💬｜general-chat

[viral atlas]

because of us lol? you stopped the peace agreements 😂😂😂

you funded the Maidan coup 😂😂😂

[rich kestrel]

who invaded ukraine

[viral atlas]

and were gonna win even with the aid

[rich kestrel]

putin keeps blaming coups

[viral atlas]

it wasnt ukraine lol it was autonomous DPR and LPR lmaooooo

[rich kestrel]

then again our presidents are even soyer

so soy president in russia, cuck prez in US

and u say the world is fine?

whoever said that must be a kid

[viral atlas]

I'm an adult

the world isn't fine I guess I take that back

but it's not heading towards total disaster either

Idc if Putin started the war I don't support starting wars at all but he had his reasons and America started many more and worse wars lol

all the leaders are trash

I'm not pacifist when it comes to wars at all

but if we're gonna permit war

then yeah

[rich kestrel]

yah I agree on this

[viral atlas]

Russia had reasons

[rich kestrel]

on this I mean

we need a cleanout

I left america knowing it would be 4 more years of biden or trump

[opal hedge]

The world is always on fire somewhere. It's just that everyone is on the internet eating overseas propaganda and the news is playing into that fear for money. It's always the same crap lol

[rich kestrel]

I aint stayin for that sht

[viral atlas]

I'm leaving America this year lol

[pearl ocean]

I wonder how close you can get SDXL to SD3

[viral atlas]

not for Russia though fuck that

not a great place either

[opal hedge]

How do you mean?

[rich kestrel]

good places are: mexico, costa rica, thailand

singapore

[pearl ocean]

In quality

[rich kestrel]

philippines if ur desperate

for digital nomads

did so and never looked back

[opal hedge]

Currently sdxl fine-tunes are better than sd3, at least what we have access to on the API

[viral atlas]

Costa Rica I've been to, it's really nice but I would go there if I was wealthier

[rich kestrel]

its cheap

all of SA is

unless u hang around rich neighborhoods

[viral atlas]

I meant like lack of job opportunities

[rich kestrel]

oh. who said anything about working there

[viral atlas]

Yeah I'm still planning to work so

[rich kestrel]

working in office in 2024

oh dear...

[viral atlas]

but if I was wealthy enough or had some tech startup I could work remotely for I'd go

[pearl ocean]

WFH it’s better

[viral atlas]

I have my reasons, and not necessarily in an office. Just different life path

[rich kestrel]

yh thats what I meant

if ur not WFH in 2024 what are you doing with ur life

[pearl ocean]

Work From Bed next

[rich kestrel]

LMFAO

if only

[viral atlas]

there's lots of careers not in an office that require an environment outside of home unless you're doing something digitally

[pearl ocean]

Working for company where the boss is a boomer

[viral atlas]

not everyone prefers to do digital marketing or programming or whatever else you do at home, although I do programming on the side it's not my only passion

[rich kestrel]

well

I have no idea what copium/hopium americans are huffing to keep going for the next 4 years

unless any of them are geniunely rooting for the geezer or the giant douche

[viral atlas]

Singapore is viable I'm looking into stuff there

[pearl ocean]

Tbh, I only know Singapore from Pirates of the Caribbean

[viral atlas]

https://youtu.be/uWGVlRQjxpM?si=0GcC2yUQ_yn9pPlQ

[sacred gull]

Hey everyone! Quick question. I want to set up a webui and I was wondering if there are any webuis set up already loaded up with some of the new PAG tech

Sorry if this is the wrong place to ask

[cedar salmon]

the guidance thing? ComfyUI has it, i dont know about others

[pearl ocean]

so far, epiCRealism XL is my fav model so far

[sage reef]

i havent tried the new juggernaut, did you try it?

[pearl ocean]

can't remember if I have

[sage reef]

i think it's the X version

[narrow kernel]

It just came out a few hours ago, so most people probably don't know it's there yet

[karmic cedar]

I’m weirded out by them dropping the L and just calling it Juggernaut X v10

But whatever

[pearl ocean]

I wonder what the X stands for, XL?

