Img gen/ stable ui nowhere to set the txt2Img denoise cfg?

[Flanker15]

I couldn't find it or I've gone blind but it seema like there's no way to change the denoise anywhere (except for img2img)? I can see it going through the console input always as .6 which is too low for low step models.
The shift doesn't seem to have any cfg either but that's less important.

[LostRuins]

You can set it here:

[LostRuins]

Unless I misunderstand what you're looking for? there's also CFG scale

[Flanker15]

I can see it but it's only on the img2img settings panel for me. I can use img2img to do txt2img and also if you set the denoise and run a job there it sets it for txt2img too but it's weird it's missing on the text2img panel.

[wbruna]

It's missing because it isn't used on txt2img: it corresponds to the amount of noise added to the input image.

[Flanker15]

On low step models like Zimage turbo (and maybe other kinds too I'm not sure) having the denoise set too low can have a bad effect.
Quick example of the same seed in Zit left is 0.6 denoise, right is 1:
https://i.imgur.com/hM9Zl0o.jpeg

I think it's to do with how the 4 and 8 step models work.

[LostRuins]

Looking at your sample image I don't think the definition of noise is the same. It looks more like the added noise of a photoshop filter.

[Flanker15]

It makes other odd artifacts too yeah.
I don't know how the internals of a low step model works but it's what allows it to get acceptable results in 4 steps instead of 20.
The only difference in the 2 images is the denoise is set lower in one sampler. I can get stable ui in kobold to make identical pics by using the img2img to change the denoise setting to something and then running the same prompt and model in text2img. So it's sort of not a problem now that I figured out the work around but other people might get stuck as more people use Zit.

[wbruna]

For img2img in ZIT, you may want to increase the step count a bit, when reducing the strength. The denoising strength is applied to the number of requested steps (so 0.5 strength roughly halves the step count), and for distilled models those missing steps may be very important. You can check on the Koboldcpp log how many steps are being actually applied.

I also had good results with the sgm_default and smoothstep schedulers.

[Flanker15]

Yeah if you run it at more steps (around 20 seems to work) it will "burn through" the lower denoise but then you've lost the turbo part a bit.
I think just changing the name of the img2img denoise setting (in kobold settings) and then having text2img inherent the value as well would work.
Right now my fix is to just run a img2img with no input pic which works by itself and then I can switch to any mode and it will remember the denoise setting still.