#💬｜general-chat

1 messages · Page 30 of 1

warm junco
#

mixed the models with these percentages

green vessel
#

imagine

warm tiger
#

maybe? i have no idea. i'd need to ask her.

#

i'm honestly pretty happy i even got this thing running on my PC. using a 980SC on windows 7 and had to go add windows libraries and install like 5 other things to get it running.. it still throws up several errors when i start it but it runs andn generates pretty nice looking photos

warm junco
#

wow good that you got it working xD

#

but still Win7 is 14 year old outdatet os, if you need something specific from windows 7 you can setup a VirtualMachine ^^

warm tiger
warm junco
warm tiger
#

yup :/

#

was very unhappy with a lot of the security stuff of 8/8.1 then 10 >.>

#

and 11 seems like it's going to be even worse

warm junco
#

security or privacy?

warm tiger
#

i directly block anything connections from Microsoft

warm junco
#

security win10 is better than 7 or 8 😛

#

but privacy gets worse

#

there are some good tools on Win10/11 to block the most stuff

warm tiger
#

any of the pluton stuff as well

fervent thunder
#

Linux desktop is brilliant nowadays I just have a new PC and a few softwares I need to use

warm tiger
#

exactly why if /whem i swap from win7 (will all th stuff i've done to itXD) it'll be to linux

#

i run linux networked on my laptop and windows 7 not networked on PC so it works

fervent thunder
#

I reccomend MINT cinnamon if you don't care about gaming performance on native programs ( not sure why but their compositor always slower than KDE for me ) or otherwise I'd reccomend Kubuntu

#

if I was being super paranoid I'd reccomend qubesOS and Whonix ?D

warm junco
#

Manjaro KDE Plasma is my go to

fervent thunder
#

I mostly reccomend MINT and KDE for simplicity and easy transfer from windows

#

plus not being rolling releases they are both stable

warm junco
#

I Love how Valve/Steam has pushed Linux Gaming forward

fervent thunder
#

me too, massive respect

warm junco
#

Steamdeck is the best Console/PC ever 😄

fervent thunder
#

I have benchmarks proving on my old pc at least even native csgo was getting 100+ fps more on benchmarks consistantly over windows 11 both fresh installs

#

and it was the only way I could hold 144fps on Apex Legends playing it VIA PROTON

#

like, they done fucked up pretty bad if a translation layer is running better on some benches

warm junco
#

haha true xD

fervent thunder
#

but also to be fair; if people here reading this are getting into Python, Machine Learning etc etc Linux is just a way more friendly place to be for it

#

if you just want new games to run best out the box and do not give a shit about privacy then Windows is what I reccomend

warm junco
#

yea thats true

fervent thunder
#

took a lot of research, opsec / infosec / security people hate themselves and their jobs and i dont blame em

warm junco
#

good links thx

fervent thunder
#

you're welcome some of it might be a little broken but theres deffo a lot of good stuff in there

#

speaking of other ai stuff, I got chatGPT to help me code in csharp a program that create a salted RSA key that can encrypt and decrypt text ;'D

warm junco
#

woah awesome 😮

fervent thunder
#

I also managed to trick it into 95% telling me exactly how to hide a process by creating low level api calls from my own DLL

#

into system32

#

and samples of code to encrypt data into memory at runtime to avoid it being scanned

#

also: I don't write viruses

#

I pentest

warm junco
#

ChatGPT is so powerfull, perfect for scripting like you said

fierce musk
#

I must say the Open Assistant 20B (Foss alternative to chatgpt) model gave me bonkers answers. Really funny. I was gonna post the screen shot at #memes channel but apparently @wise stratus or whoever has nuked the entire channel all together.

upbeat tangle
#

hello anyone can help me for TypeError: 'NoneType' object is not subscriptable

fervent thunder
fervent thunder
upbeat tangle
#

thank you i've been dealing with this for 3 hours

fervent thunder
#

one thing I've had to do when using a certain VAE is change the line in webui-user.bat to read like so

#

set COMMANDLINE_ARGS=--no-half-vae

#

and another thing is if the filename of the safetensors file is too long it can cause errors

warm junco
fierce musk
#

Anybody else's tried Open Assistant out?

fervent thunder
#

if they can create one that has enough data to speak coherantly but specialises in writing code it would be insane.

upbeat tangle
fervent thunder
#

Give carter back his server

#

for me with programming, I can read it, I understand it but I struggle with the mathmatical background when it comes to things like drawing vectors , applied mathmatics in general so I just need a kind of push in the right direction, it would take me from a tinkerer to a fully fledged programmer in a few weeks

fervent thunder
warm tiger
fervent thunder
fierce musk
fervent thunder
#

most stable diffusion now will at least get you somewhat close ( after a few hours of playing around )

fierce musk
#

most of the Open source models trained with the Pile sux big time. Regardless of the size

fervent thunder
#

thats nuts, youre on about that like hundreds of terrabytes data yeah ?

#

even chatgpt is a bit over rated imo and thats fragmented across hundreds of machines just to even store it

#

its more of a "confidently bullshitting a lot" machine

#

they need to adjust the amount of padding it does when writing too, it writes like me when I have to do enough words in an essay

#

ps: not saying ChatGPT is bad , not at all it's an amazing look at what's to come

fierce musk
#

why would you need multiple machines to store chatgpt, unless you serve big saas

fervent thunder
#

haha!

#

its a couple hundred terrabytes isnt it ?

#

I can't remember how much or if openAI said how much for that model

fierce musk
#

the dataset? or the model?

fervent thunder
#

the dataset

fierce musk
#

yeah

#

unfortunately , open source community can't use all the datasets that are there

fervent thunder
#

nope but that is also a good thing in many ways, (classical lol )artists and all these doomsday "end of human art" people need to understand we arent here to replace them that human art will always have its own special place, that some people are disabled and physically unable to create ideas in their heads onto paper, that there are so many positives to this, that the community is happy to work within FAIR legal guidelines that dont affect well written and open liscenes for creative minds to express themsevles

#

Also if I were better at digital art or drawing with my hand I'd LOVE to use AI to assist me to prototype quickly

#

sometimes I wonder if it's the super rich who use art to launder hundreds of billions of £££ every year are the ones pushing the narrative

nova lily
#

I'm assuming generation time is affected by steps, prompt amounts primarily? Res I'd image and possibly CFG level

fervent thunder
#

also things like upscaling effect is massively

nova lily
#

oh

fierce musk
#

Eleuther or Laion community can't train ,not because of storage issues but because of the legal things. Yannick has already said he couldn't train OA with chatgpt output. I believe no good Foss model will ever come out unless some underground pirate group does that. The irony is you can't terrain your model with the output of another model trained with the scraped data where no cudos nor fair cuts were delivered. How redonculous is that.

fervent thunder
#

the cfg scale put as simply as i can 0 = full creative, max setting means only takes what you wrote in prompts into account

#

A good setting in general for my uses has been 10

upbeat tangle
#

ahh its dont work again

fervent thunder
upbeat tangle
#

4 gb gtx 1650

fervent thunder
#

theres one other command line argument that should help !

#

also make sure to prune any data taht you are using

warm junco
fervent thunder
#

if youre using automatic1111 you can prune very easy

upbeat tangle
#

how much Sampling step should i use

fervent thunder
#

Start with 10 in my opinion, try 8, try 12

nova lily
fervent thunder
#

@upbeat tangle add this to your command line args

#

--medvram

warm junco
fervent thunder
upbeat tangle
fervent thunder
#

in the launch option

#

in the bat file

upbeat tangle
#

i started new so i dont know much

#

okay

warm junco
#

Commandline_ARGSs=--xformers --medvram

fervent thunder
#

what about --no-half-vae @warm junco

#

it fixed it earlier for him for a while

warm junco
#

Ah then this too

fervent thunder
#

set COMMANDLINE_ARGS=--xformers --no-half-vae --medvram

#

i think that would be correct?

