Hey, hey you..... yes YOU!
Intro
I know exactly why you are here, your friends started to tease you and called you a furry, only cause you said "Loona is hot and I would literally date her" but you forgot to say that you meant the human form. Well I think I digress a bit... long story short, you searched the web and did not find that specific Loona (ofc. in human form ;) ) pic you want. The AI boom caught you and now you are here, trying to create your fantasy pic of Loona in lingerie.
As most of you (and me) are using windows, be it for gaming and or just because you are used to it, I wanted to make a short handy guide for all of you.
AUTOMATIC1111
I will start with the UI, as you know there are some out there.
If you read A1111, Stable Diffusion WebUI or simply webUI it means all the same:
AUTOMATIC1111 - stable-diffusion-webui
https://github.com/AUTOMATIC1111/stable-diffusion-webui
I liked the UI a lot, its easy, fast to understand and if you are not planning in using SDXL or XL Turbo, its perfectly fine. But for me I got some major issues with it, I could not create a single image other than SD1.5 (RTX3080) so I started to google, and found comfy UI.
COMFY UI
I know exactly what you think right now... "WHAT THE DUCK IS THIS?!"
even tho it looks a bit messy and complicated,it is really not, well, until you look for the generate button ( its the "Queue Prompt" button ). This is basically ALL the steps that SD is doing to generate your lovely LoonyToony bikini pics.
For the usage... there are like 1000 other, better, guides out there, I just use and play with it as I go.
Back to the topic tho....
You got the interest you got the UI (in my opinion start with A1111 and try or switch whenever you want) now you are ready and you downloaded your first models. lets goooo!
Your first inputs are simple but powerfull, you look around the UI you see Lora, gfc, samplers [euler, dpm, dpm ++ 2m Karras] denoise, weights and you start to think... hmmm WHAT THE HELL DOES THIS MEAN?
I will not go into details (like I said there are like million of intelligent people, so dont take all my words for good) but for the beginning, just trust me and try some things out.
BIG SHOUTOUT FOR THIS GUIDE; HE IS ACTUALLY THE GOAT !
https://civitai.com/articles/983/insights-for-intermediates-how-to-craft-the-images-you-want-with-a1111
Prompts
Positive Prompt:
I dont give you examples, there are a lot... but some basic early tips.
DESCRIBE WHAT YOU SEE, and not what you want to see !
Create a picture in your mind (for someone who literally is unable to do this its hard) and then describe it.
What is the background? What pose does your character have? what colors do you see? how is the lighting? how is the camera angle? You see, its like painting, except I suck at painting, so I have it even harder some times.
Negative Prompt:
for the first run, leave it empty!
Some models are better than others, sometimes your negative prompts fuck up your results.
Then add the negatives, like an eraser. you dont like the umbrella, negative prompt: umbrella, you dont like her red hair? guess what, negative prompt: red hair,
Its almost the same as with the positive prompts, say what you dont want to see. Sonner or later you will have some default negatives for models, and you will use them like : young, child, kid, boy, girl, ugly, (for anime pics mostly): logo, text, artist, name, ... etc.
Normally I delete all negative prompts when I work with a new Lora or a new Model.
LORA
Did he say LORA again? Dude tell me finally what it is.
like the best guide (in my opinion) here on Civitai said:
if a picture are 1000 words, a Lora is 1000 prompts.
Loras add specifics to a model on top, you can imagine it like this:
You finally got your bb95FurryMix_v130 model but you are unable to create your desired output?
Well here comes the LORA of your choice, lets say "loona-lora-v3" you enter a positive prompt: (Loona:1.3), and BOOM, you get a loona pic. (dont worry, a short dive into lora creation will come later)
The first Loras you will use will most likely be some characters, cause reasons cough Helluva Boss cough. Play around with different Loras, different models, you will see, the character will not change, but the other aspects, do you want a anime Loona, or rather a reallife photorealistic Loona? The choice is yours, use models, play around with numbers, weights, prompts, add another lora, change the sampler etc.
Long story short, if you need clothing, styles, characters, scences, weapons etc... use loras.
UPSCALERS
So, you finally have your perfect pic, you open it up and... and... its 512x512.
WHY?! I WANT MY LOONY IN 4K !!!!!!11
dont worry, SD has you covered, UPSCALERS are the name of the game.
