I do mention in the description of V1 that it struggles a bit with hands, and I've had the four-fingers error pop up a good few times. I am working on a successor to remedy this.

I actually don't do a lot of upscaling (since so much of the generating I do is in service of testing new model merges, rather than completing individual projects), but I do use the 4X_foolhardy_Remacri upscaler from CivitAI when I do. I don't know if it's the best, but I like the way that it improves color tone.

I also don't tend to do any single upscaling job at more than 1.5-2x size at once, since I can tune the amount of detail better if I do it in two sequences of 2x than one sequence of 4x. I use Forge (not Neo), the latest version of which (that I'm aware of) includes the DDPM sampler (my favorite, and the one I use for most projects) and HeunPP2 (which is, to my eye, the most detailed, but fairly slow.) DDPM works well at over 20 steps, and I usually use 32-40. HeunPP2's sweet spot seems to be about 50 steps.

I recommend not having your base generation size at more than 1500-1600 on whichever dimension is the longest; the higher it goes, the more anomalies that will be introduced (thanks to the base training dimensions of SDXL, and therefore Pony and Illustrious, being as small as they are.)

For Euler A, I would say you don't need more than 20 steps most of the time.

For Denoising strength, think of it like this: 0.0-0.2 is practically no change, 0.2-0.4 is minor change, 0.4-0.6 is moderate change, 0.6-0.8 is a fair bit of change, and 0.8-1.0 is significant change. It's the level of freedom, on a scale of 0-100, that the AI has to rework elements in your image, and the higher it is, the more things will change. For upscaling, I don't generally go higher than 0.5, and usually aim for around 0.4.

@zxeroe Remains to be seen. Sometimes, I'm satisfied with a day or two of testing, some of my merges take me months to be satisfied with.