🏞｜general-with-images

[wispy nest]

it triggers the lyco

[wispy ether]

what is polydactyl

https://cdn.discordapp.com/emojis/986075701683838986.webp?size=128&quality=lossless

[wispy nest]

https://en.wikipedia.org/wiki/Polydactyly

[wispy ether]

https://cdn.discordapp.com/emojis/986075701683838986.webp?size=128&quality=lossless

i give up

[wispy nest]

see

no more fucked up fingers

no fingers at all

[wispy ether]

lmao

[wispy nest]

😂

[karmic sedge]

Still can't use more than 75 tokens when using TensorRT

I think it's because of this. No idea.

[wispy ether]

https://tenor.com/view/daeth-funi-periflight-dies-from-gif-22212023

[karmic sedge]

Anyone here who has a link to cudnn 8.9?

[wispy ether]

googl

nvidia website

[karmic sedge]

I don't have an nvidia acc

[wispy ether]

make one

[karmic sedge]

*dev acc

[wispy ether]

make one

[karmic sedge]

[oak osprey]

you can't use more. it's a compiled model

the way it works to use more is to hijack those functions but you can't if they're compiled so they run only on the gpu

[karmic sedge]

Okay, but why are we able to set more than 75 tokens when converting from onnx? That's a bit weird. 😄

[oak osprey]

different starting model config

[wispy ether]

@wispy nestany idea

2 arms

[karmic sedge]

Oh I get it now, thanks.

[oak osprey]

i hate tensorrt, removed it from my systems in the last cpl days because it was making bitsandbytes stop working, which means i couldn't train a model

[wispy ether]

for me it made it slower

[karmic sedge]

I would love it, if it would allow more tokens :S

[wispy nest]

well

idk what you are doing

[wispy ether]

how to not make it be 2 arms

its 2 and a half

[wispy nest]

[wispy ether]

tf

did u do

how

[wispy nest]

left hand open, right hand closed, thumb, fingers, palm, knuckles, hand, nice hands, perfect hands, 1girl, angel, wings, jewellery <lyco:GoodHands-beta2:1.0>

masking the hands

spamming inpaint

[cyan snow]

ok, got the extention working

so, i convert safetensor or somthing to onnx, then the onnx to tensorrt?

[karmic sedge]

Yes

[cyan snow]

and what do i do with minimum height and maximum height and all that?

it 512 on everything by default

[karmic sedge]

512 minimum and 768 max

[wispy nest]

@wispy ether just inpaint on ur image

you can upscale things even with just inpaint

by adapting your prompt

[strong sonnet]

Not for GPU usage?

[cyan snow]

im getting AssertionError: onnx model not found

[karmic sedge]

You have to write the full path

+filename

And wait if it seems like it is doing nothing

[cyan snow]

it says:FileNotFoundError: [WinError 3] The system cannot find the path specified: ''

[karmic sedge]

Post a screenshot plz

Of your webui

[oak osprey]

if you're downloading a model, uploading one, or a dataset, you pay $0.1/GB or so

[karmic sedge]

leave the second line blank and try again

[oak osprey]

trying 2x A6000s instead of an A100-80G this time

[strong sonnet]

Which GPU?

[karmic sedge]

Beautiful

[oak osprey]

what? all of them

wild, its workin

[strong sonnet]

Even a100

[oak osprey]

... yes...

literally EVERY machine there bills for network

and when the instance isn't running you pay for storage

[wispy nest]

https://youtube.com/shorts/LH7RG0M1rnY?feature=share

[oak osprey]

so you have to delete and re-download everything if you don't want that

[dense tapir]

no

[wispy nest]

thats some heavy machinery

[wispy ether]

what is average it/s on 4070

[wispy ether]

HP

[wispy nest]

https://cdn.mos.cms.futurecdn.net/iURJZGwQMZnVBqnocbkqPa-970-80.png

[wispy ether]

https://tenor.com/view/its-real-bad-not-good-awful-terrible-its-bad-gif-16985162

[wispy ether]

theres no 4070

[wispy nest]

just get a 4070 TI

[wispy ether]

no

https://tenor.com/view/uwu-ts-team-uwu-gif-21159211

cpu bottleneck if 4070ti

what a weird scythe

anyone got a lora for a scythe

-.-

[oak osprey]

just jammin

[wispy nest]

gotta make one

not even the base model knows what that object it

[wispy ether]

lol

i casnt make one

its cursed for me

no matter what i do i get exit code 0

even if i do everything right

[wispy nest]

i looked at lexica

SD doesn't know what a scythe is

[wispy ether]

lol

LMAOAOAOA

[wispy ether]

hmmmmmmmmmm

machete

https://tenor.com/view/asktheboywholived-ttbret-trigger-mason-knife-gif-15292903

[wispy nest]

https://lexica.art/?q=sickle

[wispy ether]

[dense tapir]

@smoky oak as well as a few others were/are perplexed at those numbers wondering where Tom grabbed them from as most feel it had to come from his ass.

[wispy nest]

i made them up

A scythe is a tool used for cutting grass or harvesting crops. It has a long, thin, curved blade that is attached to a long wooden handle. The blade has a sharp, inward-facing edge that is used to slice through vegetation. The handle is typically made of wood and is curved to allow for a comfortable grip while using the tool. The blade is typically made of metal and has a skinny, long shape that tapers to a point at the end. The entire scythe is usually quite tall and is designed to be used while standing.

[dense tapir]

You didn't make those numbers in the graph you posted up but Tom sure in the hell did.

[wispy nest]

yeah probably

its called journalism in 2023

making shit up

[dense tapir]

Oh, boy yeah

[wispy ether]

i forgot hires fix

[dense tapir]

"Reliable sources" is a great one

[wispy ether]

where can i see all arg commands

[wispy nest]

describing a scythe to SD

[dense tapir]

SD has never done weapons well or at all. Scythe was one of the first ones, if not the first one, I attempted in 1.4 then 1.5 then 2.0 to 2.1. You really want it you need to train them in.

