🏞｜general-with-images

[wispy nest]

[oak osprey]

what are you even doing now?

what's your goal here, son

[wispy nest]

just bored

[oak osprey]

man that's a real shitty thing to do, rile others up out of boredom

that's real anti-social behaviour right there

[wispy nest]

yes its a fact they dont have a brain or pain receptors, but they have a pain response in the sense of what pain is functionally and abstractly

and they are alive most of all

you cant feel pain without a central nervous system, end of discussion again

[steel yarrow]

https://www.sciencetimes.com/articles/24473/20191218/a-group-of-scientists-suggest-that-plants-feel-pain.htm

[wispy nest]

and i can link 1,000 articles rephrasing what i just said

[oak osprey]

i'm pretty sure a group of scientists means anything from kindergarteners on holiday, to, actual scientists, these days

[steel yarrow]

just following the science and correcting some misconceptions, keep an open mind righteous bros

[wispy nest]

only scientists that you agree with are real scientists

yep only peer reviewed studies written and agreed upon

im sure you could link something telling us they put a microchip in the covid vaccine too

gotta keep an open mind bro

another stereotype, one more for the collection

[boreal falcon]

lets come back more to artistic and creative stuff 😄

[oak osprey]

or that horse de-wormer actually fixed joe rogan

[wispy nest]

lol

joe rogan is the real gateway drug, hes the ultimate rube 'just asking questions' who gets fooled by the most basic shit

[steel yarrow]

you can make up all the things I could link to, doesn't mean it's real or a counter

you're just too easy to poke dude

just keep inside the house and everything will be fine

[wispy nest]

feel free to poke away franky, i havent even noitced you

[steel yarrow]

you're not interesting enough, you're too transparent

[wispy nest]

yeah im not a bigot dude

i made these earlier to day, they're 4k, what do you think ?

[oak osprey]

https://giphy.com/gifs/playstation-ps4-playstation-gifs-PhIz9UPp97nf4MetKe

you're afwaid.

[wispy nest]

love the coloring! :D
And angry at not having those muscles myself! >:I

[oak osprey]

franky's over here like "please notice me senpai UwU"

[wispy nest]

god you're so cringe

[wispy nest]

i mean its possible, just work out 3 times a week and eat healthy

[oak osprey]

cringe is actually telling you about yourself, more than the person who made you cringe. it's you remembering your own embarrassment

[wispy nest]

but who does that lol

[wispy nest]

good to know

[wispy nest]

if it was easy, then people wouldn't have issues with it. And the evidence is me, mr lazy :P

human nature strikes again

[oak osprey]

please tell us all about human nature

everything you know

[wispy nest]

people tend to think in boxes and project a lot

and most people think they're correct

I curse nature each day, but it doesn't seem to do much, like complaining to tech support for any company >:(

[oak osprey]

fuck i need to frame that comment

that made me laugh harder than deformed AI photos do

it's so meta

[wispy nest]

also everyone is stuck in their own subjective reality, thinking that shit is real, while 90% of their input is actually just virtual

and the other 10% is incomprehensible

[oak osprey]

♻️

[smoky oak]

Brooooo, my head hurts sooooo baddddd

[wispy nest]

what happened ?

[smoky oak]

No idea, just woke up and it feels like somebody took a mallet to my forehead, uhggggg

[wispy nest]

neck pain too ?

[smoky oak]

No, just headache

So strange...

[boreal falcon]

maybe some nerve

lack of water?

blood pressure?

[wispy nest]

migraines suck

[smoky oak]

Anyways, wanted to pop in and say happy pride month, love you all, and just remember you are all important and deserve the best in life

Stay safe people <3

[smoky oak]

It is probably that, though I have been drinking a lot of water

That's what it feels like ho

*tho

And it's not a migraine, I don't get migraines

[oak osprey]

if you drink too much water without eating salt, you dilute your electrolytes

[boreal falcon]

hmm

[oak osprey]

you could actually die

[wispy nest]

happy pride month dude

[smoky oak]

<3

[wispy nest]

i dont celebrate it but i agree with the sentiment

[smoky oak]

I tend to get headaches like this when I sleep for a very long time, and I'm not sure why exactly. I did sleep for 14 hours though, so that does align with my normal experience

[wispy nest]

it's easy to identify what causes headaches, it can be because 1. everything, 2. nothing, or 3. something in the middle of point 1 and 2.

[oak osprey]

im not surprised after everything you've said today

[wispy nest]

like all AI, there is bias

and self-confirmation bias is one hell of a thing

[oak osprey]

you should have a sleep study done. you might have obstructive sleep apnoea

[smoky oak]

I just hope that one day we won't need a month to feel supported/important

[oak osprey]

restricts your airways while you sleep, it could be central nervous system issue from epilepsy or similar, but mine was physical cause by tonsils being perpetually too big. had surgery to fix it

[wispy nest]

my recommendation, is to talk to qualified people if it's something serious, but random headaches probably aren't any cause of concern :P

[smoky oak]

I have had several sleep studies done, I do have occasional apnea, and buy occasional I mean like once a month where I wake up gasping for air, but that hasn't happened in a long time.

I've been hospitalized in the semi near past for how severe my sleep issues can get, however I have been regulating other chemicals and such which have allowed me to sleep better

[oak osprey]

the sleeping for 14 hours is

unless they took Haldol.

[smoky oak]

Oh I slept for 14 hours because I'm sick

[oak osprey]

haloperidol will knock you out for ~15 hours

[smoky oak]

But also, my body naturally needs at least 10 to 12 hours of sleep now after how severe I screwed over my sleeping when I was younger

[oak osprey]

that could just be apnoea

[wispy nest]

damn. im lucky if i get 6 lol

[oak osprey]

you should DEFINITELY get a sleep study done if you regularly need 10-12 hrs to feel rested.

[smoky oak]

During high school, my sleep issues were so severe I was actually dying. I never really knew how important sleep was for your body to literally not just self-destruct lmao

[oak osprey]

they're hella fun. there's no reason other than the absurd cost there, not to do one

[wispy nest]

14 hours isn't anything special if it only happened once over a long time. Might have been a simple flu, etc. If you sleep for 14 hours a day and aren't working in a extremally straining job, then it's something else though :O

[smoky oak]

Once again, I've had several sleep studies done.

That is just what my body needs to get by now, that's how severe I have screwed over my body when it comes to sleeping. I also have several sleep disorders on top of that that are contributing factors to why I need that much sleep

[oak osprey]

@wispy nest have you tried the mixing diffusion yet?

[smoky oak]

During my senior year of high school, just before my father went into the hospital, I was in the hospital for sleep issues so bad that they were actually threatening my life. It was only by some miraculous connection between doctors that I was able to get the doctor who ended up saving my life

[oak osprey]

just merged about 2 days ago but the research paper came out 5 days ago, i thought this was something old but it's very new. give it 3 prompts and it makes them a panorama and uses inpainting on the seams

but it seems to do that using latents only, all in a single pipeline shot

[wispy nest]

You think I'm smart enough to be up to date with these stuff?
If yes, Thanks! 🥰
If no, yes, you're right! I have no idea what that mixing stuff is :P

[oak osprey]

https://huggingface.co/spaces/albarji/mixture-of-diffusers

[boreal falcon]

sleep disorder that threatened your life?

[oak osprey]

you can try it out here

[boreal falcon]

Fredd Krueger confirmed?

Freddy

[smoky oak]

During my Senior year of high school, my sleep issues were so bad that I actually had to be injected in order to get sleep, because my body was shutting down.

Due to a mix of several extremely severe sleep disorders, my average weekly sleep was in the single digits

[smoky oak]

Yeah, too little sleep can kill you

[oak osprey]

@boreal falcon i realised that freddy kreuger started out appearing in people's dreams as a motivational speaker but none of them listened to him so over 800 years of that, he got tired of it

[boreal falcon]

i have sleep problems too due to panic attacks

[boreal falcon]

lol

[smoky oak]

I had Pandora's box of sleep disorders, and they were all fueling off of each other. The last sleep I got, the more severe the symptoms got, which made the symptoms from those sleep disorders prevent me from getting more sleep

Now I only have a few of those sleep disorders left over, after being able to return to a more healthy sleep cycle

[boreal falcon]

sleep hygiene is a thing

i have to do something about that

[oak osprey]

i ran into a Djinn the other day that lives inside a magic lamp, he promised me 3 wishes. i've been thinking it over carefully, and i think appearing in people's dreams is what i will request as a special power

there's just one problem

[wispy spindle]

happy pride!

[oak osprey]

the lamp is in this old, dusty antique store and the owner is real uppity

sometimes i go back but the antique store is gone 😦

[wispy nest]

I'm also not smart enough to get that locally so while it's nice to test it on their site, I'd rather wait till I can apply it to my local auto1111's webui, thanks for the tip though! :D

[oak osprey]

i doubt you'll get to. it is using Diffusers.

