🏞｜general-with-images

[oak osprey]

cinnamon roll game controller with candy buttons

[smoky oak]

sounds like a strange problem to have haha

[hearty karma]

? whats that

[oak osprey]

yeah when i read that back it does sound strange 😂

[digital hornet]

wait, the new models (deep floyd for a start) are 20gb? So if we want to keep using the cutting age stuff in a few months locally we need the 24gb gpus basically?

[smoky oak]

I am allergic to cinnamon, so please ignore my indifference haha

[smoky oak]

yup

well

[boreal falcon]

Photopea is basically the Photoshop ripoff alternative

[hearty karma]

basically yeah

[smoky oak]

yes and no

[oak osprey]

@digital hornet if you really want cutting edge make sure it's 64GB because DeepFloyd is currently only 1024x1024 and it's going higher

[hearty karma]

well i know what photopea is

whats the extension for?

[dense tapir]

Just let me die if I couldn't eat Cinnamon.

[digital hornet]

cant you do all the layer-by-layer and whatnot tricks to get equal performance but for more time on a smaller gpu?

[oak osprey]

@dense tapir Ceylon, or Cassia?

[smoky oak]

cause SDXL and Deep floyd are gonna need 24GB VRAM, but SD 3.0 will not, as the devs have told me. They said that deepfloyd and SDXL have bloated parameters, and they will take all of teh good ones with the best data set and TE they have ever had and compress it down into SD3.0

[oak osprey]

that sounds like an EMA model, Sytan

[digital hornet]

damn, 3090 is at least reachable by hobbist to have 24gb vram, but 64gb vram sounds sad expensive

[hearty karma]

so how much will 3.0 need?

[oak osprey]

@digital hornet AMD is trying to fix that

[smoky oak]

if all goes well, the same as 1.5

[dense tapir]

Ceylon is the crap one Cassia is expensive and healthy for you.

[hearty karma]

ooo damn

that would be a hell of a lot of optimization

and im sure theres going to be tons of shit to get it lower anyway

[smoky oak]

it just takes cutting out the bad parameters and refining the good ones

[digital hornet]

yeh. I've been waiting for AMD to get good support for AI for like 6 years now, despite high vram it just isnt worth

[bold cave]

finally

someone who agrees with me

team green for life

[boreal falcon]

Stable Diffusion 3.0 vs out of beta Firefly

[oak osprey]

@digital hornet i mean the effort you're willing to put in is your call but i tested my vega card last month after a new update rolled better support out and it's finally plug and play and requires zero changes from my source code to work

[hearty karma]

anyways Sytan where the comparisons at?

[dense tapir]

For me I like 7900XTX with rocm. Oh, I am only doing 30it/s instead of 40. Yep, and half the cost.

[smoky oak]

theoretically we could have multiple SD 3.0 bases

each for different things like one for realism, one for art, one for anime, all of that

they all have 890m parameters geared towards that one thing so they are better than SDXL and deep floyd, but modular and separate

That would be insane

[digital hornet]

sure, and then a new cool technique comes out that all cutting edge models adopt and you have to waste hours to get a workaround/wait months

[dense tapir]

5.6.0 will openly support it

[oak osprey]

@digital hornet not as long as they're in PyTorch. it is a first-class citizen now. similar to Apple Silicon, but you don't even have to convert tensor formats to CoreML...

[dense tapir]

What we need is for AMD, Lisa Su, to get up off their asses and really support AI with hardware not software like FSR2/3

I think they will with their acquisition of Xilinx.

[digital hornet]

it's software that they've always been behind in, they've been better performance/$ specially with more vram which is the most important part for consumer AI because that's the thing that truly constraints you and not just make you wait longer

[oak osprey]

if TSMC didn't need to make a profit we'd all be rolling in vram

[hearty karma]

is the 3060 ti a good card for uhh ai?

[oak osprey]

those bastards

[smoky oak]

yes, its a good option for image generation

its what I have, you can do 4k and 8k gens on it, and you can do good LoRA training as well

[oak osprey]

good call singling it out to image generation

[hearty karma]

wait seriously 4k and 8k gens???

[smoky oak]

yeah, i do them all the time

[oak osprey]

using upscaler and highres fix

[smoky oak]

not pixel upscaling, generative upscaling, there is a huge difference

[hearty karma]

how much vram gets used per generation?

[smoky oak]

around 7GB

could go down to 5 if you needed

[digital hornet]

damn, I've been really debating with myself whether to get a 3090/4090 instead of waiting another year as my 3080 + paid Colab has been enough for everything, but if deepfloyd and other next gen techniques require 20gb, I guess I'll start trying to come up with the 2k euros to upgrade before that

[bold cave]

get a 3090.

[hearty karma]

your probably better getting a 3090 over a 4090

[oak osprey]

deepfloyd was an nvidia plant job confirmed

[bold cave]

all ur problems will disappear guaranteed

[smoky oak]

I would say wait for the 7950xtx

[bold cave]

I can confirm

[oak osprey]

these guys think 4090s are melting

[hearty karma]

no

[boreal falcon]

2K? At that point just buy RTX 4090

[dense tapir]

No, if effing companies stopped with their madness of trying to go smaller and smaller and even TSMC said it is madness. They feel 2nm will be it as the costs, and time invested, will be too much. No one could afford a 500k USD wafer that rolls out about 200 dies.

[hearty karma]

why get 4090 when its a massive brick

when you can get a slimmer 3090

[boreal falcon]

Because 4090 is superior

[digital hornet]

..the 2k is for the 4090

[oak osprey]

performance per watt, but not everyone is on solar power like me

[digital hornet]

'oh yeh if you are buying a 4090 at that point buy a 4090' ??

[smoky oak]

or an even smaller 7950xtx thats faster than both of them for less money :>

[digital hornet]

i realized ill need a new PSU AND a new case to get a 4090, kind of pain

[oak osprey]

yeah buy an unreleased card

[smoky oak]

thats why i said wait lol

[oak osprey]

if you buy the nvidia 5090 you'll be a year ahead of everyone else

[digital hornet]

if I can get a 3090 TI 2nd hand for 850-900 and take the risk from 2nd hand purchases maybe that's most worth for running deepfloyd and similar

[dense tapir]

What the fuck is up with all this team green fanboi shit in here?

[bold cave]

u wont find a ti for that price

[smoky oak]

The 7950xtx is right around the corner, and with ROCm, it could genuinely be the best AI card for consumers

[bold cave]

trust me I tried

[boreal falcon]

Pro tip: RTX Quadro 6000

[bold cave]

for weeks

[oak osprey]

why are we even debating GPUs again i thought we all agreed to save up for a $2.5 million Cerebras CS-2 wafer

[hearty karma]

wait hold on

[dense tapir]

LOL, yep

[hearty karma]

wait nvm

[digital hornet]

no shot I'm risking wasting 100 hours of my life over the lifetime of the card on figuring out fixes for bad AMD software support

[oak osprey]

i was thinking about it and once we buy that it should probably stay with me so i can test it out first

[smoky oak]

HOLY HELL

[oak osprey]

i'd still be on AMD if they had 24GB VRAM when i bought my 4090

[smoky oak]

I forgot just how insane the cerebras is

oh my god bro

1PB/s memory

AAAAAAAAAAA

[bold cave]

is that the thing the jeopardy bot runs on

the one that beat all the jeopardy pros

[oak osprey]

i was thinking of buying two 7900xtx or whatever the hell

@bold cave that's a traditional hyperscale supercomputer called Watson

[smoky oak]

No, Cerebras is the be all end all giga wafer ATM

[dense tapir]

For those with more money than brain cells there is Nvidia ...for the rest of us there is AMD.

