💬｜general-chat

[shadow thunder]

i got my Role on the server and still i cant see #dreambot channel, what do i do

[warm junco]

Look at #1047610792226340935

[carmine cedar]

what do seeds do?

[finite sundial]

The main problem is, the memory itself simply doesn't get cleaned. Lowering usage won't help if there's nothing to use by default, right?

[cold siren]

yeah something happened to the memory usage, now i have to restart the whole thing completely after a single out of memory error ☹️

[warm junco]

Like i said make sure every Programm that is using vram is closed. When you get a cuda memory error The chances are high that it stays filled until you restart the webui again

[finite sundial]

Maybe I'm doing something wrong? It has the memory issue right after it starts, not able to generate even one image~

[cold siren]

except it wokred ~fine most of the time just a few days/weeks ago

[warm junco]

Yea its like beta testing a Programm that is bound to many other aspects, like torch, xformers, python, GPU driver, Windows updates etc.

Also the Programm itself is updating a lot

[eager cliff]

Pardon, but how am I supposed to download a model from huggingface? I'm trying to grab wikiart but there is no checkpoint in the files list and I can't seem to find any 'download model' option.

[near falcon]

is it possible to train an emebding for different char poses like kicking , fighting. I plan to use UE4 manequine to pose and render image for trainning prupose. so is it possible to just train the pose and not image ??

[chilly bear]

is there more websites like civitai

[vestal dew]

Is there an EXIF viewer that shows all the info like https://jimpl.com/ does? I have EXIF Pilot, but it doesn't show the text prompt used in SD to generate the image, like https://jimpl.com/ does.

I do not like online tools - want something that runs on the PC. I just got hacked...dozens of accounts. I hate online tools even more now.

[vestal dew]

Someone told me yes, but I haven't done it. I did see a Reddit post with images from someone who had trained a 1.2 model version for sticking out the tongue.

[clever cradle]

Is there some website where I can download hypernetworks that other people have made?

[vestal dew]

civitai

[compact stratus]

How do i generate an image

?

[compact quiver]

I'm looking for a GPU server to run SD - something cheap. any suggestions ?

[astral goblet]

They are weird yes.

They present their fundraiser as a means to make an entirely new stable diffusion that's open for all, but then the only effort they made was to refine a model

[vestal dew]

SD isn't open to all?

Who can't use it?

[compact stratus]

Hello

?

Do anyone know

[narrow sluice]

They want a filterless AI with that logic.
There's a very good reason why corporate companies won't do that.

Fastest way to getting your sorry butt sued

There's only a few things that the SD license doesn't allow, but well, I thought it's kinda something people should read when they get said software.

[carmine horizon]

Hahaha, I think I'm doing something very wrong, so I would like to ask for your wisdom once again !
I'm trying to train my first .ckpt using LORA after having trouble with embeddings (tried to read about depth maps but.. err.. I'm still too noob for this)

I'm trying to generate class images with my character name "Daelia" as Instance prompt, and "a girl" as Class Prompt (no Instance Token nor Class Token). But... mmh.. generated class images are.. landscapes, cars, roads, I feel it should't be that way.

I must do something very wrong, and I'd appreciate if someone knows what it could be :p

Thanks !

[long sedge]

anyone know why my images disappear as soon as they are rendered on a public link?

[sudden ruin]

TIL about dynamic prompts and how to use them and I've never felt more powerful

[carmine horizon]

Oh, nvm, I thought I needed to keep Instance Token and Class token empty >.<
I suppose that's solved.

[vestal dew]

https://www.reddit.com/r/StableDiffusion/comments/ynztiz/how_to_turbocharge_your_prompts_using/

All these things coming out every day like pix2pix and photoshop api, but still no solution for mangled hands

They are the bane of my existence - I get the perfect image, but the hands are mangled, and I can't get it fixed

I've tried the inpaint in photoshop, but it produces more mangled hands

these are unspeakable horrors designed to throw men's minds into insanity

then I get an image with good hands - it changes the face and other things away from whatI want

Batch count generates some number of images one after t'other

Batch size generates some number of images at the same time...requires more vram

[astral goblet]

generating more images in one batch is more efficient. you'll get faster time per image avgs

[vestal dew]

I have no idea how batch size works because my video card can't handle more than 1

gtx 970

[cloud granite]

I have a poor video card as well, GTX 1060. This is why I use Google Colab for Stable Diffusion.

[astral goblet]

2 size batch means it's twice the resolution. 512x512 would become 1024x512 image

and it'd do them as seperate images of course

[vestal dew]

I think sometimes it's so weak that instead of generating an image it does an image search - just downloads an image, then mangles the hands

[astral goblet]

the old optimization problem artificial intelligences exhibit

[nocturne mason]

I noticed mine vanished as well but made image output to folder that I could view. Was supposed to be an animation, ended up with a bunch of images.

[astral goblet]

doing what you want but not really how you wanted it

[vestal dew]

I tried google collab. I uploaded everything, but then it said I had no more cycles remaining. I hadn't even run the softwareyet

[cloud granite]

I think you should get Colab Pro like me if you had the money.

[astral goblet]

an ai that is tasked to make sure no human dies ever, could just put the remaining population in stasis and never let them create more humans

task suceeded

[white mountain]

If you want to generate a LOT of images, it is always fastest to max out the batch size as high as possible without exceeding your available vram

Batch count doesn't matter... That's the same thing as just hitting generate again

[clever cradle]

So I download a hypernetwork off of the internet. I load up Automatic1111. I go into the settings and select my hypernetwork from the drop down list.

Finally, I click "Apply Settings"

Do I need to do something more than that?

[vestal dew]

Yes

It has a trigger word or words

[astral goblet]

you're using an old version of automatic and should update. the new version treats hypernetworks like textual inversions and you put them in the prompt

[vestal dew]

you have to have the trigger in your prompt

[clever cradle]

Okay. Because even when "Hypernetwork Strength" is turned up to maximum it didn't seem to do anything. I figured something had to be wrong.

[vestal dew]

It's like an Anime. You have to call out the power you want to use

[astral goblet]

before that, people would often put the setting as a quick option at the top of their ui, so they could easily select hypernetworks that were in use without going deep into the settings tab

[long sedge]

yeah, it goes to the folder but that's only for me, other people using my link are just fucked

[clever cradle]

how do I do that?

[astral goblet]

if you originally set it up with git, you can just go into the webui root folder where your webui-user.bat file is, open a command prompt there, and type git pull

[vestal dew]

never

Do git pull by hand

[astral goblet]

that's a little over cautious

[vestal dew]

sometimes the updates aren't good, so you don't want them auto

although there is a way to go back to a previous version

[astral goblet]

don't listen to people who are afraid of updates. same old "don't update windows you could break it" philosophy, because it happened one time.

updates are a good thing. if an update breaks it, there will be another update that fixes it soon

a1111 has a stable branch now too thats worth looking at if you're afraid

[vestal dew]

If you want to do something now, soon doesn't help

[clever cradle]

Should I just open it in notepad and put the git pull command in there so it does it every time I run the program?

[vestal dew]

it's not fear. It's caution. They're different

[astral goblet]

stop spreading misinformation. updates are a good thing.

[astral goblet]

you can do that too! i was just about to say that before "don't do git pull" started up

[vestal dew]

it's not misinformation, chone

I didn't say don't do it

I said don't do it auto

[astral goblet]

weird but okay... just calm down. updates are a good thing. especially in the case of hypernetworks behaving like embeddings

you can now use hypernetworks as a negative prompt

[vestal dew]

I know

[astral goblet]

so don't tell people not to update. updates are a good thing. thye're done using git pull

[vestal dew]

Stop projecting

it's not disinformation to tell people not to git pull auto. It's good advice

[astral goblet]

and you're blocked since you're a child. best of luck to you hunter

[vestal dew]

take a deep breath.

[astral goblet]

Theres now a stable release of a1111. you're probably better off using that one if you're afraid of updates

[vestal dew]

nobody is afraid of updates

[astral goblet]

the bleeding edge of technology is not a safe place for most

[vestal dew]

it's wise practice to only update when you want to, not every time you load

places that for me are a walk in the park, flowwolf wouldn't even talk about - out of fear

[astral goblet]

stop being childish. running git pull isn't this big of a deal. if it's caused you issues in the past, that might've been pebkac

[vestal dew]

mangled hands are the bleeding edge, as in they make my eyes bleed

[vestal dew]

If it's not a big deal, then stop arguing

It's the old "it's no big deal, so agree with me" ploy

[astral goblet]

no. the bleeding edge is a real term. it comes from the textile design industry and the print design industry. the edge of fabric or a printed product. it's so fresh off the press that the bleed of the material still exists and hasn't been trimmed off yet to make it tidy.

it's not about pain or suffering. It's just a material bleed at it's core and the open source community took that as a metaphor to describe nightly builds

[vestal dew]

I know I can fix in photoshop, but I don't want to have to fix every image, so many images

I want to generate images without mangled hands

I invented the modern print industry, son

[astral goblet]

the code base has nothing to do with how hands come out. that's entirely how the model is trained. a lot of it is in the clip model at the foundation of it all

[vestal dew]

the only mangled hands in printing are if your hands get caught in a press

maybe a shredder. that happened to someone I know

[astral goblet]

how can you tell someone is approaching middle age? they start flexing this way constantly. You'll see what i mean when you're older

[vestal dew]

what do you mean by clop model?

Look up JDF, chonoe

[astral goblet]

google is a very regionalized search engine that delivers results to people based on who they are so they get relevant results for them. JDF is pretty vague and probably means something to you, but up here on the west coast of canada, it pulls up a lot of information about Juan de Fuca straight and all the local clubs named after it.

whatever you're bragging about doesn't have as much clout as you think

[clever cradle]

@astral goblet do the new live previews dramatically increase the draw time?