[smoky oak]

To name the ones that I had off of the top of my head, I had:

Excessive daytime sleepiness disorder
Potential narcolepsy precursors
Hereditary insomnia disorder
Elongated circadian rhythm disorder
Nocturnal sleep alignment disorder
Anxiety driven sleep postponement disorder
Excessive sleep paralysis disorder
Acute sleep apnea
As well as a host of other situations that were caused by other health conditions that I have that would cause me to stay up

[oak osprey]

he'll have to pilfer more code

i wish my results looked like this 😭

mine

thing is you have to have prompts that mesh well together

this is using isometric view keywords

[smoky oak]

It got so bad during my senior year, that I was losing the ability to walk, I couldn't formulate fully coherent sentences, and I lost almost all fine motor skills in my hands

It was tragic, genuinely in the stages of my body shutting down on me.

I had a sleep study done on me when I was in the hospital, and when they found that they couldn't help me in the hospital they released me back to the home.

That is when I had a different type of sleep study done on me, where I had to wear too thick wrist bands for 3 weeks, right down all of my chemical intakes, and all sorts of other crap

That is when I was officially diagnosed with all of those disorders in tandem with the first sleep study

[oak osprey]

similar thing to me but they found out i had colon cancer at one of the many doctor's visits and then things changed a lot in pace and content to what i went through for treatments

but apparently childhood cancers are quite common in chrons disease cases, so i'm nothing special there

[smoky oak]

And then out of desperation, they switched me to another doctor who basically threw all of their recommendations out of the window since they weren't helping, he had me try a crap ton of different pills to see which ones affected me and what way, then we honed in on the single one that worked for me, to which he had me take four times the recommended dose per day every day for 3 weeks in order to try and reset my circadian rhythm, and the madman was able to do it

[wispy nest]

those results are probably the best of many tests. I don't often get jealous/envious when people make better stuff than me because I don't know it those results are the norm, or a one time result. But as long as someone enjoy their own stuff, then that should be enough! :D

[oak osprey]

on the model card there, it says it took hours to make that

[smoky oak]

The greatest thing to come out of all of this? The only sleep pills that work for me out of dozens that I was forced to try, are like the cheapest and most widely available sleep pills

Literally diphenhydramine, Benadryl lol

Benadryl saved my life lol

[oak osprey]

that give me restless legs

[wispy nest]

and have you used hours of your time to make one image? :)
might be the answer why your results aren't the same? And if they are, and you don't take hours to make them, then what doesn't that tell you? ;D

[oak osprey]

i took 6 hours today to try and make one image, yeah, and it still sucked ass 😭

[smoky oak]

Benadryl ended up being the only pill that was capable of putting me to sleep, because my sleep issues were not caused from a lack of go to sleep chemicals, they were caused by a huge surplus of stay awake chemicals

And diphenhydramine / Benadryl specifically block those stay awake signals

[oak osprey]

i didn't even work

i was supposed to

everyone, all day long. "i need this", "i need that", and i'm like "shut up, i'm trying to make three oceans match up"

[smoky oak]

But yeah, long story over, that was one of the many times I've been hospitalized for a subset of my health issues.

I really am a Pandora's box of defectiveness lol

[wispy nest]

i'm happy for you that you got your sleep deprivation fixed

[oak osprey]

"it takes 10,000 hours to become an expert" yeah an expert at suckin'

[smoky oak]

Yeah, now the only ones that I still really deal with occasionally are my nocturnal sleep alignment disorder, excessive daytime sleepiness disorder, and my insomnia/sleep apnea very rarely now

[wispy nest]

btw what do you think sytan

cyborg cat

or something

[smoky oak]

But, the byproduct of all of this is that if I don't get 10 hours of sleep everyday, my body rapidly goes back to what it was like previously

My body has become hyper-dependent on a large amount of sleep, otherwise all of my other disorders and oddities start kicking back in at full drive

[wispy nest]

those things are said by the people who regret working for those hours to get somewhere while slowly realizing you don't have to spend X time to be called good at it. You're good at something when someone ask you for help! :D

[smoky oak]

It's funky, I like it :>

[wispy nest]

i wish i found the bad hands lora before i made it

getting the fingers right took so long

[smoky oak]

Also it should be noted, it was around half a year later after this hospitalization incident for my sleep issues that we found that I actually have fibromyalgia, and it is actually the root cause of almost all of my chronic illnesses that I have

But since I've been able to find that out, living with these chronic illnesses is so much easier when you're not just being fed inconclusive results all the time

It also helped immensely because I found out the reason I have fibromyalgia is because my mom has it, we have hereditary fibromyalgia, non-injured related

[oak osprey]

then i guess i'm one of the best fine-tuners cuz i had to shut off my pings for Discord because of people endlessly asking for help

more likely i'm just the loudest

[wispy nest]

I don't know if I got the right one but all my images turned really weird when I tested one of them :/

[oak osprey]

everyone else holds those secrets dearly

[wispy nest]

yeah i saw it has an impact on generation, i think i'll juse use it for inpaint fixing hands

its a 500mb model too

[wispy nest]

see! :D
why would they ask you for stuff if you weren't good at it, good enough for people to want to get as good :D

[smoky oak]

Anyways, my headache is almost completely past now. Whenever I do get them, they stay for about 30 minutes after I wake up, and then they dissipate. No idea what that is supposed to do

[wispy nest]

what's the name of it? might have the poor man's version :P

[oak osprey]

likely asking me due to ignorance, if it were about who is best, they'd be bothering nitrosocke or Thibaud

[wispy nest]

EnvyBetterHands LoCon

[oak osprey]

i went to the Civit discord to ask for advice/guidance and they're not interested in helping others. so it's likely that, too.

[wispy nest]

the best are the ones who helped the person, at least in that persons eyes :)

[oak osprey]

until 5 years later they're hyper-invested into a crappy framework and it's all my fault

[wispy nest]

hm maybe it is somethign with breathing

[wispy nest]

the eyes are the only thing not looking real

[wispy nest]

or your plan all along? 🤔

[oak osprey]

lighting on the forehead looks a bit Unreal Engine to me but it's because i'm looking for it

[wispy nest]

why do you reply to me, you blocked me

[oak osprey]

the thing about skin in AI is that it's not got a lot of subsurface scattering

[wispy nest]

it has 0 subsurface scattering because it does not use ray tracing

[oak osprey]

subsurface scattering, is a real life phenomenon

it's not tied to RT

[wispy nest]

yes i know, and it does not have it because it does not use it

[oak osprey]

there's nothing to "use". it's a feature it hasn't picked up on from images

these models are all weighted probabilities

it doesn't produce images of it because it doesn't occur very frequently in the training data

[smoky oak]

You can get subsurface scattering from AI

[wispy nest]

what i found works is to use waxy skin in negative prompt

[smoky oak]

just a sec

[oak osprey]

i can't say i've ever tried waxy in the negs

lol my result for the prompt subsurface scattering

[wispy nest]

sadly excluding concepts limits the space of possible outputs and you do not want to do that nescessarily if you want to have good looking candles

but it works

[oak osprey]

ah good point

i had to remove artifact so i could make indiana jones stealing one

😦

i was like what the hell

[wispy nest]

you can use synonyms

[smoky oak]

I have gotten it before, specifically through peoples ears

like a bright light from behind gives their inner ear a more warm reddish glow

[wispy nest]

it does have the concept mixed up with glowing

and the skin looks bad because it has the concept mixed up with so many types of images

[smoky oak]

I have also messed with SSS in blender a lot

Wonder if I could use that sort of thing to train an SSS LoRA

[oak osprey]

i'm afraid to prompt translucent skin

someone else do it

[dense tapir]

I bet so

[smoky oak]

this furry model I have has explicit support for things like that, and they look weeirdddd lmao

[wispy nest]

i did it before and it looks a bit like clearcoat

[oak osprey]

are you talking about the red hand i showed?

[wispy nest]

no in general

[smoky oak]

This model can also do iridescent, and crystal skin lol

[wispy nest]

it learned that skin is anything, 3D, photos, drawing

[oak osprey]

ah i was gonna say that's a real pic

[wispy nest]

so if not further specified it will make a mix of all of that

[oak osprey]

well the text encoder's lowest layers contain textures and shapes but the unet itself knows textures too and apparently it'll overfit on them

i don't know what "textures" mean in context of AI imagery, it seems like a pretty wide net to cast

[wispy nest]

probably what it means in general

texture, as in surface

[smoky oak]

I can get SSS by making my characters transparent lol

[wispy nest]

idk how it works in interior but i know how it behaves in practice

to get better skin it also helps adding 'smooth skin' to negative prompt

and skin impurities, skin pores, other skin features only seen in pictures to posi

[smoky oak]

this model can do weird like... translucent plastic bodies lol

[wispy nest]

bloonie ptsd kicking in rn

[smoky oak]

lmfaoooo

[oak osprey]

everything is the surface

it's not a 3d object so

[wispy nest]

thats not what i mean

[oak osprey]

to me it's harder to define the difference between the two

i know what you mean, it's not a big mystery. the mystery, is the unet. because it overfits on a lot more than just the skin type of a person

[wispy nest]

idk how it works, i'm not a data scientist. i'm an artist

[oak osprey]

i haven't tried burning a unet in very many ways, the only time i ever did it was with photos of a single subject

well i mean it's eazy to crank the LR up on a training session and kill a model in about 500 steps

[dense tapir]

How about now?