#

and thats inside the webui-user.bat file

glass lava
#

have anyone seen my favorite ketchup bottle i lost it around here somewhere, it's an heinz ketchup glass bottle from 1957 keep me posted if you find it ty very much

warm junco
upbeat tangle
glass lava
#

oh wait you have an off topic chan for shitposting, sorry

fervent thunder
#

I think for 1.5 SD based model you should use 512 x 512 or 512 x 768

#

then use upscaling to get higher res

#

otherwise that can also cause errors

upbeat tangle
#

im now using 512 512

#

When sampling steps are low, the visual is bad. Could the problem have something to do with it being too much?

fervent thunder
#

also I cant stress this enough BACK UP YOUR PROMPTS, YOUR FILES ETC ETC, sometimes this stuff breaks

fierce musk
fervent thunder
#

I had to reinstall once today because of an error trying to do outpainting

fervent thunder
#

i honestly think the way around it is to keep specific models trained from specific data because general data sets are almost impossible to truley vet

#

and the legal system is insane

warm junco
fervent thunder
#

yeah you don't want sampling too low , and also not un-necceary high depending on the "sampling method"

#

by default it will say Euler A

warm junco
#

Euler A, 30 Steps, CFG 8 is a good start

#

512x768

fervent thunder
#

oh yeah to send a pic just have to go in the general-with-images chat below

fervent thunder
upbeat tangle
#

im using protogen 2.2

#

i guess

warm junco
fervent thunder
#

the best thing I could probably say @upbeat tangle is dont expect anything to be a one tool job, you will have to make hundreds of images to find settings you like

#

which actually becomes part of the fun

#

now;' is your system at least outputting images without errors now?

upbeat tangle
#

The pictures I saw while they were forming are beautiful, but the finished picture is disappearing because of the mistake.

fervent thunder
#

could you take a print screen showing the preview image

#

and the final image

#

and put them into general-with-image

#

@warm junco knows more than me

#

but I'm sure he will take a look

#

(:

upbeat tangle
#

okay i will thanks for your help

fervent thunder
#

youre welcome mate, haha I'm sat up at night just helping in doscord, Im getting too addicted to this stuff

warm junco
warm tiger
#

hmm. any ideas on what to add to the negative prompts tho get rid of cat ears? it seems to be constantly adding cat ears even though i have animal ears, cat ears, dog ears, cow ears, bunny ears, on the negative and human ears on the positive

fervent thunder
warm junco
#

Which model and is cat mentioned in positive?

fervent thunder
#

might be too specific

warm tiger
fervent thunder
#

Anthropomorphic ?

warm tiger
warm junco
#

There you have it xD

fervent thunder
#

cat eyes will be doing the trigger for sure

warm tiger
#

dang

warm junco
#

Cat eyes should go at the end of your tags

#

That should work

fervent thunder
#

you could try to rephrase it eyes (cat like) or something and yeah like CS1o said

#

the order of tags is really important

warm tiger
#

attempting to make a pfp for a friendo and her character has cat iris eyes but no cat ears

warm tiger
#

i did all facial prompts and hair promps then body shape and size tnen clothing then background and stuff

warm junco
#

You can try slit pupils

fervent thunder
#

@warm junco bro we actually support tag team rn

#

based

warm junco
#

Haha yea!

fierce musk
fervent thunder
#

I've actually heard that Yuzu's codebase is taken a lot from Ryjinx in that way

#

which is now closed source? I might be wrong there

warm junco
#

Its the other way round also both are foss

fervent thunder
#

ahhh ty for the correction

fierce musk
#

That's the reason M$ support Linux and Adobe support Gimp

fervent thunder
#

I saw soemthing on github issues about it couldnt remember which

warm junco
#

Np ryjinx is on a good way, getting more fps in every game except smash

nova lily
#

is it possible to batch save all results, or a hotkey to save image? thank you

warm junco
#

In outputs folder

nova lily
#

No, I turned that setting of just incase I dont like the result

warm junco
#

You can right click an image and choose "save"

#

But the meta data wont be saved then

nova lily
#

all good

warm junco
#

Would leave the setting on and delete the bad images after your done

fierce musk
fervent thunder
fierce musk
#

how ironic it sounds considering I was initially post a screen shot of Open-Assistant dialog as a meme, but the #memes has been nuked all together

fervent thunder
fierce musk
#

why was that ironic or why memes was nuked or why did I want to post it there?

fervent thunder
upbeat tangle
#

When i generate face eye or mouth is seems bad how can i fix that

#

restore faces option is crashing all photo

fervent thunder
#

do i need text templates for an embed? never really made one

#

was gonna do Robert Pattinson as Batman if anyone has general advice for how many pics to use

midnight crane
#

Is it just me or is progress stagnated? I know that there is a lot of effort in each version, but maybe a 2.2? or 2.5? But yeah, maybe the next version released will be a gamechanger so I'm not really complaining, but I just expected something more of January.

fervent thunder
#

yeah it has stagnated with 2.0/1 imho but they've got a few new models cooking I guess

midnight crane
#

They have to be.

fervent thunder
#

theres a model called Deep Floyd that people seem to think will drop in early Feb but who knows of course

midnight crane
#

Yeah, I know of it, I guess we're a bit spoiled with all the progress that has been happening the last months 🙂

upbeat tangle
#

anyone can help me for generate eyes and mouth?

fierce musk
short tapir
#

How do I get my prints so it writes actual things?

#

Like I want to try and make one of a ln anime girl wearing a hat with writing on it.

waxen pasture
short tapir
warm tiger
#

someone mentiuoned a thing you could run a photo through specifically to try to fix the fingers the other day. cant recall what was said.

warm tiger
fervent thunder
short tapir
warm tiger
fervent thunder
#

if youre using 1111ui you just download them , put them into folder with the same name as your safetensors file

#

it helps with colouring and other certain details

#

sorry I can;'t honestly say more than that

#

but I reccomend both anythingv3 or v5 the vae they provide works with 1.5x

#

and also stability MSE-VAE

#

lot of these things are best if you just simply try it then compare the resulting images

warm tiger
warm tiger
#

today?

fervent thunder
#

sorry I might have said the wrong version :3

warm tiger
#

oh

fervent thunder
#

I've been awake for like 13hrs

warm tiger
#

wee

#

get yourself an energy patch

fervent thunder
#

;'D

warm tiger
#

no images in this chat, huh?

fervent thunder
#

yeah kinda annoying but it is what it is I assume there was too much spam for normal conversation

#

there is a chat below for general with images

warm tiger
#

there;'s one for images right below this one

#

yeeh

upbeat tangle
#

Is there an alternative to A1111 for stable diffusion?

