stable-diffusion / AI - PART TWO

Positive Prompt, Crashed Spaceship, Highly Detailed, 4K, Isaac Asimov. Negative Prompt None, Realistic Vision, Sample Steps 30, CFG 7, 768x512 upsacled x4 with 4x-UltraSharp, Restore faces Ticked, Euler A

Just showing what you can get with so little positive prompt information.

[Johnny Cage]

I have to say that finally getting controlnet posing to work consistently is a game changer for me.

e.g. this yoga kneeling type pose is one I would have never been able to properly create before.

Originally Posted by Johnny Cage

I have to say that finally getting controlnet posing to work consistently is a game changer for me.

e.g. this yoga kneeling type pose is one I would have never been able to properly create before.

Glad it's working

[maxhitman]

How about some "sexy Art" ?!
This A.I. gadget is just amazing at creating all sorts of art.

One particular type of art which I used to draw when I was back in
high school, was Pin-up nudes in the style of Boris Vallejo.
Years later a guy comes around doing almost the same type of art I was doing,
his names was Armando Huerta. I loved his style and his sexy Pin-ups.
So, I decided to try and teach this A.I. to create a "Art Style".
It was my first time doing one.
It came out just perfect. ANY image you want to see, just write the "prompt"
and use the words... " in the art style of armando huerta", and this AI will
create an artsy drawing of a cool Pin-up woman in that style of art.

I put it up for download at Civitai...
Enjoy it and have fun !

Armando Huerta = Pin-Up Artist Series (NSFW)
-
a Lora Art Style
--- This is Version 1 --> The Black & White pin-up art style

http://civitai.com/models/49202/armando-huerta-pin-up-artist-series-nsfw

Important Notice - This is a very NSFW type of art style.
You will only see dressed up women in 1 out of 6 images created.
So, most of the time, they will be half-naked, or completely naked.

-

-

-

-

-

Originally Posted by maxhitman

How about some "sexy Art" ?!
This A.I. gadget is just amazing at creating all sorts of art.

Very nice, I just trained a new Lora which was an utter ball ache to do but I got there by using a new method, I hope I get some pictures

Originally Posted by maxhitman

How about some "sexy Art" ?!
This A.I. gadget is just amazing at creating all sorts of art.

How are you creating your Loras or safetensors/checkpoints?

[maxhitman]

Originally Posted by Zammis

How are you creating your Loras or safetensors/checkpoints?

This is not as easy as most people think.

First you must collect some HQ/HD pictures.
They must be clear and in sharp focus. Do not use pixelated or blurry images.
At least 50 or 60 is my chosen number for a better AI training.
Some
people just use 20 or 25, but if you really want some extra face expressions
and body poses, then I would go for at least 50 to 60 images.
Then you must clean them up in Photoshop. Erase ANy and ALL letters on
the pictures. If there are extra people in the same image or alot of stuff
going on in the back of your subjct, use a grey colored paint and a brush and
slowly cover them up, or at least make it all very blurry , with just the
main subject in clear view.

Then you can either place each picture inside a 768x768 square
and save it as Jpeg or PNG, with a number (01, 02 , 03 ,04 , etc.)
OR
you can just leave the big size picture, but number it.
In the AI training process, you must click and check in the menu
where it says "enable buckets", this will allow the program to read the images
which are not in a normal size of 512x512 or 768x768.
So, it is possible to use larger images in the training... BUT, DO NOT use
super large images. Keep them in a modest size, for example, 2500x1500 size.


Then open up the LORA Blip captioning menu and add a text to each image.
The machine will do it for you,
but after that you must look at each written paper and see if it is written correctly.
Describe the image and what it shows.... EXCEPT what the main
subject is all about. Do not name the main subject, or what hair style
they have.
Just describe it as a "woman" or a "man", wearing a suit or a dress,
what type of dress, if they are holding a glass or something, and what
scenes are around the main subject. Like trees, in a park, in a beach, etc.
etc.

Next step, is to open up the Dreambooth-Lora menu and fill in the
necessary blank spaces as described in this video...

http://youtube.com/watch?v=9MT1n97ITaE&t=168s

in this video he shows you how to install the LORA program step-by-step 
very easily.
But if you already have it installed,
 just follow his training method and what selection he uses on the menu 
from the middle of this video.
It is very easy to do.
After making 2 LORA you will memorize what buttons and numbers to
write on the menu easily.

I usually select REALISTIC-VISION 1.3 checkpoint, or the main SD v1-5-pruned.ckpt checkpoint to train the A.I.
-- The old Realistc-Vision 1.3 already had a VAE included, so it is good to
use for training the AI., specially if its a person face.
-- The SD v1-5-pruned.ckpt is the original checkpoint which the 1.5 version
AI brain used. Its got all the necessary info inside it for a good training on
making a LORA. This is as easy as I can describe it, because I am no computer
engineer.

I usually select a batch of 2 images and a Epoch 1 to do the training.
This is usually enough and the machine does around 3000 to 4000 training
images on all of your pictures. Some people use more "Epochs", but so
far I have not seen any problems with my method.

