Topic: Configuring Stablemond with Krita Plugin question.

Posted under Art Talk

xtreemdirtbagge

Member

I'm trying to get the settings locked in while using StablemondAI-SDG on Krita, and I'm getting more artifacts than I was hoping for in backgrounds and similar, especially when attempting to generate large numbers of characters. (EG: Prompting for 'detailed background, multiple girls, 6+girls,')

So I'm trying to tweak all the little things - CFG, sampler steps, yes, but Krita also allows me to change 'Preferred Resolution: Image resolution the checkpoint was trained on', and I may be hallucinating, but I think moving it from 1024 to 1280 to 1344 is improving things? Is that possible? What resolution was Stablemond trained on, exactly? I can find occasional mentions of the output resolution possibly being higher (maybe as high as 1344x1344) but was that the trained resolution?

Also, enabling SAG seems to make things foggy, but, again, I might be hallucinating, some stuff DOES seem clearer... Cranking CFG way up (to like 10 on Euler Ancestral/SGM Uniform) gets rid of the fog, but seems to still have some glitches.

And it's hard to make comparisons because all of these tweaks seem to kick new randomness into the generation, so it's harder to do a side-by-side comparison on the same seed.

So, yeah.

Is there actually a way to tune things for maximum background detail in Krita (which seems to be a comfyUI frontend), or is this all me hallucinating? Are there actual known values to use? And, is there a better way to try and calibrate this than squinting into the background of pictures?

Edit: One of the numbers being quoted for the resolution, and one that seems to work for a lot of stuff with early testing, is 1384.

Updated

KalethOrebiter

Privileged

xtreemdirtbagge said:
1024 to 1280 to 1344 is improving things? Is that possible?

Yes, but there's a limit to how big you can get it before it starts to go the other way. Short explanation if the model can only remember s many pixels before forgetting what it drawn.

xtreemdirtbagge said:
What resolution was Stablemond trained on, exactly?

You'd have better chace to get answer to those quesstions on forum dedicated to Stablemond.

xtreemdirtbagge said:
Prompting for 'detailed background, multiple girls, 6+girls,'

The more complex the scene, the more numerous are the failing points. generating 6 characters is ambitious.

xtreemdirtbagge said:
Are there actual known values to use?

There's no magic numbers, nope. So you haave to slowly learn what the parameters are for either by trial and error, or by looking at guides

xtreemdirtbagge

Member

kalethorebiter said:
Yes, but there's a limit to how big you can get it before it starts to go the other way. Short explanation if the model can only remember s many pixels before forgetting what it drawn.

I know that at really high levels, it visibly goes the other way as images 'split' and double, and at really low levels it looks a little burned out. But what really confuses me is that setting the resolution higher can make refining very small details in a selected region way, way more accurate, even as larger generations get very, very weird.

I'm guessing this might be a place where the way Krita and ComfyUI are interacting weirdly, rather than a general feature of this stuff? Or does something similar happen with other inpainting/refining interfaces?

kalethorebiter said:
There's no magic numbers, nope. So you haave to slowly learn what the parameters are for either by trial and error, or by looking at guides

Gotcha. Thank you. Well, at least it's reassuring to know I'm not the only one who needs to play with all the fiddly little settings!

KalethOrebiter

Privileged

xtreemdirtbagge said:
the resolution higher can make refining very small details in a selected region way, way more accurate

Normal, because as the resolution goes up, the micro details become "normal scale" details.