[vestal dew]

it's a standardized print markup language so a digital document maintains format when printed on any printer. It also keeps a record of the print process, and parameters. Prior to JDF, this was in the 1990's, printers were still managing jobs with paper files

[astral goblet]

yup. they're optimized a bit more than the older ones, but generation does still hiccup while they're being created

[vestal dew]

I'm bragging about I need a way to get better hands

[astral goblet]

weird flex but okay. you should know what bleeding edge means then. move along

[vestal dew]

I do

It doesn't mean throw wise practice out the window. Now, I'm going back to those mangled hands

[vestal dew]

Has nothing to do with the print industry. JDF is used throughout the world in the print industry.

[astral goblet]

you've been spiraling for a bit now. Lets keep this on topic.

updating by running the command git pull is a good thing

[vestal dew]

Mangled hands is a huge issue -mangled feet as well, but hands are displayed far more often than feet are. There is no fix for it, which seems odd.

[astral goblet]

running git pull will not affect hands at all, as the code is different from the ai model

[vestal dew]

I thought it wasn't important. You seem fixated.

I didn't claim it would

[astral goblet]

just preventing misinformation wildfires where i can. smokey the bear raised me well

[vestal dew]

Then stop typing. You're the one adding misinformation

[astral goblet]

git pull being an effective way of updating the webui folder to the newest version is not misinformation one bit, ackshually

[vestal dew]

I didn't claim it wasn't

A proof that you can't dispute me is you keep arguing against what I didn't say

[astral goblet]

i fully recommend that anyone who wants to update their webui, goes into the folder in a command prompt and run git pull without fear that it will cause issues. SOMETIMES there are bugs that creep through especially when major changes come, but those are easily fixed when they happen. 99% of the git pulls i do are mundane and go unnoticed

[vestal dew]

Which is what I recommend

rather than adding it to the bat

[astral goblet]

updates are a good thing. "updates break your stuff" is something people who misunderstand software often say

[vestal dew]

I agree...when you want them

you keep arguing against what I didn't say

[astral goblet]

lol bro you got on about this in such a negative way the moment i explained the command line. go away

[quaint eagle]

where can i get the API?

[vestal dew]

I said exactly what you did

I said command line...I agree with that,. You keep claiming Im arguing against updating at all...not true

[astral goblet]

ugh. honest liars are such a dumb trend in 2020s. everyone wants to be one these days

[vestal dew]

go bully your kids...seems you're accustomed to that. It won't work on me

I said don't auto update in the bat

[fervent thunder]

grr, I updated auto1111's webui and liked the results (I've been using a version from last month as the highres fix messed up too much), now when I got another update all the results are super bad and messy 😭

[astral goblet]

woah woah projection. i've never had kids because i dont want children in my life and i have great self control. thats such a weird personal attack. i've not said anything personal towards you and have been focused on git pull being an effective way of updating.

[vestal dew]

Still going on about it

[astral goblet]

git pull is a super effective piece of knowledge for people to utilize here. it's very pertinent to this conversation

accusing me of being a child abuser less so

[vestal dew]

You claimed I am spreading misinformation, suggested I have mental issues "spiraling"

I agree. Git Pull is great. Don't put it in the bat, however

Im trying to get feedback and info on addressing bad hands

you keep pushing on this git pull - I agree, It's great to update. Do it. Keep it updated. I agree 100%.

[astral goblet]

heres where this started. when i explained how to use it in the command line, and then you said "don't do that"

stop pretending you agreed with me from the start.

[vestal dew]

I get bad handy pandies. I want good handy pandies

drop it. I am all for git pull, just not in the bat, not auto every time you run

[astral goblet]

putting git pull into the bat isn't bad either. there are caveats to it but spreading fear, uncertainty, doubt, and general misinfomraiton about it is useless advice. Instead treat it like --reinstall-xformers and remove it after you're done. for some people this is more comfortable than using the command line itself, and that's perfectly fine too

[vestal dew]

I update, I try new things. I use the Photoshop UI, tried pix2pix, try everything new if my card allows it

[astral goblet]

i actually have 3 bat files i use to launch the system, each with different switches and environment variables. i may create a dozen more

[vestal dew]

Yet you keep claiming I say git pull is bad. I never did.

[vestal dew]

good - so you can run with what you want when you want it. Great.

[astral goblet]

1 of those bat files has git pull as a line, and that's perfectly fine. there's nothing to be afraid of

[vestal dew]

Nobody is afraid

I also put on a seat belt - not from fear.

I check my tires for air...not from fear

I look both ways when I cross a street - not from fear

mostly because I don't want my pants to rip if I get hit by a car

[astral goblet]

well, not using git pull in the command line (the original statement you said not to do, as i pointed out a couple times now), is like never crossing the road, or never driving, because of those concerns.

when you git pull you should understand you've got a new version and things may have changed. That's the most caution you need.

all caution comes from fear. that's how emotions work.

[nocturne mason]

You are pretty woke, ngl

[astral goblet]

i feel so sleepy though! cold mornings suck

[nocturne mason]

I hate the cold but it's good for the immune system.

So there is that!

[astral goblet]

increases circulation which is good for everything! alright i'm going to go run around in my shorts in the sub zero coastal air

[wise flax]

🎵 Looking for best ai which creates music.

[astral goblet]

the best still isn't that great. riffusion works with stablediffusion, and google has one they haven't released to public yet.

look up youtube examples' of riffusions work and you'll see what i mean

[vestal dew]

Never said that

running is a decent warm up

I do work outs

[vestal dew]

all caution doesn't come from fear. It comes from wisdom.

If you think it comes from fear, it's projection, because you're afraid of so much

[vestal dew]

I tried on on hugginface. It was underwhelming.

it's a terrible thing to live in fear

[astral goblet]

wisdom of not crossing a road without looking comes from the fear of death from impact. survival is a fear mechanism.

being afraid is nothing to be afraid of

denying fear isn't wise at all

[vestal dew]

when you look into the abyss, you run away, then tell others they're afraid

[cold siren]

so, they removed all the image generating bots... does this mean all teh available computing power is needed for some new model training 🤔

[vestal dew]

When I look into the abyss, the abyss blinks

[astral goblet]

well that's just butcharing hp lovecraft's intention with that remix of his quote

[vestal dew]

they're using it to fix the hand problem

[astral goblet]

wait, or was that poe?

[vestal dew]

Where you are, when the wolves run, you fear the hunt. Where I am, where the wolves run, it's because they fear me.

[astral goblet]

fear is a real thing. everyone has it. pretending you don't is just delusional

[vestal dew]

Nietzsche

I have no fear

knowing I walk beneath it, the storm speeds itself, to escape me

when you walk through a storm, you bow your head, then accuse others of being afraid

[nocturne mason]

Fear is Splenic, as Motivational Color 1, Fear is the Foundation of Intelligence.

[astral goblet]

a lot of toxic masculinity comes from the stubbornness of not being able to admit when fear is affecting yourself. even in the simplest examples like crossing a road. it's a fear you've rationalized and create a habit around. that's entirely fine but is still based in fear.

[vestal dew]

This reminds me of the story of the sun and the north wind. One day the north wind bragged to the sun that it was mightier. They agreed on a contest of who could make the next man who walked by lose his coat.

[nocturne mason]

There is "true and correct fear" which leads to intelligence, and there is "not self fear, false fear" which comes from those who are Motivated by Needs.

[vestal dew]

When the man came by, the north wind took its turn first. It raged and blew, but the more it blew, the tighter the man gripped his coat about him.

Then it was the sun's turn. It smiled. The air grew warm. The man removed his coat.

[nocturne mason]

So what's the lesson?

The Sun is more powerful than Wind?

[vestal dew]

Not if the bet was about an umbrella\

[nocturne mason]

Relativity!

"you're a dumb fish, because you cannot climb this tree!" said the false self.

[vestal dew]

the lesson is that sometimes force can't accomplish what being nice does\

[astral goblet]

ask chatgpt to explain it

[nocturne mason]

kek

[nocturne mason]

Compassionate > Niceness.

[vestal dew]

chatgpt explains falsehoods with the same sincerity it uses to explain the truth

[nocturne mason]

Cuz there's no sincerity at all....

Just binary code....

[vestal dew]

sometimes it gets something right, but its explanation is wrong

it does that with math

I gave ChatGPT an SAT test

It scored 1400 on reading. Very good, not great.

[nocturne mason]

And the Math of Language.... When devoid of CSSCPSG.... It's still behind....

[vestal dew]

It scored 1200 on Math. Pretty good.

[nocturne mason]

Lol

Funny

[vestal dew]

It was a lot of work explaining graphs and charts to it.

[nocturne mason]

I was gonna get Auto 1111 on my PC and I kind of did, but after a couple hours playing with it, I feel it's just a distraction for me, not yet practical.

[vestal dew]

Well, hold on. Not 1400 on Reading. 35, which would translate to a 1400 if other secions were the same

600 on math...translates to 1200 if all sections were the same

[cold siren]

chatGPT is not so smart, in fact most of the time when people complain about it is being woke, it just pretends to be woke as an excuse

[nocturne mason]

It's just doing it's best lol

[vestal dew]

22 on Writing, which would be 880 on the test, below average

[nocturne mason]

Needs better training models....

[cold siren]

would love to have something like it locally though

[nocturne mason]

Yeah that's why I was fiddling with 1111

To have that force, locally available?

Amazing!

[astral goblet]

people can prime it to be anyway they want it ot be

[vestal dew]

1170 was the score with all the sections together...above average

[nocturne mason]

I hope that GPT-4 is like 5 bucks a month for full features.

[astral goblet]

lol no. $40 /m since it's that good of a tool that can save way more than that amount of money

[nocturne mason]

Ehhhhhhh

[astral goblet]

openai aren't in the business of giving this stuff away. they're commercializing it 100%

[vestal dew]

My offer is this: nothing

[nocturne mason]

Fuck money, it'll solve all our problems and make everything free.

[astral goblet]

well openai aren't open, and love money

[nocturne mason]

Replace Gov't with Automatic Intelligence.....

