Very nice indeed

#3
by SzilviaB - opened

This is very nice indeed, very original.

Do you have any Class 5 models ? Or even more, Class 6 +?

Thank you!
There are some class 4, 5 and possibly "6" ; unreleased.
One of the main issues: Reigning in their "creative" behavior.

As a result I am working on my own samplers right now (Jan 2025) to address / control this issue during generation in real-time.
Prototypes are complete, working on optimization / tuning.

Basically this module runs like "dry", "quadratic/smoothing", etc etc to "auto-correct" (auto detects problem behavior) model generation behavior.
This allows all class 3, 4, 5+ models to run normally without user intervention and put an end to "gibbish" , "repeats" and other issues.
Still a lot of work to do before release...

I noticed you released 3's and 4's already.

Would be really curious to see what a 5 or 6 is like, regardless of how unruly or error prone they are.

If you're looking for people for testing these models I volunteer.

@SzilviaB
Looking at late next week for first module, baring any issues.
This module will also work with any AI of any size too.

Awesome !

Oh My !

Can't wait to try this out !

BTW, we were talking about how unusual quants are and how a lot of times lower quants will be more creative and higher quants will be more dull.

People have started to make video diffusers as GGUF, I have played around with that and this holds true for video as well, for example:

Q5>Q8>Q4>Q6

BTW2, why nobody makes Q7 quants ?

RE: Q7 ; not a lot of difference here VS: Q6 to Q8 is the better fit.
It is technically possible with a modified version of LLAMACPP.

Sign up or log in to comment