#

i cant use restore faces

#

i need that

old marten
#

Does anyone here overclock their GPU? Could random errors have anything to do with overclocking? I only have a slight overclock of 100mhz processor and 250mhz memory

old marten
upbeat tangle
#

I get an error when I tick it

old marten
#

I think you need to download the model it should do it automatically though

light crystal
#

Sorry so no dreaming anymore?

vestal dew
upbeat tangle
old marten
upbeat tangle
#

and model is protogen 2.2

#

i cant use restore face

old marten
#

And yes, I use it every now and then. He can have very good results I think aitrepreneur has a video on it

vestal dew
#

extension meaning what?

old marten
#

Sorry, it's a script. Do you know what the bottom when you can select the script like X/Y plot. The one I use shows down there.

vestal dew
#

I keep hearing about it and pix2pix or whatever, and I'd like to try them

old marten
#

I'm not sure about this one, but from the previous it looks pretty similar

#

Pix2pix isnt what they hype it to be

#

You have to hit the lottery for something amazing to be generated with it

#

It feels like a program in alpha stage right now, which is OK since it's only a few days old

upbeat tangle
vestal dew
#

does the depthmap require xformers?

old marten
#

Not the one I use. It downloads a model called midas. I don't know where its stored but it's just like a plug-in.

vestal dew
#

Ok, I have to figure out how to use it, try it

upbeat tangle
#

guys is there any way to restore faces?

warm tiger
upbeat tangle
#

sorry for posting it again

kindred lynx
#

guys where you generate your ai image? Beta dream studio or where?

vestal dew
#

what do you mean restore faces?

short tapir
#

For prints. What's the difference between using ( ) and , ?

wide juniper
#

Is openoutpaint superior to what’s available in auto1111?

subtle pewter
cyan valve
#

how do you use the bot to create images?

vestal dew
#

you dont

brisk bloom
warm tiger
#

so what typically causes these 'burn' marks?

#

i'm running an image i like through img to img 'just resize' but using the same size and setting CFG scale to 5 and denoise to 0.2 to try to get different finger posistions and it's adding a burn mark in all of them to the characters face

verbal osprey
astral goblet
#

i always investigate in that order. i don't even overclock right now since these started before i tried any of that. i'm actually limiting my psu to get longer stable periods

#

i wish there was an easy software way to test psu voltages but i think you need a big contraption for it

wide juniper
#

I'm having trouble training a face through textual inversion. Would dreambooth do a better job when it comes to human faces? I have a good GPU

fervent thunder
#

the embeds seems like they can be powerful too

nova lily
# verbal osprey Use GFPGAN

I occasionally run into issues in images when I use it, codeformer more consistent for me. Any idea why? bad setup, bad installation, incorrect use?

Thanks for reading

ALSO, should I enable model VRAM to RAM setting?

fervent thunder
#

best model to use for making emojis/ simple cartoons? looking to make some custom discord reactions using img to img with some imgs I have

astral goblet
fervent thunder
#

Is there somewhere I can ask for a model? I can pay

warm junco
warm junco
short tapir
#

Had no idea

#

So the more ( ( ) ) I add the more powerful it is?

#

Do I still have to use , after ( )?

warm junco
#

Yes this or you can also make Word:1.2

#

(Word) = Word:1.1

#

Then you dont need (((((word))))) instead you can use just word:1.5

warm junco
#

You can try for example:

A girl with blonde hair, sitting at the park, its raining,

Or 1girl, blonde hair, sitting in park, rain,

#

Anime models prefer the shorter tags but you can try both

warm junco
# vestal dew what do you mean restore faces?

@upbeat tangle For Facerestore you need the Gfpgan1.4.pth file from github then copy it to the folder where the webui-user.bat is.
Then in the webui settings under Face Restore you activate Gfpgan and apply.
Now you can use the Face Restore Option.

#

I think Codeformer Files are already included in webui

lime bane
#

Whis a mod?

#

I need help

#

With stable diffusion

warm junco
fervent thunder
#

oh snap its february

warm ingot
#

Please how do i generate images on here

dapper patio
#

Question for those using DreamBooth: For training faces what has been your best settings? I'm getting inconsistent results. My settings have been 1e-6 training, 2-4k steps, with and without class images, with and without instance prompts. Wondering if you have found a config that has worked well for you

feral crystal
stark hornet
#

say, does anybody know something about 2.2 and when it will be out ?

#

and what will be different with it ?

#

maybe @wise stratus could say something to this ? wasn't it supposed to be out at the begining of this year ?

real edge
#

Where i can found bot?

#

To write promts

stark hornet
#

no bot here

real edge
#

Were

stark hornet
#

no bot

#

local install

fervent thunder
#

I want a custom one

upbeat tangle
#

What settings are suitable for making anime characters? Also, which model are you using?

warm junco
upbeat tangle
upbeat tangle
warm junco
wide juniper
upbeat tangle
#

i cant find 4.5 vae file

shadow thunder
#

So, I go to this dreambot channels but it says I dont have a right to use them, what should I do? I acquired a role

fervent thunder
#

I need a spiderman into the spiderverse style, one already exists but it is, it does not work in landscapes and only in heads which is not exact, how much would it cost me?

valid river
#

Is it a bad idea to generate images bigger than 512x512 in stable diffusion 1.5 models?

elder cradle
elder cradle
vestal dew
#

RTX 3090 has20GB Vram, 4070TI has 12GB, but is the 4070TI better for SD?

brisk bloom
#

I think the 4070ti is slower at SD than a 3080

vestal dew
#

really?

#

you meant 3090?

brisk bloom
#

a little while ago someone on reddit asked. I will try find the post

#

I compared the times to my 3080 and its slower

vestal dew
#

4070TI outperforms the 3090 with xformers, but underperforms it without xformers

brisk bloom
#

so maybe the guys posting on reddit didnt have xformers on.

#

I didnt read through but was the toms review using A1111?

vestal dew
#

"For Nvidia, we opted for Automatic 1111's webui version (opens in new tab); it performed best, had more options, and was easy to get running."

warm junco
vestal dew
#

the 4070TI is a lot cheaper

#

but the 3090 has more VRAM

brisk bloom
#

since I last saw the reddit post it looks like the guys updated torch and xformers

vestal dew
#

I think I'd go either 4070TI or 4090 , but the 4090 is too expensive, or 3090 if I can find one refurbished/used

brisk bloom
#

I was thinking of the 4090 also. Very pricey indeed

#

3090 uses a lot of power too

vestal dew
#

maybe i just have to settle for the 4070TI\

#

much better than what I have

brisk bloom
#

what about the 4080

#

that has 16gb

vestal dew
#

over 1000

#

4070TI around 800

brisk bloom
#

what do you have now?

vestal dew
#

GTX 970

brisk bloom
#

oh wow.. huge upgrade

#

I think the 4070ti will be good for you going from the 970

vestal dew
#

I think GTX 970 sends a call through time to cavemen, who make images in the dirt with their fingers

#

I had an odd error in SD earlier: "Error - no available resources. Grok kill Og with rock. You wait for replacement cave man. Him come soon."

brisk bloom
#

smoke signals

vestal dew
#

Im trying some text prompts in cave man talk

#

I just tried to make "Grok kill Og with rock". The SD showed this error: "I heard you, Dave, talking about a new video card. Why don't you ask your 3090 to make this image?"

#

I just posted the first result - very odd

#

general-with-images

#

With a 4070ti, I will be able to do upscale in generating for one thing = good

brisk bloom
#

decipher the image to wina 4070ti

west night
#

😮

vestal dew
#

there's no mention of man in the prompt. So it made an image without man

#

the second image was of a rock in a desert

#

I added caveman

#

"Caveman Grok kill Caveman Og with rock"

#

posted that in general-with-images

brisk bloom
#

missing a leather jacket 😆

vestal dew
#

I should try that with the depth model from an image with two people

#

next image has a caveman holding what looks like a knife

#

I guess Og wasn't dead. He invented the knife, and now he wants revenge

#

one more, then I have to get to work. I have Superbowl commercials to record

brisk bloom
#

hulk vampire laughing,

vestal dew
#

last image posted. Elder Zog

fervent thunder
#

?

#

How many days would it take to train on poor GTX 1060 6GB? Given that I want high quality results?

vestal dew
#

cavemen speak english, but without articles

vestal dew
fervent thunder
#

Ik know mate just being a dick 😄

fervent thunder
#

How much does the best gpu cloud for this cost?

fervent thunder
#

5 days per image?