[wild sorrel]

yea...I spent alot of time trying to do sword

[wispy ether]

https://cdn.discordapp.com/emojis/992146516921831485.webp?size=128&quality=lossless

[wispy nest]

what we need is a model that can learn on the fly like gpt

[dense tapir]

yeah, it has a hard time with a damn butter knife

[wispy nest]

i'm sad that we are stuck with 1.5 here

[dense tapir]

I left 1.5 the day 2.0 hit and each time I try it I am frustrated by it. By it I mean with 1.5

[wispy nest]

i tried 2.1 and didnt like it

[wispy nest]

like most people also did ;P

[dense tapir]

for realism I much prefer 2.1

[wild sorrel]

That was one of the best sword gens...
Didn't bother to inpaint

[dense tapir]

for all the Anime waifu freakazoids they stayed with 1.5

[wispy nest]

idk what it is but i feel like 2.1 is less responsive

[wild sorrel]

I do everything on 1.5

[cyan snow]

SDXL is by FAR the absolute masterpiece, and it's not even released

[wispy nest]

2.1 can have better details, but can doesn't mean will :P

doesn't SDXL take ungodly amounts of vram

[dense tapir]

Clip is why because you have to relearn it all. Now one thing I so dearly miss with 1.5 is all the photography stuff it knows as they trained that into it.

SDXL 12gb+

[wispy nest]

I've never seen a image from it so I can't answer that :P

i wish we had 1TB Vram gpus

[wild sorrel]

yea...like half of the community won't even be able to use it SDXL lol

[wispy nest]

just superglue a m2 to the gpu

[cyan snow]

more parameters doesn't instantly mean higher requirements

[dense tapir]

the devs said in time they will make smaller models so my 6gb can run it. I doubt that or if it does at what quality? Potato?

[dense tapir]

That came from one of the devs directly.

[wispy ether]

Total progress: 94%|████████████████████████████████████████████████████████████▏ | 282/300 [09:58<01:15, 4.20s/it]

[wispy nest]

i'm disappointed in the hardware market in general rn

[dense tapir]

Same

[wispy nest]

huge sizes, inefficient, overpriced, low Vram

[dense tapir]

700-1k mobos? wtf?

[cyan snow]

well, from my experience, on llms for example, more parameters just means that it takes more time to load, but shouldn't effect speed, or at least that is how it is with llms

[wispy nest]

we need dedicated AI architectures and hardware

not this GPU shit

[dense tapir]

I am just telling you the dev came in on the channel and flat out said that at first 12gb, maybe 10, is needed but in time they will reduce it so less vram will be needed. NO, at first my 1060 will not be able to run it.

[wispy ether]

gawd

[dense tapir]

I suspect ASIC tbh

[wispy ether]

i just realized i used 150steps

[cyan snow]

weird

well, i guess only time will tell.

[dense tapir]

Right now they are about 50-54% done

[wispy nest]

this hardware is such a massive bottleneck for the technology its insane

[dense tapir]

I am laying odds on an October release.

[cyan snow]

they were ~50% a month ago, id say atleast 60

[wispy nest]

I've learned long ago that words are cheap to make :P

[dense tapir]

it isn't linear and there are setbacks as devs said.

[cyan snow]

it's doing the same kind of training that MJ did

hand-picked dataset

[dense tapir]

Well, that's cool but so far all the stuff I have seen I really do not care for SDXL. If it remains as it is now it will be a lot of hype then all the people will run back to 1.5.

The biggest deal killer for it is can it easily train Pornhub and Anime waifu shit. If it can good but if it is like 2.x forget it

[celest mountain]

hey guys, which version of CUDA and torch should I use now?
haven't used SD for a while, still on torch 1.x, but I heard SD wants torch 2.x now

[cyan snow]

that's because they fucked up 2.1, they won't make a mistake like that again, i hope

[dense tapir]

I firmly believe they have.

[cyan snow]

i use 11.8, and cudnn 8.6

[dense tapir]

The fuck up they made is in the TE for training and that has not been changed it is the same 2.x CLIP.

[wild sorrel]

I think it's mostly about how good is it at understanding propmt and training.
If it's good at both - people will train nsfw, waifu shit and whatever else they want

[cyan snow]

idk, the generation resaults seem promising..

[celest mountain]

so the latest CUDA 12.1 isn't supported (yet?) 🤔

[dense tapir]

Thanks for the affirmation.

[cyan snow]

it is, but slightly slower for some reason

[celest mountain]

cool, thanks 🙂

[dense tapir]

We shall see. If it is like 2.1 only prettier it will be a dismal failure so I hope with them sticking to the newer CLIP they actually get up off their hind parts and release, at the same time, the training tools made for it not expect us to use the same tools we did with 1.5 as they are not good with 2.x and will not be good with XL either.

That TE actually fights you while training and why we don't have Controlnet from the original people for 2.x. Loved their open letter about this as it was 100% spot on.

[wispy nest]

whats TE

[dense tapir]

Textual Encoder it is the actual CLIP

[wispy nest]

ah

we need a multi modal enconder

[dense tapir]

Never

[wispy nest]

controlnet is an attempt at that

[dense tapir]

Sort of and the devs, basically, told SAI off and said eff you.

[wispy nest]

what

[dense tapir]

Was an open letter to SAI and just how they fucked up with 2.x and the clip they used

We still lack models in 2.x I wish I had but at least we have something thanks to someone who trained the very few we do have.

[cyan snow]

let's be optimistic and hope the next model SAI makes will fix all the mistakes, we all make mistakes sometimes

[wispy nest]

controlnet is the best input extension that exists for SD

we need more control over the model

[dense tapir]

I agree

[wispy nest]

the ability to input edges, poses, reference images, its amazing

[dense tapir]

When dealing with SAI, and emad, I lost all optimism ages ago so now I just sit back, shrug, and whatever happens happens., I am not jumping on hype trains again. Sit back, try it, and decide then.

[cyan snow]

emad?

[dense tapir]

Have we yet received any specific tools from SAI, or even an API, etc... from SAI for 2.x? I have not seen any YET emad did say, last year they were coming. Should have been here by now.

[wispy nest]

didn't they make an open source interface

of their dreambooth

[cyan snow]

they did, it's very handy!

[wispy nest]

i wish it would support all the extensions

[cyan snow]

can be combined with a1111 to support all features

[cyan snow]

it does

[wispy nest]

mhhh

[cyan snow]

if set up correctly, that is

[dense tapir]

Emad is a former hedge fund manager who wants to build SD into a "$100 billion" business (revenues? market cap?). He is the CEO of StabilityAI.

[cyan snow]

sound like a humble guy

[dense tapir]

LOL

[wispy ether]

UYS

GUYS

guess what

[wispy nest]

shouldn't the focus be to make a good product

[cyan snow]

what is it

[wispy nest]

i see the same mistake in new streamers who just want more viewers but they dont actually make content

[wispy ether]

nothing

🐗

[dense tapir]

This is just a side gig for them.

[wispy nest]

i mean the way to do it is to just hire a bunch of data scientists and artists

hire people who make good extensions

most of all people with good ideas

[cyan snow]

yeah, that was the goal of this community at first

[wispy nest]

but ?

