>>91205337Got asked to do these, so posting them here as well. Cute girls holding you at knife point
https://files.catbox.moe/671kiq.pnghttps://files.catbox.moe/u9k346.pnghttps://files.catbox.moe/rh1eht.pnghttps://files.catbox.moe/9v4yne.pnghttps://files.catbox.moe/ey001o.png>>91216000Thanks for testing. Yeah, I was actually surprised at how much better it was than 0.3.1 when I tested it, I remembered it much better in general, but really it was only a few chuubas it did better. And colors on 0.4 are much better.
Still a bit of bleed as you say, but I think that may be me stopping TE training when LR was still high, so it didn't have time to settle.
>burningI think mostly this is just cause I trained the model for longer, giving it time to settle as a whole. But I also think in 0.4, the fact that my dataset barely has any jpeg artifacts might have helped as well. The images feel cleaner to me.
>>91215178Cute Rara
>>91216046Would be pretty hype if you find a way to consistently do that.
>>91216368My wife
>>91219139kek
>>91220436Like the other anons said, the most common models are illustrious and NoobAI based. Unless you're looking to prompt indies or small corpos, those models will know basically any larger vtuber, aside from flow glow cause they are too new.
