[Stable Diffusion] Prompt Sharing and Learning Thread

modine2021

Member
May 20, 2021
417
1,389
never mind. i see u have to join before you see the prompts. kind of annoying how its setup to copy. but oh well. they all cant be Civitai..but nice site ..@ miaouxtoo

(y)(y)(y)
 
Last edited:

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
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!
NMKD Superscale Has a ton of different options. Did you try them all? I use 8x_NMKD_Superscale_150000_G very often.
From the you can find many other NMKD upscalers such as NMKD Faces. I completely disagree NMKD Superscale is great for realism, it's the main reason I use it. Try it again on other subjects and with other checkpoints and other sampler.
NMKD works great with DPM++SDE Karras and DPM++2M Karras. Also make sure that it's not the prompt that is lacking. Upscalers can only affect the out data. If the in data is not sufficient it doesn't matter wich upscaler you are using. A very simple thing is to add "detailed skin" to the prompt. If you tick all the boxes, using clarity, adding "detailed skin", using NMKD Superscale, Using SDE Karras. I guarantee realistic skin.
 
Last edited:

daddyCzapo

Member
Mar 26, 2019
242
1,500
Something i generated fiddling with . In total i created about 45 images with this checkpoint in most of them i used garbage prompts so nothing more to share :/ 00416-3695676862.png
 
  • Like
Reactions: Mr-Fox

daddyCzapo

Member
Mar 26, 2019
242
1,500
I just did a batch generation using settings:
You don't have permission to view the spoiler content. Log in or register now.
With hires upscale by 2 times, and it took ~48 minutes. I must get myself into tweaking settings a bit to make the most of my gpu without running out of vram..
Oh, and here are some of the pics from the batch:
You don't have permission to view the spoiler content. Log in or register now.
You don't have permission to view the spoiler content. Log in or register now.
You don't have permission to view the spoiler content. Log in or register now.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
I just did a batch generation using settings:
You don't have permission to view the spoiler content. Log in or register now.
With hires upscale by 2 times, and it took ~48 minutes. I must get myself into tweaking settings a bit to make the most of my gpu without running out of vram..
Oh, and here are some of the pics from the batch:
You don't have permission to view the spoiler content. Log in or register now.
You don't have permission to view the spoiler content. Log in or register now.
You don't have permission to view the spoiler content. Log in or register now.
Higher resolution on it's own doesn't do much for quality. Hires steps is what makes the big difference for details and image quality.
If the hires steps is set to 0, it does the same amount as the sampling steps I believe. It's a good idea to set the hires steps to 20-30, or to 2x the sampling steps if you wish to keep the composition and with a low denoising strength like less than 0.3
 

daddyCzapo

Member
Mar 26, 2019
242
1,500
Higher resolution on it's own doesn't do much for quality. Hires steps is what makes the big difference for details and image quality.
If the hires steps is set to 0, it does the same amount as the sampling steps I believe. It's a good idea to set the hires steps to 20-30, or to 2x the sampling steps if you wish to keep the composition and with a low denoising strength like less than 0.3
Does the number of hires steps make a huge difference in vram usage? Right now i had 70 sampling steps and 70 hires steps, as it was set to 0. And i can't upscale the 512x512 image more than by 2, even 2.05 crashes.
 
  • Thinking Face
Reactions: Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Does the number of hires steps make a huge difference in vram usage? Right now i had 70 sampling steps and 70 hires steps, as it was set to 0. And i can't upscale the 512x512 image more than by 2, even 2.05 crashes.
I have not done any tests but probably. 70 steps is overkill for both sample and hires. You get more out of it with 20-30 sample steps and 40-60 hires steps imo. Try out 20 sample steps with 40 hires and see how high you can set the multiplier.
Then you can play a bit with denoising. Set it low if you wish to retain the composition 0.3
Also don't forget to use the cuda memory "fix". It gives a small boost.
Add "set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:64" to webui-user.bat
I have not done any test with different threshold settings, 64mb seems to work well for me, this is probably dependent on your Vram.

*I forgot to say that the image ration makes a big difference for vram. A wide image means more pixels and more vram, so go with either portrait (narrow with x height) or landscape (wide with x low height).
I use 512x768 and my multiplier can go to 2.5 but is more comfortable at 2.4 .
You can use this ratio (1.5) but stay under 512, meaning w<512. Example 384x576 or 400x600 or 424x636 etc.
 
Last edited:

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
I have 10GB. End up maybe 900x1500 images. Hires 30.
You didn't understand me correct. I meant how high can you set the Hires multiplier ("upscale by") and what will the end resolution be? If you use 512x768 like me and the maximum multiplier 2.5 the end resolution will be 1280x1920.
My question is: how much will your 10GB Vram give you ? what is your max settings?
I have an old pascal card, RTX 1070 8GB.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Oh, I can't though :cry:
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.