[cyan snow]

ask them what the fuck happened

[wispy nest]

i dont know, i never used the service

i think services are the wrong way to go about it anyways

[boreal falcon]

Really?

You never used SD?

[cyan snow]

the service? SD is open source

[wispy nest]

i mean the webservice

the way they make money?

i made like 20 images, then ran out, they only take credit card with no other payment methods

basically useless for me

[cyan snow]

there wasn't supposed to be a webservice in the first place, it was supposed to be full open source, no capitalism

[wispy nest]

the better way would be to offer things like fine-tuning, lora services, 3D model like marketplaces where people can sell loras and request loras for money

offering a platform

like gumroad for AI models

then funnel the returns into development

[cyan snow]

bro. there would be SO many leaks, it would be like what happened with NAI all over again

but that's a beautiful dream, i agree on that

[wispy nest]

i mean do they want money, or do they want to change the way we make art ?

i'd like to think they want to revolutionize art

and money is just a sidequest

[cyan snow]

good question

[wispy nest]

seems its the opposite

same shit with adobe and all the other corps, everyone just trying to make as much money as possible

in the future there will be AI designing architectures for models aswell as hardware and these corps will be reduced to hardware hosts

and fabs

the sooner we get there the better for everyone

[cyan snow]

Nvidia, without them ai would be basically impossible

[wispy nest]

well

i think even this duopoly of amd vs nvidia is on the verge of collapse

[cyan snow]

amd pretty much gave up, ROCm has a date of death, CUDA and cudnn is still evolving.

[wispy nest]

i think amd will use their access to fabs and devs to make dedicated tensor accelerators

and as usual, enter the market to provide good $/ops

[cyan snow]

well, currently Nvidia is destroying them, so idk.. i guess only time will tell

[wispy nest]

but we will have china entering the market with photonic chips and resonance based hardware that will 100% turn shit upside down as soon as their fabs are finished this year, and mass producers step in with manufacturing requests

[strong sonnet]

so, when its using the gpu its around $0.3/hr and when gpu isnt in use its $0.1/hr for every GB?

[wispy nest]

and china allegedly finishes their first fabs at the end of this year, with mass manufacturing starting in less than half a year after that

[cyan snow]

what are they manufacturing?

[wispy nest]

photonic chips

ONN

optical neural nets

[cyan snow]

pretty much a TPU?

[wispy nest]

similar, i think they have something more generalized in mind

[wispy ether]

https://tenor.com/view/lucifer-wings-spn-supernatural-gif-14251133

:)

[wispy nest]

wings out the ass

[wispy ether]

the

e

e

MAN

ok no

its the

the bottom part of the coccyx

or what its called

the cross

[wispy nest]

from what i read the issue with the hardware right now is accuracy in programmability and memory

[oak osprey]

i don't know what the storage cost. the .1/gb is network bandwidth. i'm confused at which part of this is a struggle

[wispy nest]

its easy to just send shit through a photonics setup, but the memory takes ages compared to the computation

for example there are papers about CNN with ~600 picosecond computing times for image recognition

which is insanely fast

[wispy ether]

chatgpt please code a virtual gpu

:(I

with 200gb vram

[wispy nest]

but there is no memory that operates at 600 picoseconds

[cyan snow]

will a photonic chip be able to run SD?

[wispy ether]

https://tenor.com/view/d-dos-computer-virus-network-gif-14274332

[strong sonnet]

so when running its 0.3+0.1?

[wispy ether]

quantum pc

[wispy nest]

i dont think so, because it is iteratively and requires loading and unloading from memory

unless we find a way of looping back iterations inside a photonic chip and then we just read out the resulting image

[oak osprey]

it varies

[cyan snow]

So, cuda is here to stay when it comes to ai, huh?

[wispy nest]

i mean, we need dedicated AI memory and not vram

as in, vector memory that works in the picosecond range

[wispy nest]

ofc they will expand on cuda

[strong sonnet]

I mean gpu cost + 0.1(bandwith)?

[wispy nest]

it will be a gradual change from silicon hardware to fully neuromorphic hardware

as in, memory + computation in one chips, no external memory needed like in vram

https://www.nature.com/articles/s41467-021-25427-4

[sterile kiln]

Can someone remember me the command for beginners? Its for my friend

[celest mountain]

screw CUDA, ai needs some new dedicated hardware

[cyan snow]

bro where tf am i

[wispy nest]

liminal space

https://tenor.com/view/backrooms-smiler-attack-jumpscare-gif-gif-25930658

[wispy nest]

https://youtu.be/y6X6k9jwKc8?t=705

they are working on it

[acoustic oracle]

hi i am new to stable diffusion

i am using stable diffusion with the AOM3 model

my negative tags are lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name

and my positive are masterpiece, best quality

it is very hard to make it generate anything other than cp whatever i tell it to make in the positive tag

does anyone know how to fix this

[cyan snow]

this reminds me of that crazy moon from the old zelda game

[wispy ether]

@wispy nest

[acoustic oracle]

what model and prompt are you using

[wispy ether]

the fuc

[acoustic oracle]

it looks really nice

[wispy ether]

HOW DID YOU

@wispy nest

[wispy nest]

https://cdn.discordapp.com/emojis/782015386324303892.gif?size=96&quality=lossless

[wispy ether]

how so fast

betterdiscord"!?

https://tenor.com/view/ban-hammer-cinzou-ban-banido-look-gif-17786909

[oak osprey]

yeah gpu cost + storage + bw

[wispy nest]

better what ?

[oak osprey]

it's expensive

[wispy ether]

better this coc

[acoustic oracle]

unfortunately for me whatever i try i cannot generate anything which does not risk getting the fbi on my doorstep, what prompt, negative or model can fix it

[wispy ether]

jk

[acoustic oracle]

i really want to make cool art

[acoustic oracle]

but my gen is broken

[wispy ether]

https://tenor.com/view/스타레일-gif-10588812559516015292

[wispy nest]

use negative prompts and exclude what you dont want

[acoustic oracle]

ok so maybe i could add "nsfw" to negative prompt

[cyan snow]

did i have a stroke

[acoustic oracle]

ayy now it works

thanks

[strong sonnet]

whats bw?

[cyan snow]

bandwidth

[acoustic oracle]

now i will go delete the images i generated earlier before my ms acc gets termed

[wispy nest]

he just asked for your name, not your life's story! ;P

[strong sonnet]

bandwith is $0.1?
what about storage?

[cyan snow]

i have no idea what you're talking about, i just knew the acronym.

