[Stable Diffusion] Prompt Sharing and Learning Thread

devilkkw

Member
Mar 17, 2021
329
1,116
Experimenting and testing with my own created Lora. Based on SMZ-69's Character .
It's so dangerous it should be made illegal.. Worse then aged dynamite..:D

View attachment 2493610

You don't have permission to view the spoiler content. Log in or register now.
What beautiful result, nice lora and really good consistent character.

For Hires.fix i'm on 6gb rtx3060 laptop, i generate at 768x1024 and i'm able to use it at 1.1x without "--medvram --lowvram" in webui. if i add these my max is 1.4x but is time killing.
If i made a 512x768 my hires goes to 1.6x
My hires need about 30sec without "--medvram --lowvram", and abouit 3min with it.
Setting around 20-40 step at 0.4 denoise.

This is hires 1.6x at 20 step and 0.4 denoise hires-test-kkw.png

You don't have permission to view the spoiler content. Log in or register now.
I'm currently testing some upscaler with hires, but today update of a1111 made me work hard to fix ui error in some script.
 

fr34ky

Active Member
Oct 29, 2017
812
2,191
Hmm, that's wierd. :unsure: Maybe it's not only about Vram. I have 32GB of Ram and an i7 7700 CPU (4,2Ghz).
Have you tried to lower the sample resolution and then increase the multiplier? Do some testing with 512x768 and see.
Don't forget to use the cuda memory boost.
add "set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64" to webui-user.bat .
It's not only for people with lower GPU's.
At least I think it could give you a boost, since it has to do with memory and resource management.
I can't render in 1280x1920 either using that CUDA parameter, I have a rtx 2060 super.

I've tried deactivating --xformers only to do that and didn't work edither


Do you know what would be the difference between using that CUDA parameter and using other optimization like --medvram?
 
  • Like
Reactions: Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
Don't use it, but it makes your render more slow right?
I'm not actually convinced it makes it slower. Hire steps takes it sweet a$$ time regardless, and the memory fix only affects Hires fix as far as I can see. You can always try it out and if you don't like it just remove it as easy. Alternatively add it when you are planing on trying to push the Hires to the max and then remove it after.
 
  • Like
Reactions: Sepheyer

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
I can't render in 1280x1920 either using that CUDA parameter, I have a rtx 2060 super.

I've tried deactivating --xformers only to do that and didn't work edither


Do you know what would be the difference between using that CUDA parameter and using other optimization like --medvram?
The cuda parameter has to do with memory management and waste data it looks like. I'm not knowledgeable enough to understand the fine details of this. The optimization is it's own thing and I see no reason why you can't get benefit from using both. I have no idea what qualifies a GPU for medvram or low vram though.
 
Last edited:
  • Like
Reactions: fr34ky and Sepheyer

fr34ky

Active Member
Oct 29, 2017
812
2,191
The cuda parameter has to do with memory management and waste data it looks like. I'm not knowledgeable enough to understand the fine details of this. The optimization is it's own thing and I see no reason why you can't get benefit from using both. I have no idea what qualifies a GPU for medvram or lovram though.
I try to use the least parameters possible if I'm not sure what they are doing, in case stuff starts to fuck up I don't have too much customizations to blame.

I've been reading this to understand what every parameter does...



The only thing that I couldn't find is people saying if the images look better or worse, for what I've tried I can't note the difference xD

I could finish rendering in 1280x1920 with --medvram

At the moment I'm using --xformers --medvram and can render in 1280 x 1920 with a relatively good speed thanks to --xformers, so I can recommend that configuration.

Until today I was rendering at 768x1152 before the final upscale, so I almost doubled my final resolution thanks to be lurking in your conversation.

PS: don't watch youtube videos while rendering, it may give CUDA errors, as stupid as it may sound.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
Something to consider and keep in mind. You can do a hell of a lot for image quality by simply working on the prompt.
Add things like: highly detailed, sharp details, sharp focus, detailed skin, detailed face, detailed hand, detailed fingernail,
natural skin, skin texture, etc etc.. Anything that has to do with details and fidelity and texture.
Adding "sharp focus" and "depth of field " does miracles for how an image "pops". You can always weight these things.
If you don't want so much depth or bokeh effect etc add a value less than 1. Example: (bokeh:0.6)
You can also use words like "shallow depth of field", or "close focus" or "distant focus" or "subject in focus" perhaps.
The point is that the prompt is very powerful, never overlook it in the pursuit of more "anything" .
So before forking out a liver and kidney for a 4090 Ti. See first what the prompt can do after you max out your settings.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
I try to use the least parameters possible if I'm not sure what they are doing, in case stuff starts to fuck up I don't have too much customizations to blame.

I've been reading this to understand what every parameter does...



The only thing that I couldn't find is people saying if the images look better or worse, for what I've tried I can't note the difference xD

I could finish rendering in 1280x1920 with --medvram

At the moment I'm using --xformers --medvram and can render in 1280 x 1920 with a relatively good speed thanks to --xformers, so I can recommend that configuration.

Until today I was rendering at 768x1152 before the final upscale, so I almost doubled my final resolution thanks to be lurking in your conversation.

