[Stable Diffusion] Prompt Sharing and Learning Thread

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
Would anyone have suggestions about what settings will produce decent faces in img2img? I can't find the right settings, and the "restore faces=on".
You don't have permission to view the spoiler content. Log in or register now.
Grrr
00092-2202118978-photo blond wearing jeans and white teeshirt and brunet wearing jean shorts a...png
You don't have permission to view the spoiler content. Log in or register now.
 
Last edited:

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Would anyone have suggestions about what settings will produce decent faces in img2img? I can't find the right settings, and the "restore faces=on".
You don't have permission to view the spoiler content. Log in or register now.
Grrr
View attachment 2370792
You don't have permission to view the spoiler content. Log in or register now.
Did you watch any tutorial video? The denoising strength setting is very important. In img2img you focus on this setting rather that sample steps. Sett sample step high enough, above 50 and forget it and focus on cfg scale and denoising strength.
The resolution and ratio is important also. Fit the "red square" to the source image, set "crop and resize". The prompt needs of course describe the source image. Try different samplers, Euler a, Heun or DPM++DSE.

(img2img timestamp link)

If you keep having problems come back for more help.
 
Last edited:

fr34ky

Active Member
Oct 29, 2017
812
2,189
Would anyone have suggestions about what settings will produce decent faces in img2img? I can't find the right settings, and the "restore faces=on".
You don't have permission to view the spoiler content. Log in or register now.
Grrr
View attachment 2370792
You don't have permission to view the spoiler content. Log in or register now.
Practice with one girl at once to see if you can get good results, then try with that picture of two girls.
 
  • Like
Reactions: Sepheyer and Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Practice with one girl at once to see if you can get good results, then try with that picture of two girls.
Yes I agree, I think that having two subjects makes it more difficult. It's difficult enough with one subject. Also it's best to use a source image that is "simple", not having too much intricate details such as a control panel with buttons and text etc.
 
  • Like
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
fr34ky, Mr-Fox yes, gents, 1 person is no problem, finding the settings is no problem. I was interested if this is me, or if this is the current capability level of the tool - whether it can handle a "real" life scene. My message's context was if anyone found a way to deal with scene-like renderings and change them around. Cool, cool, it seems we may get there soonish, just not yet.
 
  • Like
Reactions: Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Here's an idea. You could use that image but put the red square on one girl until you get a good result. Then do the same with the other. Splice them together in photohop.
 
  • Like
Reactions: Sepheyer

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Not exactly wank material... This was a weird one. I played around with prompt and settings, it came down to trying different measurements, width and height. I have noticed strange behavior with this before. Well I gave it a go...

00115-2714297299.png 00117-1613186975.png 00128-1875397387.png
00130-2048905486.png 00132-191936010.png 00136-4091354713.png
00138-1031962338.png
 

HardcoreCuddler

Engaged Member
Aug 4, 2020
2,525
3,209
Would anyone have suggestions about what settings will produce decent faces in img2img? I can't find the right settings, and the "restore faces=on".
You don't have permission to view the spoiler content. Log in or register now.
Grrr
View attachment 2370792
You don't have permission to view the spoiler content. Log in or register now.
on top of what everyone else suggested I also noticed that negative prompts which counter the positive ones can also make the output better
as in:
say you have "butt" in your positive and nothing in negative and the butt in the image turned out a bit off
you can then put "(butt)" in positives and "butt" in negative and there's a chance it will improve things by quite a bit, for whatever reason
 
  • Like
Reactions: Mr-Fox and Sepheyer

HardcoreCuddler

Engaged Member
Aug 4, 2020
2,525
3,209
Guys
Before I go downloading other random models
Do different models need different specs to run at all? Afaik they are only the weights for the SD AI but I want to make sure before I let my pc download things for an hour
 
  • Like
Reactions: Sepheyer

Jimwalrus

Well-Known Member
Sep 15, 2021
1,042
3,984
Guys
Before I go downloading other random models
Do different models need different specs to run at all? Afaik they are only the weights for the SD AI but I want to make sure before I let my pc download things for an hour
I've discovered (through the allegedly impossible act of running SD with a 4GB model on my old 2GB GPU) that the model is loaded into system RAM. Provided you've enough of that you should be fine.
16GB should be enough, but 32GB is better and at least it can be expanded unlike vRAM.
 

HardcoreCuddler

Engaged Member
Aug 4, 2020
2,525
3,209
I've discovered (through the allegedly impossible act of running SD with a 4GB model on my old 2GB GPU) that the model is loaded into system RAM. Provided you've enough of that you should be fine.
16GB should be enough, but 32GB is better and at least it can be expanded unlike vRAM.
awesome. I have 16GB of normal RAM so that should be enough but I'll check as well, thx
 
  • Like
Reactions: Jimwalrus

Jimwalrus

Well-Known Member
Sep 15, 2021
1,042
3,984
I've been playing around with adding small amounts of LORA embeddings to tweak outfits one way or another rather than adding increasingly excessive weights to prompts "((((low cut))))" etc.
For instance, if your prompts are consistently generating slightly 'disappointing'* outfits:
00097-4132667286.png 00099-3315100988.png
and you're wanting to spice it up with a rise to the leg line and/or lowering of the neckline, maybe just make it more revealing, why not try seasoning with a pinch of a LORA trained on slingshots?
00103-4132667286.png 00101-3315100988.png

These are just 0.3 of a LORA that apparently requires a weighting of 2 or more to produce the full 'cheesewire' bikini contest special.

*"Disappointing" is relative of course!

(Prompts and embeddings etc are in the PNG info as usual)