[oak osprey]

holy shit man

[wispy nest]

https://www.youtube.com/watch?v=wrdqsMqmAig

[oak osprey]

this is frustrating, you keep asking the same questions, are you trolling?

[strong sonnet]

no, so its 0.1 for bandwith and 0.1 for storage?

[oak osprey]

i don't. know.

i'm just telling you that they charge for everrryyything

and it's not cheap

[strong sonnet]

ok gotcha

[wispy nest]

essentially its better to buy a GPU unless you want to train large models

fast

[acoustic oracle]

how long does your computer take to generate images and what gpu

my rtx3050ti laptop model takes like 10-15 sec

surprisingly fasts

[wispy nest]

someone's getting fired soon, might be both kinds of them as well :P

[cyan snow]

?

[wispy nest]

from how the ships move, it won't be a pretty sight if they don't turn :P

(it was a joke about two ships colliding and then the one at the helm will be fired from their job, and maybe from one of the ship's cannons as well from what they did :P )

[empty dust]

Hello everyone, can somebody help me fix those feet?
I tried every inpainting method, but nothing seems to work...

[wispy ether]

how to render faster than 0,01s

[dense tapir]

https://supagruen.github.io/StableDiffusion-CheatSheet/

[wispy ether]

My room temperature thanks to stable diffusion

[wispy ether]

3060 3 seconds 512x512 20 steps

5*

[acoustic oracle]

nice

desktop gpu?

[wispy ether]

3050 is the weakest rtx card

Yes

[acoustic oracle]

yeah ive heard laptop gpu is 30% worse than desktop gpu of same name

[wispy ether]

3050 is weakest in general

[wispy ether]

What

[acoustic oracle]

i have 3050 ti, is that weak

[dense tapir]

50 series is the weakest of all of them

[wispy ether]

LMAOAOAOAOAOAO

I CANT

PLS

[acoustic oracle]

i am laptop 3050 ti

[wispy ether]

hi laptop 3050ti my name is nep

[wispy ether]

@acoustic oraclehttps://www.youtube.com/watch?v=A7wwtPekY7g

[acoustic oracle]

hi nep i can keep very cool during any advanced ai image generation and kinda cool during minecraft
but i would rather be lit on fire than run osu, it makes me overheat for some reason

[cyan snow]

tf is that

[oak osprey]

@smoky oak so the terminal snr fixes have now been expanded to all samplers supported by diffusers https://github.com/huggingface/diffusers/pull/3664

[wispy ether]

[acoustic oracle]

ye i really thought a 3050 ti would be like right between 3050 and 3060

[wispy ether]

no...

[cyan snow]

that looks like a driver issue

[wispy ether]

no

3050 is just bad

• laptop

[cyan snow]

oh laptops are stupid

[oak osprey]

it isn't?

[cyan snow]

if that's a laptop then im not confused

[wispy ether]

dwarf

[cyan snow]

JESSE

WE NEED TO COOK JESSE

[dense tapir]

Yes

[oak osprey]

find problems

[cyan snow]

chain is fucked, chair faces through bike, other than that it's fine

[oak osprey]

chair?

yeah the chain is a weird one lmao

there's no deraileur hanger that i can see

[oak osprey]

but it's pretty close

2.1 without SNR fixes can't do bikes that good

[wispy ether]

cook the jesse and win a price

[oak osprey]

weird how much it resolves to apply that fix, man

[acoustic oracle]

made this with aom3

[acoustic oracle]

i like it, i will outpaint to make it bigger

[acoustic oracle]

the left one looks like walter white as a child imo

looks nice

[dense tapir]

Aye

[cyan snow]

instead of cooking a bunch of m*th why don't we cook some models

[dense tapir]

I just don't remember how to do 1.5 prompting these days

[dense tapir]

I would if I had the card to do it. Man, so many I want to train and even when I had colab it was more issues than pleasure.

[cyan snow]

yeah man, i bet when a good new base model releases people will start fine tuning again

possibly SDXL

[dense tapir]

If we can as it seems to be going more and more in the MJ direction which sucks

LOL

[cyan snow]

im not sure if that would be a good thing or a bad thing, this community already has some models that somewhat beat MJ

[acoustic oracle]

this is my outpainted image, 60 steps and its still not good

[dense tapir]

VERY bad thing as that is the cloud based, pay as you go, and no community models/training allowed.

[acoustic oracle]

i used poor mans outpainting script from sd webui

does anyone know how to outpaint better

[cyan snow]

i thought you were talking about the capabilities of the models this community has produced..

[acoustic oracle]

like does anyone have a script

[dense tapir]

That is completely the opposite to the MJ way

[cyan snow]

no, like i thought you ment that eventually people would start to over-cook models like MJ did, i wasn't talking about closed source or open source, i was talking about the sheer quality of the outputs of the diffusion models

[dense tapir]

I don't see this staying fully open is what I am getting at

[cyan snow]

that would be very sad

[dense tapir]

The $$$$$$$$$$ incentive is too great to continue open source

we are all just tools to them to beta test for them.

[cyan snow]

the goal of this community in the first place is the opposite of that

[dense tapir]

I bet a fork happens eventually

yep, community fork based on a prior version of SD

SAI goes one way the community goes the other

[cyan snow]

if stability will go closed source, then someone will take their place, open-source is inavitable.

[dense tapir]

That is precisely what I said

[cyan snow]

HF will defiantly stay i the same position, they are all about OS

[dense tapir]

the someone will be the community and paying to train new models but that shit is not cheap

yes, I like HF

Every single time I see, or type, HF I say Harbor Freight. :/

[cyan snow]

well, but so far SAI never did a closed sourced project as far as im aware of

[dense tapir]

They really don't have to tbh. Just make the model so damn hard to train you don't. They stand back and grab the popcorn while "staying true" to OSS.

[acoustic oracle]

i always think of hv (high voltage) from gregtech

[dense tapir]

2.1 is almost there

this is why I see SDXL and see MJ wannabe. Good for prompting but how will it train? We shall see.

[dense tapir]

Yeah, and MJ is Michael Jackson and I have to step back and realize MidJourney.

[cyan snow]

dude, MJ is really michael jackson

[acoustic oracle]

got an oom time to close some apps

[cyan snow]

both are horrible people

[acoustic oracle]

is there a way to use shared gpu memory

my gpu has 4gb vram but has access to up to 14gb shared memory from my ram

[oak osprey]

getting a realism model to work, is hard

[boreal falcon]

if its free you are the product

famous quote

[oak osprey]

quality 👌🏽

even the patterns on his fabric look good to me

[acoustic oracle]

cool

i just made a space image in aom3

really cool

[cyan snow]

it just makes rich people stuff

[acoustic oracle]

what model

[acoustic oracle]

what model is this one too

[cyan snow]

same model i use all the time

[acoustic oracle]

which is that

[dense tapir]

ArtiusV21

[cyan snow]

i just merged a bunch of models, if i recall correctly i already uploaded it somewhere on this server with Gdrive

[acoustic oracle]

ok

this one? https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0/tree/main

[oak osprey]

have a smoke of some Old Toby

@smoky oak that's the cleanest people face ive made through SD 2.1, ever, i think?

