/vt/ - Virtual Youtubers

>>65658611
Quantized models have less precision so they will be a bit dumber but the effect is not significant. It is not cumbersome to quantize them as The Bloke already does it for most models as soon as they appear, but I have no idea how to implement it in the script provided by pixart as my knowledge of LLMs, directly accessing diffusion models and stuff is limited. But I can probably handle API requests and saving the result to file.
The main issue is to get it to work in the first place. I managed to squish down all the errors but now it seems to be hallucinating and describing images even when I don't give it an image.

>>64526178
I think one of the problems is my base mixes are bad. I usually go weighted sum 75-60% Hluffy - 25-40% whatever else. CLIP Merge -> Only A
I didn't bother learning how to do cosine calculation in c*mfy as I just wanted a quick model to play around with.