10

[sage reef]

one thing i can think of... he didnt want to make it confusing because before that model it was like numbers, so it's easy v1 v9 etc, but then instead of v10 (which by itself is a bit ugly to look at compared to the single digit versions), he opted for the roman numbers, so X is indeed 10 and after v9, but removed the L to not make it XL X, kinda confusing perhaps, so heck, just X... anyway.. thanks for coming to my TED talk.

and it might be the last one for sdxl

[karmic cedar]

makes sense

[proud spindle]

hello

[karmic cedar]

hey I know you

[dusk tiger]

XL also means "Extra Large"

[pearl ocean]

I think you miss read, talking about "Juggernaut X"

[sage reef]

read my TED talk above

[dusk tiger]

oh my bad, i saw XL and i instantly thought SDXL haha 😅

[pearl ocean]

still better naming than Intel CPU's

[sage reef]

kek

[karmic cedar]

Fooocus seems to not like Juggernaut v10 as much as it did v9.

[sage reef]

what you mean?

[karmic cedar]

i’m getting worse looking generations using it with the default config

[sage reef]

did you follow the recommended stuff from the civitai page?

[karmic cedar]

…what? what are you talking about, recommended stuff, pffff, what a joke

alright let’s see what it says

[sage reef]

he posted a prompt guide for X specifically along with recommended settings

there is also an inpainting version released by someone else based on X

[karmic cedar]

yeah, it has to do with the way Fooocus fooocuses its prompts

so, that’ll need a new version for sure.

[sage reef]

welp

all these hello bots

[karmic cedar]

Hey I know you

Jk hi

McCrispy? yea dawg i would

[sage reef]

what's even the point of these bots.. like why do that

[fervent thunder]

anyone had luck installing triton for forge on windows? unfortunately, docker and wsl aren't an option

[karmic cedar]

To slowly flood authentic communication channels with nonsense communication

Like commercials and television 🧐

[trim magnet]

miss

[fervent thunder]

i installed it for use with kohya, and honestly have 0 idea where it was installed to or what file in kohya is pointing to it. it is not on path either..

[sage reef]

so it's purely "for fun" or "because i can" or "for laughs", whoever is doing that, they have nothing to gain from doing it tho, idk.. some people are weird i guess lol

[karmic cedar]

It’s for something!

[fervent thunder]

and for whatever reason, if i open cmd, navigate over to webui folder and activate venv, the terminal will just crash if i try to do pip install on the wheel file

[formal gazelle]

...

[sage reef]

bye

[neat hearth]

yooo

bro seriously dead chat

why dc recommends me ded chat servers

[trim magnet]

dead game

[sage reef]

hello woman

[pearl ocean]

I hear some people Purely use SD for the babes

[sage reef]

you talking to a bot

is there not a way to filter these bots? lol

[trim magnet]

yea but if u enable it most of the new accs wont be able to join the server

[sage reef]

oh you mean like some phone verification perhaps?

[opaque gate]

miss one

[spring eagle]

Where can I find people to hire to make sdxl?

[orchid vault]

hello

[sage reef]

you making sdxl?

[trim magnet]

blud payin ppl to make sdxl 💀

[sage reef]

when he realizes sdxl is already made

shocking him to the core

[half karma]

hello one

[sage reef]

im gonna accept his offer, he gonna give me lots of money and il give him the sdxl weights already available

all according to keikaku

[molten sigil]

what?

[fossil briar]

everyone s

[tame rampart]

有中文聊天室吗

[jade rivet]

It’s for something!

[trim magnet]

🤫

[stuck sun]

for anyone that cares, 4090 FE is in stock in nvidia now at 1599.99

https://store.nvidia.com/en-us/geforce/store/?page=1&limit=9&locale=en-us&gpu=RTX 4090

[clever hemlock]

what's even the point of these bots.. like why do that

[stuck sun]

me? lol

sir i am a very good and nice boy, not a bot, just sharing in case anyone has been looking to pick one of these up, i already got one and its the best price you can get 'em for new, by far

[trim magnet]

🤫

[stuck sun]

yo, you were weighing between 3090 and 4080 (1200 new) the other day, nvidia has 4090 for sale now at list price for 1599, might be outta your budget range still but just wanted to let you know since its the best price you can get 'em for new

[sage reef]

i like how this one used exactly what i said which means the owner is currently watching lol

[trim magnet]

shush bot

[pearl ocean]

McCoders

[sage reef]

what is even happening in this server

[pearl ocean]

giving Dall-e a run for its money

[trim magnet]

bots copy pastin other ppl messages 💀

[sage reef]

exactly

bots go byr

[pearl ocean]

you guys playing nintendo games on iPhone?