[bold cave]

I thought quantum computers were the end all

[oak osprey]

Cerebras is cutting-edge stuff, they spent $600 on their platform and trained an entire large language model in like 3 hours.

[digital hornet]

theres one for 840 listed as 3090 TI on the local in-person ebay from a few days ago. Though most are 3090 without TI at that price

[smoky oak]

I said of wafers lol

quantum PC's aren't wafers :p

this is one cerebras

[bold cave]

whoa

[hearty karma]

have you guys heard of this single cell yellow slime thing?

[dense tapir]

I know, the truth hurts there Sen. Fact is stop feeding that damn beast

[smoky oak]

850,000 AI cores

2.6 TRILLION transistors

[bold cave]

What do you need to cool that thing

where do u even case it

[smoky oak]

very little believe it or not!

[oak osprey]

@bold cave i explained this earlier but quantum computers have a goal called quantum supremacy that demonstrates superiority over traditional CPUs and currently they only succeed at Boson Sampling but it's "a pretty big win" there, from >9000 years runtime to half a second

[smoky oak]

its designed to have a huge surface area, so its actually really easy to cool

[bold cave]

yeah and at the same time cracking all our passwords say gg to ur bank account

[oak osprey]

they haven't demonstrated its ability to beat a simple GPU in that task yet

[bold cave]

oh really

[smoky oak]

Yeah, cerebras is a little big lol

[oak osprey]

GPUs are really friggen good for that, now.

[hearty karma]

i am doing cat images again

[bold cave]

but how do you even bruteforce a password

with a gpu

theres a try limit

[hearty karma]

i am going to compare the 16gb ultra and 2gb ultra

[smoky oak]

AAAAAAAAAAAAAAAA

[oak osprey]

you grab the shadow file or a list of salted hashes etc and just go to town locally.

[smoky oak]

THE CEREBRAS WSE

[bold cave]

wow

[smoky oak]

broooo

[digital hornet]

i wish they hadn't nerfed Colab pro so hard that I now need to upgrade gpu so hard. 100 minutes a month? The A100 has 16gb instead of all of its ram and is like 5 times slower than my 3080? suuck

[smoky oak]

the WSE has a memory bandwidth ot 180PB/s

that is

[oak osprey]

A100 is 40 or 80GB

[bold cave]

bru who cares those are never going to be open to the public

[smoky oak]

180,000,000GB/s

[oak osprey]

ohhhh

[smoky oak]

anybody can buy one

[bold cave]

if we're gonna drool let's drool over the a100

[oak osprey]

they switched to MSGs? EWEWWWWWWEWWEWEWEWWWWW

[digital hornet]

yeh if they are not shared between multiple users like on Colab and thus not having all their vram avialable to you

[dense tapir]

tile resample, what model do I use with that preprocessor?

[smoky oak]

A100 is 40/80GB

[oak osprey]

MIGs*

[smoky oak]

my friend has an A100

[oak osprey]

Multi-Instance GPUs. they divide an A100 into 7

makes me want to vomit

it's supported directly in the hardware via SR-IOV

[hearty karma]

have you guys heard or seen of this slime thing?

[digital hornet]

though yeah now on Colab it's saying 40gb for the A100

[oak osprey]

slime mold? yes

[smoky oak]

@dense tapir Little fun fact for you

[digital hornet]

but it';s like 5 times slower than my 3080

[smoky oak]

the Cerebras chips have a 100% yeild

[bold cave]

I saw that happen too

in runpod

its bizarre

[hearty karma]

i think in theory slime mold could be used to improve technologies

[bold cave]

cuz in theory it should tip thru training like a mf

[oak osprey]

depends on the current ecosystem if it's overloaded they'll cut it down. they banned SD-webui recently to fix their capacity issues ahahahahahaha

[smoky oak]

There are 2 models, thats why

[digital hornet]

that's only for free Colab, I'm on paid

[smoky oak]

there is the weaker 40GB A100, and the full fall 80GB A100

[boreal falcon]

What is never going to be open to public?

[oak osprey]

@digital hornet i am too and they cut it down once you burn through compute units

[smoky oak]

cerebras, except it is

[bold cave]

where can I buy one

[digital hornet]

well yeah, after you are out of compute units you are basically a free user till the end of the month

[dense tapir]

@smoky oak No failures? The way a GPU wafer works is there is always failures and they keep making lesser version cards from those failures until there is one that is just too far gone to make into anything.

[oak osprey]

i burnt through my compute units allocated for a whole month in like 15 hrs. in fact, if you're a Colab Pro subscriber, you can't even use the Free tier without burning credits

[hearty karma]

arent cerebras a developing technology still

[boreal falcon]

Nah, never will be xD

[bold cave]

yea google collab was a scam

[smoky oak]

They run perfected processes to make sure there is no failure

[boreal falcon]

And its questionable if they will surpass Nvidia in that area

[dense tapir]

Damn

[digital hornet]

oh shit the A100 40gb uses that much? the fuck

[smoky oak]

first gen was 16nm, second was 7nm

[digital hornet]

13.08 per hour out of 100 a month, come on

[smoky oak]

in a world where we are on like 3nm

[bold cave]

yep

[oak osprey]

@dense tapir it costs $2.5 million US dollars for one wafer system

[smoky oak]

they are on 7nm to make sure there are no issues

[bold cave]

and if I remember correctly, Google Collab doesnt even let you choose ur GPU

its a shit platform from a shit company

[oak osprey]

at 2.5 million for one unit you don't care how many wafers you destroy

[digital hornet]

Colab was so insanely good before it was paid, and then even better for $10/mo but since SD hype and everyone using it they've nerfed it soo much

[dense tapir]

For that meticulous measures no doubt.

[digital hornet]

I trained GPT2 in 2019 for a full month of nonstop Colab when it was all free

[oak osprey]

LOL you ruined it for everyone

/s

[dense tapir]

The nerfing really began with Deepfakes and the bans in 2018 then it calmed down and was alright then AI art/SD forget it.

[visual mountain]

yo guys, what negative prompt would you use in order to try to improve a text to image generation that includes words (fixing spelling and stuff like that)

[digital hornet]

nah, they had PLENTY of capacity before the recent rush of stuff - mainly Stable Diffusion and similar

[smoky oak]

@dense tapirThink about it this way

the A100 can do 312 teraFLOPS of AI

A first gen little brother Cerebras can do 1.2 EXA FLOPS of AI

[oak osprey]

@dense tapir he's trying to control the way you think

[smoky oak]

WAIT MY BAD

I READ IT WRONG

1.2 exaflops for 16 of them

[dense tapir]

What model do I use with this preprocessor tile_resample?

[bold cave]

tile

theres one for each called tile

[dense tapir]

I guess we didn't get that

[bold cave]

and just load up the image in there

[smoky oak]

which, BTW, is the fastest super computer in the world for AI tasks

ANYWAYS

[dense tapir]

no tile for 2.1 😦

[boreal falcon]

There is Colab Pro+ too

[smoky oak]

nerdgasm over

[hearty karma]

i think im gonna delete Zovya

[smoky oak]

OK

I am testing oboth zovya

[dense tapir]

I gues for 2.1 we simply do not have that model so I wonder why it is in the preprocessor?

