Title says it all. I downloaded and installed stable-diffusion-webui-forge from GitHub and Pony Diffusion V6 XL from Civitai, but no matter what everything I create is some shit, worse than the AI slop I usually seem on boorus and such. I even went through Civitai's gallery tryng to replicate stuff, using the same LoRA, prompts and such, or the prompt from AI Character Generator from Perchance, but it isn't even close. Can someone shine some light on what the fuck I'm doing so wrong?
Not entirely related, but there is any way to create lolis on Perchance's AI Character Generator? There's any other good free online generators besides that?
I'd be extremely thankful if someone gave me tips. I swear I've already tried various different prompts or searching from tutorials, but I've never found anything...
Gonna tell what's what. Don't just jump in and start with the complex stuff.
You wanna/need to learn the basics first.
Here goes:
Different base models,, or base model libraries (< would be a better name if you ask me) need/have different base setting for them to work normal and not give you a picasso painting.
So if you get what i call human meatball... your CFG scale <<<<< That one makes the AI follow the prompts you put in to the letter and strict if you set it HIGH ,,, or it will just fuck around by itself if it's low.
Start out simple and set it low at 1,,, As for the prompts,, start simple as well,, just do a "Woman", outside, beach, ...... That's it. see what it does and go from there and see the improvement when you set the CFG higher.
Don't go skipping 5 steps with it ,,, The highest i have had it was 7 so far. I went from human meatball to something like : [Redacted] EDIT,, YES,, my example pic got nuked
Yes i can't show the rest of the pic Yes i made a whole series with them,,, yes they do different poses,,,,,,,and the prompt actually is (((2girls))), 25 years old, beach, yadda yadda yadda
(
please don't kill me mods, i just wanted/needed a example kinda wanted to show where i am now 2),,,
(for the mods are reading this,,,, i told you i was practicing)
Now if the colors are out of wack or the eyes are all fucked up,,, That is your SAMPLING STEPS. < that one tries to make the picture better,, clearer,, sharper,, on all fronts: Skin wise, all colors in the picture.
There is a point that you set the steps to,,, that it goes all fucked up and the colors go bad = 2 high,, set it to low,, and the pic stays fuzzy and out of focus.
Most base model libraries are in the 20 to infinity range ... SO set it at 20 if it's not already there as the base setting.
Next: SIZE of the pic,,,, Go for 960x960 max for testing,, cause the ai WILL do some weird shit if you go higher.
But 640 wide x 960 is a good setting.
If you go smaller,,, the AI will crop out limbs and heads if your prompts are not good enough,, So as a beginner do the middle ground on size.
Now here is the fun part For figuring shit out,,, Get a small Base model library to test with. Even a actual model base library that actually only has 1 model.
It will be smaller,, and options will be limited,,, But renders will be quick,,, so quick setting change and off you go for a new test.
I would get realModelBase01_v10CLCM to do that with, it's 1.98 GB instead of the 6-7 or nearly 8 GB
civitaiDOTCOM/models/498162/real-model-base01
You will need to make a account most likely,, but all you need is a e-mail,,, and that's it,,, Unless you wanna go and find things to buy from people
Ohhh and pay attention to the comments of the poster,,, SOME MAKE MODELS WITH text, watermark, copyright whatever you wanna call it
Now for the people that noticed the (((2girls))) where in brackets,,
Stable Diffusion ,, the one i use as i stated somewhere around here ,,, or another thread
OTHER TOPIC,,,,( has the option to make things more important.. so 2girls,, (2girls), ((2girls)) or (((2grils))) are different the last one being the most important.
Also in Stable Diffusion :::::: <lora: perfectEyesXL:1> perfect eyes , <<< NO space inbetween <lora: perfectEyesXL:1> the tread makes it into
on auto
Is the trigger to use a Lora,,, or extra option / bonus effect / improvement effect ,,,whatever you wanna call it,,, it is that. In this case Better/realistic looking eyes.
So,,,,, That would make the prompt something like this :
(((2girls))), 1white, 1african, wide angle lens, ultrarealistic 4k, realistic, sharp focus, hyper realistic, lifelike texture, intricate face, intricate eyes, intricate body, high contrast, braided hair, 1 with light green eyes, perky breasts, bubble butt, lying on stomach, beach front, smiling, perfect eyes, sunshine, 1sun, blue sky, some white clouds, beach bar, seashells, crab in shallow tidal waves,
<lora
erfectEyesXL:1>, <----------------- see it did the
on auto again : p
Yes you need to become a photographer,,,,,,,,
or learn to talk like some romance novel writer to set the scene instead of doing all the ,word, word, word, like i do.
Now civitai dot com has some issues website wise,, it can't keep up with it's users and what they post on it,, but your best bet is to go get your stuff from there.
Now that you have that and understand you get to the negative prompts :
no head, no face, no legs, no arms, extra limbs, extra fingers, extra arms, extra legs, extra nipples, cropped out limbs, pubic hair, piercings, tattoos, stretching midriff, midriff stretching, deformed midriff, low quality,
AND WHATEVER YOU COME ACROSS THAT YOU DON'T WANT TOO SEE AGAIN,,, put that in there.
As for the younger women question on perchance,,, NO,, they blocked that as did all online ai gens,,, all you get at most is fully clothed i think.