#

minute

#

On gtx

#

yaa

#

Are you sure about that

vestal dew
#

You begin training with your 1060. Somewhere in space - Spock: "Captain, I am detecting a spacial anomaly opening off starboard. It is absorbing all time in the local area."

fervent thunder
#

Also don't forget I want custom image input

#

I wanna transfer me into painting

#

I had a 1660 GTX before this card and it took 30 seconds to 1 min 30

#

not days

vestal dew
#

Kirk: "What...is...happeningSpock?"

fervent thunder
vestal dew
fervent thunder
#

gpu are v. fast at doing lots of simple calculations all at once thats why they were used for bitcoin mining

fervent thunder
#

Ive done it

vestal dew
#

try it...see how long it takes

#

report back to us

fervent thunder
#

from google "They can render an image in less than half the steps needed for other samplers. I'm doing 20 steps which takes 20 seconds on my GTX 1060."

#

Yeah but why do people train on gpu clouds then and even then with like Nvidia A100 it takes forever

vestal dew
#

we await your word. If it is good, we shall surge forward to unleash hell.

fervent thunder
#

maybe they dont want to set up python enviroment on their home install or stress their gpus?

#

or dont have a gaming pc

#

GTX 1060 is rather old

#

euler a low steps is fast af

vestal dew
#

maybe on the A100 it's different training

fervent thunder
#

So how did it took so long to train for Corridor on their rendering beast

#

are you talking about training a model or using a model

#

I mean if I can use a model to still get pictures of me

#

But I suspect training a model since I want my pictures in the dataset

#

okay so wht you are talking about is more difficult and takes a lot of time, more time learning and overnight processing

vestal dew
#

wouldn't you train a TI or LORA for that?

fervent thunder
#

yea

#

exactly that youd make a lora

vestal dew
#

not a whole model

fervent thunder
#

Okay so how many days would that take on my poor gpu

vestal dew
#

then you could use it with any model. Add your tag, whatever name to the text prompt. BINGO

#

you-doom, do a LORA. It will take minutes per image, not days

fervent thunder
#

If it's too much I'll go get an account

fervent thunder
#

I honestly think the part that will take time is understanding and training it not the processing

#

Workflow:- Choose 5-10 images of a person- Crop/resize to 768x768 for SD 2.1 training- Following settings worked for me:train_batch_size=4, mixed_precision="fp16", use_8bit_adam, learning_rate=1e-4, lr_scheduler="constant", save_steps=200, max_train_steps=1000- for subjects already know to SD images*100 worked great, for subjects unknown to SD more steps or a higher LR are required- training on a 3090 takes ~20 min for 1k steps

Link to repo:https://github.com/cloneofsimo/lora

vestal dew
#

yes, you can use a lora to get a likeness

#

For example, there's a Onepiece Lora that gives you characters from the show

fervent thunder
#

there might be services offering to do it without all the technical side but Idk about that

vestal dew
#

comic...like Luffy, Zoro, etc

fervent thunder
#

What is the best amount of input images to get really good likeness?

#

I can make 50 if it will make it better

#

here is a free web demo you can try it

#

and above are the settings you can input

#

"training on a 3090 takes ~20 min for 1k steps"

vestal dew
#

why dont you look at civ ai whatever...read their LORAs, try some. They may tell how many images they used

fervent thunder
#

Also my goal is to depict me and my girlfriend in one picture, is that possible?

vestal dew
#

You'll likely have to make two separate images, then compost them

fervent thunder
#

right now stable diffusion is the absoloute purest form of "fuck around and find out"

vestal dew
#

in SD if you try for separate people, even img2img with an image of 2, you get two that look the same

#

You could do inpainting one part of the image at a time

fervent thunder
#

there is plenty of youtube tutorials, github code, online worspaces though

#

like even on the last link it says "It will take about 8 minutes to train for 1000 steps with a T4 GPU."

#

those are cloud computing gpus

#

idk why it took digital foundry days unless they meant in research and video editing time

#

aaaaand hugging face are letting you do that for free

#

What is T4

#

4 teraflops?

vestal dew
#

huggingface lets you train free?

fervent thunder
#

How many teraflops does GTX have? Does it have any at all?

vestal dew
#

when I try to copy to google collab, it says I have no cycles

fervent thunder
#

it even says to asign a t4 in the settings and that it will take 8 minutes

vestal dew
#

"Attention - This Space doesn't work in this shared UI. You can duplicate and use it with a paid private T4 GPU."

fervent thunder
#

I dont want to see myself in a virtual image cos I already exist

#

so I make booba

vestal dew
#

it says it won't work there...you have to copy it to somewhere with a T4

fervent thunder
#

Its a bit strrange they mention change the settings toa t4 then dont offer one 😒

#

does sound too good to be true though those things cost thousands

#

Bro T4 is a beast

#

My GPU is no match

vestal dew
#

it says you can transfer to HF Spaces

fervent thunder
vestal dew
#

does that cost money?

fervent thunder
#

is it possible to train models with 8gb vram?

#

yes

vestal dew
#

T4 small is 60 cents per hour

fervent thunder
#

also the t4 isnt even really a beast anymore

#

a second hand 2080ti is both faster and like $500

fervent thunder
#

3*

#

while it takes 10 minutes for other people

#

as I said a 2080ti is faster than a t4 and cheaper, but it's not specialised for cloud use at all

#

for instance; it wil shred through CUDA perfectly but gets 12fps on fortnite in 4k on low settings

#

vs 79fps on a 2080ti, which is a card 3 generations old

#

2 gens old *

#

it even has half the cuda core so like

#

I'd go find one of them second hand on ebay for $300

#

2080ti better than 3060 ti ?

#

nope

#

hmm

#

its faster but depends on your use I guess actually ?!

#

I have low heat low power setup so I can rag it all day

#

it's a beast but you won't be getting tensor cores etc etc etc

#

tl;dr if you want a fast gpu but dont have the money for one a second hand #2080ti is good for a desktop pc, if you want to rent a service at 1k steps the T4 is enough so dont get scammed

#

and if it took digital foundry a day to process then lol

#

i do have 3060 ti

#

but its slow for some reason

#

same 8gb version, thought it was a 12gb when i got it but 8gb is fine

#

pruned models all seem to sit around 4gb anyway

#

the 12 gb is the 3060 version iirc

#

its probably more suited for training models

#

6.98/it .. hang on let me load something up and see whasts up with that

#

its on 8.2 rn

#

when you are rendering diffusions , how long does it take you at 120 steps Heun

#

about 10 - 15 seconds iirc

#

not long its pretty fast when using txt2img and so on

#

i should stop the training, im not waiting 2 hours

#

so I cant tell if hugging face are wrong about 8 mins on the fairly slow compared to a 2080ti, t4, when the 2080ti is like 10-20% faster in some games than the 3060 or what

#

do you know anything about hypernetworks ? i made one but i dont know how to use it

#

here

#

obviously being able to use tensor cores for nvidia upscaling makes the 3060 really fast imo

young ivy
#

When auto tries to allocate an insane amount of memory (trying to use toolkit to import sd1.5 pt file into my model to fix broken clip) resulting in a failure and error like.
Traceback (most recent call last): File "G:\STABLE_DIFFUSION_VARIANTS\SD_Software\stable-diffusion-webui\venv\lib\site-packages\gradio\routes.py", line 337, in run_predict output = await app.get_blocks().process_api( File "G:\STABLE_DIFFUSION_VARIANTS\SD_Software\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 1015, in process_api result = await self.call_function( (bunch of errors too long to include and then:) RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 52428800 bytes.