[acoustic oracle]

i have heard 1.5 was better than 2 as the nsfw filter caused worse performance, is that true

https://www.assemblyai.com/blog/stable-diffusion-1-vs-2-what-you-need-to-know/

it says 2.1 is better than 2 but not above 1.5

[oak osprey]

also 2.1

[oak osprey]

that's a myth

some people won't let it die

[acoustic oracle]

but i can see on those images in the test that 1.5 looks better than 2 in all aspects

2.1 is better with skin than 1.5 but worse on other aspects

[wispy nest]

I would say to just look at the result you like, and not the number where it's made :D

[oak osprey]

but that's from December

also that article is about SD 2.0

2.1 was released a few months later and it's markedly improved. anyway, just use models. stop thinking too much about what party has the better music.

[wispy nest]

I'd say that it being from december is in 2.x's favour ;P

[oak osprey]

same, but it's not that interesting to argue because everyone really just wants to support what they love and not change their mind etc

[wispy nest]

I just downloaded about 1000 loras, and at least 4 of them were not of artstyles or women. In a couple months all I can say to people wanting to create something like a man is: Good luck, you'll need it :P

[oak osprey]

when you just feel the moment inspires a music and you have to squat and impress the lady

[oak osprey]

@cyan snow

chain completed and the deraileur is there now

😄

[acoustic oracle]

4gb vram is annoying

ooms everywhere

[dense tapir]

six is a little better but I have to restart constantly

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 408.00 MiB (GPU 0; 6.00 GiB total capacity; 4.95 GiB already allocated; 0 bytes free; 5.18 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

once that hits you have to restart

[cyan snow]

Why tf does it make this when the prompt is :aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa

[oak osprey]

my model can make family photos now but what the heyull lmfao

[dense tapir]

Because it doesn't know who Fonzi was

[acoustic oracle]

i got bluepastel model and it is nice

i tried artius 2.1 but it got me an oom

as i have only 4gb vram

is there a way to use shared memory so i can run larger models

[dense tapir]

--lowvram on startup

[acoustic oracle]

i have to go to bed now but i will try that tmr

[cyan snow]

you.. your!

the prompt was your fired

[dense tapir]

The hand 😦

[dense tapir]

I wonder how much longer before we get conformity and uniformity?

[cyan snow]

conformity and uniformity? sorry, im not that good with words lol

[wispy nest]

when everything starts to look the same

[dense tapir]

well, half right. How can we do a film when we can't even get a character to wear something? I had a great example of this from a Children's book author who has the main character being an elephant. They could not get it to wear a tie all the time, or the same tie, all the time and it was really screwing them over. That was just a handful of images for a Children's book can you imagine a full on short film to even a full length one?

Controlnet helps but not enough plus I am unsure how it would do in a motion video.

[wispy nest]

You lost me there, conformity is when you conform to something, like a standard. And uniformity is when you uni' toward something. And the use of 'and' brings the two together to form into a "move toward the same standard." in this example anyway :P

[dense tapir]

conform to a prompt

uniformity is wearing the tie

[wispy nest]

one day we'll be there, but it won't be for years I think. One day the ai will be so advance that it's going to be, frankly, scary! :D

[dense tapir]

bring the two together or this is never going to work right

I agree

Here is when I will know we have arrived. Take a picture. Give it to the AI. It gives you the info to spit out an AI generated image exactly the same. That day will come.

Probably faster than we realize.

[wispy nest]

for me it will be when you can write a prompt like: Harry Potter running in Hogwarts and get something like a still image from the movies. Then you can add on and on and every prompt word is 100% correct as if said to a master artist of 59 years of experience in the art-school of "wizards running in hallways" :P

[dense tapir]

I notice a lot of the time I can give a prompt, even a short one, and it skips over words that I know it knows. They simply are not in the image.

[wispy nest]

that's why I was so happy to try the neutral prompter extension, while not perfect, it is such a leap forward in my eyes! :D

[dense tapir]

Nah, I can't stand that and removed it

[wispy nest]

I'm using it in all my images now. Mostly for negative prompts, loras, etc, but sometimes I use them for normal words as well :D

[dense tapir]

if it works for you for me it was a dud

[wispy nest]

yeah, it really was a game changer :D

[cyan snow]

Regional Prompter?

i don't see any extensions called "neutral propter"

[wispy nest]

no, that one is awesome as well, but not the one where you paint regions, but where you can use BREAK, ADDCOMM, etc

no wait

I'm stupid, that's the same thing, it's the one where you ADD_SALT me know keywords :P

[oak osprey]

find problems

[wispy nest]

where is the thing you put your hands on

[oak osprey]

bingo

it usually has good handlebars and no pedals

no, pedals and no handlebars

weird

child gone 😦

[wispy nest]

steals bike and pedals to sleep 💤

[oak osprey]

thats just a used bike, m8

grandpa

[crimson dawn]

just found this nvidia datacenter commercial randomly. It reminds me of old 90s intel dotcom boom energy

https://youtu.be/DFKdU6AIseI

https://youtu.be/GMt1djPS688?t=2 height of dotcom boom energy

[oak osprey]

fixed all those issues

[dense tapir]

shit 😦

[wild sorrel]

if not hands, bent, eyes that would look like an actual decent fan art of something

eeeh...and piece of cloth connected between stick and dress

and bottom cloth thing...

okaaaay....

[earnest solar]

Like I always say

don't look too close or too long

it's got a ways to go.

[earnest solar]

better?

[wispy nest]

Can't tell if gay or based

[wraith karma]

Tried to create an image with adventurers standing on a mountain looking at foggy dungeon entrance with red eye in it, but SD wasn't having it at all

This one seems pretty cool, though

[dense tapir]

https://www.youtube.com/watch?v=hkV7FjivJ7U

[wild sorrel]

better , but previous looked more fan-art like

[earnest solar]

if it weren't for the bright colours this'd be some elden ring stuff

[wild sorrel]

cartoon version of god of war

[smoky oak]

Ok, this one is looking significantly better

[wispy ether]

Where is the airflow

[woven dawn]

LN2

[wispy ether]

?