[sage reef]

im playing iphone games on my switch

not a true story btw

or is it...

[copper adder]

Hello if I can use fooocus in mobile

For image generater

[sage reef]

il give you money if you stop the bots

it didnt work

[trim magnet]

your daddy

joe

[steep rampart]

Joe who?

[trim magnet]

bidome

[steep rampart]

https://tenor.com/view/sad-gif-7523306793289960933

[pearl ocean]

Hellcome

[pearl ocean]

Tim Cook moment

[trim magnet]

gm february bot

greetings

[pearl ocean]

greetings mortal

[trim magnet]

miss

[frozen helm]

did a whole bunch of chats get deleted

i remember there being stuff below "miss"

[regal raven]

Anyone here?

[loud solar]

No ^^

[regal raven]

@loud solar please 🥺

[loud solar]

Yes?

[regal raven]

come to #📝｜prompting-help

[pine helm]

Hi, first post here, is method 1 and 2 equivalent (as everything being up to date)?
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs

[loud solar]

1. is missing to install python and git AFAIK ... but than both should work

Maybe it will install automatically ... don't know ...

[pine helm]

I got an error with 1, so I'm trying 2

[still glacier]

method 2 is more standard and easier to manage, You just have to not be a lazy a$$ and learn how to copy paste the git clone line... just don't run it straight from the cmd as they say otherwise it will install everything in system32. Move to a correct folder beforehand

check pinned message in #🤝｜tech-support for more precise tut

[pine helm]

I'm following an Udemy course that tells me to download "v1-5-pruned-emaonly.safetensors":
https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main
Is that outdated? Where is the latest one?

[still glacier]

sd1.5 is wayyyyyyyyyy outdated, got released a year and half ago. Latest official release is sdxl (or stable-cascade depends how you count but I wouldn't bother with this one if I were you)

And there's ton of community trained models out there on civitai, huggingface, etc

[pine helm]

haha ok, what SD course do you recommend? 🙂

[pine helm]

Yeah, I see SD XL here in Core Models: https://stability.ai/core-models

[still glacier]

Depends what you're looking for. I'd say first get it running, toy with it, generate your first images. Then get lost on the Auto1111 wiki https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features and read a bunch of it. Then check out the extensions tabs in your sdwebui, install what seem useful/fun and read their readme.
ps : every label/text/button in auto1111 should open a short tooltip explaining what it does if you hover your mouse on them.

And be warry of youtube videos. They can get outdated pretty fast (and people never bother taking them out).

[still glacier]

that's where to get sdxl https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

[pine helm]

The paid licence doesn't get you anything extra, except a discord channel + licence to use it commercially? https://stability.ai/membership

[still glacier]

tbh if your udemy course is still recommending sd1.5 .... not a great sign. I hope for you that the next chapter is named "how to get a proper model"

[still glacier]

paid stuff is for online generations, commercial usage, etc. If you're a simple everyday joe trying to generate pictures you can run everything locally for free.

[pearl ocean]

CoPilot Dall-e

[still glacier]

@pine helm If you abstolutely want videos, you should be safe with those youtubers :
https://www.youtube.com/@Aitrepreneur
https://www.youtube.com/@SECourses
https://www.youtube.com/@Not4Talent_AI
https://www.youtube.com/@CreatixAi
https://www.youtube.com/@enigmatic_e
https://www.youtube.com/@lewingtonn

But again if dive first into some super complex workflow, you'll get overwhelmed. So take things slow and generate (not so good) images with the vanilla sdwebui first and once you understand what you're doing, build from that.