[vestal dew]

openai will also have to pay to license all the material used to train

[nocturne mason]

This is the shit right here

I wonder what will happen in the Homogenized Law

[vestal dew]

someone will come up with something much better

[nocturne mason]

It's already written in fiction language and thus false.....

[vestal dew]

It's embarrassing the way SD draws hands, for example, and can't do action images

[nocturne mason]

This is why I just kinda moved on atm

[vestal dew]

with multiple people/objects...

[nocturne mason]

GPT-4 gonna be dope, I bet.

[vestal dew]

it gets tiresome making portraits and landscapes after a while

I want to see two boxers fighting...

I did get a good Mike Tyson facing four cops after he responded to the Fake Deal's head butts, but it was img2img using the real photo to start, and it made everyone Mike Tyson

[twilit willow]

sooooo... is dreamer's canvas still a thing?

[vestal dew]

what is it?

[twilit willow]

competition from about a week ago

[vestal dew]

if someething came out that worked like SD but with good hands, people would abandon SD in a flash

[twilit willow]

I think it would take time because the imaginary alternative wouldn't have all the tools built around it immediately

[vestal dew]

true

but good hands

I mean, Allstate is locked out of SD basically

they're the "good hands people"

[twilit willow]

they can be improved with some TI and finetuning with SD

[vestal dew]

I am tired of generating 100 images, and the one that's good has a hand with 8 fingers

[cold siren]

cool

most of the time i get only 7

[vestal dew]

I haven't found any hypernet, vae, TI, LORA or model or combination of such that gives good hands

If I see 7 fingers, it's 8 because the thumb is not visible, but there has to be a thumb

[cold siren]

not really, sometimes it becomes a usual finger

[vestal dew]

Or has the fingers tiny, or flat, or going into the body, or hidden behind the back or head or plants or something

what about the double fingers. It looks like one. Your eye says something is wrong. Then you realize it twists into two at some point

[fervent thunder]

I get around 50/50 correctly created hands, can't say everything else is correct tho :P

[vestal dew]

I get bad hands when everything else is what I want, and other things bad when the hands are good

[cold siren]

thats the law of preservation of goodness

when something else is good, it subtracts goodness from hands or something

[nocturne mason]

kek

[vestal dew]

it's like my car

[nocturne mason]

But goodness begets goodness

It's not like we can run out.

[vestal dew]

I say draw me a knight holding a sword

If I add a lot of prompts, I manage to get him without three arms or legs, but there are 6 fingers on the sword hand

[fervent thunder]

I have more issues in getting the correct shape of a face and skull for some reason :P

[vestal dew]

you mean if you want to customize it?

or in full shots? In full shots my people look like horrors

abstracts

Ive tried things like round face, long face - seem to be ignored

[cold siren]

fun (maybe) observation: things like bad hands in the negative discourages not only bad but hands too

[vestal dew]

yes, then it hides the hands

[neon oriole]

if you don't have hands in the negs it will put too many hands

I just have (hands) and it works pretty well

the bad doesn't do anything

[vestal dew]

I want tow handy pandies each with four piggies and a thumb

two

[fervent thunder]

For me it's more to do with how it seems like the face gets more defined than the rest of the image, almost like you cut out a face from one image and paste it into another

[vestal dew]

so how do you make sure four fingers and a thumb

[vestal dew]

can you show an example?

[fervent thunder]

I can pm you about it if you want :)

[vestal dew]

Ive tried gloves and mittens in negative, to remove flat fingers, fingers stuck together

[vestal dew]

go ahead

I just got two good hands out of four generated in Photoshop UI, but the two good ones are about half the proper size

[cold siren]

should put gloves or mittens in the positive to hide them altogether 🙈

[vestal dew]

I don't want them hidden

[cold siren]

heh it won't hide them anyway

[neon oriole]

the trick that has worked the best for me to fix hands is to do a second pass with an overfitted model like AOM2

[vestal dew]

If I could tell SD - that, but make it twice as big

img2img pass?

[neon oriole]

yes, AOM2 is really good with hands for some reason

[cold siren]

yeah btw if only you could tell SD something

[neon oriole]

or you could inpaint with it

[vestal dew]

I am doing that now in Photoshop

it's giving me either one big thumb, a mangled mess, or decent hands except half the size they should be

[cold siren]

...like if you could tell the model to adjust itself when it generates not what you wanted

[vestal dew]

pix2pix supposedly does that.

You can mask, but then things like hands don't match the arm

or select in the Photoshop UI

[cold siren]

is it even possible to inpaint out all teh extra limbs anyway? every time i try it just regenerates them back

even with "latent nothing"

[vestal dew]

I haven't tried

I just throw those images away or Photoshop the extra limbs if the main limbs are complete

I got a great image of a cabin in the woods by a stream...great when I added that it was on fire...then got a three legged dragon when I tried to addit

[cold siren]

yeah that's the thing, it isn't just "bad with hands" - it is equally bad at everything, hands are just the most noticeable for us humans

[vestal dew]

In SD I get things that look more real...except for some parts, and except I can't have multtiple, and I can't have interaction, and I can't have consistent characters

In DAZ I get everything except the realistic or easily into many styles

[cerulean kraken]

Does SD run any faster on Linux vs Windows?

[cold siren]

doubt

[eager orchid]

what's a good prompt for getting the edges of an image to fade? 🤔 I'm generating black and white drawn portraits where I want them to fade into the background.

[fervent thunder]

I sent an example, don't know if you received them or if you just don't got the time to respond yet :P

[astral goblet]

a big part of the limits are because the clip model it's based on, open clip, isn't very accurate at times. i'm only just beginning to learn about "CLIP" and other models that are being used like google's T5, i think it's called. it's part of the pretraining of these models. 1.5 uses openai's trained model, and 2.0 uses stabilities trained model.

here's that project's code base https://github.com/OpenAI/CLIP

the move to openclip is a minimal viable product if i understand things right. it's a model that stability can affect better. so while it's not as mature as openai's that was used in 1.5, it's pretty much almost as good as that one. They also use a more modernized version in dall-e2 it's suspected, while not releasing it to the public. Going forward, openclip will get a lot better and fast

https://twitter.com/laion_ai/status/1618317487283802113 here's something laion put out recently, i think it's a big deal that will show awesome results sooner than later

https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k

[fervent thunder]

what's zero-shot? :O

[astral goblet]

can recreate something it's asked to recreate, with no examples of it in the data set

well, it's more complicated than that 😉 https://en.wikipedia.org/wiki/Zero-shot_learning

[fervent thunder]

it always is :P

[astral goblet]

i'm probably way off base with my ml 101 first day in class take on it

[cold siren]

vignette?

[astral goblet]

OH and the license for vit bigG is MIT! so that's a step in a new direction

honestly i think it's a huge release that the waves haven't quite reached a surface disruption yet

bomb drop deep in the academic labs of ML research that will soon create tsunamis of disruption

[cold siren]

yeah if only we could have a model that can learn as we use it

[fervent thunder]

I didn't understand any of those laion stuff. Like what it's supposed to be :P

[eager orchid]

yeah that seems like it should work, but it doesn't seem to be something the model recognizes... my model, anyway

[cold siren]

just do it in photoshop 🤷

or something

whats the best free alternative these days btw?

offline only*

somehow i don't understand krita like at all 😐

[neon oriole]

gimp is probably the best open source photoshop alternative

[astral goblet]

the checkpoints we use are trained on images thhat go through another model, pretrained on tons of image pairs and words so that it builds a natural language understanding of what words mean what images. Contrastive Language-Image Pre-training. CLIP

1.x uses openAI's released clip model for training. 2.x uses a new one trained with OpenClip, a new implementation of CLIP created by Stability. LAION just put out their own trained model of OpenCLIP, released to the public. They used stability's massive super computer cluster to train it

more accurate clip models lead to better diffusion training

[fervent thunder]

so, laion is the name of a company?

[astral goblet]

laion yeah, they're a non profit that does research, funded by a lot of different interests. they're mostly known for the image dataset everyone uses for this stuff, Laion 5b

[fervent thunder]

is that why there's a laion vae? I never liked that one tho :P

[astral goblet]

also another use of clip models is image classifiers. recognizing images. more accurate ones lead to better interrogations of images. prompts that could potentially get very close to a recreation

laion releases a lot of different stuff

[gaunt coral]

F

[fervent thunder]

thanks for the info, I thought laion was the name of something technical :P

[clever cradle]

I got a Lora from the internet

It doesn't seem to have any noticeable effect on the rendered images

Yes I did use the correct keywords

I even put the keywords in parenthesis as it was in the example images I was given

[vestal dew]

I see what you mean. It's as if it's a different style, different coloring. I get that sometimes, especially with full body

[rancid sail]

What kind of projects are you all working on with your SD generations?

[clever cradle]

I got this Lora from the internet, but it doesn't seem to have any noticeable effect on the image generation

I made sure to use the proper keywords. I even put them in parenthesis as it was in the example prompts I was given

[rancid sail]

What are you trying to generate?

[hollow bison]

Hi, I'm trying to use a .safetensor to train a model, but TheLastBen's collab doesn't read my checkpoint, can't they be used for training? Or do I have to look for another collab?

[clever cradle]

Will that knowledge help you help me figure out what the problem is, because otherwise I'd rather not say

[rancid sail]

Ah, I wasn't trying to help you debug per se, more of just curious what folks are working on 🙂

[violet dawn]

Have all the dreambot-1 etc channels now been archived ?

[formal coral]

I’m running meet-ups in London, any ideas for how I can get the word out? https://lu.ma/LondonPromptEngineers-13FebPromptJam

[hybrid sphinx]

Como usa?

How to use?

[vivid portal]

@karmic brook can you reply to this?

[neat condor]

guys something very weird happened to me, so basically i've been using windows 11 on aorus elite x570 and i never touched bios flash for 4 years i have it, today i was reinstalling windows and decided to what the hell and go with it, so i flashed to latest version (F37). my generation times went down by HUGE amount - to 19 seconds from 30 seconds, so little more than third. so what can be the case? i am thinking maybe old bios didn't supported windows 11 properly?