[oak osprey]

you can learn a lot

[oak osprey]

whut the hell is it lmao

is it just me or does the tight focus kind of feel unreal

[smoky oak]

also foudn this model is trained on 100's of celebs and fictional characters

[oak osprey]

i'm going to just put one image of every celeb into my next training run so it can make bizarro versions of them more easily

[wispy nest]

do all generations have fake teeth and 2cm makeup?

[smoky oak]

like, fine trained on them, not just the base data on them

[oak osprey]

technically yeah all generations have fake everything

[wispy nest]

[smoky oak]

thsi translucent skin prompting is interesting lol

[oak osprey]

i went back to basics and removed all my negs and started from scratch, i came up with like, duplicate teeth was the first one i added. dear lord. oh, and red skin because why the hell does it want to make people look like demons

[oak osprey]

quick, end it before it assumes its final form and assimilates

[wispy nest]

i think that the language we have might not be the best way to handle creation of skin

[smoky oak]

hopefully we won't have to, for much longer

[oak osprey]

@smoky oak tried to make family portraits and forgot i'd removed all my negatives

[wispy nest]

since one word that describes skin, also describes the texture and look of many other things

[smoky oak]

we really need an AI layer that translates what we want into what the generators actually know

[wispy nest]

in general language we have is very interwoven and with many meanings

[oak osprey]

that likely won't be done at 40 it/sec

😛

BLIP2 uses a 175B parameter language model as its translator and was heavily trained by Salesforce and all i hear about it is how much it sucks

[smoky oak]

I don't like how hes looking at me lol

[wispy nest]

that smugness

like he's just waiting for you to say something

[oak osprey]

he's smug because he's waiting for the perfect moment to drop his hot line:

"You jelly?"

[smoky oak]

this model has fantastic control over expressions and poses, so I can do a lot with him lol

[smoky oak]

SHUT UP LMAO

[wispy nest]

for poses i just use controlnet

[oak osprey]

jesus wept, what the hell where's the fingies

[smoky oak]

oh watch this

this model has 45 predetermines materials you can make basically anything out of

its super trippy

[wispy nest]

[oak osprey]

you do that by training the early layers of the model on images that show just the texture with its captions

you likely also have to provide images of basic shapes eg. kitchens and stuff lmao

[wispy nest]

i think that our models are very rough and limited and that they did not fully understood the concepts we tried teaching them

[smoky oak]

the model I am using is wayyy ahead of any others I have ever used in terms of giving what I ask for

[oak osprey]

the models weren't trained to full convergence, though. they were trained sequentially and only froze a layer as the loss sort of reduced to a certain level

you can retrain it using something like SmartFRZ which is an attention-based predictor that looks at the current performance of the predictions in the latent space and then decides how to train the thing, adaptively, rather than sequentually from beginning to end

[smoky oak]

iridescent skin

[wispy nest]

what about the training data, its all 2D images

what if we used volumetric data and sterescopic images

[oak osprey]

they found that in a 40 layer model, the 30th layer converged faster than the 20th layer. i tried to reproduce this yesterday and went so far as to find the original author and try to reach out to him, but no reply yet. i don't know how they put together their training data or how they trained their predictor

[wispy nest]

then just use some kind of learned orthographic projection to get the 2D image

because what i see is that it thinks like humans when trying to put something on a paper

[oak osprey]

that's pretty much how controlnet depth works

they trained it on pairs of depth maps with images

i think they use OpenCV to generate the depth map from an image. and OpenCV is pretty incredible, honestly, how much it does without any weights

[wispy nest]

because when i look at the results they all look like images from someone who has no understanding of 3D made them

[oak osprey]

what do you mean by that

[smoky oak]

this model is insane, almost any decently known animal you can think of, this model can do

[wispy nest]

like i learned drawing because of SD

[oak osprey]

capybara

[wispy nest]

and i noticed that i had to think how things look 2D

[smoky oak]

easy

[oak osprey]

easy he says

even nature can't figure out what a capybara looks like

[wispy nest]

basically the model has no understanding of 3D

or only a very limited form of it

or maybe it is because the training data contains so many images from people who dont

becuse humans put stuff on paper and convert their 3D imagination to 2D image, and it comes out all weird

so the training data reflects that

[oak osprey]

sorry man you basically repeated the thing you said earlier, to explain it. i still do not get it lol what is weird about the results

[wispy nest]

its just an idea, what if you make a model that has sterescopy images, two times the same text with a pair of images rendered in 3D

or one text +2x images

just render it in 3D, simulating eyes

[oak osprey]

that works already, Sytan can do it

[wispy nest]

and then finetune that model to other images

[oak osprey]

well the dual image thing

[wispy nest]

maybe it will get rid of all the weird perspective artifacts

[wispy nest]

thats a hyena

[oak osprey]

that's a fawkin horse

or that

[smoky oak]

It would look better if I used the capybara LoRA

[oak osprey]

that'd be cheating

[smoky oak]

they do provide hundreds of additional LoRA's with their model

[wispy nest]

the technology itself, is it possible to do this reverse-diffusion process in 3D ?

it should be right

[oak osprey]

why not combine textual inversions with LoRA and a hypernetwork on top of 7 controlnets with the base model

AND THEN you shall have your "Capybara"

[smoky oak]

I mean, it is taking some artistic liberties for facial proportions, as they do for all species lol

I have a LoRA I made on my own that makes photo realistic capybara. Pretty proud of that one

[smoky oak]

now THAT looks bad lol

[oak osprey]

babybara

wat, i think you've been staring at anime dongs for too long 😛 that looks incredible

it might not be, like, proportional

[smoky oak]

That looks like a groundhog, not a capybara lmao

[oak osprey]

at least it didn't give him human teeth this time

[smoky oak]

Thats my Capybara LoRA

thats what capybara actually look like lol

[smoky oak]

not this

[oak osprey]

omg they're the same animal but mine is straight-on shot

😦

and smol legs

[smoky oak]

no, not even close

they have a way different head shape

[oak osprey]

mine's cute

[smoky oak]

Their noses are much wider, mouths are much lower set, and overall head shapes are much blockier

They don't have cheeks, and have a much longer body down to where their legs are

What you sent is a groundhog lol

[oak osprey]

its just a baby

they all look like that when smol

[oak osprey]

that's a beaver lmao

[smoky oak]

no, thats a capybara

[oak osprey]

wait, are beavers groundhogs?

[wispy nest]

when it comes to the ai understanding depth in the image, I've noticed it can be tricked into painting about 3 levels of focus. But that's just the general understanding I have, and it can also be changed sometimes depending on X, etc :P

[oak osprey]

omfg what the hell is a prairie dog then

[smoky oak]

you're trolling, right?

[oak osprey]

no

sometimes two things are the same. have you ever investigated what a Yam is?

good luck with that one

we have these things called Elk here in Canada which are called something else in Europe, and Elk down in the US are yet another different thing. we've got Caribou as well, which are known as Elk somewhere else.

wouldn't be surprised if capybara translates to groundhog

do they have teeth like that?

[smoky oak]

capybara, beaver, grundhog, prairie dog

They are all 4 different creatures lmao

[oak osprey]

and they're not muskrats, either?

[smoky oak]

They don't even inhabit the same regions

[oak osprey]

so they could have been the same thing at one point then. if they started out at one spot

[smoky oak]

muskrat

[oak osprey]

oh that's totally different, damn

like an opossum

which is, not the same thing in europe vs usa

[smoky oak]

not even close to an opossum, you're trilling, right? lmao

[oak osprey]

i'm saying they have names that are used different regionally

[smoky oak]

thats cause they aren't the same word lol

Possum and Opossum are not the same word, and are not the same creature lmao

Opossum left, Possum right

[oak osprey]

you are fighting with merriam webster on that one

i'm with you, they're two different things

but the whole world is fucked up on animal names

[smoky oak]

They are spelled differently, even if they are pronounced the same

[oak osprey]

that's my point

[smoky oak]

all that AI cares about is spelling, which is why it can tell the difference between a Possum and an Opossum lol

I think Opossums are super cute, personally

[oak osprey]

The word "elk" originally referred to the European variety of the moose, Alces alces, but was transferred to Cervus canadensis by North American colonists

like if i went to Europe. and someone told me, "don't go over there. there's Elk in there." i would somewhat laugh and thank them, and then disregard their advice and go into the woods

but they're going to be referring to a fucking Moose. a huge, vehicle-size piece of danger animal

i just hope i never discover that Hippo are called Puppies in some land i visit, i will absolutely be eaten

[smoky oak]

Yeah, but thats then when you use scientific names

The furry model I use has support for colloquial and scientific naming, which is supper cool IMO

[wispy nest]

Id not wanna fuck with an American Elk either, they get huge

[oak osprey]

i have video of me trail building and scaring off some roosevelt elk with a chainsaw

[dense tapir]

Possum becomes Opossum when you see it and "Opossum done got kilt by a car."

