stable-diffusion / AI - PART TWO

Nice one max, hey have you merged Realistiv V 1.3 with 2.0, gives some interesting results lol

[maxhitman]

Here is an interesting Textual-Inversion you can try out on your images.

Read this below before using it...
As we all know by now, that it will take a little while longer until Stable Diffusion or
any other A.I. program can really create better fingers, hands and toes on a photo realistic
Ai-created picture.
These A.I. programs still struggle a little making them properly.
Some people have found a good way to make images with better looking
fingers by using "Anime Style Art" Checkpoints, and other Art related Checkpoints.
However, using an "Anime" style checkpoint will often give your model (female/male)
an Asian look.
Their faces will look more Asian and not Western people.

Someone over at Civitai has created a great little Textual Inversion (T.I.) which you
can use with ANIME art-style checkpoints and your models will not
look Asian !

Test this with your A.I. program.

Style Asian Less

http://civitai.com/models/50755/style-asian-less

What version of Automatic1111 are you using?

I am on Version 1.3.0 as when ever I use git pull for an update it always seems to screw things up, even when I update a clean unused folder of SD.

I wish SD was ironically more stable.

I keep getting this error with Realsitic Vision 1.3, any ideas? as in why should this start happening now when it was fine yesterday, pain in the ass.

NansException: A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[maxhitman]

Originally Posted by Zammis

What version of Automatic1111 are you using?

I am on Version 1.3.0 as when ever I use git pull for an update it always seems to screw things up, even when I update a clean unused folder of SD.

I wish SD was ironically more stable.

I keep getting this error with Realsitic Vision 1.3, any ideas? as in why should this start happening now when it was fine yesterday, pain in the ass.

NansException: A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

I have Stable Diffusion version 1.5
python: 3.10.6
torch: 1.12.1+cu113
xformers: N/A
gradio: 3.23.0  •
with ControlNet v1.1.201



I do have in my cmd.exe the words "git Pull" command and it usually updates and upgrades everything.
It often does make some changes to the menu items/new settings
now available, but does not change anything else with my S.D. version
or how it works.

Have you removed any other files from their places?
In my models folder....
C:\stable diffusion\stable-diffusion-webui\models\Stable-diffusion
I still keep the checkpoint "model.ckpt"
and in the VAE folder ....
C:\stable diffusion\stable-diffusion-webui\models\VAE
I still keep my file, " vae-ft-mse-840000-ema-pruned.safetensors"
These are important to make your checkpoints work properly
(so I have heard) Without this special VAE some checkpoints will not work
correct.
If you are missing it, go here...
http://huggingface.co/stabilityai/s...inal/tree/main
Download the last version (not the Pickle version) - click on the download arrow button
to download it and place it into the S.D. folder...
C:\stable diffusion\stable-diffusion-webui\models\VAE

....................................................

I have never used any command-line for the NansException
Since I have no problems with my S.D. so far, I will not use such thing.

....................................................

I keep getting this error with Realistic Vision 1.3, any ideas?

What problems are you getting with Realistic Vision v.1.3 ?
Image distortion ?
Badly created images?
The Checkpoint seems to not be working properly ?

Are you using the Checkpoints special trigger words in your prompts?

Some times it could also be the way you write your prompts.
If its for Realistic Photographs you write the prompts in a particular way,
if its for art or Anime-style-art , you write the prompt in another way.
I will have to go into an in-depth tutorial to explain this better (later).

Although Realistic-Vision version 1.3 does work very good all by itself,
it is always good to include an extra set of 3 other checkpoint mergers to give your
A.I. more available information for creating the scenes.

Try this settings version I sometimes use....

Your MAIN Checkpoint is this -- Realistic Vision version.1.3
--- triggers = PHOTOREALISTIC

In Checkpoint Mergers include these checkpoints . . .

A - ProtogenX3.4
--- triggers = SAFETENSOR

B - Deliberate
--- triggers = DELIBERATE

C - Clarity
--- triggers = PHOTOREALISTIC

Where you see the words I wrote "triggers" you NEED to include these
words into your positive Prompt , so the A.I. can use that checkpoint !
Some people forget that some Checkpoints also use special trigger words
which will make the A.I. use that particular checkpoint.