[vestal dew]

Old bios might not have supported some driver in windows or some hardware on the motherboard or video cards

I tried 2 TI for fixing hands...Dreamartistbadhand and bad_prompt: still getting mangled hands

[warm junco]

Yea bios updates also support new gpus aswell RAM and win11 always a good idea if your reinstall Windows especially if its after 4 years

[orchid wadi]

The OpenClip in 2.x was trained by Stability. The CLIP in 1.5 was pretrained by google. The switch was done - far as I know - because google refused to allow stability to train it further. Problem was that CLIP was trained with different pre-selected dataset of image and text pairings, and it had content in it which was not in SD's training material.

[neat condor]

figured it out via help from another discord, it was xamp. before it was turned off by default but it's turned on by default now on second level so instead of 2166mhz my ram runs at 2667mhz so that's most likely to be the case.

[astral goblet]

oh it was google not openai? mb. this new one is released by laion

[warm junco]

Yea this was exactly my problem, my RAM was on 2600 and xmp was not working. After BIOS update its working woth 3200

[vestal dew]

Someone train it to make good hands

[warm junco]

Merge models give you better hands

[orchid wadi]

The MODEL of clip that was used was trained by google. CLIP as a system belongs to OpenAI.

[neat condor]

oh wow so it's a common problem. still surprising how RAM could improve generation times by a whole third though. didn't knew it was that important.

[vestal dew]

merge models...what does that mean?

[astral goblet]

everything by deepfloyd exhibits good hands if you've seen any of that. deepfloyd uses stable diffusion, but a model trained with a more accurate clip model. that's a big part of the key

[orchid wadi]

You take two models and you merge their weights together using an algorithm.

[vestal dew]

I've tried a few TI's, a few VAE's. They say they fix hands. They don't

[orchid wadi]

You can't fix hands like that!

[vestal dew]

how does merging two models fix hands?

[warm junco]

In the webui you can merge/mix models under the checkpoint merge tab

[orchid wadi]

It doesn't. However another model can have more refined training on hands overall.

[vestal dew]

ok'

So tell me what model has better hand training

[warm junco]

I mixed an anime model with proton and Now i get 80% good Hands

[vestal dew]

So in one of those alone you get good hands too?

[orchid wadi]

The problem is that try to imagine how many configurations of "hand" can you come up with:
Open palm up, open palm down, fist, thumbs up, holding to someting, fingers speard, fingers together, pinkey out. Then 5000 different angles of images.

[warm junco]

You should try ProtogenX2.2

[vestal dew]

No way

how about face? Mouth open, mouth closed, frowning, smiling, eyes open, eyes shut - faces are fine

I will try protogen

[orchid wadi]

To "fix hands" you'd need to make a model like GPFGAN that recognises hands and the angle they are being show from.

[vestal dew]

And now in the new UI they can be merged, for A1111?

[orchid wadi]

Are they thought? They are better, but if you look at them clsoely you will see akll sort of issues.

[vestal dew]

I see many faces without major issues

not like with hands

[orchid wadi]

The reason faces are easier because we really can see faces from only one set of drections and there really aren't that many variables.

So do I see faces that are good. Especailly in 2.1 especially with a style embeddings I made.

[vestal dew]

so mixing models and protogen really don't fix hands?

I see good faces in 1,5

[orchid wadi]

Assuming it has been trained more with different more refined images, the all it does it ass more refined UNET weights.

You can fix hands by just DB training with good labels.

But just imagine how many different hands you'd need.

[vestal dew]

so where are the models with fixed hands?

[orchid wadi]

Try this out. Try setting you fingers to as many different poses as you can. And count them.

[vestal dew]

5

[orchid wadi]

Try first with just looking at your palm.

Really?

[vestal dew]

5...5...5...5...5...5...5...5...5...5

every pose: 4 fingers, one thumb

[orchid wadi]

because I can count to 20 with one hand.

[vestal dew]

you count 20 fingers?

[orchid wadi]

With 5 fingers I can count to 20.

[vestal dew]

but there are only 4fingers 1 thumb

look at most photos of humans showing hands...4 fingers 1 thumb

my right thumb was nearly ripped off, put back on. Guess what? I still have 4 fingers, 1 thumb

[karmic brook]

The winners will be announced this week

[orchid wadi]

Ok. Lets say thumb is leftmost:
1----
-2---
--3--
444--
-5555
-66--
-777-
---88
99---
--10--

That is to 10

Then for 10 modifier we use thumb to hold down pinky and 4th and then count the variations from there.

[warm junco]

Yes under checkpoint merge, use Model A and B, Slider to 0,5 and then Weighted Sum, then ckpt or tensor and check "Dont"

[orchid wadi]

Then for 100 you actually you flip palm.

And you doo how may 100 then how many 10s and then from 1-10

Then most fingers can be hooked individually. You can hook your whole palm.

[vestal dew]

Ah I wondered what that was

[orchid wadi]

My point is... There are so many fucking variations of hand poses that it outmatches faces. How many variations of face there can be in compostion? Like 1 eye, 2 eyes, no eyes, nose, no nose... You can't really have a face without mouth.

[vestal dew]

Yet you still have four fingers, one thumb

it mangles hands in every pose except hands hidden

show me a hand pose that is never mangled...one that shows the fingers

[orchid wadi]

I have elbow and wrist that makes it possible to turn my palm 180.

Well actually I can go way past 180 because I'm double jointed-

[vestal dew]

I get an image of a person with hands on hips. One hand is good, the other is mangled. Why?

they're both in the same pose. One is good

Or both are perfect, except on hand is backwards...has thumb on the wrong side

[orchid wadi]

Because the AI has seen images where in that pose one hand might on the side, behind back in a pocket

[vestal dew]

everyone can turn their palm 180

[orchid wadi]

Once again THE AI DOESN'T KNOW WHAT A HAND LOOKS LIKE.

That is the thing

[vestal dew]

yet it gets one right

[orchid wadi]

It has an average mess of "hand" in its "mind"

Well it did "a hand".

[vestal dew]

on some images right one,...on some left

[orchid wadi]

Try prompting hand twice.

I do that

sometimes helps

[vestal dew]

so why isn't there any model with that trained

[orchid wadi]

You need a massive dataset of hands that have unique labelling for every variation.

[vestal dew]

I put my name in there...it gave me a picture of me. How can it know me but not hands? I'm a nobody.

[orchid wadi]

Like it is doable. I got a draft in my notebook for it

[tender wave]

(human hand)
negative prom. extra fingers extra limb, bad anatomy, deformed

[vestal dew]

Isn't that what people are doing when they make TI's for hands?

[orchid wadi]

Yes... You make a TI for specific hand poses.

Like it isn't hard!

[vestal dew]

I've tried all of those types of prompts, TI's, models that claim to do hands well

I still get horrible hands

[orchid wadi]

I don't generally...

So it sounds like you issues

[vestal dew]

I just tried a guy holding a bat

half the time, the bat is like a finger, or coming out from a snake like finger

it's me issues?

Someone posted three images of just hands to show hands could be done well. One of the three had a hand with six fingers

[orchid wadi]

Ok so... Tell me. If you show a picture of hand holding a bat. How is the AI supposed to know where one ends other begins?

[vestal dew]

if that's one of the proof photos of good hands, then it's not just a me issue

[orchid wadi]

Well I fix hand with PS or in img2img if I am desperate.

[vestal dew]

right, I can fix all of this in PS

But I wouldn't have to fix it if the thing generated ggood hands

[orchid wadi]

Well why dont you fine tune a model with good set of hands then?

Like with everydream? or whatever it is

Clear set of very well imaged hands, with clear labels, many variations. Clearly defined from other things.

You just need to train the AI to get a clearer idea of the thing.

[vestal dew]

https://neuroflash.com/blog/stable-diffusion-examples-3/

[orchid wadi]

What is your point?

[vestal dew]

An article about terrific SD images that inspire people. Oh look, the astronaut has six fingers

in an article bragging about how great SD images are.

[orchid wadi]

So?

[vestal dew]

Whoever wrote that picked the best images. Imagine what the bad ones looked like.

[orchid wadi]

What is your point?

[vestal dew]

It's a me problem?

[orchid wadi]

Yeah? I can manage to fix it.

I use TI embeddings

[vestal dew]

My point is it's not a me problem

SD has problems with hands

[orchid wadi]

Well yeah not fixing it is.

Yes it has.

And there are tools to fix it

[vestal dew]

what tools?

[orchid wadi]

TI, DB, LORA, HN.

All can be used

Hell you can fine tune the whole model with hand model images.

[vestal dew]

show me 100 images generated with your tools...with even just 50 with good hands

Which ones?

[orchid wadi]

Remind me on wendnesday.

I'm going to bed for it is 1am and I got work tomorrow

[old marten]

People are still using the bad hands argument for AI? Lmao.

[orchid wadi]

Need to drive 2hrs to fuckoff nowhere.

[vestal dew]

what loras, TI's, Hypernets, models have good hands for even 50% of images?

[old marten]

2022 called. They want their dumb arguments back.

[orchid wadi]

I don't know! I MAKE MY OWN!

[vestal dew]

Why don't you upload so others can use it?

[orchid wadi]

Remind me on wednesday. Because I probably wont be home till very late tomorrow... well today

[old marten]

Bruh. You don't need any model with good hands you just need one and use inpainting.

[orchid wadi]

because they are for very specific uses and don't work universally?

[old marten]

Even the standard inpainting model 1.5 will probably get it right if you create a batch of like 10 pictures.

[orchid wadi]

Like I got one for hand with palm up. But it will stylise it to Icon painting.