[smoky oak]

got distracted by thefastest wafer chip in the world 😅

[oak osprey]

if you wanna get scammed while you're gettin scammed

[hearty karma]

yeah i could tell

[oak osprey]

https://giphy.com/gifs/yo-dawg-xzibit-ToMjGpM4h8DAsgL0fEA

[hearty karma]

i asked you about the comparison earlier 🥹

[wispy nest]

@oak osprey post shrek

[smoky oak]

Sorry, I am foing it right now lol

[wispy nest]

@oak osprey what did you mention about the p40

[smoky oak]

You have signed away your rights to live

[oak osprey]

p40? the nvidia? 24gb? it's cuda 5.7 or so pretty old

[hearty karma]

i probably shouldnt have used high res fix for a comparison

[smoky oak]

huh

[wispy nest]

This what SD is ment for

[oak osprey]

your future is in nvidia's hands at that point

[hearty karma]

because the results will be different anyway then 🤔

[smoky oak]

now this is a weird phenomena lol

[oak osprey]

@hasty nova https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py

[smoky oak]

HMMM

[dense tapir]

P100 at only 16gb is better

[bold cave]

boo i dont wanna comapre dudes

post boobs

[smoky oak]

not sure why it put a picture of them in the forest on their hoodie in the forest

[hearty karma]

1 is the 16gb zovya

[smoky oak]

nope

[hearty karma]

and 2 is the cleaned up 2gb model

[oak osprey]

i love how his sweater has a representation of the child that lives within all of us

[smoky oak]

the cleaned up model is 5.6GB

[wispy nest]

P100 > P40 just speed wise?

[dense tapir]

speed, and I believe uses a newer cuda

[boreal falcon]

Why 😄

[hearty karma]

no it cleans up to way less that that for the ultra model

[boreal falcon]

I mean i dont use Colab anyway

[wispy nest]

Ahh I thought had same cuda cores

[hearty karma]

there was tons of junk for the ultra zovya model

[boreal falcon]

Dunno what i would use Colab for

[oak osprey]

@boreal falcon easy to burn up through compute credits in less than a day and when you want to use Free Tier Colab with a pro+ subscription, it deducts from your credits

in other words you'd need two colab accounts to continue to use it the way you've always done without eating compute credits

[dense tapir]

Cuda 6.0

[smoky oak]

cuda six point bro

[oak osprey]

cuda60 jesuz crust

[hearty karma]

what card is that about?

[wispy nest]

That's better than not having cuda cores like amd cards 🤔

[dense tapir]

P100

[smoky oak]

4096 bit bus is more insane IMO lol

[oak osprey]

just buy AMD

[dense tapir]

200 bucks and a fan cooler

[oak osprey]

Sytan, memory bus doesn't work the same way now

[smoky oak]

why would it have changed in the last 24 hours? lol

[bold cave]

what cfg

lower it, theyre a little too bright

[hearty karma]

drop it to 4

it works best on 4

[dense tapir]

It isn't a bad little pascal card

[oak osprey]

it was like a slow change over time. the NVIDIA 285 card had a higher bus than the 1080

[dense tapir]

I almost went for it

[smoky oak]

I am beginning to think you just have a poorly calibrated monitor, cause they are nearly perfectly exposed in a histogram

[oak osprey]

they're exposed? i don't see anything

[smoky oak]

They look perfect on my color and gama calibrated monitor

[hearty karma]

kitty cat

[smoky oak]

har har lol

[oak osprey]

huhuhuhuhu

[dense tapir]

Two issues stopped me. The cooling solution is horrible and deafening, and it lacks tensor cores

[smoky oak]

essence of gato

[wispy nest]

Typical Shrek response

[hearty karma]

i think i want a cat now

[oak osprey]

you don't want tensor cores. no one cares about tensor cores. you want to run DeepFloyd. you want vram. tensor cores? pfft. it'll just take an hour longer

[dense tapir]

you do you

[oak osprey]

DeepFloyd is the future, and so, we must go backward to NVIDIA P40 24GB

[smoky oak]

I mean, you can give a slower card more time, you can't give a low VRAM GPU more VRAM

[bold cave]

nah g. wait for xl

xl is where its at

[oak osprey]

you don't understand, Marty! we have to go back to the past, to get, to the future!

[smoky oak]

SDXL is parallel to deep floyd lol

they are basically the same thing at base

[bold cave]

oic

[wispy nest]

Did snoop just copy an image mid render

[hearty karma]

where?

[bold cave]

those look hella dope

[hearty karma]

i dont remember copying an image mid render :-)

[bold cave]

the ones on the right look a tad more coherent but thats it

[smoky oak]

I am obsessed with this dress lmao

[hearty karma]

i am going to redo cat image

and put high res fix on this time

i need ultimate cat

[smoky oak]

Gato^2

[dense tapir]

Nope, 2.1 didn't get that model.

[oak osprey]

the ones on the left are holding a secret they want to tell me. the ones on the right, i have met their family. we ate dinner together at their cottage in the Swiss Alps. there are no secrets, we know each other too well. it is a boring. I lust for the left side, the memories of the past. the discovery phase. the secrets.

[bold cave]

come join 1.5 then

[dense tapir]

Nope

[smoky oak]

[hearty karma]

hey sen do you know what tile resample would do if you were to use it for a gen?

[bold cave]

perhaps its for a reason no one bothered to create a 2.1 model :kek:

theres only 1 wym

[wispy nest]

cat! :3

[oak osprey]

i am divorced

[hearty karma]

not not good enough

[hearty karma]

the ultimate cat will arrive in like 30 years

[smoky oak]

These colors are gorgeous

[hearty karma]

so which is which?

[smoky oak]

I shant say yet

[oak osprey]

tbh we're not there yet for peak Cattiness

[smoky oak]

more tests!

[oak osprey]

women named Kat are always so hot

[smoky oak]

the same model is on the same side each time, so make sure to keep track of how many times you like the left or the right more

[oak osprey]

you can't make me

[hearty karma]

welp it always seems to be the right for me

they are more coherent

[bold cave]

agreed, right looks a bit better

[hearty karma]

anyway my beautiful gato generation is coming along nicely

[smoky oak]

good to know, good to know

[oak osprey]

right side you can really grab a handful of, if you know what i mean

[hearty karma]

currently in the high res fix stage

[wispy nest]

left loos better

[smoky oak]

just make sure you keep notes so you know which model you visually prefer

[oak osprey]

oh yeah we're all writing notes down

for sure

[smoky oak]

I am gonna test a range fo subjects

[wispy nest]

right looks like jeff bezos wife after purging her reptilian skin

[hearty karma]

do a cat!

specifically maine coon

please

[smoky oak]

I can try

[oak osprey]

didn't Grimes say we can use her for AI?

[hearty karma]

thank you 🥹

[oak osprey]

you know what that means

[wispy nest]

yeah but who would want to

[smoky oak]

the AI got a little too boobish lol

[digital hornet]

how risky is it to even get a 2nd hand gpu (3090 or better 3090 ti)? finding a bit hard to quantify how much more 'expensive' it works out due to the risk

[bold cave]

not risky

[boreal falcon]

ok i laughed a bit too loud...

[digital hornet]

like 1 in 15 chance of it turning bad?