[oak osprey]

but a moose wouldn't be scared of that

[smoky oak]

Like to get a bat, you can use pteropodid to not get confused with the object; bat

[oak osprey]

a baseball pteropodid

[smoky oak]

lmfao

[smoky oak]

I am now gonna generate a bat playing baseball lmao

uno moment lol

[oak osprey]

that went from 0 to 60

[oak osprey]

i dont remember bats looking like that

[wispy nest]

that guys def got rabies

[oak osprey]

is it the with human teeth throwing it off?

[dense tapir]

Well, I am not gonna tell him

Here we go

[oak osprey]

my therapist says the only way i can unsee these is by sharing them

feels better already

[smoky oak]

@oak ospreyBaseball pteropodid

it has been done lol

[oak osprey]

that's A KANGAROO

[smoky oak]

I am convinced this dude has no idea what animals are what lmao

[oak osprey]

weren't we trying to make a bat

🦇

not a 🦘

[smoky oak]

yeah, it is a bat lmao

[oak osprey]

look, it even resembles its emoji namesake

Kangaroo

[oak osprey]

ever seen Roger?

fuckin 'roos'll gut ya

with that insane foot of theirs

roger was so handsome

[oak osprey]

wild to watch his handlers try and like, pet him/scratch him, and him try and unalive them

[smoky oak]

Kangaroo

[oak osprey]

those ears are more jackrabbit

[smoky oak]

and now hes gonna say its a rabbit

yup, knew it

[wispy nest]

just be happy the ai isn't in the Finnish language

[smoky oak]

I'm done here lmao

[dense tapir]

Well

[oak osprey]

look his small Rogers' are

hmm maybe donkey ears on yours actually

[dense tapir]

Bear is having some me time I think

[oak osprey]

make someone ride him

[smoky oak]

This model could generate that exactly, but I would get banned lol

[dense tapir]

I know

[smoky oak]

lmao

[oak osprey]

[north monolith]

Hey All,

I'm brand new here, and as of about 2 weeks ago, also brand new to AI generated art as well 🙂

I received an early access invite to Leonardo AI recently, and basically starting generating images, and learning as much as I could, prompt engineering, how Stable Diffusion actually works, checkpoint models, safetensors, Lora's, Textual Inversion, Hypernetworks and the list goes on.

I then set about installing Stable Diffusion onto my new PC, latest Intel core i9 CPU, 32GB RAM and a 4080Ti, so I knew my PC was up to the job lol. I opted to install Automatic111, and have been generating so many images, and just learning how the sampler, prompting, steps, CFG, denoising etc can affect the output image.

I have mostly been using the RPGv4 model, as it's one that I was using in Leonardi AI. I have a question though. I created an image in Leonardo AI, and for the life of me I can't recreate it in Stable Diffusion?

Everything is the same:

Sampler: Euler Ancentral i.e. Euler A
Steps: 30
CFG: 7

Positive and negative prompts are the same, I'm using the same seed, as I grabbed it from Leonardo AI
RPGv4 resolution according to the models PDF is 512 x 768, so that is what I have set in SD

The images just come out different and I don't get it???

One thing of note is that Leonardo AI sets itself up with a generation resolution of 640 x 832? How can that be? As that's above the training resolution, and would likely result in bad generation e.g. double images, limb issues, etc

Any idea how I might be able to get SD to replicate what I created in Leonardo AI. Theoretically I should be able to do so

In the mean time, here's a couple of images that I have created in Leonardo AI, when trying to create a pirate girl, scrubbing the decks, on a 17th century galleon, in a storm. Came out looking ok, however location isn't right, as I wanted to be long range, and see her on the galleon itself. (WIP)

Leonardo AI has a new option called Alchemy AI, that produces drastically different pictures using the same settings!! No idea what they've done, but it's a huge difference, and one that I prefer. Not sure if behind the scenes they;re just using a different model / sampler, but it's quite the difference

I'd love to get that look in the last 2 pics in SD. Leonardo AI uses SD, so no reason I can't

[cyan snow]

it's the model, you said you got early access to a specific AI, means that the model is probably closed sorced(my assumption)

[oak osprey]

reading

[wispy nest]

"just another model" isn't really a "just a" thing ;P

[cyan snow]

yeah the model is the most important part

[oak osprey]

as I wanted to be long range, and see her on the galleon itself

from a distance
scene
cinematic

could help

[north monolith]

I did have cinematic and long range in my prompt, but it didn't help 😦

[north monolith]

There are so many models around. I've got a bunch installed already. I'll have to try and find one that gets a look close to what was in those last two pics I posted

[wispy nest]

probably not the same thing, but using free stuff awesome people have uploaded can do a lot, you just have to fist find it :P

[cyan snow]

you can also merge models, that is what most people do

[dense tapir]

I never had much success merging

[north monolith]

Yeah I've seen a few I've downloaded off of civitai are merged ones

[dense tapir]

always tears the face up

[sterile steeple]

is there a name for fire clothes , i tried fire costume and burning clothes , both didn't work , is there a solution ?

[cyan snow]

you can even merge merged models, it's all just a bunch of math

[north monolith]

Yeah model weights in the neural network or something...

[cyan snow]

well, that is if the models have the same architecture, if they don't then you have to train a model

[north monolith]

One thing I'd appreciate some advice on is the setting of the "Sampling steps". I'm mostly using the default of 20. I know what it does and the higher you set it the more passes of the AI at improving the image to match the prompt, so refining it. Well it seems that way to me, and is what I got out of reading about it.

What kind of step setting is best to use, or do most people just "play around" and see what works?

[cyan snow]

there is a limit, you can over-cook images.

i'd say 50 tops

also, id say the sampler is just different techniques. the smplers that have _a in their name are adaptive, meaning they are suitable for pretty much everything

[north monolith]

If the model used by Leonardi AI, and the one that produced the last two pics posted is closed source, won't that model be based off of the base SD 1.5 or SD 2 models? If so, then I would think they'd have to open source their model?

I can't believe that Leonardo AI have trained an entire model from scratch!! Not given how long and how much it cost SD to make theirs

[wispy nest]

why would they have to open source the model?

[cyan snow]

not from scratch, ALL text2image models use the diffusion architecture, they all start training from the same point.

[north monolith]

License's? If a model is trained from SD 1.5 base, does that then not mean the trained model that is a result also has to be open source like SD 1.5 is

I might be wrong..

[wispy nest]

have you read that in a license agreement?

[cyan snow]

no, that is not how it works, in that sense then MJ would have to go open-source. and here we are

[north monolith]

Is MJ based on SD?

[cyan snow]

very likely, they modified the architecture and trained it in a specific way.

[wispy nest]

I think MJ has their own thing, didn't they have their own stuff before SD?

[cyan snow]

nope :/

[north monolith]

Ah ok. I thought MJ was a completely proprietary model from scratch, same with DALL-E

In that case, I get why people can create new trained models from the SD base models, and keep the closed, even if people are paying for access. Which is what Leonardo AI does

Another question if I can..

One thing I'm struggling with is AI upscaling. I have an RPGv4 512 x 768 image, and I can upscale it by two times, and do that again, and it looks ok, but artefacts start to creep in. If I do a 3rd pass, it looks bad.

I've got a few additional UI scalers installed into Automatic1111, and I've read up on what to do, and do multiple passes etc, but with photo realistic images I just can't get past 2 iterations

[wispy nest]

I checked mr google, but no one said it used any SD thingies. But that was from searching and reading for as long as it took me to answer your post :P

[cyan snow]

bruv you can litteraly see it going through the diffusion procces. it IS a diffusion model, just a well trained one.

[north monolith]

Well here we are...same as Leonardo AI does

[wispy nest]

isn't stable diffusion as diffusion math thingie from "stable?" meaning that both MJ and SD uses it but not that MJ uses SD's diffusion?

[dense tapir]

Leonardo AI I never heard of then went to it a couple of weeks back. It was too restrictive for what I do, and felt inferior to me. Almost felt like a con job.

[cyan snow]

that's debatable. they will never confirm anything or deny anything about their model

[oak osprey]

depends on the upscala

[north monolith]

Works really well, for the 2 weeks I've been using it 🙂 What kinds of things won't Leonardo let you do? I guess they are trying to make it incredibly user friendly, so limited feature set, and less controls

[oak osprey]

i blame the fuckin people trying to sue them for that tbh

[wispy nest]

the thing I mostly asked was if MJ uses the same model as SD trained, not if MJ used a diffusion calculation thingie :P

[oak osprey]

maybe once they have legal protections for their existence they'll be more open, idk

[dense tapir]

Can't use my Lycoris/Lora/TI Embeddings/Hypernetworks just to name the first thing

[north monolith]

I've tried many. remarci, nmkd-siax, esgran-4X and the other built in ones. No joy past 2 iterations, image just get's worse

[oak osprey]

what they've done is really cool, even if i can't run it locally

[oak osprey]

MJ doesn't need any rinky-dink SD LoRAs

[north monolith]

Ah yes, that is true lol. No use of your own stuff, just what they give you. They do let you train with images, but that's it

[cyan snow]

unknown, just wait till SDXL goes open-source. i bet sh*t will go mayham

[dense tapir]

Yep, that is far too restrictive for me.