For example, if you look into Civitai for Realistic Vision version 1.3,
on your right side of the screen in the information section, you will notice
that the creator has used 4 different words as this Checkpoints "trigger words".
They are...
ANALOG STYLE
MODELSHOOT STYLE
NSFW
NUDITY
You just need to use one of those words in your prompts , so your A.I. will get its needed data-info from that checkpoint.

- Some Checkpoints also need trigger words , and some do not use them.
.........................................................................................

Here is something interesting,

New news from OLIVIO (That cool YouTube A.I. Guru ) in which
he says your Automatic 1111 has new updates which no one noticed before.
He also suggests to delete the folder VENV folder !

Just watch this video to know what to do and what is new news

the new updates news start at 07:09 in the video

You Tube

You Tube

NansException: A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

is the error which is really odd with RV 1.3.... Gonna check some shit

EDIT

I have no idea why it has now decided to work unless using Ccleaner to clear the cache fixed something, hey max when you installed the VAE into the SD/Models/VAE folder did you go to settings in Automatic1111 and use the SD VAE that is in your VAE model folder, I didn't see you mention it so thought I had better just ask.

[maxhitman]

Very Cool and Useful Tips !

BEST Automatic 1111 Tricks from my Community !
You need to watch this!

http://youtube.com/watch?v=V4oEY4Y44Ws

-----------------------------------------------

Turn boring into WOW ! - Model Switch Trick A1111 + Vlad Diffusion

http://youtube.com/watch?v=kKfdwhQR_ZE

-----------------------------------------------

I reported about this a few weeks back and now Olivio Sarikas has 
made a video about it.

Detail Tweaker Lora - Massive Details even with slow GPUs

http://youtube.com/watch?v=DMFP8ZHb5-8

Link for download below... 
These 2 models are both GREAT, but USE only One at a time in your image creations!

model by Lykon - http://civitai.com/models/82098?modelVersionId=87153
model by CyberAIchemist - http://civitai.com/models/58390/det...aker-lora-lora

Great for ANY type of image or art

-----------------------------------------------

Step-By-Step Tutorial

Map Bashing - NEW Technique for PERFECT Composition - using ControlNET A1111

http://youtube.com/watch?v=z6Xwh9G24uw

-----------------------------------------------

Very useful ! Step-by-Step instructions
Add Stable-Diffusion A1111 in Photoshop ! - One Click Install Plug-in

http://youtube.com/watch?v=Y3KJli8ohKI

-----------------------------------------------

ULTIMATE Upscale for SLOW GPUs - Fast Workflow, High Quality, A1111

http://youtube.com/watch?v=3z4MKUqFEUk

-----------------------------------------------

Originally Posted by maxhitman

Detail Tweaker Lora - Massive Details even with slow GPUs

http://youtube.com/watch?v=DMFP8ZHb5-8

Now as per usual he has missed a step and it may confuse a few people if it is left out.

After selecting Prompt S/R at 3.55 you will need to put in the X-Values but you cant do this from a value of +2 to -2 as it wont work using low Vram

[Johnny Cage]

Decided to reuse some old prompts to try the more detail LORA you guys promoted recently. Interesting results. Mostly seems to add a lot of detail to the dresses.

[maxhitman]

This is a heads-up to for those who may not know.

I recently discovered this website

h++ps://sd4fun.cn/

which has some Chinese written words on screen, BUT it is an exact copy of EVERYTHING that is also posted at the website CIVITAI ,
where you usually download free models for Stable-Diffusion A.I.

A VERY Interesting thing about this "Copy of Civitai" is that you can sign up
for Free , and you can Download ALL the free models which were
previously DELETED from Civitai
because of their recent NEW RULES !
LOL

So, with that being said, if anyone was missing a certain model
Now is the time to download it once again.

If you do not know Chinese, do not worry, if you are familiar with Civitai
This website is half-Chinese and also english

[maxhitman]

Originally Posted by Johnny Cage

Decided to reuse some old prompts to try the more detail LORA you guys promoted recently. Interesting results. Mostly seems to add a lot of detail to the dresses.

BEAUTIFUL !
I like your women and I like the way you write your prompts !

If you see too much wrinkle clothes, you can always test using in your
prompts some type of word, example, " a smooth tight dress "
or " tight evening dress ".

Perhaps this could fix the problem (?)
I have to test a few more words.