[bold cave]

if u can inspecting it beforehand and buy from someone not shady

[smoky oak]

I have owned multiple used GPU's, never had an issue

[bold cave]

it is much better to buy used

[oak osprey]

@boreal falcon glad someone else found it, that's what i meant when i said "you can really grab a handful from the one on the right"

[digital hornet]

I mean at minimum it's risky in the sense that a new one will have warranty and sometimes shit randomly stops working, so not being covered by warranty increases the risk

[bold cave]

it does, but in practicality it doesnt happen

[digital hornet]

but also risk from the GPU having a not immediately obvious issue and that's why it's getting sold

[bold cave]

GPUs just...sit there, untouched, after u install them

if there is any issue, it wont boot

[oak osprey]

so if you can test the GPU in person you should definitely put it through the fuzzy donut test

[bold cave]

it will crash immediately during a stress test

[bold cave]

exactly

[digital hornet]

nah, there's sometimes issues with e.g. overheating later because something isn't soldiered amazing

[bold cave]

if it can run it for a few minutes then its 99% just fine

[smoky oak]

great benchmark to prove health

[hearty karma]

fluffy cat!

[oak osprey]

yep it'll surface coil whine, and other BS

[smoky oak]

my old GPU had hella coil whine lmao

[digital hornet]

damn maybe ill straight up message the guy who listed 3090 TI for 840e 2 days ago in case it's still available then

[oak osprey]

it's mostly a problem when you LC it

[hearty karma]

:O

[smoky oak]

Gatos

[hearty karma]

what sampler did you use?

[smoky oak]

I always use DDIM

15 steps

*20 steps

lmao

[oak osprey]

would you believe 35 steps

[hearty karma]

the perfect breed of gatos

[smoky oak]

I almost always use 15

[hearty karma]

largest domestic cat

i think

[smoky oak]

Chonc-kat

[hearty karma]

this is the same breed 💀

[smoky oak]

Gato Supreme

[hearty karma]

yis

ultimate gato

[boreal falcon]

Id overpaint it or photoshopped

[wispy nest]

https://www.ebay.com/itm/204327422153?mkcid=16&mkevt=1&mkrid=711-127632-2357-0&ssspo=vPncx0qoQa2&sssrc=2047675&ssuid=&widget_ver=artemis&media=COPY

[oak osprey]

@digital hornet https://youtu.be/Cfy0cr9VorE?t=3355 you reminding me of this scene from the mel brooks film where the guy tracks these chairs down to a couple in Siberia and begs to buy them

[smoky oak]

bruhj

[smoky oak]

stupid Clyde

it was giving NSFW warning lol

[digital hornet]

damn in 2020 I've bought a 5700xt for 370, then 2021 Ive gotten 3070 for 720 and now already I need to pay 900-2k for a 3090 or 4090, I swear just gpu expenses work out more than buying and owning a car

[hearty karma]

:O HUSKEY

[oak osprey]

great deal. buy a liquid cooling block too.

[smoky oak]

sorry its so low quality, can't zoom in or it flags for NSFW

[hearty karma]

but why th o

thats just kinda odd

[smoky oak]

no idea lol

[hearty karma]

stupid clyde bs

[smoky oak]

Clyde always does that

[oak osprey]

Clyde sucks ass

[wispy nest]

Clyde sucks ass

[smoky oak]

Anybody have any suggestions to compare?

[hearty karma]

the eyes keep getting buggered

[oak osprey]

@smoky oak jack black riding a bear in the swiss alps, is an oddly difficult prompt for most models. give it a whirl.

[bold cave]

u always have to inpaint eyes

[oak osprey]

eyes getting bugged out could be the medvram options for a1111

[bold cave]

cant wait for this training to finish just to see torch 2.0 screw it up

[wispy nest]

https://civitai.com/images/661380?modelVersionId=60341&prioritizedUserIds=10921&period=AllTime&sort=Most+Reactions&limit=20

try this prompt haha

[bold cave]

yo I like it

[smoky oak]

trying Jack Black ATM

[oak osprey]

it took a LOT of code changes to make the jack black prompt work. needed to disable tiling VAE, switch to MSE, use Karras, and then, and only then, did it kind of work out regularly with Stable Diffusion 2.1

[wispy nest]

its cuz jack black is most generic name ever

[smoky oak]

OMG BRO @oak osprey

THESE ARE GREAT

[digital hornet]

ugh actually fuck the 2nd hand 3090s, ill just spend the money on and time on a new case and PSU this month, install them to fit the damn 4090 and then buy it next month

[smoky oak]

LMFAOOO

[oak osprey]

i see some of the facial oddities but mostly resembles my fixed stuff, so i guess that's really as good as it can get

[wispy nest]

damn that was fast

[oak osprey]

[wispy nest]

mine would of taken 4hrs to do that

[smoky oak]

Alright, trying the blue girl now

[oak osprey]

blue gal is hawt

[smoky oak]

The provided prompt isn't exactly that good, so reserve expectations IMO

[wispy nest]

i feel like we taken nerdy basement dwellers to another level

[hearty karma]

i better not be getting lumped into that

[oak osprey]

speak for yourself, i live in an RV and i'm homeless

[hearty karma]

all i am doing is making cats

[smoky oak]

An RV is a home :>

OK YUP

[oak osprey]

officially in the canadian census it is not

[smoky oak]

CAN'T SEND THE RESULTS

SORRY GUYS

[oak osprey]

F

[hearty karma]

oh

oh no...

[wispy nest]

idk if that is worse or better

[hearty karma]

it went lewd didnt it?

[oak osprey]

👮🏽

[smoky oak]

there is a lot of boobage here

[oak osprey]

[hearty karma]

just genuine cats

i swear

[digital hornet]

CPU bottleneck doesnt really happen as easier for SD with a 4090 as it would for gaming right? As my 5800x3d would likely bottleneck a 4090 hard in gaming

[smoky oak]

lmao

[oak osprey]

idk i have a 5800X3D with 128GB RAM and a 4090

[wispy nest]

damn

[oak osprey]

what the hell you playing that will be bottlenecked

[wispy nest]

i like the fact it looks like she doesnt have makeup on

[smoky oak]

remember to keep track of which one you like more

I will be sharing which model is which soon

[digital hornet]

do you end up using the RAM much? I only have 32 and dont want to get more ram before next year when I'll switch to the new RAM+mobo+CPus?

[oak osprey]

@smoky oak the odd way you masked those reminds me of when i tried "modest nsfw turtleneck"

[smoky oak]

I wanna do a few more comparisons, any idea?

[bold cave]

I can already tell which is which bc of the freckles

[oak osprey]

@digital hornet yeah but i run Vicuna, LLaMA, and quantize models like it's going out of style. it can quantize LLaMA 65B and run it in Llama.cpp without running out of memory, though for quantizing you do need about 250GB of NVMe swap space

[bold cave]

those chickenpox freckles are zovya's watermarks

[smoky oak]

which do you think is which?

[hearty karma]

im like 90% sure its a comparison of both zovya models

[smoky oak]

it is

[hearty karma]

so you really got sytan there sen 💀

[smoky oak]

Its ultra vs pruned

[oak osprey]

with 128G RAM once I load LLaMA 65B i have about 58G RAM left

[digital hornet]

I dont really care for running language models locally anymore but if people start making huge diffussion models andusing the same tricks...

[bold cave]

yeh I was about to say... they both have them loooool

If I were to vote tho

I would vote right

just a tiny bit more depth

[oak osprey]

@digital hornet well DeepFloyd has a text embedder. a whole-ass language model called T5

[smoky oak]

Seems to be that most say right, interesting

[bold cave]

but its minor

[oak osprey]

@digital hornet i can say i would have been fine with the 64G RAM i had if i never cared for language models

the ram was cheap enough i figured, that'll be really nice to have.