Basically another model which tries to upscale and even increase the overall quality.
Take this my friend and you will see that the creation time is literally nothing:
4x-AnimeSharp
4x-UltraSharp
4x_NMKD-Siax_200k
and yes your read it right, this step will take WAAAAAAAAAAY longer than the initial creation, so, open you...tube ;) and relax, depending on your Hardware, the batchsize, or the upscaled resolution it will take a lot of your free time. (if your GPU wasnt really working until now... it will be on 99% now)
Alrighty, slowly but surely your "homework" folder is filling up with... scientific papers of werwolfs...
SHARE them, I know I dont do it that much either, but creators of Loras, Models, Guides, will love critique and praises. Overall, the whole CivitAI community is super nice and friendly. Go give that good H-Model a nice rating!
Back to work tho, after (how I call it) shopping on CivitAI (downloading loras, models, etc.) you can try to use the same input with different models, or with a different lora.
But you finally said, Hey, I undestood Loras (he said knowingly that he still has no idea) I want to create my own lora, how do I do this, what do I need?
Well, sorry but the guide is in another castle:
https://ko-fi.com/post/EDGs-tutorials-P5P6KT5MT
https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake
But I can help you with different things, and the reason why I wanted to create this guide:
WSL
YES; WSL! Windows subsystem for linux, what is it, why do I like it, and how much does this cost?!
Its Linux, well, its a linux shell in windows, with all of your drives mounted already.
I like it cause for me its much easier to bulk download, rename, create files in for example ubuntu than in Powershell.
Also, its free real estate:
Open "Terminal" on your windows 11 machine.
type: wsl --install (or if you want a specific distro you can get a list whats available and download even more distros)
we stick with ubuntu but, feel free to use any available distro as you like.
congratz after the initial setup you are able to use it as you like.
Here you can see your mounts:
gallery-dl
for bulk downloading of pictures.
https://github.com/mikf/gallery-dl
renaming your downloaded pics from 1 to X.
"ls -v | cat -n | while read n f; do mv -n "$f" "$n.jpg"; done"
this is needed for the Lora creation.
I create all the text files at once with:
touch {1..X}.txt
with this method I can download a bulk of images form various sites and then rename and create .txt files accordingly.
or if you want to remove a certain part of each file in bulk:
" rename 's/1920x1024-//;' * " the part in the middle is what you want to delete in my case if you download a lot of pictures and all of them have the resolution in it, you can remove it easy and fast!
Loras loras loras, can you maybe stop talking about them for a second?
NO! Cause they are cool, and if there is nothing for your needs you will sooner or later create on for yourself! (As you might guessed it, Im a DotA Fan :D ) But I could not find some good Loras, so I created one myself. And you can too!
EDG Tutorial
Some tutorials and dont forget to buy EDG a coffe (or 5!, he deserves it)
https://ko-fi.com/post/EDGs-tutorials-P5P6KT5MT
KOHYA
so yeah, now you got all you need to create your sweet Loras, WRONG! you should also get the UI for it, I use (for now) Kohya:
https://github.com/bmaltais/kohya_ss
Play with it, and dont forget, you cant fail, you just make happy little monstrosities with too many hands,feets,eyes and literally nightmare fuel.... But hey, just delete it and do it all over again!
Words are hard, new words are harder:
Checkpoint: Is a Model, normally people will say model
Lora: basically a specific model that sits ontop of your Model
Sampler: I suggest this guide: https://stable-diffusion-art.com/samplers/
CFG: is the imagination of the AI -> 0 = basically not using your prompt 20=strictly using it
Upscaler: Scales (and improves) your image
Weights: written in (((furry tail))) or (furry tail:1.3) is how much this prompt will be dominant
Seed: You do not create "random" pics, but rather with a seed, copy it for reusing the image
embeddings: simply said, a word list
Prompt: Your input, be it positive or negative
Steps: as you might guess what this could mean... less is more in this case. stick with 25-50 steps
prompt position: the position of your prompts are important, basically the earlier the prompt the more weigth it gets.
Link List:
https://supagruen.github.io/StableDiffusion-CheatSheet/ <- found it on reddit LOVE IT!
https://blenderneko.github.io/ComfyUI-docs/ Comfy docs
https://huggingface.co/docs/datasets/index Hugging Face dataset docs