CFG scale on Pony XL-based finetunes
What I personaly love about Pony XL and finetunes is not only their great knowledge of anatomy but also their ability to add extra stylization on high CFG scales.
Yeah... 120 pictures, 6 samplers, 1-20 CFG scale. And I tested a few Pony models over some time and all of them work more-less the same.
Main points that I noted:
The higher we go - the crazier colors will be. So, if the base model loves bright or vivid colors - the results may be overturned.
The same goes with... well stylization. More details, more crazy stuff, sometimes changes stylization completely, more chances of weird anatomy. But Pony handles most of this VERY well.
Non "a" (Ancestral Discrete Scheduler) samplers are almost the same at lower CFG scales (1-3). Same with "a" group.
Some non-Pony finetunes can also benefit from that, but they mostly turn into a mess after CFG 14.
I didn't notice much difference in generation time comparing CFG 3 and CFG 20.
LCM can compete in detail with typical samplers at CFG 8+ but will be much faster. But also there is a lot more garbage. But also LCM works in an interesting way if we include LORAs into the mix, pushing the base model very close to LORA's style.
DMP++ has the most boring stylizations so far.
"a" samplers offer much more variety on higher CFG.
Heun still offers the most detailed results but also the most slow one.
Euler, Euler a, and DMP2 a offers interesting stylizations in almost 1/3 of the time of Heun. But DMP2 a jumps into way crazy results. But some of them are still very usable, hah.