[woven dawn]

Liquid Nitrogen

[wispy ether]

What

[woven dawn]

Jibberish

[wispy ether]

Gibberish*

[woven dawn]

gazuntite

[smoky oak]

The edges are all open, and the fan creates negative pressure that draws it in, then forces our of the top holes. Pretty standard design lol

[wild sorrel]

Lease stomp braying gumball blings, your knot bleven taking tense

[wispy ether]

Shut the fu

TOP STORIES FOR YOU

Quora
 
I caught my son playing his Xbox at 12:00 in the morning on a school night. As a result, I broke his console and now he won't talk to me. How can I tell him that it is his fault?

[oak osprey]

@smoky oak i expanded the dataset to include a few thousand high qual national geographic photos

they took me A WHILE to download, not many seeders. like a month i waited

i can provide to you if you like

i'm hoping to keep things like instruments, bikes, landscapes in tact. i had to turn off learning rate scaling because i'm using multiple GPUs to train now and it really goes wild

[smoky oak]

I am not sure I have a need for them, but it might be nice to have them on hand

[oak osprey]

yeah and i'm not sure how well they work either

[smoky oak]

I'm gonna be messing around more with tensor RT today. I really hope I can somehow get it working with batch sizes greater than 1, that would speed up my generations massively

I could only imagine how fast it could go if I could get a full BS8 working lol

[wispy ether]

https://tenor.com/view/i-asked-gif-23588960

[oak osprey]

on my cluster training i can do batch size 128 now with no gradient accumulations LOL

@cyan snow

[oak osprey]

@smoky oak those images are captioned at least so you can search for bits out of it, which is what i've been doing

[oak osprey]

i love how mindless these robots are usually

[crimson dawn]

i hope the standard robot personality is paranoid android

[oak osprey]

lmao

my wizard lookin like tommy chong

[wild sorrel]

what's this magic, how do you fit so many details here

Why am I always refactoring my promts to have as low amount of things as possible lol

[astral flax]

I would have expected more smoke 😉

[smoky oak]

Oh nice, that's actually useful

I just need to get BS 8 working with tensor RT, and I will be going faster than a stock 4090

I wonder if running the command on its own will help. They recommend doing that, but I'm not sure what they mean exactly

[dense tapir]

I just prompt but do not forget to use the word (lowercase) "AND".

[dense tapir]

I can imagine that on an actual 4090

[smoky oak]

Have somebody here who did it, I actually got a considerably bigger perf uift than they did

*uplift

[dense tapir]

well, that makes zero sense

[smoky oak]

They went from 35it/s to 53its/, and my 3080 went from 15.3its/s to 27.7its/

[dense tapir]

like how an AMD 7900XT trains 4 times faster than it generates

[smoky oak]

I saw like an 87% speed increase

[dense tapir]

for training?

[smoky oak]

Tensor RT doesn't do anything for training, only generations from what I knoe

[dense tapir]

well, that fucking sucks

[smoky oak]

That's not at all what it's made for lol

[dense tapir]

it can though just not implemented yet

[smoky oak]

Tensor RT is just a cut down unet

It's shark, but better

[oak osprey]

GPT wants me to sing you a song called Salad Fingers

[smoky oak]

Cause you don't have to make one for every single BS and dimension

[dense tapir]

there are tensor cores in all 2k+ cards and those can be used for anything as they are so damn fast, but looks like no one has yet

[oak osprey]

that man looks so happy

[smoky oak]

My current trt unet works for anything between 512-768 on either dimension, and up to 375 tokens

But I can't get the batch size range to work :<

[dense tapir]

wait, what are you calling unet? Unet and TE for training is the only unet I know

[smoky oak]

So try makes a modified unet, you apply that when generating and that's what speeds it up

*trt

[dense tapir]

sounds like it isn't the real deal yet.

[smoky oak]

You go into stable diffusion and then change it from automatic to the specific one you just made, and it only supports the ranges you specified

[smoky oak]

NVIDIA's official version should be even better still

[dense tapir]

I bet you so

tensor cores is what they use for gaming now

[smoky oak]

If I could get the batch size optimization working, I would be as fast as a stock 4090

[oak osprey]

he's poppin'

[smoky oak]

Oh wait, I have an idea

[oak osprey]

wednesdays on nopeTV

hilarious that it lost the wizard dude but his stuff is on the bench??? unsolved mystery

[dense tapir]

I am becoming VERY ill at ease with how cozy Nvidia is with Intel and working together.

[oak osprey]

oh, okay. i updated my prompts and changed the seed i'm testing with

definitely did a good job on that one

you might think those hands look unnatural, but i encourage you to look at what Paul Gilbert's hands look like

[sterile temple]

morning all

[smoky oak]

@sterile templehowdy!

I have a quick question for you, if you don't mind

I'm trying to figure out how to run the onnx to trt out of SD to free up VRAM, and potentially allow it to do BS 8, but I am not too sure how to run it outside with the command they provide

[dense tapir]

Even on my card there is something running onnx but I forget now as I have seen it error out before.

Something in auto

[smoky oak]

I just need to figure out how to get the command running outside of SD, cause I really wanna get this working with batch size 8 to get the real speed increases

[sterile temple]

the command to create the trt file?

[smoky oak]

yeah. I wanna run it with Sd clsoed like they suggest, I am just not too sure how

[sterile temple]

what if you saved the commands into a bat file and then ran that?

[smoky oak]

this is what it gives me to run

[sterile temple]

like this

[smoky oak]

just put it all into a bat and run it?

[sterile temple]

yep

[dense tapir]

Is ONNX faster than TensorFlow?
Overall, the study found that ONNX outperformed TensorFlow on all three datasets. This suggests that ONNX is a more effective tool for deep learning models than TensorFlow. As such, it may be a better choice for developers who are looking to build and deploy deep learning models.Apr 9, 2023

[smoky oak]

alright, that worked, but it had the same error

so annoying ._.

IDK why I can't do greater than 1 batch size

[dense tapir]

Isn't onnx using the diffuser directly?

[smoky oak]

this speed benefit isn't really important for single image gens, its much more important for multi image gens

but it looks like its broken for multi images, even though it has a setting for them

[dense tapir]

I am not caring about tensorrt just more for the onnx

[smoky oak]

not sure why onnx is worth caring about

[dense tapir]

it is better than tensorflow even on my card

https://developer.nvidia.com/blog/speeding-up-deep-learning-inference-using-tensorflow-onnx-and-tensorrt/

[sterile temple]

looks like the product of the image size and batch size can't be more than 92160

[smoky oak]

any idea on how I could try and fix that limitation, or is it a hard limit?