Is it a borked venv?

warm junco
young ivy
#

8gb vram, but it's not vram that it complains about. It's regular ram

warm junco
#

how much ram? xD

young ivy
#

32gb

warm junco
#

what did you do?
(trying to use toolkit to import sd1.5 pt file into my model to fix broken clip)

young ivy
#

not enough vram to merge 2 of the same model in merge block

#

trying another method now

warm junco
#

ah okay thats a very specific problem

young ivy
#

gonna try to remove the venv and have it redownloaded

ruby scaffold
#

Guys, who can tell? On which version is it better to train the model on pruned or regular?

young ivy
#

EMA if you're gonna train

cold siren
#

got a strange problem here, under certain circumstances my generated images look like they are sort of badly upscaled 😕

#

now seeing the same with inpainting

fervent thunder
#

can you train on ema only to save on vram ?

shadow birch
#

is the bots to generate the images down? when do they come back?

warm junco
#

From official SD 1.5 HF Site:

v1-5-pruned-emaonly.ckpt - 4.27GB, ema-only weight. uses less VRAM - suitable for inference
v1-5-pruned.ckpt - 7.7GB, ema+non-ema weights. uses more VRAM - suitable for fine-tuning
fervent thunder
#

i think maybe anythingv3 or v5?

warm junco
#

there is no v5

#

i can check my models for large pruned

fervent thunder
#

that would be cool thanks, do you know say if I take waht I have now and merge on 0.3 will it hugely change the style or just give it more points of data?

cold siren
#

have to find out yourself 🤷 nothing stops you from merging several version to see which one works best

fervent thunder
cold siren
#

i suspect results will be super different depending on the model you merge and/or on how different it is from teh bas model 🤔

#

i had some nice results with just 0.1 of some realism mixed into anything v3

warm junco
#

Yea it depends on the model. If its like a model ArcaneStyle or Anime you would know that this has influence but stuff like dreamlike you maybe wouldnt notice baceuase its has less specific data

fervent thunder
#

hmm i'll try adding in 0.1 or 0.2 Protogen x3.4 to my current mix!

#

something I keep wondering, if I mix and add in a new vae will that replace the vae that was embedded or overwrite the old?

warm junco
#

Proto will give you more realistic images

rotund hemlock
#

How to use dreambot channels

fervent thunder
#

thats alright, basically I'm going for anime but highly detailed for futuristic / scifi, right now I get some good results but it's not varied enough

warm junco
cold siren
fervent thunder
#

I don't WANT to but for whatever reason Automatic1111 is refusing to load VAE even with the same name and in the correct folder so I'm having to bake them in rn

warm junco
#

no wait

#

go into the webui settings

fervent thunder
#

I tried that but when I clicked apply setting it broke my whole install of automatic1111

#

and I had to redownload it

warm junco
#

under User Interface, Quicksettings add , sd_vae,

cold siren
warm junco
#

yes that shouldnt be the case

fervent thunder
#

yeah its super annoying but like, it is what it is, im on a fresh fully updated windows 11 install with fresh up to date version of automatic

cold siren
#

damn web apps, always find a way to break in the most ridiculous way

fervent thunder
#

Lol true

warm junco
#

did you tried it again with apply?

#

does it still break?

fervent thunder
#

I am scared to try it now :L

warm junco
#

backup your folder without models then try, but it shouldnt crash or destroy anything xD

fervent thunder
#

I'll give it a go

#

okay it seems fine now strange

#

maybe beacuse I've now used the --no-half-vae

warm junco
#

My biggest pruned Model after SD1.5 and 2.1 is AnythingV4 and 4.5

fervent thunder
#

which do you reccoemnd from anything 4 or 4.5

warm junco
#

4.5

#

even the creator likes it more xD maybe becaue he merged orange mix in

fervent thunder
#

so I'm thinking, use v3 VAE, moistmixv2, dream, proto and anythingv4.5

warm junco
#

i tought you now dont inbake the vae

fervent thunder
#

yeah! I'll redo the mix from the beginning and if i can just apply the vae from settings I'll just do that!

warm junco
#

okay

fervent thunder
#

just had a smart thought, time to prune moistmix before trying to merge it so its not 7.something gb

#

i feel dumb now 😄

warm junco
#

where is the tut for pruning, someone posted a link a few mins ago

#

ah nvm found it

fervent thunder
#

no idea I'm using a plugin to do it

warm junco
#

yea its an extension

fervent thunder
#

this is where my head explodes

#

dreamshaper mixed with grapefruit 0.3 then mixed with Moistmixv2 0.3, mixed with protogen 0.1 mixed with anything4.5 0.2, vae from anythingv03
@warm junco this is what I'm trying to do

warm junco
#

there is moistmix v2?

cold siren
warm junco
fervent thunder
warm junco
slim quarry
#

does anyone know why my deforum video only stays true to the prompt on the first frames and then just disintegrates into neon chaos?

fervent thunder
#

yep!

fervent thunder
hard osprey
#

I looked on civitai and some models have VAE files, what do those do?

warm junco
#

its fixing the pale output of anime models without vae

hard osprey
#

Ah, so generally speaking it's better to use models that include VAE?

fervent thunder
#

you can use any VAE with any model if you put it in settings

#

some have VAE baked in but it limits what you can mess around with

hard osprey
#

Okay because I assume the more things a model has the better it must be

warm junco
fervent thunder
#

oh unless the vae is meant for 2x and you have 1x SD

hard osprey
fervent thunder
#

nah just means you can load another one beacuse it gets spooked

#

generally just seems to cause issue when they are included built in

#

did you select ema-only ?

#

thats the part you don't want!

#

that is just the training data

#

I think hang on merging so cant check ;'D

warm junco
#

in 3 minutes i can give you all an comparison in vae + non vae

fervent thunder
#

you don't want ema or a sd model that is ema only

#

when pruning I always select no-ema

#

no problem!

#

the reason they provide the ema-only 1.5 is for people to use when training , it needs less vram to train if you arent loading data used only for outputting prompts

vague coral
#

hi everybody. where is the dreambot located here?

fathom kettle
#

Hi all, I've made a web application for creating custom T-shirts with Stable Diffusion.
Check out pearstop.com/app and tell me what you think!
Right now I added some auto-styling using ChatGPT because I thought some people might find it difficult to get all the promptcrafting right.
Next up: all over printing. I saw some models for creating tileable outputs, could be very cool to place this on physical products - do you think it's possible to train models with any kind of custom topology?
Anyways, all comments welcome, or if you wanna help out in any way send me a DM!
Love, Seb 🙂

fervent thunder
#

basically just close any programs besides SD, close the tabs open, if its fully stuck restart automatic1111 youre probably pushing how much ram/vram is being used

#

laptop HDD ?

#

No idea, I'm keeping all my stuff on an old slow hdd beacuse of how large it all is lol

#

I think there might be multiple probably unrelated issues here

#

check your cpu and gpu temps?

#

chec process monitor

#

and try running latmon to see whats causing the lag too

#

sorry mate I have no idea what bert even is 😄

vivid portal
#

Anyone who knows how to extract lora’s out of ckpt’s?

fervent thunder
vivid portal
#

Nice to be recognised. No one remembers my name most of the time 🤣

#

Too Swedish

fervent thunder
#

haha I'm a big fan of embeddings and only see a few people doing them! keep up the good work, sorry I don't know about extracing the lora from ckpt ):

#

@warm junco holy gods the mix is made, pruned and working

warm junco
fervent thunder
#

5 models into one

old marten
fervent thunder
#

gpu fans go brrrr

old marten
#

Easier even if you know that your graphics card has a max voltage of 450 W or so, and that applies to spikes as well. Honestly, I think unless people have a really shit PSU errors like these are pretty rare.