[digital hornet]

damn, I really dont want to buy RAM now just because I know that next CPU upgrade will require the ddr5 ram anyway so any ram bought now will be used for a year

[wispy nest]

@oak osprey how much is your setup

[oak osprey]

lost track

[smoky oak]

DDR4 is super cheap ATM

[dense tapir]

DDR5 is slow right now too

[wispy nest]

give estimate

[digital hornet]

I mean with the new CPUs you need the new motherboards and the new ram you just can't use ddr4

[oak osprey]

2800 gpu, 500 mb, about 450 mem, 500-1100 storage, prollly $5350

[digital hornet]

that's why I only upgraded to 5800x3d end of last year so i wont have to bother with new ram/mobo

[oak osprey]

i'm paying canuckistani rubles, so

[smoky oak]

ah lol

[oak osprey]

3994 USD

i already had the motherboard for a few years, and the case.

and the storage.

[digital hornet]

hm actually 32gb extra of ram to bring me to 64 is only an extra 90euro, but even then it's like 1750 for gpu + 200 for new big case to fit it + 200 for new PSU to power it, it adds up to a lot

[oak osprey]

but it has two 16tb disks in a RAID1 and 1tb of nvme

@digital hornet that's exactly the mindset i had

lets see if you have more self control than i do

[smoky oak]

Yeah, DDR4 dropped massively in price these last couple months

[oak osprey]

because of the "i need the latest and greatest" clowns that do us these favours

[smoky oak]

Stupid ass clyde

FFS

JUST SEND IT OH MY GOD

here

https://cdn.discordapp.com/attachments/836162857644720151/1105624534083698688/image.png

There

god

have to send it in a different server and copy the link

[oak osprey]

they're amish so the nsfw barrier is much lower

i think i see an ankle in there.

[smoky oak]

LMFAOOOO

[oak osprey]

those boobs look a lil too big for amish i think

[smoky oak]

Alright, I am gonna do one last test, any suggestions?

[smoky oak]

They are supposed to be medieval pesants lol

[oak osprey]

try to get a "World's Best Boss" coffee mug

[hearty karma]

do spider-man

[smoky oak]

I'm going for realism, not memes lol

[hearty karma]

👀

[oak osprey]

oh yeah indonesian hotbody woman spiderman

[bold cave]

its pretty conclusive at this point

[oak osprey]

malformed outputs galore

[bold cave]

right keeps winning

[hearty karma]

no not indonesian spider-man

[bold cave]

I vote right again

[smoky oak]

alright, everybody cast your picks between right and left

I am generating the last comparison for now

[hearty karma]

right

[smoky oak]

Sen likes right more

snoop like right more

@oak ospreyhow about you?

[bold cave]

dont do so many portraits tho

just do 2 for each side so we can zoom in better

[smoky oak]

Alright, I can do that for one last comparison

[hasty nova]

Don't really know what I am supposed to do with this...
I am not well versed in what diffusers are and stuff.

[bold cave]

right again

[smoky oak]

that is incredible lmao

[bold cave]

You can see more depth in the shadows

yes they are all damn good

the shadows are insane on the right

[hearty karma]

right gets more stable spider-logos too

from what im seeing

[bold cave]

that ones the Ultra model isnt it

[smoky oak]

alright, I am gonna finish off with a one to one portrait of a pretty woman

[bold cave]

math doesnt lie

[oak osprey]

As an AI Language Model, I do not have opinions or personal preferences. However, I can agree that these models likely have their strengths and weaknesses and are likely to compliment each other in synergistic ways, moving forward.

[bold cave]

fuck off with that censored bs

stuff like that made chatgpt a joke

why anyone would still use it for creative outlets is beyond me

[oak osprey]

hey, chatgpt taught me to write censored python code in less than a couple weeks

[bold cave]

for coding its good

creativity, nty

[oak osprey]

it's good if you literally have no ideas and need to burst the creative block bubble, but beyond that it pisses me off when people try and write with it without editing its result, and sharing. it just writes TERRIBLY BAD.

[digital hornet]

there's a ton of creative stuff that aren't nsfw for a start, anyone daeling with those wont be particularly hindered

[bold cave]

In case you guys are interested in dreamboothing Im also gonna post my results and determine whether or not A) Dreambooth is broken bc of Torch 2.0, B) Its broken bc of A1111 or C) It's broken bc of the model itself

[dusk sorrel]

Is there a general chat for video sharing?

[oak osprey]

@bold cave if you want to see the only good thing it can write, have GPT4 write you an imaginary episode of I Think You Should Leave, with Tim Robinson it's a surreal comedy show where nothing makes sense

[bold cave]

#🎥｜animation

[digital hornet]

like it's dumb how limited it is but people on discord thinking that everything that can't porn isn't 'creative' is kind of lol

[hearty karma]

people think that?

[oak osprey]

well i don't care that it is censored. i hate how repetitive it is and how "on the nose" its writing is.

[digital hornet]

like 5 lines above yeh

[wispy nest]

I asked for ideas about a haunted island last week :(

[oak osprey]

if you don't know what i mean, you tell it to write "a story where X happens" and then at some point in the story it'll blurt that out almost verbatim. at least it does for me

[bold cave]

SEE

its not just porn

the whole thing is just brainwashed

[hearty karma]

ok but its not just porn... it censors lots of shit

[digital hornet]

it's brainwashed but you can def still use it for plenty of things

[oak osprey]

don't worry guys even when you jailbreak it, its writing is still awful

on my discord server, the bot uses Anarchy as its default persona. it sucks hard

[smoky oak]

Alright, about to finish the last comparison gen

[bold cave]

Help me cook rice

As an AI language model, I cannot provide instructions on cooking rice. Cooking is an extremely dangerous activity, where people can get seriously hurt.

[smoky oak]

here it is people

cast your final votes

[oak osprey]

can you write a script of an episode for I Think You Should Leave where the character is trying really hard not to talk about hot dogs
this is the best writing prompt for GPT

[hearty karma]

right still

[wispy nest]

it's not dead yet though! :P

[bold cave]

Ok that one is a tricky one

[hearty karma]

unless flower is meant to be in hand

[bold cave]

I see more detail fo' sho on the right, however the increased depth of the shadows backfires on the eyes

[hearty karma]

which in that case left

[oak osprey]

i pick left side because she's more modest

[bold cave]

still, I like the details on the right more

[smoky oak]

I will say, I did specify flowers in the background, not her holding them

[hearty karma]

ah ok then

[smoky oak]

alright, so right it is again?

[bold cave]

it just has more intricate shadows although wouldnt necessarily say "better"

[hearty karma]

ye

[smoky oak]

AND THE WINNER ISSSSSS

https://tenor.com/view/lafuddyduddy-drum-duck-cute-drum-roll-please-gif-16950134

[bold cave]

watch him troll us

[smoky oak]

Photoreal ||pruned||

[oak osprey]

"All of you with varying backgrounds, eyesight, and mental eptitude, please, develop a universal answer for me on a subjective truth ❤️ "

[hearty karma]

💀

[bold cave]

He trollin

[hearty karma]

hey guys can i input a 1024x1024 image on controlnet for a 512x512 generation?

[dense tapir]

With controlnet which preprocessor goes with the color model?