[smoky oak]

so I have to make sure that none of the numberxnumberxnumber parts are over 92160?

[dense tapir]

@smoky oak Do you have a link on installingthis and what to do to try it?

Damn thing even works on maxwell based cards

[smoky oak]

Humble mikey was the only one that was able to help me and everybody else with the install

[smoky oak]

yeah, cause its just a cut down unet

[sterile temple]

yeah, seems pretty limiting

[smoky oak]

its not using any special tensor cores or anything, just optimizing the unet

[smoky oak]

that is the maximum size you can do? ._.

[dense tapir]

seems it takes 16, and 32bit into 8 bit realm

[smoky oak]

there is also 0 quality loss too

really cool stuff

[dense tapir]

Now I wanna tryit as I thought tensorrt demanded tensorcores

[smoky oak]

but it appears like it may be far too limited to be particularly useful where speed matters lol

[dense tapir]

for me any speed is a good thing

seconds per it sucks

[sterile temple]

just reading through the issues, trying to make sense of it 😄

[dense tapir]

I would be stuck at 768x768?

[smoky oak]

yes

thats what it looks like from my math

looks like the absolute max is about 842x842

[dense tapir]

DOA for me

[smoky oak]

well, RIP you I guess lol

even tho it says how to fix it

[dense tapir]

only say it MAY

[smoky oak]

thats the price to pay for 2.1

[dense tapir]

besides I use weird rez (half FHD)

[sterile temple]

from someone working adding the option for batch conversion, they have been playing around with the sliders trying to figure out the max settings possible

[dense tapir]

No way in hell will I go back to 512x512

[sterile temple]

but it does limit you to batch size of 1

[smoky oak]

damn, this is super super disappointing

its almost useless ._.

no high res fix, limited to one image at a time, low res

its like, they chose the already fast part to speed up lmao

[dense tapir]

Oh, and no controlnet

[smoky oak]

I don't really care about controlnet, but it sure is a lot of drawbacks

[dense tapir]

too many to have it be viable for most

[smoky oak]

yeah, huge shame cause damn it is a lot faster

[dense tapir]

Maybe the Nvidia one will work around these limitations

[smoky oak]

if I could get batch size 8 out of it, I would be at stock 4090 speeds

I hope somebody new comes along in the SD webui space, cause I don't like having to pick between two assholes in order to have a decent webui lol

[dense tapir]

This is insane

[sterile temple]

it's only a new thing. hopefully it will be improved

[smoky oak]

yeah

it is cool to see it so much faster, but its not really usable as is

[dense tapir]

Do a google search and find TensorRT isn't really new.

[smoky oak]

well yeah, but support for stable diffusion sure is

[dense tapir]

That is just a couple of dicks trying to out script each other. Once Nvidia's new ui happens we may then see stuff change.

[smoky oak]

You know what would actually be a massive time saver? Faster post processing

[dense tapir]

What is with that long ass pause anyway? Done with the steps and another 5-10s pause

[smoky oak]

cause with tensor RT, the slowest part at 512x512 for me is post processing

I get 27it/s on a 15 step image, yet it takes 1.2 seconds, cause of post processing

[dense tapir]

on larger images that pause can be up 20-30s

[sterile temple]

btw, @wispy ether you can disable by setting this option to None in the settings. no need to uninstall or switch branches

[oak osprey]

why pixels

[sterile temple]

freaking out for a minute there. my outputs were suddenty complete shite. thought I broke something. turns out it was just that neutral prompt extension doing something weird with my prompt

[rain gazelle]

evenin'

[dense tapir]

Hello

There we go

[sterile temple]

https://tenor.com/view/ru-paul-much-better-ru-paul-uk-drag-race-drag-queen-gif-15287257

[smoky oak]

Man, i really wanna find cool and new models, but I am so damn tired of seeing sexualized female EVERYTHING all over the damn place

[sterile temple]

I've tried to make mine non sexy

[smoky oak]

like its all just hypersexualized and fetishized women

I just want a cool model FFS

[sterile temple]

hypersexualized and fetishized

[smoky oak]

._.

[wispy nest]

a very cool model!

[dense tapir]

ugh

[wispy nest]

hope she don't get a cold ;P

[sterile temple]

wow a singular arm of the milky way, not crisscrossing

[wispy nest]

my brain! 😵‍💫 🙃

[smoky oak]

I really wanna play around with a model that is more abstract and freeform

but like, still looks good lol

[wispy nest]

I've noticed the ai like to create arms at the end of the character's hair if you prompt in something like long braid :P
Can sometimes look rather cool! :D

[dense tapir]

Never tried that, lol

[wispy nest]

I don't know, it somehow look like it'd fit :P

[acoustic oracle]

does low vram affect quality

[acoustic oracle]

yeah yesterday I was having a very hard time getting anything other than lolis out of aom3

then I added nsfw to the negatives tag

and it fixed it

[smoky oak]

genuinely so damn tired of the blatant hypersexualization of these models

[dense tapir]

loli loli loli get your adverb here.

[acoustic oracle]

so yeah its quite easy to fix

[dense tapir]

Hyper sexed children all Homer Simpson drooling over their waifus so I am glad to have 2.1 tbh

[acoustic oracle]

some of the outputs can be slightly sus even with the negative tag but not too bad and easy to fix by just regenerating, most are ok if you add nsfw to negative tag

[dense tapir]

nsfw didn't help me yesterday until I gave up after 5 or 6 models

[acoustic oracle]

was it nsfw or just slightly suggestive sometimes

[dense tapir]

even a Disney one had nipples and tits. Clyde would not let it pass

[wispy nest]

I most often get the result I want when it comes to output. Can't say I get chocked by, "why did it create this I told it not to do." ever :P

[acoustic oracle]

i thought maybe it was just slightly sus (like mine seem to be 20-50% of the time) but how did it make that so blatant when you had the negative

strange

[dense tapir]

I have no idea. I even tried nude, titties, nipples. That is when I threw in the towel when even Disney had it.

[wispy nest]

well, the first one is a color, the second one is slang'ish, and the third is a detail :P

[dense tapir]

Well, fuck it nothing worked so I went back to 2.1 without all the damn nudity.

[wispy nest]

it's all about the words, some are more, let's say, "fine tuned" than others :P

[dense tapir]

Some 2.1 models are all nudes with a dash of sfw in them

[dark flare]

The other day I saw a "peeing" model, that was ... interesting ...

[dense tapir]

WHUT? LOL

[rain gazelle]

Nice tree

[dark flare]

It's uh ... yep ...