[pine helm]

Thanks a lot! ❤️

[nova zodiac]

you missed matt wolfe, sebastian kampf, nerdy rodent, and olivio sarakas

[pine helm]

I'm using install windows method 2: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs
I got some errors:
fsspec[http] 2024.3.1 depends on aiohttp!=4.0.0a0 and !=4.0.0a1; extra == "http"
But got past it by doing (thanks to chatgpt):
pip install pipdeptree
pipdeptree -v
Now it's running:
Downloading: "https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors" to C:\git\stable-diffusion-webui\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors
So I guess this is the old model, so I just add the latest model (dl from you link and choose it in the drop down in the software):
sd_xl_base_1.0.safetensors

[still glacier]

that's a topic for #🤝｜tech-support
Also don't copy paste random stuff from internet or chatgt if you don't understand them please

[loud solar]

You can also try: https://lykos.ai/downloads

Kinda central to manage different A.I. User Interfaces ... but go your way first!

[still glacier]

it adds another layer of complexity and potential bugs / mistakes. I wouldn't recommend using tools like those first.

[pine helm]

It's working! http://127.0.0.1:7860/

[still glacier]

that pipdeptree error and the way you solved it worries me but okay then :p

[nova zodiac]

127.0.0.1 points to the computer you are on - like a home address ( what you see on your 127.0.0.1 will be different to what I see!!!)

[pearl ocean]

127.0.0.1 > localhost

[nova zodiac]

there's no place like 127.0.0.1

[pearl ocean]

😛

[lusty beacon]

SD3 is available on ClipDrop by Jasper AI

https://clipdrop.co/stable-diffusion

[alpine hull]

bruhhhh why does SD3 suck so damn much

[pearl ocean]

czo its not out yet 😦

[alpine hull]

???

stability ai released the api like a few days ago 😭

[pearl ocean]

not on local

¯_(ツ)_/¯

[alpine hull]

thats not even the point

its not gonna be any better on local

[unkempt hatch]

how involved are the Stability Developers involved here? Do they read our chats?

oh man, that sentence. Off topic but I feel like I've lost 20 IQ points in the past year. I know stress can do that but it just feels too much

[alpine hull]

SD3 was hyped up as this big model that's going to beat out everyone else by a mile even before it's fine tuned by the community...

Dall-E 3 is arguable more coherent than this

And Ideogram 1.0 is undoubtedly the king

[versed goblet]

hi bro

[alpine hull]

man i'm just disappointed

[wicked bolt]

hello,bro

[potent spire]

king of what? 😄

[potent spire]

people never gotta learn the lesson in this case

[unkempt hatch]

Stability's strength is the community. Onсe it gets a Pony/Think/Humans fine tune, it'll be king

and once the community makes good controlnets etc....

[loud solar]

I think MJ is pretty boring. SD way more flexible ...

[potent spire]

depends on several circumstances and factors that makes one tool better than another

[alpine hull]

not imo

[potent spire]

MJ definitelly beats SD in terms of image quality for example

[alpine hull]

once a bad model, always a bad model

this isnt llms where good data can fix it

[potent spire]

DALL-E is the king in its own way as well

[alpine hull]

SD3 is fundamentally flawed

[potent spire]

i mean they all have their flaws

[alpine hull]

no like

SD3's biggest "features" are completely flawed

text generation is actually ass

way too many cases where there's artifacts, flawed objects, stuff like that

[potent spire]

what made people even think SD3 will blow out OpenAI, Midjourney and Adobe out of the competition? Delusional as always

[vale steeple]

clearly you didnt see the #🆕｜sd3

tbh people expected it to be a great leap from SDXL

but no

for API SD3 it is just a better version of SDXL.

you may also met website that advertised SD3 but in fact it is SDXL / Ideogram

"blowing out competitors" is seriously an exaggerating dream

there will always be an competitor, even just around OpenAI and Midjourney

[potent spire]

exactly

and in reality a bunch of people would simply use multiple of those instead of just one

[alpine hull]

it's supposed to be

you don't release a research paper showing that you've beaten out basically every other competitor by a long shot

and then the model is sdxl 1.1

[vale steeple]

evidence of them saying that?

[alpine hull]

https://stability.ai/news/stable-diffusion-3-research-paper

performance graph

[granite topaz]

I have a rtx 3060 12gb and ryzen 7 7700 processor.
Which AI can I run?
I used to run Fooocus.

But
I don't like it

I am new
Can anyone suggest me a better something

[unkempt hatch]

base SD1.5 isn't nearly as good as others, but fine-tuned SD1.5 is arguably better than any competitor. As evidenced by the fact that many still use SD1.5

[vale steeple]

you see the point is that SD3 are not made to "beat up" closed source AI, of course that is still a goal. But not long shot especially when you read the entire news.