Just do exactly as the guy does in his video and you will create a cool Lora.
Doing this part is easy.
My computer usually takes about 15 to 30 minutes to create it, depending
on how many images you have chosen to use.
I have a 15 year old computer.

So far,
I have tested and done LORA on a few women, one guy (Indiana Jones),
an alien-creature (Predator), a car, and an art style. Just to know how to
do these types of Loras. Because if you never done one, you must try and
do a few on diferent subjects , so you can learn which is the best way
to do them.

The most dificult of them all was making the Predator Lora. You have to
explain to the AI in your "Blip-text-files" what the hell that alien-creature is
and what its supposed to look like, WITHOUT actually using its name
or the word "Predator". Otherwise , the AI will think its a lion or a leopard
or some type of animal predator.
This A.I. has never seen an "Alien" Predator before and does not know
how to create such a creature. But after a few hours thinking about it, I figured out a way to do it and it worked pretty good.
Everyone is now happy making some cool images using this LORA
and training their personal AI to create it.

Originally Posted by maxhitman

This is not as easy as most people think.

First you must collect some HQ/HD pictures.
They must be clear and in sharp focus. Do not use pixelated or blurry images.
At least 50 or 60 is my chosen number for a better AI training.
Some

I am doing the exact same thing but with some iffy results and I just found why in your text, I watched a video months back that said use around 30 images, but you say 50-60 so I will increase the images for training

But I do use the LowVram Settings as explained here.

You Tube

You Tube

[maxhitman]

Originally Posted by Zammis

I am doing the exact same thing but with some iffy results and I just found why in your text, I watched a video months back that said use around 30 images, but you say 50-60 so I will increase the images for training

But I do use the LowVram Settings as explained here.

http://youtube.com/watch?v=70H03cv57-o

I decided to use double the images as presented in that video, so I can
have MORE facial expressions and body positions. I do not want to be limited
to just a few faces in my images later on when I use it for creating photos.
The AI will also train on more face/head positions and facial expressions to
recognize them better AND to create them wih even more details.
You decide which is better for you to do.

The LowVram Settings is good option for those people who may need
to use it and it also works very well.

It all depends how much your graphics card can handle.
I bought a new NVIDIA 3060 series a few months back, so I am
comfortable in using the regular settings, even for an old computer
such as mine.

---------------------------------------------------------------
Lets make a new LORA !

Right now I am doing a New LORA with colored images painted by
Armando Huerta.
I will post right now all the steps I am doing, otherwise I may skip some
parts if I write this for you later. LOL

OK pay attention and do exactly what I do.
- I hope you already have the Dreambooth gadget installed into your PC
in the "documents" section, as was explained in that video.

STEP 1 - Collect 50 images with very good details. If one is blurry or pixelated, throw it out and get another one. Make sure its a very good picture!
Look at each one in Photoshop , fix any color issues if possible and sharpen
them just a little (not too much) , and place each image into 768x768
square box.
Save them as Jpeg and name them 01 ,02 ,03 ,04 , etc etc.

STEP 2 -Once you have done all 50 images, go to your "documents" section of
your PC and go into that folder named "Lora_TrainingData"
my example... C:\Users\Maxxx\Documents\Lora_TrainingData

Inside this folder , I have another folder which I have re-named
arhuertacolor_v1 (this will be the name for my LORA)
- Name your training LORA with a name the AI does not understand. make
it a really dificult name to understand. In this case, I am creating a ART STYLE LORA
so it is not that important. BUT if you do a real person, do NOT name the person.
Example for my Harrison Ford Lora... I named it " 1nd1anasolov1 "
I used numbers into my name for traing the AI to never understand what
a Indiana solo is, or who Harrison Ford is.
Because if there are images of Indiana solo or Harrison Ford
in the CHECKPOINT you are using, the A.I. will mix the faces all together and you
will not have a proper face of a younger or older looking person.
So,
name your training LORA or Textual-inversions with weird and bizarre names !

Open it and you have 3 more folders...
image
log
model

Open "images" folder and create one called " 100_test ".

Place ALL 50 images into this folder for Blip-text-captioning and
later AI training..

STEP - 3
Go out of this "Lora_Training Data" folder and go into another folder
in the documents section named " Kohya"
open the next folder.... " kohya_ss "
and go down to a file which says "upgrade " (Windows Power shell)
Clik on it to open the Windows Power Shell to run that program.
A blue window will pop up and it will automaticly download ALL available
new information and data you need to upgrade and update your
Dreambooth program.
Do this EVERY TIME you use the Kohya !

Leave the window open and dont touch anything!
It will update itself and will then automaticly close the window.

STEP - 4
Once the window is closed...
go up the file list and select the one named " gui_user".
(Not the other Windows-Power-Shell" version but the normal file)

A new black pop-up window will open.
It will configure itself and then a few seconds later will open up
the Dreambooth Menu on your screen.