Reason I do not like MJ too

[north monolith]

Which is why I installed SD about 5 days ago 🙂

[wispy nest]

I actually don't like MJ's images all that much, they are too smudgy and "thick" with color lines, or how to explain it. As well as their "centering" is not that good looking for me :P

[oak osprey]

when i trained my model on 22k MJ5.1 images i was worried it'd adopt that, but it did not

[dense tapir]

That chin, lol

[north monolith]

Ooohhh I didn't know about SDXL https://stablediffusionxl.com/

I've been reading tons on SD, and everything around it, but never came across that one.

Nice!

[oak osprey]

if you surround a url with < >'s it will not preview

SDXL is really cool but i wish there were an outtakes channel on here where we could concentrate all of the deformed shit it makes

so i could make a training set and make a completely deformed version

[wispy nest]

that grill is killing me! :P

[cyan snow]

it's doing the same process MJ does while training. the giant advantage is that it will go open-source(i hope)

[oak osprey]

i have finally succeeded at making a bat @smoky oak

[north monolith]

I've been using SD 1.5, any reccomendation to use 2.1? I hear bad things about 2.1, but as I've not used it I've no idea, so appreciate any advice

[oak osprey]

Fluffy vampire bat++ in pink pyjamas++ holding ice cream--

i love 2.1

[north monolith]

I didn't know about the ++ and -- prompt syntax? I've been using word:1.2 etc

[wispy nest]

and I'll let people who like 2.1 talk about 2.1 because it wouldn't be fair otherwise :)

[cyan snow]

let's just all agree that SDXL takes the cake, there is no way after it releases that people would still be training SD1.5

[smoky oak]

I personally wouldn't recommend it, unless you are willing to give up coherency and consistent quality for more randomness and goofing around

2.1 is good for doing random and wacky things, but it is just not good for the level of specific quality or consistency you can get out of 1.5

Although, I have recently seen a 2.1 model that may be promising towards a future where 2.1 is more of a tool unless of a plaything

[wispy nest]

I don't think I've seen a image from that thing so I can't comment on that :P
But I won't say one model is better, I just want people to have as many tools as possible to make what they really want :D

[oak osprey]

why don't it make him a human

😭

take Sytan's suggestions with a grain of salt lol they use a furry model

[smoky oak]

I mean I'd say it's a pretty general consensus that if you want something specific, use 1.5 and you'll get better results, but if you want to screw around and mess with weird results, then use 2.1, as it's less prone to consistency and repetition

[north monolith]

Sounds like they've gone backwards with 2.1

[smoky oak]

In a sense they have, but mostly because they never gave out good information on fine tuning 2.1, so the models available for 2.1 are really bad compared to the fine tunes for 1.5

[smoky oak]

However, that new artius model I was shown isn't too bad, it lessons a lot of the glaring issues with 2.1, and I think it's actually kind of fun to mess with

[cyan snow]

they said one of the reasons SDXL is taking ages is because they make in a way that can be easily fine-tuned

[smoky oak]

Just keep in mind, if you're looking for a specific thing, like realism, or specific styles of digital art, or anthro character creation, a really anything that fits into specific model subset, you'll have a better time with 1.5

But if you want to screw around and have fun and just get weird and wacky results, then I would suggest 2.1, it has a lot more creative liberties

[wispy nest]

and this isn't even the real batmobile >:(

[smoky oak]

2.1 is basically 1.5 rebellious brother who's always on acid lmao

[cyan snow]

and just the base SDXL model even currently(it didn't even finish training) is better than most of the finetunes that currently exist

[north monolith]

I also like that it can deal with adding text

[oak osprey]

no, OpenCLIP has higher FID scores and accuracy, better prompt comprehension

[smoky oak]

Oh yeah hands down, I've been following the development, 50% trained SDXL makes even some of the best fine tunes for 1.5 look like jokes

[oak osprey]

i've fine-tuned 2.1

Sytan has not 😛

[smoky oak]

2.1 has potential to be better than 1.5, but it's not yet

There are no good fine tunes out there in my opinion, but I would say that the best one I have used is definitely artius

[oak osprey]

it's already better. its baseline resolution is 768x768

[smoky oak]

That one felt the least artifacty out of all of the ones that I've used

[oak osprey]

does much better at resolutions like 1152x768 native gen

[smoky oak]

I mean it is, but I also have this anthro model, which hasn't even higher base resolution

And it's 1.5

[oak osprey]

@north monolith this is 2.1

[smoky oak]

Also, the fact that 2.1 base resolution is 768x768 is a drawback in my opinion

Takes longer to get large amounts of images out to select the one you want to fine tune, and on top of that, it's not consistently 768x768 so you can expect a lot of the deformities that 2.1 is known for because of that

[oak osprey]

requested a cross-animal hybrid of a gorilla, lion, and tiger

can't get that kind of prompt comprehension in 1.5

[smoky oak]

All it did was give it a couple extra arms and make it to you lol

Like yeah, 2.1 can comprehend things much better, but that's not a good result by any mean

[oak osprey]

that's subjective as hell and that's why i say to take Sytan's opinions with a grain of salt

[smoky oak]

I'm sure there are much better examples out there

[oak osprey]

i love that picture

[cyan snow]

yeah, id say SDXL will blow up

[smoky oak]

I don't know, I personally don't think that giving a tiger a few extra limbs and making it green has anything to do with mixing it with a lion or a gorilla

[cyan snow]

i hope they won't disappoint

[oak osprey]

Snoop Dogg and Lex Luthor starting a cannabis company

2.1 has compatibility with a new noise schedule called terminal SNR right out of the box, that 1.5 will take a lot of fine-tuning to get

[smoky oak]

SDXL is all ready so much better than even some of the best fine tunes for 1.5, that I am extremely confident that when it comes out, it will blow up

It has prompt comprehension that makes 2.1 look like a joke, and consistency and reliability that makes 1.5 look like a joke lol

[smoky oak]

That is a true benefit, I will admit that, but I just use offset noise on the rare occasion when I want a really bright early dark image, works for me

Now the fact that SDXL has it built in, that's dope

[oak osprey]

2.1 is hella consistent and reliable. my friend who just shows up once in a while to generate images, will type some truly stupid non-prompt that would get him totally broken results on 1.5 (which he still loved, because he's weird) but in 2.1 it just works

[cyan snow]

when will they release it? they are on ~50% for almost a mounth now

[oak osprey]

the one thing it can't do for him is the celeb stuff. i want to fix that, but i don't have a good dataset

[smoky oak]

I've agreed with you with a lot of your points up until now, but consistency and reliability is probably the worst thing about 2.1 IMO

[oak osprey]

my problem with 2.1 is that the baseline res being so high makes it hard to find good training sets. everything is curated for 512x

[smoky oak]

Looking at how long it took them to get from 0 to 25 to 50, it's looking like potentially another two to three months in my opinion

But the fact that it looks as insanely good as it does I only 50%, that just gives me massive hopes for it's final release

[oak osprey]

i mean, maybe that's an a1111 thing. because everything my bot makes is ❤️‍🔥 it's hard to decide what to love, when you love all of it

[cyan snow]

yeah, i don't think it will get overcooked

[oak osprey]

oh but before the SNR fix was trained into the model, it's less reliable. like my v1 model has some issues that the v2 resolved there

[smoky oak]

The truth of the matter is, if there were any good 2.1 models out there that actually seemed useful and not more of a pain in the butt for me, I would hands down switch over to using them, but they're just aren't any, at least that I have seen

[oak osprey]

it's crazy, man. i was talking about it with a coworker and the brightness avg being .5 was so restrictive that it overly-generalised the image it produced into a mediocre result

when you apply the tsnr fix, it can average the pixels of an image from 0 to 1, which means you can make anywhere between a pure black or a pure white image

[smoky oak]

Yeah, the averaging also really affects people with fair skin and people with really dark skin as well

[oak osprey]

that changes the reliability of the model in an insanely positive way

[north monolith]

Sorry another question. Using xformer with Automatic1111

I was using it (as I have a 4080Ti), by specifying it in the cmd line params, but then read earlier today that I don't need it, as the latest version of Pytotch+cu118 (that comes with latest Automatic1111) already uses Cude.

