[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/vt/ - Virtual Youtubers

Search:


View post   

>> No.69586790 [View]
File: 250 KB, 832x1216, 24880-3311828113-source_anime, score_9, score_6_up, by jack dempa, by muk _(monsieur_),_BREAK( rosemi lovelock, rosemi lovelock 1st costume_1.1),.jpg [View same] [iqdb] [saucenao] [google]
69586790

>>69581528
>doesn't seems like we getting XL hll anytime soon.
SDXL is just too slow
It takes 13 hours to train on 10k images - https://litter.catbox.moe/ewkzcx.png
Training on ~ 1 million (on 1 gpu, like in hll-fluff) would take ~ 1300 hours = 54 days
Even if i rent 4x a100 for $5/hour at https://lambdalabs.com/service/gpu-cloud#pricing it would still take about a week

Remember that the guy who made ponyXL used 3x A100 and trained for 3 months on ~ 2 million images. And he bought those A100s for ~ $35000. Even if you rent instead, it would still cost about $8000-10000 to do the same.

Since this is just for fun, i don't want to spend too much time or money. The most I'm willing to do is something like a vtuber-only lora, 300-400 img per vtuber = ~ 50k images. Should be good enough for prompting different vtubers without switching loras, and it should be possible to train it in 3 days

>>69583950
idk how it's supposed to look
https://litter.catbox.moe/ty1t3s.jpg
https://litter.catbox.moe/buaeqa.jpg
https://litter.catbox.moe/t1g6g1.jpg

Navigation
View posts[+24][+48][+96]