>>2124
>>2125
These are awesome

>>2126
>Not sure if that question was for me or for the guy(s) having trouble to get the Lora working at all, but yes,, I'm using Karras. I've been using nothing but DPM++ 2M SDE Karras for over a year by now. Haven't really kept up to date in that department. Only tried DPM++ 3M SDE but somehow got worse results. Allthough I've not tried other samplers with pony yet. any recommendations?

Samplers are a bit complicated to discuss because they largely don't affect gen quality, it's just the denoising algorithm.

I don't wanna send you down a rabbit hole, but there are specific situations where switching the sampler helps or hurts.
I don't know all samplers and their behavior, but I know these 3 are good to cycle around:
Euler
Euler A
Karras

Euler A is considered pony's "default", as recommended by its creator. It's good all around, the algorithm "zones into" the target gen but doesn't converge, this means, to name one thing, increasing steps over 30 is worthless, since it will keep slightly changing things back and forth, oscilating over the target, so to speak).

Euler is like Euler A but supposedly better (although theory isn't practice). Unlike Euler it converges. This means your gen will get more "precise" the more steps it has.

Both Euler and Euler A calculate composition during the first steps and polish details on the latter ones. This mean if it decides there's a ball shaped object, it won't change shape after the first steps, then it likely won't change what it thinks the object is, and during the last steps it will barely change detail at all.

So for example, using keyword blend as: [watermelon:basketball:0.8] won't do much at all on Euler, and you'll end up with a watermelon, maybe with some vague hint of basketball lines.


Karras, in contrast, has the unique feature of feeding noise into the gen inbetween steps. This can be a good thing or a bad thing. Since it means it can correct stuff at random by "renoising" random areas, or screw everything up by destroying instead of fixing.
If your prompt is full of contradictory tags, or is just really long, chances are Karras will just fuck things up more often than not. It can for example add random limbs or body parts.

But the reason Karras can be better with this lora is because this renoising helps to breakthe "welding" process where for example, if a severed head is drawn a foot over the neck, Euler would slooowly try to reconnect it, either by bringing the torso up, head down, or elongating the neck.

I thoroughly recommend that you lower preview quality and make it refresh every 2 steps or so, so you can visualize how denoising works. It's a much better explanation than I could ever give.