, bad news and amazing news for 0.4
Disappointing: I played around with more approaches to training since learning that just unet does work and got basically no dataset work done
Worrying: Tested learning artists/styles alongside lokr, cause I did want to add stuff like oil paintings etc. Was only smallscale, but results don't seem too promising, with the learned characters not really working well on learned styles. could have many factors, so only worrying, but still makes me hesitant to call Lokr the way forward, cause abstraction is the thing it's supposedly notoriously bad at, so it might be the reason.
Bad news: Because of the above, I wanted to test finetuning again. I quickly found out that I can't seem to get vpred working for OneTrainer. Supposedly it's automatic when rescale noise schedule is active, but that's not right, and manually setting the flag in the config file also didn't seem to help, and it seems like there's generally an open issue around vpred, seeing as a lot of people can't get sampling to work.

....However the amazing news is that I then tried it on sd-scripts where I then had to really figure out how dataset prep works, and now it works. And boy does it work. Was able to basically use my lokr config, without the lokr part and it not only didn't explode or not learn at all like eps did on OneTrainer, it learned better than lokr in my tests so far in the same time. Remains to be seen if it was just luck, but as it's going well for 1/3 of the 0.3 dataset I'm hopeful.
So yeah, 0.4 will just be a straight up finetune, no workarounds.

>>90468534
Afternoon TWL, cute bear!
>https://files.catbox.moe/djf87s.png
Kinda impressive multitasking here
>>90463592
Godspeed anon! I know the feeling now
>>90453956
I admit I don't know a whole lot about yoga, but isn't that exactly how yoga works?
>>90446211
Cute Orca
>affiliate
I didn't watch her, same as with ame, so I'm not that affected, but I feel with all the other anons.
Funny that I've watched more dooby now than Ame ever
>>90445390
Goated song