[vestal dew]

Ive yet to see one that for any pose, any style, any images is consistently good hands

[orchid wadi]

Again... Remind me on wednesday

[old marten]

This guy is trolling

[vestal dew]

I am not

[orchid wadi]

Ill be on my tower then

[vestal dew]

hands are my bane

I get terrific images for what SD does, except for hands

[old marten]

Then either learn to inpaint or learn to Frankenstein

[vestal dew]

I'd love to be able to make action shots of people fighting...SD can't do that

I'll settle for good hands

I can never get inpaint to do good hand replacement

I can easily fix this in Photoshop

[old marten]

Once my PC is free, I'll make a good video tutorial just to prove you wrong lol

[vestal dew]

You wouldn't be proving me wrong

[old marten]

I will pick out a random topic no cherry picking and I will fix the hands in less than three minutes no Photoshop

[vestal dew]

You'd be teaching me

I don't want to be right aboutit. If there's a consistent fix, I want to use it

[warm junco]

I can make good hands with merged models, tomorrow i can show some

[old marten]

Merged models are the shit

[vestal dew]

what models if I merge them will give me good hands?

[warm junco]

Protogen

[old marten]

They seem to exponentially complement each other

[vestal dew]

and anything>

[warm junco]

If it fits your style yes

[vestal dew]

I want to make this clear. I am not trying to be right or prove anyone wrong

If there's a way to get it, I want to learn it

[orchid wadi]

Use TI.

[vestal dew]

what TI?

[orchid wadi]

Textual Inversion

[vestal dew]

I've tried today three TI's that claim to fix hands

they don't

I know what it is. I want to know what TI has it.

[orchid wadi]

You need 5 CLEAR AND GOOD IMAGES OF THE SAME POSE. In no more that 45 degree variation. Different background, preferable skin colour.

[warm junco]

@vestal dew bad-artist, bad-prompt, bad-image, deep-negative

[orchid wadi]

500 steps at 0.001 usually does it for me.

[vestal dew]

I see. You're saying if I want one good hand pose, like hand holding sword, then I can make a TI of that specifically

[orchid wadi]

Well I wouldn't train a sword specifically

Unless it is like 2handed sword

[vestal dew]

Holding something shaped like a sword

[orchid wadi]

But like "holding to a cylider".

[vestal dew]

right

[orchid wadi]

Well generally the pose is just enoug

[vestal dew]

So I have to wait until I get a better video card to train

[orchid wadi]

Then make filewords to ELIMINATE EVERTYHING ELSE! Do not use CLIP captions

Filewords should be like "Person's [name] a bat" or such

[vestal dew]

I can't use CLIP captions because I don't know what they are

[orchid wadi]

It is the default caption maker if you use Autos' repo to preprocess images

You TI train on just about any auto's repo Colab

Or just a specific TI colab

Up to you

I recommend random sampling. Use same resolution images.

Have the hand at centre of the image or in the general sector of the image you want it to be in.

[vestal dew]

Ive tried training on google

it says I have no cycles ...it said that the first time I tried

[orchid wadi]

I don't train on colab so you have to contact the person upkeeping the notebook about that

BUt contact me torrow evening my time and I see if I can help you or even make an embedding for you for a pose

[vestal dew]

I can make hands in whatever poses are necessary

I have DAZ

[orchid wadi]

I'm off to bed.

What?

[vestal dew]

I can make a character in daz with the hand posed however, holding whatever

[orchid wadi]

Wtf is Daz?

I draw them by hand myself

'Becasue I can do that

[vestal dew]

it's a 3D figure software

I can't draw

my little profile picture...DAZ

[orchid wadi]

Posted an image on general with images

That is how I make material to train with.

[forest aspen]

I've been working on something, i hope this helps people who are starting with AI, as well to help to others who are trying to get specific results with their designs https://www.youtube.com/watch?v=-jCuP8NYQgA

[vestal dew]

What are we all doing fundamentally wrong? Your video says we're doing it wrong.

I will try protogen for now, then wait for a better video card so I can train

Protogen 2.2 or 3.3?

er wait.,..now all kinds of other ones popped up

x5,3 realism

[warm junco]

There are multiple Versions, can recommend 2.2, 5.8, 8.6(Infinitiy),

[vestal dew]

x3.4 photorealism, 2,2 anime

[warm junco]

2.2 and 5.8 best for humans

In my tests

[vestal dew]

I will try 5.8 then

[warm junco]

Infinitiy 8.6 good for realism

You can also try 3.4 i skipped it

[vestal dew]

I see a world in which AI has advanced a great deal. Through the use of AI and nano technology, what we do today on a computer people of the future do in the real world

5.8 is downloading

will it work? https://www.youtube.com/watch?v=7KeFYXtCPUM&ab_channel=UnfamiliarPlace

[warm junco]

You can also try 3.4 and compare it with the same seed, then merge the better model with other good models

[vestal dew]

goos?

good

[warm junco]

Goooos

[dapper patio]

Question: in my trained models close up headshots look perfect, but full body photos the face gets absolutely ruined... Any ideas why

[vestal dew]

I already have 3,4

[vestal dew]

Make more detailed text prompt

re: the face.

[warm junco]

The closer the face The better the Details

[fervent thunder]

new stable model when? 😭

[vestal dew]

yes, but even far shots...wasn't it you who told me, more detail re: the face results in better faces

I have protogen 22 and 34

[warm junco]

It just depends how close the face is and the image resolution. The closer the face The more Pixels it gets for better results

Face fix can help for far shots

But Portrait is my go to

[dapper patio]

So generations waste up look perfect, any prompts the include legs the face gets mutilated

[warm junco]

Yea for example dont connect highgeels blue eyes and standing in one prompt xD, sitting will fix it

[sick barn]

Why can't I use the dreambot?

[warm junco]

Dreambot is shutdown #1047610792226340935

[vestal dew]

Right, but what if you want a clear face in a full body shot? Can't get it?

[uneven heath]

Speaking of which when is Emad doing QnA? There are a lot of us here now

Maybe dealing with a lot of stuff, its all in wildfire right now

[warm junco]

Only with face fix or realy high res

[vestal dew]

I see

so res on initial generation matters

[warm junco]

Yes when you generate 512x768 every stuff it generated is sharing this little space. With more res stuff can get more space to get more Details (can recommend highres fix upscale)

Need to go to sleep now to get 6 hours hopefully xD

[spring thunder]

Help..... I am currently trying to install stable diffusion and I have been stuck for hours. While running "webui-user" (windows batch file), it says python not found. Python is installed and there is a reddit post recommending that I delete the "venv directory". Does anyone know how I could do that or know what i'm doing wrong?

[rich sand]

Dream bot channels closed?

[waxen pasture]

So i have done a clean reinstall of the webui, as well as downloaded a version from the same day as the time I got a good result, and neither worked. I still have the issue. There is another thing that shows up in the console that might be related to the issue, I am not sure, but it happens on both tthe new reinstall and the old version, and I have never seen it before, so I don't know. I think I have it posted in my issue in #🔧｜finetune . I am getting really frustrated by this :(.

[brisk bloom]

that sucks. not sure what that could be

[waxen pasture]

I'm sure it's some dumb little thing that is a super easy fix, especially if no one else has said anything yet.

[abstract ice]

3.0 ready yet?

Need that 1024*1024 magic

[fervent thunder]

I wonder if Deep Floyd is 1024

[abstract ice]

Hope so

Imagine the quality

[obsidian moon]

Hi, I am downloading 2.1... how long does it take for you all to download it? my cmd stays at "installing open clip" FOREVER... ugh

[wide juniper]

How do you think corneo’s 7th Heaven is compared to Anything V3?

[neat condor]

Okay guys bout to drop interesting data:

So since i'm all about saving time anywhere i can and since i was reinstalling windows anyways, decided to test SD performance on 10 and 11, both installs were fresh with only SD on them - here's the results:

System: GTX 1070, ryzen 5 3600x, 32GB ram 2667 Mhz corsair dominators, M2 ssd, Aorus elite x570 mobo

i did X samples on each windows and rounded up then compared against eachother

first time SD boot up time after restart - did 20 samples on each windows: windows 10 wins by 5 sec
second time boot up after restart - again 20 samples - windows 10 now wins by 7 seconds
image generation on 512x512 - 100 samples on each windows - windows 10 wins by whooping 9 seconds
checkpoint load time - 20 samples - windows 10 wins by 2 seconds
UI reload time (from settings) - did 20 samples - windows 10 wins by 5 seconds

at this point i stopped testing and decided to roll back to win 10 since i use SD almost 90% of my PC time. Hope this data will be usefull for you too.

[brisk bloom]

did you test w11 with hardware accelerate GPU scheduling option disabled? I know that can cause many slowdown issues. Not sure about SD but its mostly GPU usage. Im still on w10 so cant test that

[neat condor]

i didn't activate nor deactivate anything on neither of them, just ran SD to get as vanilla results as i could get.

[brisk bloom]

w11 has it auto enabled. its not avail in w10

so its something you need to change

[neat condor]

well i'm all out of the testing spirit for now and decided to stick with 10 but in few months on next reinstall, definitely will be keeping that in mind. was using win 11 for last few months and didn't even know about that, who knows how many times it caused problems for me lol.

[brisk bloom]

yeah from what Ive read, in games w11 is pretty much the same as w10, as long as you disable that option. So its the only thing I can think of that would create the large difference in times you recorded

[neat condor]

yeah in games i haven't noticed any big difference during its use, however these are just statistics, now windows 10 has to prove to me that it actually "feels" faster, if you know what i mean.

numbers on paper is one thing, using it for months and "feeling" the speed is another.

[brisk bloom]

I havent upgraded yet as MS keeps changing a lot of things and they break too often. w10 is stable for me so I will wait it out

[neat condor]

yeah w11 wasn't that bad in stability department maybe had max 5-6 BSOD's during what 7-8 months

not bad for 1 year old system

[fickle crane]

how to use SD

[neat condor]

check aiterpreneur on youtube

he has all the info u need

[muted parrot]

I have 3 monitors, totaling 7680x1440 pixels. What would be the best approach to generate a prompt to fit that pixel ratio? obviously prompting a 2048x384 image and upscaling could work, but I noticed a significant difference in prompt outputs between, say, 2048x384 and 1024x1024 resolutions.