#

Although I did once get my computer to shut off during a heavy load, but that just happened like once out of hundreds of trainings

maiden crystal
#

what is some good prompts to get better quality images?

warm junco
old marten
#

That's what she said

fervent thunder
#

bro look in the general-with-images I swear by Heun as the best sampler

fervent thunder
# warm junco how big is it?

the crazy thing is the way ive pruned it at least it's the same size if not slightly smaller than anything v4.5 !!!!!!!

#

4.156 GB

warm junco
#

wow crazy

#

how big was the pruned moistmix?

fervent thunder
#

4,4gb

#

I pruned them at each step of the mix i think thats why

#

all 5 things I mixed together I'd pre-pruned

#

( if they werent pruned already )

#

that is probably how it should be, I think either the pruning methods used previously or the mixing done didint fully prune some of it, and since it's mathmatical I thought that ( unless random vae mixed in etc) they will all be the same size give or take a little

warm junco
crystal remnant
#

how can i start ??

#

im new user and i wanna try it !!

#

any one can help

#

??

warm junco
#

you want to try it local or online?

crystal remnant
#

here in discord can i ?

warm junco
#

not on this discord since the bots are offline

crystal remnant
#

it's free on local?

#

and how can get it

warm junco
fervent thunder
#

this discord needs a pinned post agony

warm tiger
#

I seem to have downloaded a few extentions i thought useful off the web UI and now the thing is ignoring all of my prompts and just generating a "picture of a donkey" nomatter what i put in as settings. ignored promts, resolution, batch sizes, everything. any idea what extention that is/how to remove it?

warm junco
astral goblet
warm tiger
warm junco
astral goblet
#

randomly installing an extension to replace all your prompts with a donkey doesn't sound useful, so i'm guessing you installed stuff from a weird source and it was a "virus" script somebody made to troll ai users like yourself

fervent thunder
astral goblet
#

maybe you're just not describing the problem well and i'm a runaway train on this idea

warm tiger
#

ah found it. it was diffusion defender. since i literally had 0 prompts set in it as default it triggered no matter what i entered into the promts XD

#

you'd think with 0 trigger prompts it wouldn't trigger at all

astral goblet
warm tiger
astral goblet
#

I wouldn't be surprised if DD had code submitted that was janky on purpose. The anti censorship people in the community do some pretty mental things.

fervent thunder
astral goblet
# fervent thunder I mean I guess :3

community resources exist for a1111 and this is the official server. i don't think pinning guides for his webui on here would look really great for stability, considering his inner circle are mostly supremacists

warm tiger
astral goblet
warm tiger
#

oh no

#

i thought automatic1111 was the default one? XD

astral goblet
#

nope. it's what everyone recommends because he's done a lot to make it compatible for a lot of people. it's been great work on his part, but it's tainted because of legacy

warm junco
#

its the one with the most features

warm tiger
#

ok i was going to say that's the one everyone of my friends recommended me

astral goblet
#

until recently, invokeai didn't even offer 2.x support

warm tiger
#

well 'meh' I won't let personal ideology stop me from enjoying a good product

#

especially if i didn't need to buy it

astral goblet
#

most people don't talk about automatic's race supremacy past. it's swept under the rug a lot. but i get it. this is ground breaking tech and we all want to play with it without getting left behind

#

its good to be aware of why stability the company might not want to help his efforts along so much though

warm tiger
#

considering their moves to go more PC with everything it'd make sense too

warm junco
#

maybe they just want to promote Dreamstudio more than local installs

astral goblet
#

promoting their own product makes a lot of sense too, but also not promoting products from rabid race supremacists is likely a big part of it

deft bronze
#

I like Automatic1111's work :3 though maybe not all his mods ;p

warm tiger
warm junco
astral goblet
#

before he worked on the UI, most of his public development work was towards rimworld mods that were very anti black lives matter

#

mostly just rimworld mods that were super racist, and also being a total ass to people on servers when he finds out they're not white

warm tiger
#

example. there was a video of a jewish dude being attacked by 5 people ( trying to kill him) he shot one of them and the rest ran away. I defended his right to defend his own life and i LITERALLY had people calling me a "nazi" for defending a jewish person.

#

that's how absurd a lot of people are these days

warm tiger
astral goblet
#

he spent a lot of work going through his old artists.csv file before removing it from the repository, making sure any artist name that generated black people as an image would be labelled with N. there's only two categories that he made for that list. N and C. c shows up on artists that do children work.

warm tiger
#

was it stallaris?

#

XD

fervent thunder
astral goblet
#

i focus more on individuals than what a small minority on reddit say

#

"people are so absurd" usually boils down to two people on the internet

fervent thunder
#

I've done an image in 512 x 768 and want to expand it out into widescreen

warm tiger
warm tiger
fervent thunder
astral goblet
#

these days public opinion is just corporate opinion. they own all the media. i'd rather just use my own brain as a keel instead of public opinion

fervent thunder
#

I just have no idea what settings to use :3

warm junco
astral goblet
#

i want to start using invokeai. they've made a lot of feature progress over thhe past month

#

their outpainting infinite canvas is way more supreme

fervent thunder
#

I liked the look of invoke but its just too annoying for using custom models (mixes) for me

#

the thing is my model is a mix of 5 others and I need to outpaint with it for consistency right ?

warm junco
fervent thunder
#

"feature in the img2img tab at the bottom, under Script -> Poor man's outpainting" ooo

#

no, It is called outpainting mk2 now I think that git is outdate

warm tiger
#

could someone give me a simple explaination of what the various default sampling methods are/do? or link to a wiki on them many?

#

am attempting to run through a few prompts to see the results comparable to ones i have already done. on PLMS atm but it seems it need a lot more VRAM than the EULER a one and crashes with the same settings i use for that one

fervent thunder
warm tiger
neat condor
#

guys after a fatal HDD crash i've restored around 1.5mil images, there's no way i'm going trough them one by one, can you reccommend any good duplicate image finder?! (paid or free doesn't matter)

#

problem is after restore hashes and file sizes are different so i need something that compares visual similarity with good quality (so it doesn't delete very similar but not 100% duplicates)

warm junco
neat condor
warm tiger
#

ah there you go

#

i use a program like duplifnder

#

so seems PLMS gives some more 'depth' color and detail to photos compsared to the eiler a but it needs more steps

fervent thunder
fervent thunder
neat condor
#

@warm juncothis czkawka is VIP , thnx again, i was looking for this exactly.

warm junco
thorny canyon
#

evening guys...I need to ask you all about can you make an images fom text using stable diffusion discord? what command do i must use..and is there a tutorial fot it, thanx..

fervent thunder
#

lol

#

the discord bot to do that is not online

thorny canyon
#

haa thank you @fervent thunder

fervent thunder
#

all good m8

wispy jasper
#

Any news updates on promised (new models/tools) I remember @wise stratus

#

Saying we were gong to have almost weekly released, but haven't seen anything in months?

#

🙏

proper karma
#

I literally had this message typed out lmao:
anything interesting to look forward to from stability in the near future in terms of image gen, or is everything frozen atm with no expected future models being released for a few months?

wraith sky
# wispy jasper Saying we were gong to have almost weekly released, but haven't seen anything in...

Considering the lawsuits and papers like this: https://arxiv.org/abs/2301.13188 @wise stratus might be focusing on stopping his company from being bulldozed by lawyers rather than SD v3
This latest paper claims thousands of original works in the training set can be recreated in SD which goes against the "training only captures generalised info about the original works" defence.
I am hoping the paper is somewhat wrong, because we shouldn't be able to pull exact copies of the training data set out of SD, and if the paper is right then seems there are loads of overfitted examples in SD which will make the argument against ai stronger.
Now more than ever we need SD and Midjourney to move to only using ethically sourced data training sets.

proper karma
#

They are probably dealing with a majorly huge amount of nondescript hate and probably making a several-month transition into being a major AAA-level company, I would imagine they have several updates and interesting features in development, it's probably going to take a little while before they are ready to withstand the onslaught after releasing another major version.

proper karma
fervent thunder
cold siren
#

since when a low res artefact ridden image is an "exact copy"

fervent thunder
proper karma
#

anti-ai folk will be walking around screaming about how evil it is while the rest of us are walking around immortal, beautiful, disease and cancer free just existing 🙄

cold siren
#

well big corporations dont have to worry about aging & any diseases, why would they care

wispy jasper
wraith sky
# fervent thunder wouldnt using photoshop to do that be transformative anyway ?! I thought that is...

