/vt/ - Virtual Youtubers

Apparently I just don't know how to actually quit something since I tried yet another anya embed. At this point it's less because I want the embed and more just because I wanted to know what was going wrong, and I was pretty sure it wasn't any of the settings (triple checked those) or the dataset (it has some limitations but wasn't awful and no one said anything about it when I linked it here.

So I figured it might be the graphics card, bought a bit of time on google colab premium, and trained it there. It's still not amazing, takes a lot of extra descrption tags to really get an Anya and seems to have a high rate of generating extra legs, but I think it's good enough to say that it actually did work this time.

Which I guess confirms that graphics card is more than just generation speed. Just because your gpu CAN handle the training at some super slow pace doesn't mean you'll get anything good out of it if you wait that long.

I might try training an 8 token one on the same data just to compare them and then link whichever one worked better? I don't know I'm still not totally satisfied with it