[bold cave]

yes

[bold cave]

u can, but u shouldnt

its outdated

thats how we used to upscale before CTU

[hearty karma]

oh so i should use a 512x512 then

[bold cave]

oh wait

u mean just to get the preprocessor

in that case yeah

[hearty karma]

ye

alrighty great 👍 thanks for the answer

[bold cave]

Yo im gonna test to see if what sy says is true

cuz if it is, then that might be good news for my dreambooth training

a smaller model might be easier to train

if he's right, then maybe the added junk in the 16gb file makes it worse?

[hearty karma]

it probably does yeah

[smoky oak]

As you can see, photo real pruned was on the right every single time

17.442GB

[digital hornet]

what's it based on

[smoky oak]

OH

pruned

sorry

[hearty karma]

apparently the clip thing is also buggered on ultra

[digital hornet]

because yaeh looks pretty good

[smoky oak]

5.539GB

Welp, I have my new model

[digital hornet]

is it a sd2.1 finetune or? I missed some of the episodes of this chat

[smoky oak]

1.5 finetune

I will now say that Zovya is the best model I have ever used

the pruned version

[hearty karma]

i will switch to zovya pruned then

since it seems to be better

yeah ik but still

[digital hornet]

presumably 2.1 can be better than 1.5 if as much time was spent on it by as many people but 1.5 still soaks up most of the resources

[hasty nova]

I'm kind of proof of that - my model is decent, almost kinda close to a nice 1.5 model (without nsfw stuff though, I didn't feel like training that in because I am not into it), except I barely know what I am doing and I don't have the skills a lot of good 1.5 trainers have.

[smoky oak]

2.1 is inherently flawed because it has a worse text encoder, meaning that it has a worse understanding of what your inputting, so while it may have a better data set, it's not capable of retrieving the details from that data set is effectively as 1.5

It also makes it a huge pain in the butt to train, because you need drastically different training parameters based off of what kind of things you're trying to train into it

The devs in the server themselves said that, 1.5 utilizes the open AI clip text encoder, Wallace 2.0 and 2.1 utilize an in-house text encoder

*while, not Wallace lmao

[hasty nova]

That to. Hopefully v3 if it is anywhere near combines the strengths of both 1.5 and 2.1, though I kinda doubt it because of where a lot of things seem to be headed.

[smoky oak]

As I have been talking to the devs about it, 3.0 is supposed to have a considerably better text encoder than even SDXL does

[digital hornet]

ngl I assumed you can just switch text encoders out when retraining/finetuning and at most just need to train a bit longer once

[hasty nova]

Ooo great! I'm excited. Hopefully it releases in the same ways as 1.5 and 2.1, where it's still completely free for anyone.

[smoky oak]

That would require retraining the entire base system of the model, which is the hardest part of AI. Are fine-tuning models on top of the base of stable diffusion is nothing in comparison. I believe stable diffusion 1.5 took something like 180,000 hours on it a100s to train

Sorry, voice typing is tripping me up at the moment

My bad, 1.5 was around 79,000 A100 hours

[hearty karma]

id assume they mean its bigger but not necessarily better

[smoky oak]

Yeah

It has a lot more information, but not all of that information is high quality

More than likely, the referring to the fact that the model is not just a one trick photorealism pony, but you can still get digital art and stuff out of it. Which in the grand scheme of things is not important when we have several different specialized models

[digital hornet]

I havent looked into it obviously but at least with previous techniques and models it would be the kind of thing that you can switch, freeze everytthing but the most relevant layers for the encoder so you train it to use the n ew one quicker, and then train a bit all of it and then finetune the rest, but if it's not done I guess it's not as simple here

so it shouldn't be as bad as training from scratch but I guess it can still be much worse than just finetuning the part people typically finetune anyway

[hearty karma]

👀

[smoky oak]

Very interesting, I have experienced almost 0 noise offset from kt

[oak osprey]

you can do better than that now thanks to pytorch2 and xformers and accelerate.

compiled graphs give something like a 46% speedup when training on an A100

so in theory you could hit 40k hours and get SD 1.5 now if you do everything correctly. wonder how many of those hours were spent fixing issues

[smoky oak]

Yes, but you also have to factor the compiling of billions of images

[oak osprey]

i assume the dataset prep is separate

the dataset is reused across projects etc

[smoky oak]

The data set changes with every major release of SD

[oak osprey]

yes and no

[smoky oak]

1.5 and 2.0 have different data sets, as well as 2.1

[oak osprey]

the previous work still contributes to the new dataset

[hearty karma]

man openpose sucks for complex poses

:-)

[smoky oak]

I remember when we couldn't control poses at all lol

[digital hornet]

tbf that part matters much less when people finetune enough

[smoky oak]

I generated a hot Samoan guy, and tell me why it made the background the American Flag lmao

[digital hornet]

we are all living in america

[dense tapir]

Cause it smelled what the Rock was cooking?

[hearty karma]

so wait sytan which model did you download?

the 5gb or 1gb

[smoky oak]

There is no 1GB

There is 5 and 17

Oh, I didn't realize, my bad

[hearty karma]

i was on about the pruned model

[oak osprey]

LAION 2B (and subsets) were used for both SD 2 and SD 1.5, while LAION 5B is used for SD 2 and 2.1, which is a superset of LAION 2B: https://laion.ai/blog/laion-5b/

it was also done by a separate group than the folks who trained SD, though SD provided compute (again)

[smoky oak]

Testing something about this pruned model, hmmm...

[oak osprey]

man it costs like $12,000 just to store the data for LAION 5B with hard drives, locally, not counting electric use

[smoky oak]

Nope, nevermind, still can't do men properly

[oak osprey]

i hate how the SD devs cropped all the images in the training set when they were too large

good stuff if you want out-of-frame half-missing subjects

[smoky oak]

Yup, cropped rather than down sampled, very annoying

[smoky oak]

WAIT WHAT

👁️

Bro is 6 miles away

[acoustic moth]

any way to disable vae in text to img? outputs been ass on certain models since installing

[smoky oak]

@bold cave membersense 2021, 47 items sold, 18 five star reviews 👁️

I might have to get it lol

[smoky oak]

Honestly, I can probably sell my 3060 ti for around 300

[oak osprey]

im hindustani and i find this quite funny so anyone offended just go away 🤣

[dense tapir]

I never saw this before - WARNING:fvcore.common.config:Loading config D:\stable-diffusion-webui-ux\extensions\sd-webui-controlnet\annotator\oneformer\configs/ade20k/oneformer_swin_large_IN21k_384_bs16_160k.yaml with yaml.unsafe_load. Your machine may be at risk if the file contains malicious content.

[oak osprey]

it's facts, bro

[digital hornet]

you can just set VAE to None, next to model

if you dont have the VAE option you can enable it so you see it from settings first

[oak osprey]

set vae to none?! what the hell does that do

does it stick with default vae?

the 'hard baked' option

[hearty karma]

i think that just turns vaes off

[oak osprey]

you can't lmao

that's what make the img go brrrrr

[hearty karma]

really? 🤔

[oak osprey]

as i understand it they're responsible for changing the weighted noise into the img, which is why the VAE matters so much

[acoustic moth]

where in settings can I toggle the vae option

[digital hornet]

hah I was just checking if I can get a 300 for my 3070 when I get the 4090 and related crap

[oak osprey]

hmmm apparently VAE weren't in use by default before 1.4

[digital hornet]

idk they asked about disabling vae

[oak osprey]

i guess you can revisit and explore the terrible outputs from 1.3 era.