Depends. Many digital arts are digital paintings where the artist paints from scratch similar to how a oil painter starts with a blank canvas. Not all digital artists copy others works into their art.
I'm a digital painter. I start with a blank canvas and paint my art with brush strokes. No copied art involved.
Of course my imagination has inside many components of all my experiences, and all the photos and artworks I've been ever seen. Some stronger concepts in my mind are those they have many sources to create strong generalisations.
In a similarish way one could argue stable diffusion works the same. It shouldn't be able to recreate 100% the things it has seen during training. But all the generalisations it has witnessed during training allow it to then create unique original works without copying chunks of images. Similar to many digital artists including myself.
If it's not clear I support these art image ai and think they will serve to empower artists.
I'm just worried how the anti ai camp are possibly misconstruing the nature of latent diffusion to make their case.
Tl:dr: no, digital painters build up art via brush strokes similar to a traditional artist and are not copying actual artworks into their images. The people that do copy photo and art into their images are "photobashers" which is what they are claiming stable diffusion is. A photo bashing database. Which is not true imo. But that is their argument.

burnt wadi
#

Hey everyone, in the hieararchy of AI , would Stable DIffusion fit into its own bubble inside Deep Learning? or is there another level before?

burnt wadi
#

Artificial Intelligence> Machine Learning > Neural networks > Deep Learning > Latent diffusion & Stable diffusion?

wraith sky
# wispy jasper I think they are VERY wrong. You cannot pull exact copies. Sure you can sort of ...

Yeah I studied quite a bit of AI. Built a small one ground up myself. So am very interested how this paper claims they are pulling copies of works from the training set out of SD.
I feel they are lying like their original lawsuit claim, the claim the spiral image example is evidence of copied feature. When in fact the spiral image is presentation of thousands of images exhibiting a similar generalised feature which is what the model learns. Always generalised features.
Otherwise how would a 4gb mode contain all the images of 2.5 billion in the training set. I feel the anti ai camp are lying. This is my opinion as an artist and ai enthusiast.

wraith sky
burnt wadi
#

is there an easy resource to explain latent diffusion? i feel like theres a lot of subjects that need previous introduction for others to get it

wide juniper
#

For music, is there an img2img version of it to generate cover music?

burnt wadi
#

Magenta from google has tools for that

#

there is Riffusion but i find it pretty weak

wraith sky
#

I feel the anti ai camp are going to play dirty to attack stable diffusion and co.
It's a shame because this technology can help artists.
I myself use ai to help make reference images that have helped me grow making my own paintings. In fact my skill has progressed faster as a painter since using ai art gen to make the references I need for a artwork. I don't use the reference in my painting but put them up on another monitor and use them by eye to help guide my own painting.

This technology has the chance of enabling more new artists than ever before.

It's a shame it's being attacked like this.

burnt wadi
burnt wadi
fervent thunder
burnt wadi
#

only until the tools show transparency and active encouragement of the arttists that they were trained on, there wont be a resolve

#

no adoption no understanding for it

fervent thunder
#

I do understand that but they are also missing the huge amount of gatekeeping saying you have to draw "real art" is aswell because like some people have physical anid mental disabilities but still want freedom to express images they see in their head!

fervent thunder
wraith sky
# fervent thunder I saw an image that explained the reaosn the pruned files are always similar siz...

The danger is if they make a compelling case in court and that judgement isn't taken seriously and fought against - with the aim of having it overturned, then it'll create precedent and more cases will build on that.
It seems even at this early stage they are being sneaky and misconstruing parts of the stable diffusion white paper, so it's clear they are willing to play dirty. I hope SD and co are taking this seriously, the last thing we need is new limiting laws based on some shoddy knee jerk foundations. 😦

fervent thunder
burnt wadi
#

👀

fervent thunder
#

have you ever tried to extract an image from a safetensors file?

#

its literally just numbers

burnt wadi
#

ehhh sure

fervent thunder
#

Not saying there arent models that have copied work from places but not only copyrighted art contains watermarks

burnt wadi
#

i did not say that

#

nor even hinted that

slate rain
#

why i cant type / dream

burnt wadi
fervent thunder
#

so therefore the logo being generated unless you mean one that is exactly an authors name with copyrighted work? I've just seen alient looking logos in bottom right on some of the very first SD

slate rain
#

??

#

why

ashen atlas
#

Hey there 🙂 Out of curiousity.. what sampler do you guys usually use and for what occasion? Is there like a rule of thumb to use a certain sampler for a certain style of image?

fervent thunder
#

but like the ++ are at least better on lower steps, euler-a is supposedly weirder but i get just as weird results with the others

burnt wadi
#

use an X/Y plot to do a comparative visualization of how the different samplers affect the choices for your specific prompt

fervent thunder
#

get blocked dipshit

#

are there any actual mods here? i never see one

flint fiber
shrewd zephyr
#

Hello everyone, I would like to use Stable Diffusion to generate images for a personal non-commercial project. I would like to know what the limitations are for this, in case anyone is aware. Preferably, what is the license/copyright on the images generated by the software?

fervent thunder
#

go wild

#

buck wild

shrewd zephyr
#

ty

cloud granite
#

I noticed when I check out a certain channel on this server, I see the white dot on the left but I don't see any new messages. What is going on here?

warm tiger
# shrewd zephyr Hello everyone, I would like to use Stable Diffusion to generate images for a pe...

Pretty sure there's both a ToS for the program itself and then again for each of the models you may use. I suggest reading both, ESPECIALLY if you're personal project will be making you money. And you need to keep in mind even if you make something with this tool you don't 'own' what you make. Part of the rule is that anything made with this is fair use for anyone. Unless I'm mistaken when i read it.

#

Also keep in mind that depending on what country you may live in your various laws may be different in regards to something like this, so something to keep in mind just in case

young ivy
#

aaaah finally got Dark Gemini 2 out there and uploaded. I hope people find it enjoyable :o

warm tiger
#

I'm kinda snapping up anything good i can find atm. only have 3 models

young ivy
#

there you can find links to hf and civit. I'm not done uploading to civit

#

full version is on hf, pruned on civit. Others are being uploaded

warm tiger
#

well that's sexy

#

100% DLing that

#

now can you explain what the difference between full and pruned is?

#

any significant differences in usage?

young ivy
#

I know training or merging with pruned versions is hell on earth

#

but just prompting should be fine. It's my first pruned model though that I self-pruned

warm tiger
#

I see

#

I'd eventually like to train off a few models. some stuff refused to generate certain things and being able to train it with images to do so would be good. but 0 idea how to do it and pretty sure my GPU cant handle trying to train a larger model anyway

#

am DLing it. your example photos look amazing

young ivy
warm tiger
#

I'll DL the full. I have something like 40TB of free space so might has well

#

just running on a VPN so it'll take a little longer than I'd wish XD

upbeat tangle
#

which location should i paste vae files?

warm tiger
#

probably in the VAE folder

#

models/VAE would be my guess

native jetty
#

yO

#

hOW do I use bot

warm tiger
native jetty
#

WHere

young ivy
#

there might be better ways, but this is the best I found so far

upbeat tangle
warm tiger
#

if you're downloading it somewhere would it not say where to put it from that spot?

upbeat tangle
#

@warm junco can you help again : D Also, if I put all the model and vae files in the folder at the same time, will it cause a problem? Also, does vae have to be the same as the file name?

warm junco
#

put vae.pt files next to the model in models/stable-diffusion

#

make sure both files have the same name before .filename

fervent thunder
#

What does the sampler do? What does it sample and when?

upbeat tangle
#

so i can place more than one model in same time

#

but its name should be another

warm junco
#

?