STEP - 5
In that menu go to the " Utilities " section... "captioning" and
choose "Blip captioning".
In that menu write the previous folder whre you have placed all your 50 images.
C:\Users\Maxxx\Documents\Lora_TrainingData\arhuertacolor_v1\image\100_test

in the area where it says "Prefix to add to BLIP caption"
write your LORA name... arhuertacolor_v1

Dont touch anything else!
Just click on CAPTION IMAGES.

Wait 1 to 3 minutes and keep looking at that little black window
which runs the program (the cmd.exe) to see the details of what its doing.
After a while it will say " ... captioning is done "
COOL!

Leave this Dreambooth program running and don´t shut it off.

Now go back to your PC "documents" section and open that folder
where you placed all your images.
They will ALL have a new text-file next to them with the same number
as the images.
Go look at each one and read them carefully.

Make changes to the words if you see it should describe the images better.
DO NOT NAME THE Faces or the main subject person in the pictures !

Make believe they are invisible to your eyes in teh pictures.
Name ONLY what is around them , or what they are wearing, like a hat, or a coat
or a dress.
DO NOT EVEN NAME the hair style they have ! This way, later on, when
you create a new image with that person, you can change their hair styles
or even make them look bald if you want.
DO NOT even say they are using lipstick or face make-up !
Make believe that person is invisible.

OK, Now I am going to look at MY blip-text-file pictures.
I will be right back.

Hey , I am doing this "LIVE on the Air "

Sorry for my writing/spelling mistakes, I am writing this really fast

[maxhitman]

OK, I am back,
I looked at each of the 50 images and re-wrote into them the proper words
to describe them as "a woman", and not as A PIN-UP as the A.I. labeled some of them.
I have also included the words... on a white background, in the color art style of armando huerta
Because I am training the AI to recognize the artworks as being in the art style of
a particular artist.

This is what one image caption has written on it...

arhuertacolor_v1 a woman is posing naked, showing her ass, on a white background, in the color art style of armando huerta

A very simple description of an artwork done by Armando Huerta. LOL
But take notice that these will become "trigger words", for when you write your
positive prompts.
If ANY of the above words are included into your prompt, it will trigger the A.I.
to create an image in that art style... as an example...
on a white background,
in the color art style of armando huerta,

With these 2 sentences, the A.I. wll create the image in this style and none other art style.

Now this step is done, lets move on...
Where was I ?
Oh yeah ... STEP 6

Drop-down these folders and go back to your already opened dreambooth program
MENU which has been running in the background...

Select
" Dreambooth Lora"
and in the Menu, select a Checkpoint to which to train your LORA.
In the section "pretrained model name or path" - click on the little white paper icon
and go to your main stable-diffusion program Webui folder where you usually keep your
working checkpoints.
I will choose the V-1-5-pruned.ckpt ...
C:/stable diffusion/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned.ckpt
... and don´t touch anything else on this menu.
This original SD-checkpoint has all types of good training on images and art,
so its cool to use.


Next step...
click on " Folders "
Write in here the names of the folders which you need to place the LORA training info.
They are the 3 folders inside the LORA_TrainingData folder.
image
Log
Model

C:\Users\Maxxx\Documents\Lora_TrainingData\arhuertacolor_v1\image
etc.

For " Model output name" I place the name of the LORA I named...
arhuertacolor_v1

Dont touch anything else !

Next step...
Open the next selection which says "Training parameters"

The only things I am going to write here on this Menu are...
Training batch size = 2
Epoch = 1
Caption Extension = .txt
Mixed Precision - I leave it at " fp16 "
Save precision - I leave it at " fp16 "
Seed = 1234
LR Scheduler = select " Constant"
LR Warmup = write " 0 " (number zero)
Network Rank = 128
Network Alpha = 128
Select the box where it says "Enable Buckets"
Max resolution = 512,512


Click the Menu below where it says " Advanced Configuration "
Go down to where it says "Clip Skip" and write the number = 2

Now go back to the first folder where it says "Source Model"
Make sure the Checkpoint name path is selected in the window.
The "model Quick Pick" is written "Custom"
The "Save Trained Model as" is written "Safetensors"
Nothing else needs to be selected or clicked on.

I am using version Stable-Diffusion 1.5 , so I do not need
to click on the check-box where it says "V2"

Everything is READY TO BE COOKED !

Click 2 times on the button "Train Model" .

And go look at the progress in the little black pop-window (cmd.exe).

Go grab yourself a beer and watch the A.I. do its training runs on the images
you selected.

On my little black window it says its going to train the images 5000 times in
2500 steps.
That sounds good to me. It ought to be enough to make some
proper good looking nude pin-up artworks in the art style by Armando Huerta.
It also show me how long it will take to do this... 20 minutes.

OK, I am going to drink my beer slowly.
After this is done, I am just going to go into my
folder "LORA_Training Data" and go inside the "models" folder where
I will see my new LORA file.
I copy that and place it inside my Webui folder of my Stable-Diffusion
where it says LORA.

I close ALL of these opened windows of the Dreambooth
and then start up my stable-diffusion as usual, which will be opened
in another screen on my PC.
Then I will test the LORA making ALOT of images for the next hour or two.
Testing all types of prompts and checkpoints to see what would work best for
the Ai to create a nice image for me.