Guide Overall basic guide to using Stable Diffusion WebUI

  • Welcome to the Fakes / AI / Deepfakes category
    When posting AI generated content please include prompts used where possible.

Handyfan

Han of Fandies
May 4, 2022
254
5,074
1,302
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes
I'm afraid it's just the nature of the beast. Couldn't begin to tell you the technical reason for it, only that it's a known trait of large resolutions. I'm afraid the hi-res fix is your best option. It's generally better than upscaling after the fact.
 

dkdoc

Superfan
Mar 11, 2022
63
519
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

The software is made to create images at 512x512 but can create other image sizes and aspect ratios. It technically can make larger images, but there is a limit to how large it can go, that is why the software includes an "Upscale" option inside it.

Through a LOT of testing on my end I've learned that if you want 1:1 images your best bet is to try 768x768 as your largest image size. If you have a beefy GPU you can make a 1024x1024 image but it's going to have errors in it, or even one image duplicated and split up.

I like to make 4:3 aspect ratio images, and the largest that I've been able to reliably make without errors is 704x872 resolution.

Anything larger than 704x872 produces weird results on a constant basis. Over the course of 45 image results of "woman standing in front of a wall" it gave me separate images 31 times. It would either create 2 similar images with a split-screen effect, or just make multiple people. Typing "group, duplicate bodies, duplicate faces, multiple people" in the negative prompts did nothing to fix it

There are some model checkpoints out there that have been trained to make larger images, but even then they're not going to be what you're looking for if you're wanting HD resolution photos. I think the biggest I've seen is 768x768 checkpoints, and selecting any other resolution than that gives you errors all the time.

I've ran tests with the following resolutions, (each generating a batch of 45 images):

512x672: no errors
520x680: no errors
528x688: no errors
536x696: no errors
544x704: 3 errors, 42 good images
552x712: no errors
560x720: no errors
568x728: no errors
576x736: no errors
584x744: no errors
592x752: no errors
600x760: no errors
608x768: no errors
616x776: no errors
624x784: no errors
632x792:: no errors
640x800: no errors
648x808: no errors
656x816: 11 errors, 34 good images
664x824: no errors
672x832: no errors
680x840: 27 errors, 18 good images
688x848: 8 errors, 37 good images
704x864: 11 errors, 34 good images
704x872: 31 errors, 14 good images
712x872: 22 errors, 23 good images
720x880: every image generated higher than this features multiple people and/or split screen duplications every single time

So in short, it's how the software performs. Make the image small then upscale it. Take your desired image size and divide it by 2 (or 4 for big ones) and use those numbers for the SD resolution, then when it's done generating upscale it by 2 or 4.

**It's also important to note that the software has a clear preference for either Portrait or Square images. Trying to produce a landscape orientation image (no matter the resolution) will result in weird results about 50% of the time.
 

WHITEX

Lurker
Dec 18, 2022
2
0
46
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

Damn, boy, your attempts are crazy)
I was able to make good images at 768x1024, with a pass of 50-100 steps, I'm fine with that, the result is very good and detailed, but unfortunately most of the time there are defects.

But when I zoom in 512x768 upscale I get no detail and sharpness, unlike the generation of 768x1024, that's what I don't like unfortunately, well what can you do
 

dkdoc

Superfan
Mar 11, 2022
63
519
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

This is what my WebUI gives me any time I try a 1:1, 16:9, or 4:3 ratio image that is larger than 704x872
Please, Log in or Register to see links and images

The prompt for this image was, "Photo of a woman standing in front of a blank wall"


This is also what happens at larger resolutions:
Please, Log in or Register to see links and images

The prompt for this one was "Woman standing while posing in an art studio, even lighting, charcoal sketch, golden ratio" (this is a batch grid save not an individual generation, though some did give me a split screen effect)


If I scale the resolution down to 704x872 or smaller, it gives me exactly what I ask for. It'd probably do better if I had a newer GPU with more Vram (1080) than I currently have, but eh for what I use SD for it's more than enough. Then again I'm not looking to create hyper realistic 4k image resolution images with the software.
 

bbackagainn

Superfan
Mar 12, 2022
36
580
829
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to see links and images

i think you need a github acct now for this but it fixes mutant faces and upscales pretty well, it can only do so much for really bad faces tho, although you can run then re run that result, but its pretty good if something is just slightly off and you dont wanna inpaint. one click fix.
 
Last edited:
  • Like
Reactions: Handyfan

bbackagainn

Superfan
Mar 12, 2022
36
580
829
0fya082315al84db03fa9bf467e3.png
is it possible to run through a variety of camera angles all in one prompt using some combination of instruction?

like photo of a dog, close up, medium shot, wide angle, birds eye?


I tried just listing a bunch of different camera angles but it seems like every thing is just a medium shot always, i know the obvious thing is do a close up batch and then a mid range batch etc but im impatient and my rig isnt that powerful to rip through these.
 

duke_ellington

Bathwater Drinker
Mar 17, 2022
426
12,470
1,402
0fya082315al84db03fa9bf467e3.png
wish i could do trans models with urpm.ipynb . even with inpainting its super challenging/ impossible. where my futa lovers @
 

wmw418

Tier 3 Sub
Mar 12, 2022
8
383
603
0fya082315al84db03fa9bf467e3.png
How to solve the problem that the two eyes are not the same?
 

Beakwood

Bathwater Drinker
Mar 11, 2022
656
5,706
1,217
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes
Please, Log in or Register to view quotes
It's really hit or miss for me with them. Nothing comes out to my liking. Still going to give it some more playtime.
Please, Log in or Register to view quotes
Usually one of two ways, either A: Run it through InPaint, and try drawing method or B:Use similar method of A bit of PAINT entirety of the face and in the prompt. Put like a 'celebrity face' or train some random face of your liking and put that in the prompt. But 'celebrity face' and change maybe of the cfg in InPainting can also help. It's really about fine-tuning and using many of the tolls available to come with the right finish.

These what I got so far, using various models. But, as I said, can't get it to my liking. So many misses, very few HITS
Please, Log in or Register to see links and images
Please, Log in or Register to see links and images
Please, Log in or Register to see links and images
 

Bropx025

Superfan
Sep 2, 2022
25
530
734
0fya082315al84db03fa9bf467e3.png
I get a good pussy image when its building the image and then in the final one, it cuts it off nearly always. How do I stop that happening?
 

LittleJ

Fan
Oct 29, 2022
13
77
232
0fya082315al84db03fa9bf467e3.png
When using Dezgo does anyone know how to get the AI to show the entire body? For example, from head to toe. I want the AI to create some women in high heels.
 
  • Like
Reactions: Uamiugi and Xjmsxxx