NFSW
https://prompthero.com/ai-models/realpiss-v1-download
NFSW

[dense tapir]

OMG

Hahahahaa

[dark flare]

I was trying to get info on "realistic-vision-v1-3 (by SG_161222)" and unfortunately found that along the way

[dense tapir]

The dprevity amuses me

[dark flare]

There's a market for EVERYTHING!

[dense tapir]

Well, the image surely didn't

I was think Calvin pee not that image

[sterile temple]

One of the things I did to my mix was calculate the difference between beautiful realistic asians and realistic vision 2, and subtract 0.9 of that from my mix. Boom, bias towards drawing pretty asian girls gone!

[dense tapir]

too bad it is for 1.5 I would like to see how it does for non pee

I still haven't had a success like that

MBW/Super Merger as the classic 50/50 type is meh

Like the nude ones Walach clued me in on a trick of adding Corset and sure enough it works.

[wispy nest]

it's easier to trick the ai into making what you want, then telling it what you want sometimes :P

[smoky oak]

@sterile temple uhhhgggggggg

[wispy nest]

that's how I actually "see" most artist names used as prompt words. Like a shortcut to get the artstyle instead of writing it out in detail :P

[smoky oak]

the Sd update broke controlnet hard for me

[dense tapir]

I am not sure why I am getting this comic strip like white bars now

[wispy nest]

I sometimes add things like panel to the negative when I get too many "comic book white border effects

[dense tapir]

oh, good one

[sterile temple]

you okay there? lol

[dense tapir]

Alright I just got straight up gay porn. NGL, I dare not even post this as clyde would freak

[sharp saddle]

has anyone else just slapped this script with random stuff and just let it run overnight?

[smoky oak]

the SD update destroyed my controlnet install

[sharp saddle]

[smoky oak]

none of any of it works now

[wispy nest]

but if you use artist names then it might not be enough as I've noticed they use different things there so start with one, then add or change it with similar stuff like border, bar, strip (also work against lewd stuff by the way :P ), frame, etc

[smoky oak]

all sorts of file errors, and other issues

[dense tapir]

I don't update due to fear

panel didn't work 😦

[sharp saddle]

[wispy nest]

are you using the same seed?

[dense tapir]

No, random

[smoky oak]

same here, I updated to pytorch 2, and its been a little better now, but the extensions don't like moving on from a 4 month old SD install lol

[sharp saddle]

oh i've got a question, when you guys inpaint or latent upscale, do you paste the same seed

[dense tapir]

Well, panel no workie

[sterile temple]

I couldn't get the tile diffusion working with the controlnet tile model either

[smoky oak]

no, thats a good way to get seed issues

[wispy nest]

aww, too bad, maybe my "wonderful," but mostly lazy skills don't work in 2.1! 😭 😱

[wispy nest]

what's the error?

[smoky oak]

this, along with many MANY others

I just updated all of my extensions again, hope that fixes it

[wispy nest]

ah, gradio, that one is always like that :(

[sharp saddle]

so just leave it at -1

[sterile temple]

I'm gonna try that too. See if I can get upscaling working again

[wispy nest]

seems like the issue might be the placement of the entire prompt as you get those bars more constancy than anything else in the image :P
The ai reads each word in more than one way and the word after it might transform it to something else entirely as well. That's why it's hard to get more than two colors named one after the other because a lot of people just detail the first two colors :P

[sharp saddle]

[wispy nest]

if I get many errors at one time, then I just delete (after backup!) the entire venv folder. Has worked 100% of the time so far :P

[woven dawn]

[sterile temple]

Barrylin

[smoky oak]

nope, still ruined

probably gonna have to gut out all of my extensions and redo everything like I REALLY didn't want to have to

[sterile temple]

weird, updated extensions. now when I click generate, it opens up the file browser and I get some javascript error in the terminal window

[dense tapir]

I had to do that last time I updated so I don't update though I am soon going to need to as extensions are bugging out lately

[smoky oak]

they at least seemed to fix the issue of pytorch 2 being slower for me, so that isn't holding me back anymore

[wispy nest]

I learned my lesson about updating when auto1111's webui changed how their high res fix worked :P

[smoky oak]

you can still use the old version, tho the new one is much more reliable, and controllable

[sterile temple]

looks like taking out neutral prompt extension at least lets me generate images now

[smoky oak]

@sterile templeI reinstalled all of controlnet, still fucked

[sterile temple]

tile diffusion with controlnet working again

[smoky oak]

also, I have pix2pix errors too

of which I have literally never used pix2pix

[smoky oak]

can I just delete it or something?

[sterile temple]

yeah, if an extension is playing up you can just delete the folder

I have a timeout folder I move them to 😄

[smoky oak]

yeah so it looks like I can't use controlnet at all now

dope I guess

Glad to see my concerns about updating were valid lmao

[sterile temple]

What are you trying to do, i'll test it on my instance

[smoky oak]

reference only

[wispy nest]

ahh, after 20 cups of coffee I'm finally ready to take on the world! or at least until my body starts shaking from withdrawal again, but then I just fill up again! >:D

[wispy nest]

yeah, I've seen those errors lots of times before

[smoky oak]

I have no gradio file

what the hell happened

I am so damn tired of everything breaking lmao

This is the exact reason why I was SO against updating, cause nothing good ever comes from it

Let me see if I can just restore my old install

[oak osprey]

i told you not to do it

[wispy nest]

I'd delete the venv folder and let it build it back up. But I don't offer much help because I know that one person's settings often isn't the same as someone elses :(

[oak osprey]

not true once we merge with the singularity

[smoky oak]

I cached my old venv and extensions, but knowing how A1111 is, it will still be broken

[wispy nest]

yeah…wait—what do you mean with "we?" 😱

[oak osprey]

it's not automatic1111's fault

it's the bad code

inside automatic1111

[sterile temple]

I'm using gradio: 3.32.0

[smoky oak]

revrted to old venv

🤞

[smoky oak]

nope

its perma fucked

[smoky oak]

it deleted my gradio folder @sterile temple

I think going to dev branch ruined everything, its all unusable now

[wispy nest]

version: v1.3.1 • python: 3.10.6 • torch: 2.0.1+cu118 • xformers: 0.0.17 • gradio: 3.31.0

[oak osprey]

[sterile temple]

you could try installing it manually

[smoky oak]

on my old install, still fucked

I knew I should have never updated, nothing good ever comes from it

[wispy nest]

did you install it by a file, or by git?

[oak osprey]

ugh i feel that, i had 2 days of not being able to train because i installed mixture diffusers and it reinstalled bitsandbytes