I mean they do the same thing with SDXL right?

[vale steeple]

try generate a dog laying on grass shine under sunlight

[potent spire]

beating closed source AI is a very ambitional goal to say the least

[vale steeple]

in base SDXL and 1.5

[potent spire]

you mean competitor within the SD?

i mean competitor SD models?

[unkempt hatch]

For many use cases (obviously not all) SD1.5 is the best.

But my point there was that "once a bad model always a bad model" is ridiculous when you see the difference between base and tuned SD1.5

[granite topaz]

Little help from the AI whales would be appreciated.🫤

Suggest me something better than fooocus
Pls🐥

[vale steeple]

some said that the real value of SD3 is its adaptability, optimization and finetuning

you can see what image you generated through base 1.5

but if you using its finetune like Dreamshaper or Anything, it doesn't perform like its base at all

If we can replicate the same thing on SD3, in 8B to even 800M, then it will be a godsend and it will also be the time when everyone starting to move to SD3

[potent spire]

majority uses 1.5 or?

[vale steeple]

Forge by the same dev or A1111

[potent spire]

or did the majority meanwhile migrate to SDXL

but due to hardware limitations i guess many are still using 1.5

[vale steeple]

half-half

i think

[granite topaz]

Thanks brother/sister 🙏🏻

[vale steeple]

some people will recommend you ComfyUI though, I think you could be overwhelmed by node noodles

[potent spire]

i simply cant get along with Stable Diffusion tbh

starting with the setup, ending with the prompting itself and the performance (latter one wont be a issue as soon as i get my high-end PC)

[vale steeple]

So you want something that could just YOLO without any installation or something. Like DALL E?

Stable Assistant is kinda that.

[potent spire]

ideally yes which is why i mainly use Adobe Firefly + DALL-E via Microsoft services

[craggy bramble]

Is AI like snapchat AI easy to make?

[unkempt hatch]

I use both, even though I got good hardware, 1.5 has so many loras and models not in SDXL.

[unkempt hatch]

I would rather they'd made the 8b accessible. The Cascade idea was better, with a small and easily trainable latent, then a larger model B for scaling.

[rough tartan]

how can i donlowd anaconda

[vale steeple]

Google Anaconda or anaconda.com and try to find the download link. Alternatively you can download Miniconda for less space.

Why do you want to install Anaconda though? any popular Stable Diffusion webui dont required you to manually install Anaconda

[rough tartan]

i am going to train a Dreambooth model

[vale steeple]

I would say fair, I realized that people just throw-away refiners (C?) anyway

[unkempt hatch]

Because people tuning 800m won't help the 8b and vice versa

[vale steeple]

hm OK

[unkempt hatch]

No, cascade was completely different. It was like a centre unet plus outer unet in different models

[rough tartan]

when i try to train the dreambooth model an error code appers i belive it is error33

[vale steeple]

oh yeha it should be A. I know it is different

[unkempt hatch]

The outer unet would get refined with high resolution images and require a lot of vram to train, while the inner would compress to a 24x24 latent and be very easy to train

[vale steeple]

what kind of Error 33 and where do you get it? Suggest moving your issue to #🤝｜tech-support

[unkempt hatch]

Anyhow, hopefully they'll just optimize the 8b version and everyone will be able to train it with 8gb cards and it'll get a huge community like SD1.5

If not I'll offer my card to help people train

[vale steeple]

woah that is pretty generous ( normally I would expected people to give money to borrow card for trains )

[unkempt hatch]

Still, I'm sure some hero will optimize things to help everyone train in 8b, they just need to release the weights

[vale steeple]

... If you asked me no I don't trust anyone who offer me to train my LORA, could be a serious security risks when someone approach you.

[unkempt hatch]

How?

[vale steeple]

You never know, although .safetensors are fairly secure than its predecessor .pickletensor (.pt), maybe there was one person who zipped up both the trained model and virus altogether and you extracted it

then bam.

... again Windows Defender is pretty gay of detect them.

I am not a cybersecurity guy but such trojan virus could exist if someone is knowledgeable enough

[rough tartan]

plsa can someone look at my problem in tech support

[magic warren]

Hi i just downloaded stable diffusion webui

added models to models folders, sdxl and juggernaut, but i cant see it under stable diffusion checkpoint

any help?