@digital hornet i tried to set mine to None and it blew pytorch up but i probably did it too naively

[digital hornet]

I think it was only 2.1 models you can do without VAE? not sure, I just use a custom one mostly

[wispy nest]

I sometimes wonder if the ai know what a "headwear" even is :P
also collars :P

[smoky oak]

The reason I think I can get 300 for my 3060 TI is because it's an FE

Which is basically like liquid gold

My only gripe about this 3080 is it's MSI

[ripe cedar]

How can a face be turned into a face made out of something else? For example, a famous person's head in the clouds made of cloud

[oak osprey]

buy Zotac ahahaha

@ripe cedar double-exposure diffusion

[smoky oak]

Yeah see, $600 for a 3060ti FE used lol

[oak osprey]

for 2.x you have to enable the upcast final layer to float32 or some shit

[ripe cedar]

I looked at that, I don't think that's what I'm going for. It's more like the entirety of the face is made out of it rather than a face among it

[oak osprey]

wut

[ripe cedar]

I tried shuffle and canny ControlNets but it just made a face among it and made outlines of the canny line while parts were transparent

Something like this is what I'm trying to go for, like jabba the hut made out of ice cream

(not mine)

[hearty karma]

why 300 for a fe 3060

[smoky oak]

Cause FE's are extremely hard to get

[dense tapir]

SMH at those prices.

[smoky oak]

I am 99.9% decided on buying this GPU

[dense tapir]

which? You already have those

you have a 3060ti

oh, you mean a 3090?

[wispy nest]

Sytan going to do Sytan things

[dense tapir]

a 3090 for 599 is get it

[smoky oak]

@dense tapir somebody a couple miles away is selling a 3080 for $350

[smoky oak]

That's for the 3060ti :/

[dense tapir]

3080 10 or 12?

[wispy nest]

i can get a used 12gb 3060 for 300$

[smoky oak]

Not sure

Asking right now

[dense tapir]

a new 3060 with 12 is 320

[smoky oak]

Alright he said it's a 3080 MSI 10GB for $350

[dense tapir]

ugh

the 12 I would have said to grab it

[smoky oak]

Still, even just 2GB more is huge for LoRA training

I am confident I could sell my 3060ti FE for $300, so it's really a small jump

[dense tapir]

3080 12gb tears the 4070 12gb to shreds until DLSS3

for our purposes no competition

with 12gb you begin to get into the dreambooth realm

[wispy nest]

can the p40 24gb do dreambooth?

[smoky oak]

yes

[dense tapir]

the problem is none of those cards have any cooling to them

they are server rack cards

[wispy nest]

thats not a problem. have 3d printer and extra fans

[dense tapir]

the fan adapter sticks out the back with a blower fan is as loud as a damn jet

so loud you can barely talk over it

[wispy nest]

that matters on which fan you get

[dense tapir]

smh, you will soon see

it requires a shit ton of air pushed over it as it has no heat pipes.

Ever been in a server room? There is a reason they now have ear protection before you open the door.

[smoky oak]

I wonder how much faster a 3080 would be than my 3060ti in gens

[dense tapir]

A lot

[wispy nest]

how much ram would you need to run a p40 though, 16gb work

[smoky oak]

and those 2 extra GB VRAM would allow me to run vicuna 13b as well...

[wispy nest]

with just a cheap ryzen setup

[wispy nest]

thought the 3060 was 12Gb or you just have the 8gb?

[smoky oak]

I have a 3060ti

[hearty karma]

that says ti on it

[dense tapir]

ti is 8gb

[oak osprey]

that's because they have more than one GPU in that room

[dense tapir]

Don't rationalize with me. The fucking fans are hideously loud

[oak osprey]

@wispy nest basically you're going to get all of these lazy guys telling you not to do X or Y because "it'll be hawd" but it's all fine, people have done it. get a liquid block for the P40 and be happy you saved like 2 grand

[dense tapir]

Now if you stick it in a garage then that would be fine

[oak osprey]

stick it into a mineral oil bath

[dense tapir]

You and that damn mineral oil

[oak osprey]

it keeps me alive

[wispy nest]

haha

[oak osprey]

don't let him pour me the fuck out

[wispy nest]

shrek sticks theirs in the swamp water

[dense tapir]

I tried to find a water block for a P40 but it is so old I couldn't find one. I think I did finally and it was 300-500USD. Then I would need my open loop too and that isn't cheap.

[oak osprey]

you just need a fish tank

truly open loop

[dense tapir]

There is a channel I watch that uses all those old things.

[oak osprey]

i already have EKWB fittings, water blocks, radiators, high SP fans

[dense tapir]

M40, but don't do K40 it is two 12s

M40 is pretty eol so don't touch it either

[wispy nest]

as long as you get something on the die you should just use some clamps and call it good

[oak osprey]

yeah make sure the clamp is metal, and have it properly rubbing on the contacts on the backside of teh PCB so that it holds it as tightly as it can. it's okay if you crush the PCB and the solder joints a bit

[wispy nest]

worst that can happen is you break a 200$ card

or you fry your whole system

hahaha

[dense tapir]

Oh, and the P40 takes a special cable for electric as the standard is for GPU not Accelerators. Cheap around 14 bucks on Amazon

[wispy nest]

yeah i saw that

my only issue i can think of is if itll fit next to my other gpy

gpu

[dense tapir]

No idea with the big bastards they are now

[wispy nest]

maybe ill just take it out and use the integrated display

[dense tapir]

P40 is slim

[wispy nest]

i have a x570, its absolute shit motherboard

wouldnt mid frying it tbh

haha

[oak osprey]

i have a soldering iron

and i knows how to uses it

[oak osprey]

the issue is vram, not performance, @dense tapir

if something is going to crash rather than run, any performing at all is an improvement

[wispy nest]

i like intel name for their cards, its sexy

[oak osprey]

what a deal to get a 3% performance loss from Intel

[dense tapir]

Just letting them know.

[wispy nest]

shrek literally lives in a swamp and has a better PC than 99.9% of people

[dense tapir]

See it runs 250w now let's say I got a 3090ti that is 450w. the P40 takes 5 times longer to do the same job. I will gladly pay for a 3090ti because I will feel that with as much training as I used to do.

[wispy nest]

honestly should just move my pc to spare bedroom and buy a 20ft HDMI cable to my monitor lol

[oak osprey]

friggen

colour lost

[wispy nest]

3090ti is 8gb also

?

[dense tapir]

24

3090 is 24 as well

[wispy nest]

damn that is like 5x as expensive as a P40

[oak osprey]

i was looking at 3090s and decided the $800 difference to the 4090 was worthwhile

[autumn moss]

Please generate an image of a 3-year-old girl with realistic features. The girl should have white skin with freckles on her face, and her hair should be styled in two ponytails. She is wearing a white dress with red polka dots and is walking on a street with a reddish sunset in the background. Please ensure that the image quality is excellent, with sharp details and realistic colors.

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐｜pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610⁠ for more details! You can also stop by ⁠#1025467151206854736 for any issues you experience while using DreamStudio or ⁠#🤝｜tech-support for any problems you encounter while installing it locally!

[wispy nest]

why cant we just hotglue 4 1080's to get a 4320 GPU

[oak osprey]

@autumn moss yikes on trying to generate a 3 year old

[hearty karma]

💀

[wispy nest]

prob an fbi agent looking for easy targets

[autumn moss]

Estoy aprendiendo a utilizar vamos a probar

[oak osprey]

nada mas no entiende, sorry

[wispy nest]

si

[dense tapir]

@wispy nest P40 is 183 GFlops 3090ti is 40TFlops. That is about 214 times faster give or take.