Does anyone have experience using SD to generate these long/wide aspect ratio images? if so, what is the best approach to ensure representation of the prompt in the final result.

[granite stratus]

hello there

[peak orchid]

The 4x upscaling model if you have a good enough GPU

[muted parrot]

I have an RTX 3080 and have tried upscaling 2048x384 into 7680x1440 (3.75x) but always run into CUDA out of memory exceptions

[brisk bloom]

try the ultimate upscaler for A1111, it doesnt rely on gpu memory as it generates tiles

under extensions

[granite stratus]

i have a question i can put the bot on my server?

i will just use to play with some friends.

[waxen pasture]

The bot is no longer active anyway

But I don't think you could even when it was active

[granite stratus]

F

[peak orchid]

I guess you'll have to use tilers, and not the 4x model: https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler

[muted parrot]

thank you @brisk bloom and @peak orchid for the suggestions, i will try those

[brisk bloom]

for reference, Ive used the ult upscaler to generate a 7kx7k image . I also have a 3080

[wide juniper]

What's the point of forward diffusion process if it's just about adding noise? Why can't diffusion models just begin from plain noise?

[astral goblet]

its for training if i'm right

[muted parrot]

I don't see this under extensions, is there a specific github link i need to point to? google also doesn't provide very many useful results for that search term

[peak orchid]

This one? https://github.com/Coyote-A/ultimate-upscale-for-automatic1111

[brisk bloom]

ah right. you need to install from URL, for that one I think

[wide juniper]

Training? But I’m just generating art

[waxen pasture]

I see it under the extensions tab

(Although I already had it installed I had to uncheck hide installed to see it)

[wide juniper]

Are you implying that the first half of the steps is me donating processing power to train the model?

[muted parrot]

ok yeah, got it installed, and then how do I implement it? I don't see it in my scripts dropdown

[brisk bloom]

if its not in img2img script dropdown then reload SD

[waxen pasture]

If I am understanding what you are saying, no. Just generating art does nothing to a model.

[short tapir]

question about one of the options.

The option to do an initial image. What does that do?

I tried using a picture of a person for it but it doesn't generate anything even remotely like them. Am I doing something wrong?

[obsidian moon]

..

[short tapir]

Shouldn't that be in #🤝｜tech-support

[obsidian moon]

oookay thanks~

[short tapir]

Also looks like low memory if I'm reading it right

[neat condor]

@obsidian moonu are trying to run it with 2GB gpu, i'd drop it

trying to make it work with anything below 8GB is just plain masochism

[obsidian moon]

but my pc says I have 8GB

[neat condor]

well in that log it said u have 2GB lemme look at it again in tech support

[waxen pasture]

I need someone to test if an embedding I am currently trying to make actually does anything to your image, anyone up for it really quickly? Just need like a few prompt s/r or prompt matrix images

Because I have been having an issue where embeddings aren't actually training for some reason

And I would like a tester other than me to see if it's working

[waxen pasture]

forget about this, I found my issue on the issues page on the github for a1111, trying to fix it now

[wide juniper]

Whenever we do a text to image prompt, does the noising process happen at all? It's just the denoising process, right?

[waxen pasture]

Yes, it is denoising from random noise

[wide juniper]

https://civitai.com/models/66/anything-v3
It says updated five days ago. But if so, why is it still V3 and not V 3.1?

[waxen spruce]

just wanted to say thank you to the devs, this is working great and the install was way easier than some online made it seem

[astral barn]

Специально для русских.

🤖 ** Dreambot сворачивается ** 🤖
После **2 месяцев ** непрерывной генерации, почти 2 миллионов ** изображений и некоторых действительно выдающихся результатов Dreambot пришло время сделать заслуженный перерыв!
Каналы Dreambot будут отправлены в архив на выходных!
Перейдите на канал #1002292387051212860, чтобы получить роль ** Dreambot-archive, и вы по-прежнему сможете просматривать их вместе со всеми своими творениями как каналы, доступные только для чтения.
По мере того, как одна глава подходит к концу и Dreambot уходит, нас ждет еще много захватывающих сюрпризов ...
** ОБНОВЛЕНИЕ: ** Все каналы Dreambot теперь закрыты! Спасибо всем за то, что мечтаете вместе с нами!

🤖 Dreambot winding down 🤖

After 2 months of non-stop generation, almost 2 million images and some truly outstanding results, it’s time for Dreambot to take a well-deserved break!

The Dreambot channels will be heading into the archive over the weekend!
Head on over to the #1002292387051212860 channel to pick up the Dreambot-archive role, and you’ll still be able to view them along with all your creations as read-only channels.

As one chapter draws to a close and Dreambot takes its leave, many more exciting surprises will be heading our way...

UPDATE: All Dreambot channels are now closed! Thank's for dreaming with us everyone!

[waxen spruce]

this will be very useful for creating handouts for my dnd group

[astral barn]

А пока занимаемся саморазвитием, сказать точно не могу, когда обновят ботов. Вчера обещали, что сегодня. Если нет, значит ботов все таки навсегда закроют.

In the meantime, we are engaged in self-development, I can't say for sure when the bots will be updated. Yesterday they promised that today. If not, then the bots will be permanently closed after all.

[warm junco]

Would recommend using lower res + highres fix upscale then upscale in extras

[wide juniper]

I am working on a visual novel right now with around five central characters. I’m thinking of using dream booth or textual inversion to create consistent characters, especially their costumes. I have a 3090. Should I use dream booth or textual inversion?

[halcyon granite]

I was wondering if there's a difference using using upscale in GUI (e.g. auto1111 or nkmd) Vs upscaling the image in post (e.g. cupscale). Will the upscale be handled differently if it starts during the AI generation process?

[warm junco]

Dreambooth for sure

[warm junco]

There is a significant difference between upscaling with highres fix and upscaling with Extras tab.

[halcyon granite]

can you use DB with a single characters image by the way?

as far as I'm aware people upload a small set of images of the subject

[warm junco]

Idk, only heard someone made this With an an embed

[warm junco]

Yea it should be around 10-50 images

The more the better

[vestal dew]

Someone told me model merging does a good job of fixing hands if Protogen is mixed in. Mixed or on its own, Protogen does little to stop mangled hands or feet. More than half the images have extra fingers, and many have mangled hands including fat fingers, noodle fingers, merged fingers, no thumb.

[warm junco]

For me its working good with the merge

[vestal dew]

It hides hands often. But when shown, hands are mangled

[warm junco]

I dont have much images on my Smartphone but i can post one in #🏞｜general-with-images

[vestal dew]

I can post one too. The point is what %

Using a blended model with Protogen, two TI hand fixes, a vae for good hands - still getting mostly mangled hands. Also, once the first image with mangled hands appears, the frequency of mangled hands increases - then almost every one has mangles

[warm junco]

Can you post some pics ?

[vestal dew]

The images I am making are kids playing soccer, so no. When I tried to post before, I was banned for a day.

[warm junco]

Ohh

[vestal dew]

Im trying to make something I could post to fiver or shutterfly, maybe people would buy the images or similar - caricatures of their kids in their sports.

like people buy the team and player photos

your kid hitting a HR in the world series. Your kid scoring the winning goal in the world cup - like that

I posted a sample somewhere...the only comments were "what's with the hand". Even when it was a cartoon..."what's with the hand" or "why is the foot like that?"

Im making some of man holding bat now. I'll post some of those. I tried to do a recreating of some WS celebrations after the winning HR. They didn't come out good.

[warm junco]

Ah okay, Holding stuff is hit or miss but yea would like so see some samples

[grave storm]

Guys, What's the best up scaler for Realistic images ?

[fervent thunder]

Hi. Any chance you guys can help with Playground AI? I am trying to use their new inpainting and I am getting very close to results I want but I am not sure if Playground AI uses Stable Diffusion for inpainting or where to find their Discord server if they have any. I would be much thankful for guiding me to proper server or channel on your server.

[shy thicket]

Hey all is there any news on the next updates? 2.2 or whatever they call it?

[supple zephyr]

Hey does anyone know if there's a script/tool for creating an image of any given size using tiles? Like, for example wanting to create an image that's 4096x4096, normally a GPU would melt trying to do so, but if it was done in tiles and smaller batches it could be done and within a reasonable amount of time. Anyone know if a tool like this exists already?

[supple zephyr]

Just use Latent

[supple zephyr]

Why not use a webui & a model of your choice to be sure you're getting exactly what you want?

[fervent thunder]

I am not exactly too experienced in this and Playground AI does offer easy to learn Web UI. They also mainly offer Stable Diffusion 1.5 and 2.0 and DALL-E 2 but DALL-E 2 as paid option

Say, could you possibly point me in direction of where I could get help with that inpainting feature or Playground AI's Discord if they have one? I am getting the results that look really close to what I want. In fact, they are almost perfect, well some of them at least are.

[supple zephyr]

I did some quick research and didn't stumble across a discord. Best place to get help with that kind of stuff would be here anyway.

[supple zephyr]

It's never too late to learn, and besides... having much more control over everything is better in the end; perhaps you'll find a use case scenario not possible with what you're currently using. I've found plenty of things to do and much fun using a stable diffusion webui :)

[fervent thunder]

Thank you. Can you point me into a channel that could help me with both a little bit of prompt crafting and inpainting at once here?

Unless you would be open to try helping out as well with the issue I have. It's fairly simple but yeah, a bit of pain in y'know, regardless

[silent dome]

guys, what is the simplest and best lora training guide you know? I also want to start simply, a lot of ideas)

[slim shadow]

so we can not use the discord channel for creation anymore?

[ornate field]

can any one tell me how can i use commonds

[light crystal]

Doesn't look like it, hopefully, this might mean we are getting something new!!

[fervent thunder]

Hello, I have a lot of issues with my images. I don't know if I can report it there or not ? That's mostly a mistake from my side so no technical issue but need help :/

there is a

tech support channel

but I think mods are kinda chill so just ask lol

but if you put it in tech support its easier it doesnt get lost !