[vale steeple]

If you are installing Comfy
your models should be in ./models/checkpoints/
If you are installing Forge/A1111
It should be in ./models/Stable-diffusion/

get it to refresh too

[magic warren]

ah thanks! i just found out what was wrong

i had 2 stable diffusions, i added to wrong models folder.

Fixed it, put it in right one now

[cloud lantern]

https://mp.weixin.qq.com/s/tz6iKxHQqGfvYWzf_lslRg

[edgy moth]

123

123

123

[carmine glen]

哈哈哈

有没有微信交流群呀

[obsidian viper]

ruinedfooocus vs fooocus, can somebody please explain the difference to me

[fervent thunder]

is the bot ever working

my god

it's been months

[magic warren]

How to fix speed issues with stable diffusion

Had it generate an image for 2 hours on M2 chip apple

[karmic cedar]

we talking a desktop or a laptop?

[magic warren]

oh shit apple does has a desktop, didnt know

mac

laptop

[karmic cedar]

https://youtu.be/mbhipHCuOEw

probably your settings

[magic warren]

graci

ffs where was this video when i was downloading stable diffusion. but it was just a guide for downloading it. i have it downloaded already

or are you telling me to uninstall it then reinstall it?

[karmic cedar]

hit up #tech-support

[unkempt hatch]

I'm on Linux, but it'd be me training it for another person... but that wouldn't be any more dangerous than downloading any model on Civitai.

[pale latch]

there's never been a proof of concept attack of ckpt attack through stable diffusion libraries

there's a ton of panic about this all the time, but it's basically a nothing burger. custom nodes for comfyui workflows are a bigger risk vector. all the major UI's don't have ckpt execution on by default and it has to be forced on. Unlikely that an embedded script would run in the apps we use.

[storm shard]

Looks like I keep saying this, if you're on a mac use DrawThings
It's free, no ads and regular updates
https://drawthings.ai/

[lunar spire]

Hey. Is this a place to ask questions about training lora using kohya?

[magic warren]

Looks cool, i'll try it, thank you.

However ideally would master stable diffusion and how it works - end goal of mine is to train my own models

[sterile raven]

Will SD3 have style ref

Its such a good feature to rip styles

[pale latch]

there will be extra networks made for sd3, if we ever get weights.

[karmic cedar]

I recommend finding a good summary feed of all the new developments in image models and even beyond, since there are constantly new things happening all the time now

That’ll keep you primed for when something brand new drops

[charred mesa]

if IPAdapter or some clip vision thing exists for it then yes

these will come sometime after release 🤷‍♂️

[sterile raven]

Is it easy to code, so i dont have to wait

[charred mesa]

no idea

these have to be compatible and trained

[storm shard]

if you have 16gb you can start with training loras

[sick lily]

anyone have any resources for how to use SD efficiently? I've been able to setup both WebUI and ComfyUI, and get basic generation happening with models i download, but i'd like to learn more about schedulers, upscaling, refining, etc...

[knotty fog]

My new app is crude but functional now. I'd appreciate any retweets as I'm still a nobody with no visibility.
https://twitter.com/Dan50412374/status/1781790992318042428

[civic grotto]

Whats the best inpainting interface lately? I'm specifically looking to inpaint on existing photos, not generated images

Need to be able to import masks, and ideally do some batch prompting

[sly sinew]

hi

[sick lily]

anyone know the difference scaling latent space vs image space? for example i have an upscaler that works on image space, but i could also just upscale the latent space and sample again. how exactly is this different?

[narrow kernel]

Use differential diffusion in comfyui. It's a node built into comfy and works with any model. You don't need inpainting models for it.

Just make sure you blur the mask so that there are gradients, since it works on them to make the blends more seamless

[karmic cedar]

I wonder if there could ever be an upscaling model built upon the principles of convolution, but pitting image space versus latent space

As it stands, the likes of SUPIR remain the more immediate future of super resolution upscalers.