I checked my GPU usage, and that does seem to be the case, though unless it's a placebo effect, it feels just a bit slower. Maybe 1s. Not that that will make much difference in the grand scheme of things

[oak osprey]

if you're on pytorch2 you probably don't want xformers but you'll have to experiment

[north monolith]

Pytorch 2.01

[smoky oak]

I stay away from any installs that use pytorch 2, as I don't think it has any benefit, and it definitely has a lot of drawbacks

[oak osprey]

this shit changes daily but for a while the recommendation was to uninstall xformers because it would conflict with the new pytorch stuff but, they've released a new xformers since then and it now seems to tell you it's ignoring it rather than conflicting

[smoky oak]

of curse, eventually I will be forced to switch

but by then I hope that its less of a problem

[oak osprey]

pytorch 2.0.1 is superior to 2.0.0 because the pipeline compiling works so much better but from the arguments on the issue tracker it doesn't look like sdwebui will ever use that feature

if they converted over to diffusers (i know, i know) it would Just Work

so when i compile a model on my A100 i get like 80 it/sec, i don't even get a progress bar

the slowest part is the saving and uploading of the image

[smoky oak]

yeah, 2.0.1 is better than 2.0.0, but it is still worse than 1.3 for me

[oak osprey]

as a developer, i don't want to support 1.13. that shit's broken

so i can understand wanting to force you

[north monolith]

80!!! I'm on a 4080Ti, and looking at my log window, I see around 8-10 it/s

[oak osprey]

an A100 is a $25,000 GPU

don't feel bad

[north monolith]

Bloody hell!!!

That's some serious kit

[oak osprey]

the 4090 is about as fast as the A100, if not faster

[wispy nest]

as a dev, I'm obliged to show everyone what coding really is all about:
https://www.youtube.com/watch?v=Uo3cL4nrGOk
mandatory watch for all devs! >:)

[north monolith]

So my poor 4080Ti seems a lot slower

[oak osprey]

well it's all a CPU bottleneck at a certain point, honestly

look at your single thread use

[north monolith]

GPU was about 95%, CPU overall was 25%, but those cores were being hit hard

[oak osprey]

the 25% is of all cores

right?

[north monolith]

Yes

I have 32 cores

[oak osprey]

yeah so it's peggin' at least one core then

[dense tapir]

Pytorch 2 I can't use or I spend weeks doing a simple gen

[north monolith]

yep

I can live with generations taking 4 - 5s

Upscaling is where it takes a little longer

[dense tapir]

Ahem, most of my gens are already 2-6s/it

upscaling I do not do at all as it takes 10+ minutes

[north monolith]

Yeah I'm only doing that on images I like, and even then just to play around and learn as a noob

[smoky oak]

I still do not quite know why they are so slow for you. Like that is wayyy slower than it should be

[dense tapir]

Way faster than it was and I asked other 1060s I am right where they are

[oak osprey]

all of the despair: as a beach-ball

[smoky oak]

it just doesn't make sense to me, as 1080ti's are still several times faster

[dense tapir]

WHen 1.4 came out I was Euler_A 10 steps was 50 steps 5 mins, 100 steps was 10 mins, 150 steps was 15+ mins

[north monolith]

My 4080Ti I'm only getting 5 -6

[dense tapir]

1060 has emulated fp16

[north monolith]

Can't believe a 1060 or 1080 is the same

[wispy nest]

using auto1111's webui and having the arguments --precision full --no-half halves the it/s in my experience. So there's a lot that changes the speed

[smoky oak]

4080ti? That GPU doesn't exist

[dense tapir]

I don't have the ram for that

[north monolith]

Thankfully I don't have those set

[oak osprey]

use bfloat16

[north monolith]

Sorry 4080

[smoky oak]

do you mean 3080ti?

ah, 4080

[smoky oak]

what res?

[dense tapir]

Wait a min

[wispy nest]

do you use --xformers?

[north monolith]

Currently generating at 512 x 768

[oak osprey]

bfloat16 is mixed-precision, it'll keep the text encoder at float32 so that it does not suffer numeric instability and it keeps the unet at float16, to simplify

[smoky oak]

ok, something is wrong there

[dense tapir]

that 4080 gens at the speed the 7900XT trains loras 768x768 at.

[smoky oak]

yeah, that 4080 is way slow for some reason

let me check my GPU

[oak osprey]

it's probably the sampler.

[dense tapir]

7900XT is unsupported still so surprised it works

[north monolith]

No, as I found out today, Pytorch 2.01 + cu118 will use cuda anyway, so read to turn off xformers

[smoky oak]

yeah, it get 9it/s at 512x768 😅

sorry, 10.5

[dense tapir]

once rocm supports 7k cards 7900xt should be 12-18it/s for training

[north monolith]

I'll turn xformers back on

[smoky oak]

that seems reallyyy slow for a 4080

[dense tapir]

yes

[smoky oak]

something is def off here

[dense tapir]

extremely

[smoky oak]

thats slow for my 3060ti too

[wispy nest]

I hope nothing bad happens then because I never read documentation before it's already too late, cross my fingers my card won't explode or something then ;P

[dense tapir]

I can understand the 7900xt speed due to no support but the 4080 has full on support

[smoky oak]

yeah, the 4080 should be doing like 18it/s minimum at 512x768

[dense tapir]

yes

[smoky oak]

I never realized how fun it is to type minimum lol

up down up down up down lol

[north monolith]

With xformers on:

[wispy nest]

I'm "only" at 11it/s without using those precision stuff anyway :P

[dense tapir]

768x768 is about 15-17it/s

[smoky oak]

that still seems way slow

[dense tapir]

something is way wrong

[smoky oak]

very very weird

[smoky oak]

@north monolithdo me a test please, try 512x512 with batch size 8

[oak osprey]

it's just your CPU bottlenecking it, bro

32 core CPUs aren't good at single thread

[north monolith]

What model. I'm currently using RPGv4

[smoky oak]

what CPU specifically?

[smoky oak]

any

512x512 at BS 8

[north monolith]

I've almost top end core i9 too 😦

[oak osprey]

any 32 core CPU is going to be like this

[smoky oak]

thats false

[oak osprey]

i know because my A100 is hampered by one too

[smoky oak]

which one specifically?

[kind quartz]

@north monolith
--opt-sdp-attention
Have you tried this instead of xformers? It is only for 3000 and 4000 gpu cards.

[oak osprey]

he already showed that his CPU is bottlenecking

and 8 it/sec is fine, quit making him think he needs to scratch until it bleeds so that he gets another 3 it/sec

[smoky oak]

dude, thats less than 1/2 what that GPU can do lmao

my power limited

[oak osprey]

nvidia song begins playing

[north monolith]

Well that's worse in the stats. I get 8 images + the grid in Automatic1111, but the stats show the following. each line is a batch of 8

[dense tapir]

wth?

[north monolith]

batch of 1 does 7 - 8

[oak osprey]

a batch of 8? god dayum

[wispy nest]

I used to get 2it/s because an error I can't for the life of me remember now :/

[oak osprey]

you're running 8 concurrent gens

[dense tapir]

which sampler and step count?

[oak osprey]

its gonna be slower with bs=8, try =4 or =1

[north monolith]

Check this. I just changed it to a batch size of 1:

[oak osprey]

yeah.

[north monolith]

10 - 12 now

at 512 x 512

[smoky oak]

So that's 512x512 at a batch size of 8?

[oak osprey]

and what scheduler, DDIM?

[dense tapir]

that is 100% shite, no offense

[oak osprey]

UniPC is fast but i have not messed with it much

[dense tapir]

something is messed up for sure

[smoky oak]

If you got 1.9it/s at 512x512, that is much closer to where your GPUs performance should be

[north monolith]

Yes, the slower speeds are 512 x 512 batch size 8, but that's for all 8

The faster 3 speeds at the end is when I switch it to a batch size of 1

[smoky oak]

I'd still slow for a 4080, but it's a lot closer to the target it should be at

Let me check my GP really fast

Never mind, I take it back

[dense tapir]

told ya

something is wrong

[north monolith]

Heun

[dense tapir]

oh, lord

[oak osprey]

nooo

[dense tapir]

ditch Heun

[oak osprey]

use ddim

[north monolith]

I normally use Euler A

[oak osprey]

ddim is the benchmark

[dense tapir]

ddim, euler_a, dpm++ SDE Karras

[north monolith]

Errr DDIM...

Errr

[dense tapir]

there we go

[north monolith]

lol

[smoky oak]

There you go

Lmao

I knew something was wrong lmfao

[north monolith]

See told you I'm a noob!!

[oak osprey]

heun is pretty cpu-heavy

[smoky oak]

DDIM is the go-to speed demon

[dense tapir]

I saw heun lost a nut

[wispy nest]

You haven't insulted bill gates too many times, right?

[oak osprey]

i think sytan was about 10 minutes away from having you reinstall your OS

[smoky oak]

Glad we were able to figure out what was wrong, that was abnormally slow for such a beefy GPU lol

[north monolith]

I played about with the samplers, read some stuff, and some website just said "use what you think looks good".

They never told me to use DDIM, and that the others were dog slow

[smoky oak]

I don't know why the hell you always have to insult me

[dense tapir]

what is your Euler_A at now?