[oak osprey]

i can make you a 3 year old, but they ain't gonna be human

i wish the A100 weren't so slow. its only upside is 80GB VRAM

all these people raving about how fast pytorch 2 goes on the a100 are only because it's so god awful slow

[wispy nest]

i should really just upgrade to a 4080

[static tusk]

what it/s do you see with A100?

[dense tapir]

Only 16gb. 😦

If the 4080 were 20GB I'd say go for it

[wispy nest]

4090 than?

[dense tapir]

I am saving for one, yes, or the 7950XTX if I see some real world people trying it

Fuck used

Unless you know the person then cool

nope, DOA

Titan is

Seems Nvidia has scrapped the 4090ti and is concentrating on the Titan and 48gb

[wispy nest]

why the 7950XTX if no cuda

what

[dense tapir]

I am sad because a ti would have pushed the 4090 cheaper but titan will do nothing to the consumer grade cards.

[hearty karma]

i think nvidia hates its consumers

[dense tapir]

it does and

tbh, most expect it to go full on AI/ML and leave the consumer gpu market. Let Intel and AMD have that in the next 10 years

[wispy nest]

thats why you get p40 and just go sloth mode

[dense tapir]

SLoth mode, hehehehe

[dense tapir]

I hope they do it as well so I can finally kiss their ass on the way out. Here ya go Jensen.

[wispy nest]

tbh yeah why would you even need 4090 unless you generating cat pics

[hearty karma]

the hand went a bit buggered but

[dense tapir]

Games are getting stupid cause of the lack of any optimizations

Indie games are much better at it

[wispy nest]

do people actually game in 4k

[hearty karma]

games are getting bad anyway

most games recently seem to be launching as bug ridden messes

[wispy nest]

i dont see the point of 4k gaming unless its a fpv

its cuz i dont play games and just want more cat pics

[dense tapir]

big screen 140+ hz 4k yes

why I said 140hz plus

as I said why I said 140hz plus

[wispy nest]

some QLED

yeah 4k

i dont even own a tv tho

just bought a definitive audio surround sound system and doesnt have tv lmfao

7.2 has bluetooth

100w per channel

blacks look like shit on any LED tv

was trying to find that tv with dual screens, but almost impossible to get it

[oak osprey]

jeez i came back to all of you Boomer'ing out on me

"wah video games arent fun anymore" man we started out with god damn pong

you can go back to that. it likely still runs in a museum somewhere

😄

[oak osprey]

wonder if a cat is cheaper than a gpu

if we're using a frozen moment in time at purchase, the cat wins out. over the factor t (time) we're going to see the cat spend a lot more of your money

speaking of whom, she slept all day and will keep me up at night now

when you guys finetune* a VAE, what is your goal and how does the dataset look?

[smoky oak]

@dense tapir on my way to buy it

[oak osprey]

the WSE 2?

[smoky oak]

Lmao

No, a 3080 haha

[oak osprey]

dont text and drive

[smoky oak]

Somebody is selling one for $350, I'm gonna buy it and sell my 3060ti for around the same

[oak osprey]

[smoky oak]

I can't drive, had to get my mom to take me

[oak osprey]

👶🏽 ?

are you an AI Language Model?

[wispy nest]

[oak osprey]

[smoky oak]

Lmfaoooo

[oak osprey]

bro i been chattin with a kid 💀

[smoky oak]

I'm 19!

I just can't drive lmao

[oak osprey]

im raggin on others for talking to young kids and you're not telling me "uhm mister" 😭

[smoky oak]

IM NOT A KID BRO

I can vote

[oak osprey]

ok true

[smoky oak]

:C

Lmao

[wispy nest]

Mister shrek

[oak osprey]

you are at the age where i can start to begin blaming all of my problems on your generation

[smoky oak]

Sir Shrek's alot

[oak osprey]

that's Mister Shreks-a-lot to you

[smoky oak]

[oak osprey]

i got big jars and i cannot lie

[smoky oak]

I do have my driver's license test in 3 weeks tho

[oak osprey]

guys if the Dreamworks or Pixar guys come in here, THAT IS NOT SHREK. ok? it's just some green ogre ok no it's not even that, it's just a fantasy character

[smoky oak]

Well. For my permit

[oak osprey]

did you guys hear they're adding Shrek to the next Final Fantasy?

[wispy nest]

Same

[oak osprey]

i cant wait to be 69 years old

[wispy nest]

I'm old enough to hate the next generation also

I'd rather live in the woods with green ogre than deal with them

[oak osprey]

that's MAH swamp

[wispy nest]

I bring gallons of peanut butter and waffles

[dense tapir]

Let me know your it/s after you get it installed.

[smoky oak]

Will do

[wispy nest]

Just raise chickens

[smoky oak]

I am filled with anticipation

[wispy nest]

[dense tapir]

So am I

[wispy nest]

So you were rendering this fast just on a 3060?

[smoky oak]

3060ti

[wispy nest]

That's still fast

[smoky oak]

I get about 11.5its/ on my 3060ti

The slowest part of the comparisons was the model loading

[static tusk]

this is my 3080 as a comparison for benchmarking

[dense tapir]

Drive speed is a lot of it as it takes me 2-3m the first time around due to 100% HDD activity.

[smoky oak]

I was having to load that 17.4 GB ultra model, so it was taking like 30 seconds a Gen just for that

[dense tapir]

once in my ram it is 2-3s

yeah

I had to go to CL16 from CL14 when I added 32GB more. I was at 18 but managed to tweak it. Man, at 3200 CL18 vs 3200 CL14 I felt that

[static tusk]

Model loaded in 2.3s (create model: 0.1s, apply weights to model: 0.6s, apply channels_last: 0.3s, apply half(): 0.5s, move model to device: 0.8s)

[dense tapir]

I am not going to mess with NVME as M2_2 takes all my slots away from me so I will raid 0 a few SSDs.

I have it in my UEFI

even windows10 does software raid

a tad slower but not by much

[static tusk]

I tried raid ssd for a while . its slower for speed, you just get more storage

[dense tapir]

Well, that is what I am going to do since I will be damned if I do another nvme and lose all my slots

[wispy nest]

What

[static tusk]

I meant more storage on 1 drive, like C: drive

[dense tapir]

I saw a crystal benchmark 5 seagate HDD in raid zero. The same one I use (2tb) and he had 10tb and was as fast as my gen 3.

NVME M.2

[oak osprey]

i use RAID-1 on ZFS

[wispy nest]

I use m.2 for all software and boot, and hdd for storage

[hearty karma]

haha yes they are...

[oak osprey]

2x 16tb mirror for /archive and 1tb nvme for /

[hearty karma]

i would know since im using a hdd as my os drive

[static tusk]

I also used RAM drive for a while. then NVME came out and its basically the same speed

[hearty karma]

yes it does suck

[oak osprey]

hdd as OS drive is okay if you have enough RAM to store your "working set" of data in the cache.

[hearty karma]

no its not ok

[oak osprey]

and never reboot

[hearty karma]

it takes years to bloody load my computer

[oak osprey]

it'd be way faster if you ran Linux.

[hearty karma]

and im stuck on it until i can get an ssd to replace my old ssd

[dense tapir]

5 2tb seagate in raid 0 sata HDD is about 3.3-3.5GB/s. That is my nvme m.2 gen 3 speed.

[wispy nest]

M.2 and ssd are same speed unless you on different bus ports 🤔

[dense tapir]

who cares, make backups

[hearty karma]

welp im going to disappear now