[twilit pendant]

what is this and how do I use it?https://civitai.com/models/118811?modelVersionId=128941

[fossil phoenix]

it is likely a json for the comfyui interface, download, drag and drop to your browser window with comfyui loaded, install whatever missing nodes are there (for god's sake use managaer if you don't already) and that's it

[vale steeple]

that is SD1.5 -> SDXL inpainting workflow

[twilit pendant]

what do you mean, can you explain like i'm a noob

[vale steeple]

basically like a double pass.

You generate SD1.5 first as the base for the final result, then you put that base into rebaking with SDXL as inpainting.

[twilit pendant]

do i put that into comfui?

also where tf do people learn this shit

[vale steeple]

you haven't see those workflow with at least 5 ControlNet combined together and IPAdapters.

I mean that's the purpose of ComfyUI, to create sophisticated workflow for SD

and yes

put that .json file into ComfyUI

[twilit pendant]

where do those kinds of people learn their craft

[vale steeple]

eh. I think they learn every single nodes and their functionality. Kinda like electrical circuit designer, you must know how a transistor or fuse or diodes worked

in order to create much more complicated workflow or circuit

[fervent thunder]

"torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 7.20 GiB. GPU 0 has a total capacity of 19.99 GiB of which 0 bytes is free. Of the allocated memory 27.27 GiB is allocated by PyTorch"
i finally did it , OOM. i feel like ive accomplished something today

[twilit pendant]

do you think the image in #🏞｜general-with-images was created with a complex workflow

[fervent thunder]

not especially so, but not beginner friendly either.

[vale steeple]

workflow is the entire circuit. workflow is made up by tons of nodes or component

[fervent thunder]

they did a pretty good job of keeping it readable

[vale steeple]

simple inference of SDXL and SD1.5 is pretty simple though in ComfyUI

at least you could find the specific workflow made by someone else in Civitai

[twilit pendant]

wait what do you read on the image

oh wait the knight thing

yeah

[fervent thunder]

actually, after actually zooming in and reading through, a bit of it is over my head. looks overly complicated and convoluted to be honest

rube goldberg machine headass workflow

[knotty rivet]

for upscaling an image is it better to crop the image or resize it so thats its like the recommended model size?

[shell tendon]

that absolutely dependns on a number of factors

#1 by far is... will it get you to an aspect ratio that sdxl is good at

#2 is... are you cool with the image changing a bit during upscaling? adding new details maybe?

[undone pendant]

why my embeddings cannot work

whatever the checkpoints,lora or T.I Embedding cannot display when i run stable difussion

it have trouble me for two days, and I have no idea which makes me frustrating

[pearl ocean]

new 5090 coming
https://www.youtube.com/watch?v=eD7R3chkRQ0&ab_channel=Paul'sHardware

[stuck sun]

so is forge dead or what

[vivid valley]

Hello

Spirits

[gray wolf]

waiting for weights

[vale steeple]

at least wait for 2 to 5 weeks

[gray wolf]

sounds good,marking my calendar rn

[pearl ocean]

I like generating images, its fun.
but sadly haven't found a use for generating images other than just fun 😦

[graceful sparrow]

hey there everybody, I just got nmkd sd gui running for local generation, I'm wondering if anyone knows where I could find models to download?

[silver steeple]

civitai

just google civitai

[silver steeple]

?

sd webui

are you running it locally

[foggy halo]

guys when will SD3 release on HuggingFace?

[graceful sparrow]

oh nice, thanks! lots of stuff there

[jovial tiger]

image

how can use it

[silver steeple]

nw

you could also look at huggingface for controlnet models/AD models

[charred mesa]

probably 2-3 weeks

[pine helm]

Hi, let's say I have a real landscape photo. How would you generate + place an AI Model in it so it would look real? What tools would you use and how? Would you suggest a selfie or a full body model?

[pearl ocean]

I say, if you got a real photo, there is no need for A.I

[unkempt hatch]

do you think training a model on same-images, but photo and cartoonized version, would teach a model how to turn cartoons into photos?

[silver steeple]

nope

it could help but not very effective

it just shows the model the textures and etc

[unkempt hatch]

for example, I take a movie screenshot (Say, Malcolm in the Middle), I add 20 screenshots of the show into my dataset, tag them all.