[north monolith]

Thanks all!! Glad I joined this Discord 🙂

[smoky oak]

Like bro, just keep my name out of your mouth, please

I'm glad we were able to figure out what was going on

Enjoy your dramatically faster generations haha

[north monolith]

Euler A

[smoky oak]

I use 15 steps of DDIM, as I found that going any higher doesn't really benefit at all, and it allows you to get extremely fast images

[dense tapir]

Oh, sweet

[north monolith]

I switched to Heun earlier for some reason

didn't switch back lol

[smoky oak]

It is a little strange that you're having so much variation in generation speed, but hey, it's still way faster

[dense tapir]

now, please kick it up a notch and do dpm++ SDE Karras

[north monolith]

Yes way way faster

[oak osprey]

pytorch2 on Ada ftw

[dense tapir]

20 steps on the sde

[smoky oak]

DPM++ SDE karras is a more intense sampler, so I'm curious to see how it does

My guess is probably 16

[dense tapir]

15-16

[smoky oak]

Remember when we were trying to figure out what was wrong with my GPU, and why I lost so much performance? Lmao

[smoky oak]

Oof

[dense tapir]

damn

[smoky oak]

That's a much bigger hit than I was expecting

[dense tapir]

Yeah

[north monolith]

Back to DDIM for me

[smoky oak]

DDIM is what I use always

[dense tapir]

I don't like DDIM if I am after photorealism.

[oak osprey]

DDIM doesn't support tsnr yet

[smoky oak]

It's the fastest, needs less steps than most of the other samplers, and I've also have observed that it is better at not ruining high values on LoRA's

On average compared to Euler A, I was able to go about 15% higher on the drive value before I started seeing any artifacts with DDIM

[dense tapir]

For me I found the sweet spot for ++ SDE is 40-50 but my word it takes my card forever

at 50 the caustics was out of this world

[north monolith]

RIght I gotta go. It's 3:34am here, and I've work tomorrow.

Thx for the help all!! I'm learning more all the time

[dense tapir]

glad we could help

[smoky oak]

Even these images are all 15 step base, 10 step upscale

[smoky oak]

No worries! Get some good sleep, glad we were able to help :>

[dense tapir]

Now try them with ++ SDE Karras at 20 steps

[wispy nest]

when it comes to the use of sampler, then all I know is that they are like different flavour of coffee: same taste, different name! :P
Except for LMS, that one just don't do anything good! ;P

[smoky oak]

The only time that I use more steps with DDIM is because I am using several Laura's, and more steps allow them to be a little more coherent

[dense tapir]

UniPC is the worst

will not even run on any card without tensor cores

[smoky oak]

I don't touch unipc ever

It's just worse DDIM lol

[wispy nest]

I…I used to love it for several weeks…runs away… 😶‍🌫️

[smoky oak]

UniPC is legit just slower DDIM

Like the results are almost identical, but it's about 25% slower on my GPU

[dense tapir]

Has the same restrictions too

[smoky oak]

And it also has several values that you have to tweak in order to get it properly optimized, but then changing certain aspects can change what values are optimal for it, it's just a pain in the ass

[dense tapir]

I got it to work one time, ngl. It was so slow and meh in results but after that never again

[oak osprey]

controlnet recommends unipc for upscaling

[wispy nest]

yeah, I thought it was great, 99% because it was a new one, then I tested using one of the old, and smelly, samplers and noticed how…not good unipc really was :P

[oak osprey]

it's good at certain things, but it's hard to generalize improvements in this space

[dense tapir]

A whole lot of stupid going around in this and the Jurrasic Park syndrome of see if they could instead of thinking if they should.

[oak osprey]

try unipc with controlnet tiles

there's no such thing as a stupid research project

[dense tapir]

Loras are like that now where even the master lora trainers are saying enough is enough and stop the shit

[wispy nest]

I have actually no idea what the samplers really do, I know, shocking revelation about me! :P

[dense tapir]

I still have not ran across a DyLora.

[oak osprey]

We present high quality image synthesis results using diffusion probabilistic models, a class of latent variable models inspired by considerations from nonequilibrium thermodynamics. Our best results are obtained by training on a weighted variational bound designed according to a novel connection between diffusion probabilistic models and denoising score matching with Langevin dynamics, and our models naturally admit a progressive lossy decompression scheme that can be interpreted as a generalization of autoregressive decoding. On the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and a state-of-the-art FID score of 3.17. On 256x256 LSUN, we obtain sample quality similar to ProgressiveGAN.

[smoky oak]

There are so many variations, and most of them are unneeded

[oak osprey]

there, now you understand

[dense tapir]

precisely

[wispy nest]

https://www.scs.stanford.edu/~dm/home/papers/remove.pdf
Not really what was talked about, but when would I ever get to show this obscure knowledge I have?!

[dense tapir]

locon is as far as I could train with 2.1 and Loha was utter failure for everyone on 2.1

[smoky oak]

Also, I'm curious, I'll try your SDE karras and see if I see any improvements @dense tapir

I used it in the past and didn't see any reason to continue using it, but I'm much better at prompting now, so maybe

[dense tapir]

Try on colur images that normally shows noise more

[oak osprey]

i didn't get great results from the Karras pipelines in sd-webui but the same Karras stuff used directly, works fine

so i'm not sure what's going on there

[dense tapir]

for me I can take the same prompt and sde ++ 20 and 20 ddim and the ddim looks like a painting

[wispy nest]

are you trying to say that I'm of high quality? Yes, must be, I'm totally not totally confused by the language in the text…yes…me smrt! 😶‍🌫️ 🤣

[smoky oak]

You do use mostly 2.1, so perhaps it interacts that way. For me, even with the few images I've already tested, I'm not really seeing an improvement

Just a much slower sampler

[oak osprey]

idk man that song the robot sung at the nvidia keynote has been stuck swirling in my head for 3 days

[dense tapir]

Possibly

[oak osprey]

i just want it to stop

[dense tapir]

if I had a 4090 I would use it at 40-50 steps

[wispy nest]

same settings, one use DDIM, the other use uniPC, which is which?
This should be one of those captchas? :P

[oak osprey]

the 2nd is most likely the workhorse DDIM adding tons of awesome detail. the left is weak kiddo UniPC?

[dense tapir]

Same prompt you tell me which is which

[oak osprey]

you have to test the samplers across a series of steps. some of them converge earlier or later

[dense tapir]

one is ddim the other __ SDE

++

[wispy nest]

left is DDIM, right is uniPC (I had to double check because overe the 10 seconds from me posting it, I already forgot which was which :P )

[oak osprey]

too bad you guys don't have DDPM

[dense tapir]

I was spot on for yours then as Uni drops colours

changes stuff too

Anyway Sytan, as you can see ddim vs sde ++

[wispy nest]

the right one had me laugh out loud because I imagined a bar where the one behind the desk was a paper cutout and no one noticed :P

[smoky oak]

For me, SDE is worse

[dense tapir]

one looks real the other painted

for 2.1 it has always been this way

[wispy nest]

is it contagious? ;P

[dense tapir]

for me

[smoky oak]

These are close ups of DDIM Vs SDE, which do you think is which?

[oak osprey]

if he's using 2.1, karras will work better because it essentially turns on some SNR fixes.

[dense tapir]

the first looks painted so in 2.1 that would be the ddim

[smoky oak]

Yeah, that's what I figured was the cause of it working better for him and not for me

[smoky oak]

This is digital art. So yeah lol

The first one is DDIM, and the second one is SDE, but as you can see there's way more crunchy artifacts in the SDE one

[oak osprey]

it's also why i don't like any of a1111's sampler implementations. most of them are not implemented correctly, they look far worse on 2.1 than their Diffusers counterpart implementations.

[smoky oak]

Yeah yeah, we get it, you hate everything A1111 lol

[oak osprey]

@smoky oak try skip clip 1 with SDE

[dense tapir]

That has to be due to the base model

[smoky oak]

Done. Results above lol

[dense tapir]

I mean 1.5 not the model you are using

[oak osprey]

wait, is it 1 or 2

mine's always done it for me

[smoky oak]

Yeah, I think that SDE is better for a 2.1, but not better for 1.5

[dense tapir]

in 1.5 I always used euler_a

[smoky oak]

Same, until I switched to DDIM as it was faster for me

[wispy nest]

I noticed that one didn't apply the background as often for some weird reason I never learned, and probably never will :P

[dense tapir]

I am just not that into DDIM as I will take more time if it looks better.

[wispy nest]

the one without the a is better for me

[smoky oak]

Yeah, I can see the benefit for 2.1

[dense tapir]

I actually like the ancestral samplers

I was never a fan of just Euler

[smoky oak]

Here, let me try them for realism

[oak osprey]

i hadn't had any opinions on any of it beyond the code quality or the interactions i'd seen, until i experimented with it locally, and determined it has a lot of issues, but i mean, do you want me to just never talk about them? i thought it was useful context, i'm not mercilessly shitting on the platform

those issues can be fixed

i thought 2.1 was way worse than it is because the documentation on which samplers work best and why they work, is next to zero

it was adding the karras sigmas pattern to my sampler that fixed it for me, the first bit anyway. that makes baseline 2.1 generate really nice and clean images - as good as it'll get, without fine-tuning, that is

[wispy nest]

I was one of the 12 ( ;P ) people who made embeddings for 2.1 when it came out. But I never got many good images from it that I liked :(

[oak osprey]

https://giphy.com/gifs/maudit-maudit-lord-of-the-rings-peter-jackson-RlZO2mKipB3u8

one of twelve, eh

[wispy nest]

was a joke, there were at least 20 of us! :P

[oak osprey]

whew

that was nearly devastating

[wispy nest]

huh, wait, this is at clip skip 7 :O

[smoky oak]

Realism comparison

Which do you think is which, General?

