[Stable Diffusion] Prompt Sharing and Learning Thread

[jarman Kell]

Could you give a link so we know what you are talking about?

You must be registered to see the links

Did you do a full restart or only a webui reload? You need to close the cmd window and restart webui-user.bat .

And yes I did a full restart closing cmd and Firefox. Then once back inside of stable I restarted it using the ui

 

[Mr-Fox]

You must be registered to see the links

And yes I did a full restart closing cmd and Firefox. Then once back inside of stable I restarted it using the ui

I would recommend going with a different upscaler from the

You must be registered to see the links

. I'm particularly fond of NMKD Superscale that you can find

You must be registered to see the links

, NMKD have many different upscalers with different purposes, I would also recommend getting NMKD Faces. NMKD is ESRGAN based.

 

[devilkkw]

In search of some extension, i discovered

You must be registered to see the links

UI mod extension for a1111.
Tested and work like a charm.
Do you use some different UI? what's your favorite?

 

[jarman Kell]

Spoiler: Prompts

You don't have permission to view the spoiler content. Log in or register now.

Spoiler: Nprompt

You don't have permission to view the spoiler content. Log in or register now.

I would recommend going with a different upscaler from the

You must be registered to see the links

. I'm particularly fond of NMKD Superscale that you can find

You must be registered to see the links

, NMKD have many different upscalers with different purposes, I would also recommend getting NMKD Faces. NMKD is ESRGAN based.

thanks i will have to give that upscaler a try.

on a off note. ive started messing with everything v4 and cant seem to replicate the results of this photo

even when i copy the image straight from the meta data mine comes out gray out like this. what could i be doing wrong besides not having the exact same upscaler, i wouldnt think that would matter. everything i have generated with v4 comes out looking grayed out like this. Thanks again Mr fox!

 

[jarman Kell]

Spoiler: Prompts

You don't have permission to view the spoiler content. Log in or register now.

Spoiler: Nprompt

You don't have permission to view the spoiler content. Log in or register now.

thanks i will have to give that upscaler a try.

on a off note. ive started messing with everything v4 and cant seem to replicate the results of this photo View attachment 2607141

even when i copy the image straight from the meta data mine comes out gray out like this. View attachment 2607142 what could i be doing wrong besides not having the exact same upscaler, i wouldnt think that would matter. everything i have generated with v4 comes out looking grayed out like this. Thanks again Mr fox!

hey everyone i figured it out. i had to manually select the .vae file for the everything v4. thanks again

 

[Mr-Fox]

Improved VAE:

You must be registered to see the links

 

[Jimwalrus]

hey everyone i figured it out. i had to manually select the .vae file for the everything v4. thanks again View attachment 2607232

I was about to suggest the issue was not having a VAE enabled.
Tip: I'd recommend to tick the box to always use a default VAE (84000 as recommended by Mr-Fox above seems the best general one) which it will then do unless the model comes with one already baked in

 

[Nano999]

Does anyone have this lora?

You must be registered to see the links

It's archived now, can anyone share?

 

[Elefy]

Does anyone have this lora?

You must be registered to see the links

Look in author folder first.
LoRA/Artstyles/CuteScrap

 

[Nano999]

Did something happened to SD with the last update?

Scripts > checkpoint name seems bugged now, it doesn't switch checkpoints like it should, instead it just re-generates one checkpoint several times.

#3 and #5 were only applied, 1,2 and 4 were never changed

From left to right: #5, #3, #3, #5, #5 (the order is also weird):

I tested it manually and the result is different each image

 

[devilkkw]

is it x/y script?

 

[Kalpod42]

Thanks for the responses to my last question!

A new thing is popping up. Well, two new things. The main difference is that I've added a VAE. I've started using

You must be registered to see the links

, because that's the one recommended for the models I'm using. The backgrounds have gotten all wacky along with anything even remotely detailed on the body, lots of artifact noise all over, and colors are all over the place. Is there a setting I'm missing? This is txt2img.



Wasn't doing this before, and I'm not sure what I've changed to cause it. Appreciate the help!

EDIT: It seems to have fixed itself. I pulled in the data from a previous image that didn't have these issues. Still not sure what was causing it though. Perhaps the seed? I also switched to vae Mr. Fox linked above.