PS: don't watch youtube videos while rendering, it may give CUDA errors, as stupid as it may sound.
Of course anything that can take resources from the GPU needs to be shut off. You have a phone with wifi don't you? Use it to watch Sebastian Kamph or what ever while generating. :D
 
  • Haha
Reactions: Sepheyer and fr34ky

fr34ky

Active Member
Oct 29, 2017
812
2,191
Something to consider and keep in mind. You can do a hell of a lot for image quality by simply working on the prompt.
Add things like: highly detailed, sharp details, sharp focus, detailed skin, detailed face, detailed hand, detailed fingernail,
natural skin, skin texture, etc etc.. Anything that has to do with details and fidelity and texture.
Adding "sharp focus" and "depth of field " does miracles for how an image "pops". You can always weight these things.
If you don't want so much depth or bokeh effect etc add a value less than 1. Example: (bokeh:0.6)
You can also use words like "shallow depth of field", or "close focus" or "distant focus" or "subject in focus" perhaps.
The point is that the prompt is very powerful, never overlook it in the pursuit of more "anything" .
So before forking out a liver and kidney for a 4090 Ti. See first what the prompt can do after you max out your settings.

I totally agree, adding lots of deliberate details to the prompt is a gamechanger. And I super agree on that final one, FFS don't throw your money to compensate lack of knowledge, I see people doing that on every field.
 

daddyCzapo

Member
Mar 26, 2019
242
1,502
Hmm, that's wierd. :unsure: Maybe it's not only about Vram. I have 32GB of Ram and an i7 7700 CPU (4,2Ghz).
Have you tried to lower the sample resolution and then increase the multiplier? Do some testing with 512x768 and see.
Don't forget to use the cuda memory boost.
add "set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64" to webui-user.bat .
It's not only for people with lower GPU's.
At least I think it could give you a boost, since it has to do with memory and resource management.
Oh yeah, i used the set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64 when i was restarting the web ui some time ago, and the results are here. I can upscale by 2.5 but for 512x512 image it takes fucking 15 minutes XD I stopped further testing for today tho
 
  • Like
Reactions: Mr-Fox

Nano999

Member
Jun 4, 2022
173
74
How to install locon?


I went to extensions and installed it there


But when I was genereting thing cmd send me some missing files when using


Then I pasted the source code into lora folder:
stable-diffusion-webui\models\Lora\a1111-sd-webui-locon-main


And tried to generate something, but it was again not working
 
  • Sad
Reactions: Mr-Fox

miaouxtoo

Newbie
Mar 11, 2023
46
132
I can do about 1140x1900, from 600x1000 upscale by 1.9
I believe I've read that when you choose an aspect ratio, it should be a 0.5 multiplier increment of 512x512 on either side.

E.g.:

2x on one side would be 512 x 1024, or 1024 x 512
1.5 on one side would be: 512 x 768, or 768 x 512
1.5 on both sides might be 768 x 768

I think that's because of the way models are trained on 512x512 images (for v1.5, I think 768x768 on v2 Stable diffusion). Maybe it's fine to do 1.2x on both? That kind of thing?
You might get better speed or aspect if you change from 600 x 1000 maybe.


Max resolution on mine I've had with highres fix 20 steps,LDSR is I think 1920 x 1920
It's slow to generate though lol
 
  • Like
Reactions: Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
Oh yeah, i used the set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64 when i was restarting the web ui some time ago, and the results are here. I can upscale by 2.5 but for 512x512 image it takes fucking 15 minutes XD I stopped further testing for today tho
Yeah, I'm in the same situation. It takes 10 minutes plus to render each image at my current max. It sucks, but it just means I have to be more deliberate with my settings and prompt while using this addition, tag, flag what ever. I actually never remove it, I simply switch between normal and Hires fix. We don't need to use our max at all times, only when we have an image we really want the max result for. So today for example I'm running a plot script to compare some lora's, so I have set the multiplier low at 1.5. We still get the benefit from the Hires steps with low multiplier.
 
  • Like
Reactions: Sepheyer

devilkkw

Member
Mar 17, 2021
329
1,116
quite good result. i'm actually testing something about skin. your image look like as photo-retouched skin (my impression). is this done by lora you using?
 
  • Like
Reactions: Mr-Fox

daddyCzapo

Member
Mar 26, 2019
242
1,502
Guys, do you have some pro-tips on how to incorporate domino mask? Robin/green lantern style. Cause at this moment the best i got, without messing up the face too much was this:
You don't have permission to view the spoiler content. Log in or register now.
I tried to delete some/all of the negative prompts, but then the output was just awful. I'm using clarity checkpoint as it responds the best to my shitty prompts, without any loras
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
quite good result. i'm actually testing something about skin. your image look like as photo-retouched skin (my impression). is this done by lora you using?
No, these are a basic prompt without the lora. I was comparing different iterations of my Lora with plot script s/r and it was necessary to have images without any lora. Most of them came out very good so I decided to share them.
 
  • Like
Reactions: devilkkw

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,804
Guys, do you have some pro-tips on how to incorporate domino mask? Robin/green lantern style. Cause at this moment the best i got, without messing up the face too much was this:
You don't have permission to view the spoiler content. Log in or register now.
I tried to delete some/all of the negative prompts, but then the output was just awful. I'm using clarity checkpoint as it responds the best to my shitty prompts, without any loras
I suspect that you need either Lora or embedding for something this specific. I only found this .
Though it is in anime style it might work if you use it with a low value 0.2-0.4 .
 

fr34ky

Active Member
Oct 29, 2017
812
2,191
Guys, do you have some pro-tips on how to incorporate domino mask? Robin/green lantern style. Cause at this moment the best i got, without messing up the face too much was this:
You don't have permission to view the spoiler content. Log in or register now.
I tried to delete some/all of the negative prompts, but then the output was just awful. I'm using clarity checkpoint as it responds the best to my shitty prompts, without any loras
If you don't use a lora, the only chance you have is to generate the picture in txt2txt, go to photoshop and sketch some mask, or even copy paste a mask from some picture, and then in img2img add the ((mask)) to the prompt. If you are lucky, the AI will understand and it will render a perfect mask on your subject.

In this process you will be replacing the 'hi res' step with the img2img, so you have to do the img2img in the usual resolution you use in hi-res.