Every models that I use return to me a strange painting art even when prompting realistic, anime or something like that

[flint fiber]

I'm here for help too lol but I can try to help

[fervent thunder]

Can I go DM Jingo ?

where can I ask for help with goal I am trying to achieve? to make it short because that's not the channel for it: I am trying to change color of part of an image with Playground AI, Stable Diffusion inpainting, but I am also needing help with prompt a bit

[flint fiber]

sure if you want but asking here will have more people potentially know the solution

[sleek osprey]

Are there any other companies/orgs making open sourced diffusion/art models?

or is stability.ai the only one

[flint fiber]

can you give example of the model you using?

[fervent thunder]

I can answer some of your questions guys hold on

[fervent thunder]

I am working with a project that includes women, I am using the babe model. It works good but all my images are saturated

But not for only people, it can be enviromnents as well

[fervent thunder]

yes there is loads on huggingface and civitAI, you will need to look up automatic1111 and have a compatible PC with certain amount of gpu ram (vram) to load different models some are small some are larger

It is like increasing the contrast to maximum with a screen

which channel is best suited for inpainting help with Stable Diffusion as hosted on Playground AI?

here is a good starting point if you dont want a stock SD 1.5 / 2x https://civitai.com/models/4384/dreamshaper

[flint fiber]

never tried that one what's the prompt you using?

also you're using negative prompt too right?

[fervent thunder]

either support or under AI generations theres an inpaint-outpaint channel where people doing that stuff are

oh, thank you. I didn't notice it

youre welcome!

[fervent thunder]

I am not using negative, but already tried and same result. I use some of the trigger prompt for the model, it works but always having this saturated thingy

[flint fiber]

does optimized mode lower the workload overall or only the minimal requirement? like would I be able to generate bigger image if I use it? like it say it'll be slower and I'm fine with that

[sleek osprey]

those are finetunes of sd not new models

[fervent thunder]

so most models are trained in 512 x 512 but many work also 512 x 768, what we then do is use AI to upscale the images!

[flint fiber]

have you tried increased the cfg scale?

[flint fiber]

how do you upscale it?

the hires, fix?

[fervent thunder]

yeah they take a base and then add additonal training data, so it is a new model, the newest bases people use are mostly from openai's 1.5 though

[fervent thunder]

Yep I thought that was the issue I even increased to 100 😄 lets show you an example

[fervent thunder]

actually I wouldnt put it to 100, 10 seems to be a sweet spot for me on most models, the lower the number the more creative and better quality tends to output, the higher the more it sticks to what is written precisely

say captain, are you maybe any experienced with SD inpainting? I would be thankful for some help

[fervent thunder]

that is one way yeah! I reccomend the ESGRAN4x setting and then set the scale accordingly to what you want to do, sometime this might eat up too much vram to do it all in one go so often i do 1.25x scale then upscale it with another piece of software afterwards called Topaz AI

[fervent thunder]

Oeps confusing with the steps, I use 15 for the steps but same results 😦

[fervent thunder]

sadly it's literally the only thing I don't really do I just had a play around with it last night

I have moved to that channel already just in case

oh

that's no problem

much thanks regardless!

[fervent thunder]

what sampling method are you using?

[fervent thunder]

youre welcome!

[flint fiber]

ok thank you so much that really helped my issue

[fervent thunder]

if you guys want to see some images I did yesterday go into the favorite-images channel i put a couple in just so you know I'm not talking complete rubbish lol been at this for 6 months now but still a beginner really spend most of my time making cutom mixes and prompts so just starting to do training and embeddings

[fervent thunder]

no problem mate enjoy 😄

[fervent thunder]

do you wanna see out of curiosity what I am dealing with?

I think you might find it interesting at least!

[fervent thunder]

yeah! if it doesnt let you post it go into the chat below "tgeneral - with - images"

I hosted it on the channel you recommended

AHH ok

[flint fiber]

I still wanted to see that example tho maybe go to channel below

[slim shadow]

how do we use gobots? Or we dont?

[warm junco]

To high CFG Scale

[fervent thunder]

I sent you a DM

[flint fiber]

oh didn't noticed

[fervent thunder]

I think some one said earlier that it's not currently working in here I think the old bot was retired

[soft vapor]

Someone share midjourney discord link please

[fervent thunder]

you can find it easily on google but linking to it isnt allowed here, I tried :p

[soft vapor]

DM me. I googled but turned out to be invalid link.

[drowsy crown]

Does anybody have a working tutorial for installing ROCm in ubuntu 22.04

[fervent thunder]

https://askubuntu.com/questions/1429376/how-can-i-install-amd-rocm-5-on-ubuntu-22-04 I use nvidia but looks like some one got it working!

[warm whale]

[fervent thunder]

hey so, it turns out Playground AI does have a Discord server

it just popped up for me on their website asking if I want to join

you want me to share you a link on DMs if you are interested in them?

[drowsy crown]

i have probably seen this before. nothin works. E: Can't find a source to download version '15.0.0.22465.50401-84~20.04' of 'rocm-llvm:amd64'

[fervent thunder]

did you check the repo from the tutorial? it might have an updated version so the filename no longer can be resolved?

https://repo.radeon.com/amdgpu-install/22.20.1/ubuntu/focal/amdgpu-install_22.20.50201-1_all.deb

that's the right link just checked it for you!

[fervent thunder]

alternatively there is this official guide for it https://docs.amd.com/bundle/ROCm-Installation-Guide-v5.3

[drowsy crown]

i will check that out after this one https://askubuntu.com/questions/1429376/how-can-i-install-amd-rocm-5-on-ubuntu-22-04

to think i tried to get this working on windows, tried too hard to get it running off cpu, never got it

[fervent thunder]

Not sure why but the support for nvidia seems to be better atm which is a shame beacuse I like open source software a lot ( this gpu I have was a gift )

[drowsy crown]

there was only support for cuda

i found one fork compostable diffusion iirc that would default to cpu if there was no cuda devices

[fervent thunder]

what interface are you trying to use?

[drowsy crown]

i had tried 2 or 3 SD forks multiple problems with windows and or something else like ROCm for example

[warm tiger]

Hey there. Where's the appropriate channel for asking noob questions if they don't fit with the technical support channel? stuff like "will increasing sampling steps reduce messed up photos?"

for stable diffusion

[waxen pasture]

here is probably ok for that

And no, not necessarily

[drowsy crown]

Holy crap, Setting up rocm-core5.2.1 (5.2.1.50201-79) ...

[fervent thunder]

I can give you that answer here, it totally depends on the sampling method used, for instance DDIM is pretty impressive at lower steps, whereas the nueral network based Heun I run at 120 steps

[drowsy crown]

victory drinks at 8:30am

[fervent thunder]

hell yeah glad youve got some progress mate

[warm tiger]

I'm running a local instance and don't know what DDIM means

[drowsy crown]

you thought it was that easy.. rocm-llvm5.2.1 depends on rocm-core5.1.1; however:
Package rocm-core5.1.1 is not installed.

[fervent thunder]

dependency hell

inst their a command to install all depends

It's been a while since I've had to manually install something like that

try sudo apt-get install -f

I think that -f will install dependends

[warm tiger]

could someone potentially link me to a thing explaining what the various sampling methods are/which one would be best for me to use?

i've been running it on the default Euler a and not tinkered with the others as of yet

[drowsy crown]

no i had to add another sources

[buoyant cargo]

What's the best stable diffusion client available atm?

Haven't updated in a couple months and I figured I would update to the latest n greatest

[fervent thunder]

I'd say it's still automatic1111 theres some new slick UI's but they are a pain to get many models working on

[muted parrot]

upscaling with highres fix always led to these CUDA out of memory exceptions when I tried them (scaling 2048x384 into 7680x1440)

[drowsy crown]

Anybody tried this on HBM graphics card? I got a R9 fury X

it might be a lemon for this 🙂 time will tell

[warm junco]

Yea like i said try lowres gen not 2048 and then upscale

[fervent thunder]

as long as it has enough vram itll still probbaly be faster than CPU

Not sure though on support or anything like that i'm using the nvidia 3060ti 8gb

where can I share my image results? any sort of channel for that?

[waxen pasture]

#🏞｜general-with-images and anything in the "dreamer communities" category.

[silk coral]

Guys where are the dreambot channels?!?!?!?

[muted parrot]

so something like 1024x192 and then upscaling? I'm worried a bit about the quality of the composition at such a wide-angle. Isn't there a way I could do tiles of say 720x720 until I hit 3840x720 and then 2x upscale to 7680x1440?

[buoyant cargo]

what's needed to get many models working?

I read that it's possible to combine models together with different weights

if that wasn't a thing in the past

Is that possible in automatic's?

[fervent thunder]

the way the models are made on huggingface and other popular places are often made using automatic1111 just beacuse it's really popular and theres no guarentee it works on the other UI, I did try one I forgot the name but it was really a pain besides the automatically installed ones to get others workign

[buoyant cargo]

I like that in automatic I just drag+drop into the folder and it works.

But yeah, I guess to my other question, I never used to have an issue running a single model. I didn't know it was capable of blending models. Is that something that requires more work? Or is it just setting certain parameters i.e. the models and weights?

Or, do you have to run it through each model once?

[fervent thunder]

basically in automatic1111 you can go to the checkpoint merger tab and then select model a and model b then make a merge and it's pretty much that simple

[warm junco]

Yea thats something you can try or the SD Upscale in img2img this works with tiles. There is also an extension for big upscaling

[wide juniper]

I have around 15 textual inversions now and I can't tell which ones are trained for SD 2.0 or 1.5. They stop working without error when using the wrong version, right?

[waxen pasture]

if it's for 1.5 and you load a 2.x model, look in the console and see which embeds are loaded and which were skipped. Skipped ones would be for 1.5 in this scenario.

[wide juniper]

Thanks

My SD gets stuck once in a while, usually when I overload it, forcing me to restart SD. Any way to interrupt its task? auto1111's Interrupt doesn't work

[waxen pasture]

Anything other and the iunterrupt button in the ui, not that I can think of.