 

[Nano999]

is it x/y script?

Yes, I recall that previously this field was looking in a different way, just a casual prompt field where you drop names of checkpoints, but right now they added a feature with this (x) thingy at the end. Maybe this new update screwed it up? Or because I have over 300 checkpoints in my folder

 

[devilkkw]

Yes, I recall that previously this field was looking in a different way, just a casual prompt field where you drop names of checkpoints, but right now they added a feature with this (x) thingy at the end. Maybe this new update screwed it up? Or because I have over 300 checkpoints in my folder

The problem of x/y script, isn't fixed, so as many user switched to

You must be registered to see the links

. it work's like a charm( it's old but still working with last update).
You can use it for test multiple checkpoint, i use it and don't needed x/y anymore.

 

[Mr-Fox]

The problem of x/y script, isn't fixed, so as many user switched to

You must be registered to see the links

. it work's like a charm( it's old but still working with last update).
You can use it for test multiple checkpoint, i use it and don't needed x/y anymore.

x/y script is the old one and x/y/z is newer improved and makes x/y script obsolete.

 

[PandaRepublic]

I've been looking into what is affecting some of my character LoRA's it turns out it was the denoising strength. I usually go 0.33 on the strength for most of my character LoRAs it works fine but for others it doesn't. I tried lowering the strength to get more accurate results but it lowers the quality of the image. Face restoration is off. Is there a way to maintain the accuracy still using Hires.fix?

Spoiler: prompt

You don't have permission to view the spoiler content. Log in or register now.

reference image no Hires.fix, same seed

 

[Mr-Fox]

I've been looking into what is affecting some of my character LoRA's it turns out it was the denoising strength. I usually go 0.33 on the strength for most of my character LoRAs it works fine but for others it doesn't. I tried lowering the strength to get more accurate results but it lowers the quality of the image. Face restoration is off. Is there a way to maintain the accuracy still using Hires.fix?

Spoiler: prompt

You don't have permission to view the spoiler content. Log in or register now.

reference image no Hires.fix, same seed
View attachment 2614317

View attachment 2614330

Yes.. Sample Steps 20-30, for hires steps multiply the sample steps by 2, meaning 40-60 hires steps. I use ap 0.3 denoising strength one or two tenth more or less sometimes. I use postprocessing GFPGAN and sometimes in combo with face restoration GFPGAN. Wich sampler and upscaler you use is very important. Btw, looking at your prompt it looks like you know these things already.. Only thing is to change the sampler, I use either DPM++ SDE Karras for photo realism or DPM++ 2m Karras for semi realism.
You never specified wich "effect" or problem you are talking about. I noticed a space "<lora: DaenerysV1:0.8> ", it probably doesn't matter that much but to be safe, whenever you are changing the prompt go through it and correct punctuations and spaces etc after. Also cfg scale is something to adjust in order to get the best desired result. cfg scale determines how much SD will follow the prompt and how much it will invent itself.

 

[Nano999]

The problem of x/y script, isn't fixed, so as many user switched to

You must be registered to see the links

. it work's like a charm( it's old but still working with last update).
You can use it for test multiple checkpoint, i use it and don't needed x/y anymore.

My bad, was using x/y/z script, didn't know about the other one.
I don't even have x/y script there

Might it be something wrong with the checkpoints?
I just tested the other 5 checkpoints and see no such issue

 

[Nano999]

My random seed (-1) is not working now, no matter what I do
It generates the image and then the second image +1 to the seed, no matter of the settings

I've pressed Unload SD checkpoint to free VRAM and renamed some checkpoints before doing it.

what's wrong with it?

 

[Jimwalrus]

My random seed (-1) is not working now, no matter what I do
It generates the image and then the second image +1 to the seed, no matter of the settings

I've pressed Unload SD checkpoint to free VRAM and renamed some checkpoints before doing it.

what's wrong with it?

View attachment 2615087 View attachment 2615088

I thought that was how batches have always worked? Certainly that's how it's always worked for me i.e. the first image is a random seed, every subsequent image in the same batch is +1 to the seed.