Cause I don't think it makes a big difference for 1.5 models, certainly not worth needing more steps and much more time

[wispy nest]

I'd say the left is 2.1. Why? Because the fingernail! :P

[oak osprey]

i always thought the sarenghetti was a type of pasta

[smoky oak]

Neither are 2.1 lmao

It's DDIM Vs SDE karras

[wispy nest]

I…I knew that…I was just, erm... look over there! a rainbow tiger! 🚀

[oak osprey]

i prompted as an ai language model, show me where you want to die

pretty good choice

[dense tapir]

I made a lot of TI embeddings I released for 2.1 but they no longer work with other models only on the base.

[wispy nest]

wait, is that true for all embeddings on 2.1, or just yours?

[dense tapir]

ALL

[smoky oak]

@dense tapir which do you think is SDE?

[dense tapir]

with embeddings it is model specific and might work, might not or do something completely different on any other model

[oak osprey]

who are you?, the image model makes a self-portrait

[dense tapir]

Not sure but the one on the left has a different lighting scheme so hard to tell.

[smoky oak]

Yeah, I think that alone shows that I have no need to go SDE karras and wait 2x as long for results 😅

[dense tapir]

the one on the right looks more natural

[smoky oak]

But I will keep it in mind for 2.1

[smoky oak]

That's DDIM haha

[oak osprey]

asking model for self-portrait

[dense tapir]

yes, as I figured because it has a dull apparence while I actually preferred the left one

[dense tapir]

the lighting in SDE is just superior

I see even in 1.5 that remains true

[oak osprey]

DDIM's looks better there for sure

[smoky oak]

See, both of you guys are on exact opposites lol

[dense tapir]

at least in 1.5 they don't look painted (digital or canvas)

[oak osprey]

i like the orange coming through the hair

why are the images so different?

[smoky oak]

They are two completely samplers lol

*different samplers

[oak osprey]

they don't seem to have much of the same composition, sometimes when switching samplers you see pieces stay in one spot

[smoky oak]

Voice typing skipped "completely" lmao

[dense tapir]

tbh, the SDE looks like one taken on a smartphone and the ddim one on a more professional setup with rim lighting, etc...

[wispy nest]

this was made with 2.1 + a billion embeddings, but I was never able to recreate the "painting style" in 1.5 :/

[smoky oak]

I specifically prompted for a professional grade DSLR photo :p

[dense tapir]

LOL, glad I am on 2.1

[smoky oak]

I don't think either look better, however I do not like how much longer SDE takes

Not only is it considerably slower at the same samples, but it needs more

[dense tapir]

I suspect you would need a model of just that for 1.5

[smoky oak]

I'm pretty sure I could replicate that painting style pretty easy just using prompting lol

Let's try

[wispy nest]

I'm not sure, the level of small details is just something that I've seen happen with 1.5 models. Or if so, then it's beyond my knowledge anyway

[dense tapir]

If you took a 100 samples from 2.1 and dreambooth trained (not that damn kohya turd thing) it would do it for you

[wispy nest]

NEVER seen*

[dense tapir]

for dreambooth my 1.5 go to was Shiv's. Damn great work

got to 2.1 and nothing. TheLastBen I never liked in 1.5 or 2.0/2.1

Now Kohya has a dreambooth but considering how screwed up, and half baked, his 2.1 Lora stuff is I would never touch it

for 1.5 kohya is good

[wispy nest]

I had to remove dreambooth (if I'm thinking about the same dreambooth) from my webui as it kept bugging out the entire webui

[dense tapir]

As Kohya told me he only cares about 1.5, and Anime so my realism and 2.1 if it worked great and if not oh, well.

Yeah, that extension is a no go as the author despises auto with a passion (with valid reasons) so Auto would break his shit all the time. I knew how frustrated he was until he cracked and posted on his github.

[wispy nest]

I think I still got your embeddings pre new year somewhere. My folder structure is very, very messy :P

[oak osprey]

by bob eggleton lol

[dense tapir]

my embeddings are on here and HF as I refuse to put anything on Civit. Doesn't matter since I don't train at all now (can't as a free colab anyway).

[dense tapir]

too many abused it

[wispy nest]

did you? ;P

[dense tapir]

Running services for pay while the backend was colab

[smoky oak]

All right, I didn't match that specific painting style, but I do really like this result

I may save this prompt for future reference

[dense tapir]

I wish as I could afford a 4090 if I did, but no

[oak osprey]

try adding brushstrokes

[wispy nest]

ah, I thought you talked about that everyone had 24 alt accounts for each hour of the day :P

[smoky oak]

Already in there

[dense tapir]

yep

[oak osprey]

add more

O NOOOO, EGG MANNNN

[dense tapir]

Man, I ran into people with over 200 alt accounts

I never figured out why but now I know why

Considering the scum I ran into when SD released over on the lexica discord I can understand it

Same prompt only Euler_a

[oak osprey]

Sytan, what does yours do if you add by bob eggleton to the end

[smoky oak]

To the prompt I already have, or to a different prompt?

[oak osprey]

that one

the painted gal

i think bob will make her more "UwU"

[smoky oak]

I just did this by tweaking some weights and adding a couple extra tags

I actually quite like the style

[oak osprey]

or maybe intergalactic

that's so 80s

[dense tapir]

sde++, ddim, euler_a, euler all are 20 steps

[wispy nest]

I love it! I like more, and stronger, color. As long as I don't see much blurry purple/pink that is >:P

[dense tapir]

as you see ddim is the loser

[oak osprey]

i think it tried to make a motorcycle

[dense tapir]

unless you are after the digital artwork style

[smoky oak]

Yeah, I can see why you say that for 2.1, definitely

[oak osprey]

hahahaha yes.

[dense tapir]

I wonder why that is?

[wispy nest]

My eyes are too harsh when it comes to photo realism, it has to be a lot better than if I saw a painting etc. :(

[oak osprey]

i'm guessing darryn eggleton's art is mostly skulls

[wispy nest]

probably because it don't know what eggleton is and confuses it with skeleton from he man ;P

[smoky oak]

Okay, I am very happy with the way this came out

I will be saving these as prompt templates for the future

[oak osprey]

oh baseline 2.1 definitely knows both Bob and Darryn eggleton's art but if you don't have one or the other's name in the negatives, it melds them

[dense tapir]

btw, we have a new blip coming

[wispy nest]

try android jones that artist has the worst name ever to try and get the ai to recreate :P

[oak osprey]

harley quinn in her more explorative, creative years

the 1920x1080 deformities are so fun sometimes

[oak osprey]

ohhh you might be right, there's eggs appearing now

[dense tapir]

Here is ddim of the above

[wispy nest]

quick, grab the goose! you'll be rich!

[oak osprey]

i wish

[wispy nest]

I wonder if I can apply strengthening on the loaded VAE the same way as a lora/embedding 🤔

[dense tapir]

++ SDE

[wispy nest]

I also hate it when I get rice in my hair ;P

[oak osprey]

emphasis on bob++

[dense tapir]

SDE

[wispy nest]

clip skip 9, and I kinda like it :O

but there was supposed to be a image of a car sooooo :P

[oak osprey]

9 on 1.5 or 2.1?

[wispy nest]

2.1 doesn't have clip skip if I remember correctly?

but, it was from 1.5

[oak osprey]

it's mandatory

but you can put it higher

[dense tapir]

DDIM of the above

[wispy nest]

I was sure I saw one of the devs say it didn't have it and then showed an example image set 🤔

[oak osprey]

1.5 has only 12 layers in its TE

so you're basically seeing one of the earliest layers of the model. check that result on stable-diffusion-1-5 checkpoint

i'm curious what it's like, because it should theoretically be more basic than that unless the whole text encoder was fine-tuned

[dense tapir]

Euler_a

[oak osprey]

i dig the textures and i like their style but i would absolutely squish these with my shoe if they showed up outside, sorry

[sterile temple]

hello peeps, happy friday to those in the + timezones 🙂

[oak osprey]

shouts in Japanese, releasing egg

[dense tapir]

I downloaded a 1.5 model today that everyone was raving about but I don't see anything great about it

Ahem, this is from it

[wispy nest]

"insert joke that it's not a 2.1 model" ;P

is that image bad you mena? :(

[oak osprey]

2.1 is literally 10 times greater than 1.5

wait

er

https://giphy.com/gifs/rodneydangerfield-thinking-math-rodney-gEvab1ilmJjA82FaSV

[dense tapir]

that is euler_a and the same prompt from 2.1

Here it is in ddim

[oak osprey]

when i tried 1.5 again yesterday trying to get it to listen to my prompt was like wrestling a greased up pig

[wispy nest]

default 2.1 vs 1.5 I'd say yes—ish :P
but INSERT_FINE_TUNED_MODEL_HERE 2.1 model vs INSERT_FINE_TUNED_MODEL_HERE 1.5 model for me is almost always in the 1.5 favour :P

[dense tapir]

That one is wearing Kabuki face.