[buoyant cargo]

got it -- so it makes a new model file/package and I select the combined one?

[fervent thunder]

yep! just head back to the txt2img tab once its done merging and click the refresh symbol and it will be in your list

[drowsy crown]

ROCk module is loaded
Segmentation fault (core dumped)
What now >.<

[buoyant cargo]

You're the real MVP. Made it super easy to understand. Now only if I could figure out how to keep a clean install. Appreciate the help!

[fervent thunder]

all I could find > Most of the method from this thread wont work with newer drivers. AMD changed their driver packaging method recently.

I suggest you try to find older drivers. I have installed amdgpu-pro-20.40-1147286-ubuntu-20.04
Everything newer either will not install or will not work properly, at least on Debian. On PopOS it should be similar.

To be clear. This will give you OpenCL only, so you can forget about using Blender 3.0 and newer if you want GPU rendering.

AMD does not release PRO drivers for PopOS nor for Debian. If you want to render on GPU with Blender 3.0 and newer your best bet is to wait until ROCm is packaged in Debian, so it can finally trickle down to PopOS. When it happens - its hard to tell.
Alternatively you can compile ROCm from source, but it can be complicated.

https://devtalk.blender.org/t/guide-install-amd-opencl-on-pop-os-some-ubuntu-derivates-linux-amdgpu-amdgpu-pro-rocm-rocr/13458/109?page=4 they are talking about opencl but for blender on odlder AMD hardware

There is also this but its a debian based OS https://forum.garudalinux.org/t/trying-to-run-stable-diffusion-with-rx-590/22898

[fervent thunder]

no problem !

new model today i hope (from sd i mean)

[wide juniper]

I just put a burger in my prompt and got this. Why? File G:\Stable-Diffusion\stable-diffusion-webui\extensions\stable-diffusion-webui-wildcards\wildcards\burger.txt not found for the burger wildcard.

[fervent thunder]

did you write burger

?

"__ " word "

Lol discord refuses to let me type it

two undescores before and after something indicates a wildcard

so it tried to look for a wildcard ( text file with lines of things like artist names or styles of art )

[wide juniper]

Nope I didn't

I refreshed my browser tab and it started to work properly again

[fervent thunder]

huh, random glitch I guess :3

[drowsy crown]

I have made a mess of of things with this "echo 'deb [arch=amd64] https..."

where can i edit this list

oh wait, i see it now. yeah this is gonna be a headache

[empty barn]

Which GPU do you have, and did you try the official installation guide @drowsy crown ? Which distro are you on?

[drowsy crown]

okay, i am on 22.04 i am trying the official 5.3 guide

and i am not sure if it works yet

[empty barn]

Use latest unless you have a specific reason not to.

Which GPU?

[drowsy crown]

fiji

R9 fury X

[empty barn]

Oh boy.

[drowsy crown]

yeah lets try it 😈

ohh rocminfo did not shit the bed 😄

[empty barn]

Huh, that's gfx803, same as Polaris. Weird.

[drowsy crown]

hm? explain

[empty barn]

AMD gives codenames to its GPU generations.

[drowsy crown]

yeah fiji threadripper and such

[empty barn]

Polaris (and Fiji, it seems) is gfx803, Vega is gfx900, Vega 2 is gfx906, etc.

gfx8 is not officially supported, but may work.

[drowsy crown]

Agent 2

---

Name: gfx803

[empty barn]

Yep, looks good.

Then just install ROCm version of PyTorch and good luck.

[drowsy crown]

4096bit better be worth something somewhere

ya know i had to shrink windows and add 20GB to kubu cuz i ran out

how much more space am i going to need? anaconda3 is 11.7GB

[nova lily]

New here hi, thought there wasnt a Discord yet as per the reddit

[waxen pasture]

Welcome! It's been around for quite a while

[nova lily]

oh, well i'm new to SD and this Ai art as well.

[waxen pasture]

It's super fun and a great community :)

[nova lily]

Yea confusing and daunting on initial setup but its worth after, kinda of like modding games, like Skyrim

I for sure have some stuff incorrectly configured or small stuff missing, I just know it.

I get great results that I'm happy with so its probably for the most part set well

[waxen pasture]

using a1111 ui?

[empty barn]

The entire setup needs quite a bit of space, yeah.

[wide juniper]

Do I need --nohalf? I'm using 3090

[nova lily]

oh yesir btw

[nova lily]

I watched/followed two tutorials when first installing SD, one didnt mention Anaconda but the other did, I have it installed anyway

[forest apex]

I find anaconda pretty heavy when a python venv will work, some tutorials recommend them though

[drowsy crown]

most tutorials (for 22.04) dont work 🤷‍♂️

and 20 is reaching end of life in about 80 days

[warm tiger]

is there any way to prioritize the text prompts? like if one thing is critically important or more important than others is there a way to mark that?

like if i pit 'pink hair' in it but it keeps making the people blonde?

XD

[nova lily]

STill new, but I found that making a long, cohesive comprehensive at the start is usually leads to best.

[warm tiger]

i sort of noticed that. like most important goes first and seems to do better

i have seen some where you can add ( ) around it or (( )) and the more ( ) the more important the prompt is. or adding a number at the end like pink hair5, and the higher the number the higher the priority

but idk if this one runs that way

[nova lily]

Yea i dont know what () do, Here's what I do for characters

amount (a, 2, goup) - look (beautiful, etc) - species/gender - Outfit (white shirt)

Hair i put inseperate promt, and I put more clothing details in sep prompts. If you really wanted tattos for example or they don't seem to come out you could put in the first main prompt.

The initial main prompt you should be concise and what you desire specifically, but be careful not to confuse ai by add too much @warm tiger

[warm tiger]

what models do you use? mine cant seem to do any more than 1 single person and have issues with anything other than human

[nova lily]

I like to primarily make simple human people, so I discovered/use realistic vision12.

[warm tiger]

or any idea how to use multipe models at once ? someone mentioned you could set % based of each model to run it at but i dont see any options for that

[warm tiger]

is that for RL human art?

[nova lily]

yes real ppl

[warm tiger]

huh.

is that model on the official website? forget the name of it

might DL it and poke around on it for fun

some AI rendered kitties or ducks could potentially be neat

[supple zephyr]

It's fine, I've never used it so I won't find much use in the server. Thank you for the offer though <3

[tender wave]

any one can use multi-subject-render?

[strong brook]

Sorry for the noob question but here goes. The images I am getting are perfect except it keeps adding text over them like a movie title. Is there a way to specify not to add any text but keep the rest the same? Sorry Ive searched but google keeps giving me results.

[fervent thunder]

Hello, I am using the 2.1 model and I am always having drawing style in every of my generate images even when saying "ultra realistic, 8K, etc." and putting "drawing, painting" on the negative prompts ... I have the feeling that my software is broken and does not listen my prompts

[cobalt cliff]

it's over https://twitter.com/eric_wallace_/status/1620449934863642624

[warm tiger]

pffft. means nothing. as long as someone isn't selling the AI art they generate, using it to defame or do anything illegal to someone or harming their ability to make money off their own content it's fair game.

also looked like that person just took 1 image to make a model on that one or something

[nova lily]

I;m still wondering what it exactly does, does itindiviual images using two different groups of prompts?

[tender wave]

yes

join us on diffusers 4

[warm tiger]

is there a way to take 1 image you like that it made and then ask it to remake the image with slight variations a bunch of times as a batch run to then pick your favorite?

[nova lily]

You mean re using the seed

[warm tiger]

i have no idea.

I'm running the text2img with randomized seed

[nova lily]

hit green recylce icon, it will reuse the same attempt, and then you can generate at varioussapmling step levels

[warm tiger]

i basically have 0 clue what i'm doing and probably shouldn't even have been able to get it running on my system

[warm tiger]

how would this work when you're generating images in batches of 60 and it seems to be usig a different seed for each?

[nova lily]

In a single image scenrio what I said applies

But it should from a batch of a seed as well

[warm tiger]

hmm. so is it not possible to have it say. if i liked a specific outfit design it generated or pose it generated i couldn't do something to make it more likely to keep said pose or background or outfit? just the seed thing with +1 or such to it?

like img2img then... something with that?

[nova lily]

Yea its impossible to recreate something again, except just keep generating same prompt setup

[warm tiger]

i see

that makes things a bit difficult but i suppose i'mgood enough with gimp i can correct some stuff if i really like a photo

[warm junco]

you need img2img for that

[warm tiger]

i looked at it and tried a few images but it seems to only be able to do 1 at a time and generates things not related to the original image at all. but i'm probably just not doing it right

[warm junco]

press the button Send to img2img of the image you like, then it will take over all settings

[warm tiger]

oooh yeah see. screwed up fromt he start by uploading it from the folder

[nova lily]

havent started to mess with img2img at all

[warm junco]

it has so much features, even i dont get what all stuff is doing there

[warm tiger]

promising

i'm mostly just poking around and screwing with stuff with very little idea what i'm doing

[ancient mango]

Anyone know a good cpkt for realistic pictures I could train an ai with? Im using analog diffusion

[warm junco]

dreamlike photoreal, or dreamshaper

[warm junco]

thats good, everyone startet this way ^^

[ancient mango]

Ill check them out. Thanks!

[warm tiger]

someone earlier suggested realistic vision12

[warm tiger]

realistic vision12

[nova lily]

yes and Ill try those two others

[warm tiger]

have you guys messed with that "arcane" one?

it seems really cool

[warm junco]

yes, not much but the style is very cool, watched the series last month

[warm tiger]

one of my friends has a lot of artwork of her character she designed and attempted to recreate it with some of the models she found/used through SD. came pretty close on a lot of them.

she apparently trained it as well using art she had of herself.

and said something about being able to use certain percentages of each model to blend an image together

but i've not seen anything like that in the webUI

like 30% Model 1, 60% model 2 and 10% model 3

[warm junco]

maybe she used the checkpoint merger