#

you only can load one model + one vae at the same time

upbeat tangle
#

okay i get it

warm junco
#

if vae setting is on automatic, it will pic the right vae for the model with the matching name

hallow wolf
young ivy
#

or just use kl anime2 from WD 1.4

#

best vae

upbeat tangle
warm junco
warm tiger
warm junco
upbeat tangle
warm junco
#

web

warm tiger
#

am interested in any and all good quality anime/furry things as well as good stuff for maybe generating 'fantasy' style portrait for DnD or some other stuff

hallow wolf
#

#dreambot-1

upbeat tangle
warm tiger
#

Heun takes quite a lot longer to generate images even with the same sample steps but the results seem to be a lot more... detailed/in depth/colorful than euler

warm junco
spark dust
#

how do you use stable diffusion ai

young ivy
upbeat tangle
warm junco
#

with a space too

#

word, sd_vae,

warm tiger
young ivy
#

download the config.yaml too and rename it to kl-f8-anime2.yaml so it matches the kl-f8-anime2.ckpt

#

other than file extension

upbeat tangle
warm junco
#

now you should see the vae selector

warm tiger
young ivy
#

sure, just don't forget to read the introduction text where you downloaded the model. I try to explain how this particular model responds

warm tiger
keen lark
#

best anime models?

#

is waifu diffusion better than anything v3?

warm tiger
keen lark
warm junco
young ivy
young ivy
warm tiger
keen lark
#

is there any downside to using .safetensors compared to .ckpt

warm tiger
young ivy
#

It's not. Anything 4 and 4.5 aren't by the same creator as anything3

warm tiger
#

huh. alrighty then

fervent thunder
#

I only have anything mixed in with other 5 but I'm using anythingv3 vae and it looks good to me :p

#

also using Heun ( 140 steps )_

#

cfg 10

young ivy
#

140?

warm tiger
#

wow 140 steps

fervent thunder
#

yeeep ;'D

young ivy
#

why on earth would you use 140 on a model known for only needing 20-30 steps

fervent thunder
#

Huh ?

warm tiger
#

i stick with 20 steps until i find what i want then bump it up to 50 for a few then maybe 150 for just a very few

fervent thunder
#

Heun works best on higher steps 70 - 140

warm junco
#

he merged

warm tiger
young ivy
#

then just use DPM++ 2Sa

warm tiger
#

i've noticed bumping it up to 50-100 steps most hands come through okay

#

rather than completely ruined

fervent thunder
#

A lot of people say to use DPM++ but it just doesnt look as good to me

warm tiger
#

this tech definetly needs a lot of experimenting XD

#

i'm tempted to rent a server from google and just run all my stuff off that

fervent thunder
#

I'll try again with DPM

young ivy
#

SDE karras only needs like 10 steps to start producing good images

#

closer to 15 than 10 but still

fervent thunder
#

time isnt an issue, accuracy of input is

#

and consistancy

young ivy
#

indeed

#

depends what you're rendering too

warm tiger
#

this is quite fun experimenting with

fervent thunder
#

I'll show a comparison for my specific use on a specific seed I like lol so its not a great comparison but

warm tiger
#

i mayu have generated over 1400 images int he last 24 hours

fervent thunder
#

i'll put something in the chat with images below

young ivy
#

I run all my first time prompts with DPM++ SDE Karras at 15 steps, cfg 7, 1024x640 (my model Dark Gemini v2 works on that resolution 99% of the time)

warm tiger
#

lastnight i just set it to generate 200 as a batch and just went to sleep XD

young ivy
#

without hires fix

warm tiger
#

i ddid notice that different sampling methods use more VRAM thanothers

fervent thunder
warm tiger
#

also any idea why the AI tends to completely bork when you change the dimentions of the picture? like with anything 4.5 i took the exact same prompts that were giving REALLY good results at a 800x600 px and i changed it to 1200x 800 and it completely destroyed every single result

fervent thunder
#

also im trying to do 5 things at once so cant do a comparison rn sry ;@D

young ivy
#

Euler A changes the image with steps rather than refine it

warm tiger
#

everything was just a splotch of colors or woven threads together

young ivy
#

so it's not very trustworthy

warm tiger
#

or it's make like.. a human centipede with the background being all colored fog. simply by changing the dimentions of the rendered photos

warm junco
young ivy
#

not on Dark Gemini 2!

warm junco
#

i can post an old comparisson

young ivy
#

hawhaw self-promotion

warm tiger
#

ah. well anything 4.5 does really well ay 600 width 600 height

#

but renders basically all nonsense at 1200 width 800 height

young ivy
idle storm
#

hey does anyone know a server where i can buy models?

young ivy
#

why do you want to buy one?

idle storm
#

so that i can build off of it

young ivy
#

most use openrail which means you can build off of it

#

and if you change it enough that it's no longer the former model, then do whatever you want with it

idle storm
#

thank you!!

young ivy
#

mine you can sell images generated, but you can't resell the model unaltered

warm tiger
#

anyone have a link to a tutorial on how to properly train a existing model or if it's even training you want or embedding or hypernetworking?

#

example is like.. on anything 4.5 if i want to add in a dog or a cat or specific patterned clothing. if i add "dog" in the prompts or "cheeta fur jacket" or something it'll just add dog ears/cat ears to the character rather than the clothing

#

or an actual animal it wont do except for like. 1/200

#

like "black cat sitting on table" it just adds black cat ears and no ac tual cat except very very rarely

idle storm
#

@young ivy thats sick im looking at it

#

can i dm u later if i have questions?

young ivy
warm tiger
#

heh. hires fix isn't really viable for me i guess. takes too much VRAM

#

yeah cant even do 2x upscale on a single photo

#

is there a way to reboot the script after an error without having to close and restart it?

idle storm
astral barn
#

I used Synthetic Playground

warm junco
warm tiger
#

and that might fix it even from a CUDA "out of memory" error?

warm junco
#

You can try, i didnt test it yet

fervent thunder
#

you put it in the launch parameters in the webuiuser.bat

warm tiger
waxen pasture
fervent thunder
warm tiger
#

xformers wont run and i have several other issues i'm dealing with

#

using --lowvram

fervent thunder
#

ahh... UHH well one thing if youre using VAE I found recent build needed me to use the no half vae launch option too!

warm tiger
#

idk what thatmeans

fervent thunder
#

set COMMANDLINE_ARGS=--no-half-vae

#

you could try that with low or medvram

#

just a thought, not sure if it will help you but did me!

warm tiger
#

what is that even supposed to do?

#

found some stuff anything 4.5 has a real hard time with. rendering dark elves in particular. it REALLY doesn't want to and then when you manage to get it to they are almost always wearing skimpy black and gold leather armor. regardless of what type of cloithing you try to prompt it to wear. it like, ignores everything and just does that RPG elf armor style thing.

#

like black turtle neck and jean shorts and it's just like "nope RPG leather armor"

#

XD

warm junco
#

Did you tried ()

#

Or shorts:1.3 ?

warm tiger
#

can you explain this to me? i asked before about setting priority of prompts and people tended to just say "put things more important first"

#

but i saw this one website that uses SD and it had a system where you could put things in () or (()) or mark it with a number to set priority