[Stable Diffusion] Prompt Sharing and Learning Thread

[Nano999]

yes, you have to add "--no-half-vae" in set COMMANDLINE_ARGS and remove " --disable-nan-check "

Still having the same issue, it generates black solids in some cases.

 

[Nano999]

btw can we put this setting to the main page? As we did with SD VAE selection?

 

[Nano999]

Is there a way to lower the strength of image change when applying hires.fix?
Because I like the style and design of the image more than hires.fixed image version in some cases.

 

[Mr-Fox]

btw can we put this setting to the main page? As we did with SD VAE selection?

Yes. Add "save_images_before_highres_fix" in settings/User Interface/ Quicksettings list .
" You can put any and all options from the settings tab up there, you just need find its id/key. you can find them here"

You must be registered to see the links

 

[Mr-Fox]

Is there a way to lower the strength of image change when applying hires.fix?
Because I like the style and design of the image more than hires.fixed image version in some cases.

Not that I know of unfortunately. I have had this thought too at times.

 

[daddyCzapo]

Is there a way to lower the strength of image change when applying hires.fix?
Because I like the style and design of the image more than hires.fixed image version in some cases.

As far as i know, the change made by the hires fix to the art style depends heavily on the upscaler you use. The fix aims to preserve the original composition, but the overall style can, and in 90% of the cases will be changed no matter what you do.

 

[Mr-Fox]

Still having the same issue, it generates black solids in some cases.

Make sure that you don't have the NSFW setting enabled. I know that I have seen it somewhere but I can't find it atm, I must be blind..
Read this

You must be registered to see the links

and see if anything they suggest can help you.

This is the main solution:
" Open “webui-user.bat” in any text editor.
Write Text after [set COMMANDLINE_ARGS=].
・For RTX16 series users, enter [ “–precision full –no-half” ].
・If you have a low performance grabo, enter [ “–medvram” ]. "

 

[devilkkw]

My first posts using

You must be registered to see the links

by our own devilkkw.

View attachment 2491154

You can see more in this post.

Thank you for posting result, i like tath and appreciated take some time to made it. Good result for my, and first time i see how it work with hire.res upscale.
I've have to ask: you linked kkw-fantareal, but in png i see kkw-ideal. is there a mismatch?

Is there a way to lower the strength of image change when applying hires.fix?
Because I like the style and design of the image more than hires.fixed image version in some cases.

For keep image use low denoise strenght, and double step of image generation.

 

[Mr-Fox]

Thank you for posting result, i like tath and appreciated take some time to made it. Good result for my, and first time i see how it work with hire.res upscale.

My pleasure.

I've have to ask: you linked kkw-fantareal, but in png i see kkw-ideal. is there a mismatch?

Yes you're right.. Sorry about that.

 

[devilkkw]

No problem, thank you for fix.

・If you have a low performance grabo, enter [ “–medvram” ]. "

I run without it and without -lowvram. these slow down generation.
After some test on torch: 2.0.0+cu118 i restored old torch: 1.13.1+cu117  •  xformers: 0.0.16rc425 because is faster.
I have made 2 venv folder for each version, so i decide what version run without reinstall all.

Waiting for xformers for cu118.

 

[Sepheyer]

Stealing ppls sht:

You must be registered to see the links

by: TobeyGG

 

[Mr-Fox]

Stealing ppls sht:

View attachment 2492291

You must be registered to see the links

by: TobeyGG

This is essentially an advertisement for his Margot Robbie Textual Inversion (embedding).
Very nice image though. I have to admit that I have a small crush on Margot.
On civitai there are 2 lora's and 1 textual inversion for free.

You must be registered to see the links

 

[fr34ky]

Experimenting with the various control net settings, using a photo as original and own created LORA for the face: nothing beats early morning yoga! View attachment 2490820 View attachment 2490821

Is her face based on a pornstar? It reminds me a lot of 2 pornstars, I don't remember their names xD but they are european.

 

[fr34ky]

Is there a way to lower the strength of image change when applying hires.fix?
Because I like the style and design of the image more than hires.fixed image version in some cases.

I don't use hi res fix, my workflow is sending the first txt2img pic to img2img and make a higher resolution picture there.

But, I tried hi-res a couple of times, and you can choose the hires steps and the denoising streght, if you set a lower denoising strength you'll keep a picture very close to the original, but with higher resolution. Hi res steps are like sampling steps, you can play around with them too, but for what you want to do the most important is the lower denoising strength.

I've tried before posting, start with a low denoising streght like 0.22 and you get a picture almost exactly like the original but with higher res.

 

[Mr-Fox]

How long do you take per image?

That's the big drawback with hires fix. It really takes a long time. For me it takes 6-7minutes usually, sometimes even more but I always try to squeeze everything I can out of my GPU.

 

[azrazr]

How long do you take per image?

4 mins ( gtx 1060 6gb) ,,

 

[miaouxtoo]

Don't forget to check out

You must be registered to see the links

.
I would recommend

You must be registered to see the links

and

You must be registered to see the links

.
For Ultrasharp, choose to download in zip form and then unpack in Stable-Diffusion\stable-diffusion-webui\models\ESRGAN.
NMKD should be put in this folder also.

I tried NMKD superscale - but it doesn't seem as good for the realistic skin (but my preferences I guess). I'll give Ultrasharp a try!

 

[miaouxtoo]

How long do you take per image?

This might depend on the size of the upscaled image, and the highres upscale sampler used I think.

For me, upscaling a 512x768 by 2x (1572000 pixels squared total?) takes only a few minutes (5 mins?), but scaling it by 2.5x can take about 13mins.

 

[miaouxtoo]

There's an amazing site with comparisons of models here:

You must be registered to see the links

 

[modine2021]

There's an amazing site with comparisons of models here:

You must be registered to see the links

how do you copy the prompts there? its to much going on