Then using Stable Diffusion and controlnets, I make cartoonized versions of every image. I use the exact same tags, only replacing "photo" with "cartoon"

so now it has a concept of "Malcolm" and know how he looks as a "photo" and as a "cartoon". Same with everything else in the scene (couch, lamp etc)

but then later, I take something it only knows the cartoon version of (e.g. Bart Simpson). And I say "a photo of Bart Simpson".

would it be able to "logic" that backwards? Maybe that's something SD3 could do since it's got 8billion transformers?

[silver steeple]

maybe? SD using those tags could get close but true backwards logic is harder than you think since it might generate a photo like image based on cartoon features but it would not be a perfect photorealistic version

[vale steeple]

parameter doesn't mean its prompt understanding it better some kind.

you can try it as your own experiment in SDXL

[unkempt hatch]

wouldn't more transformers, especially now with diffusion transformers, mean higher chance of emergent understandings?

[silver steeple]

SD3s transformers might improve results

but its still under development for this task

[unkempt hatch]

I would really like to try that exact experiment

[silver steeple]

yeah sure

[unkempt hatch]

like, there was a research paper showing (it may have been SD1.5) had an emergent understand of 3d space, even though it was never trained

[charred mesa]

yeah it understood depth somewhat

[silver steeple]

despite it being never trained on 3D spaces at all 🤣

[charred mesa]

yes

[silver steeple]

it was 1.5

proved itself that it had capabilities in observing 3D spaces

[unkempt hatch]

so I'd assume even more likely in SDXL/Cascade/SD3 with more transformers

[silver steeple]

true, there is a strong possibility of encountering more emergent understandings

[unkempt hatch]

I've also noticed that training cartoons is a lot easier, and improved composition for realistic images. Like, it's more easily able to understand the boundaries and what a cartoon person is built from... so hence I feel training both photos and cartoonized versions of the exact same images at once could help understandings?

[silver steeple]

they have simpler details and nothing very specific just the textures and etc which is why its easier

[unkempt hatch]

maybe even more than 2 versions... like, have fristly the photo version, then a kind of semi-cartoonized version, where the textures are just eliminated and lines are more defined (like what happens when the image goes through an anime upscaler). The a genuinely cartoonized version, where you've got something like triangles to represent spiked-up hair.

[silver steeple]

yep yep

all of that helps

training it in multiple stages is not a bad choice

youve got the vision

just built it and see what you could come up with

[lofty sphinx]

Hey, is there a way to report easily smt to mods?

[pine helm]

Landscape photos are boring. Bringing an AI model and place i side the image would make it more”alive”. Think Instagram. Would it be possible?

[amber cairn]

so, not been on this server for a while (just in active)
where do I contact a mod / admin / modbot?

because there are users (user bots) in this server scraping messages and selling it

[trim magnet]

how can you sell a message

[amber cairn]

heard about spy.pet already?

[trim magnet]

is it that site that lets you find usernames?

[amber cairn]

not only that
the person behind it is someone that is part of kiwi farms (homophobic etc. hate groups)

[trim magnet]

oh ok but how can they sell messages that we send here

[amber cairn]

because you need to pay them, to get the messages
and see info on the user

[trim magnet]

ah you mean they sell information about users of this server

[amber cairn]

if I search on your user id, I can see you (but not the messages you have send etc, since I didn't pay for it)

[amber cairn]

as I said, they also sell it off to people from kiwi farms
they are quite.... homophobic from findings online

https://twitter.com/PirateSoftware/status/1782014084499792119

[trim magnet]

oh thats prob why theres been a lot of bots joining the server the past week

[amber cairn]

some deeper info on the whole operation: https://www.youtube.com/watch?v=ktxbXlF6UQE

again, if there is a better channel for this, tell me

but I do think this is a concern for this server and that it should be dealth with

[trim magnet]

try to send a msg to one of the mods maybe they can track the bots doing this,the bots all have their acc creation date at february 2024

[pulsar venture]

Hi, there is someone that could help me with sampling method and hires fix?

[light granite]

hi

[pulsar venture]

Can you help me?

[echo aurora]

Hi anyone can advice me, sometime generate photo while using lora, the generated photo is much more suitable when the genereted halfway (for example 70%), how to get the result in the middle of the generation instead of wait the photo 100% generated?

[sage reef]

you mean from the preview at about 70%, you see that it looks better right?

[echo aurora]

yes, but if i stop at 70%, i would not get the 70% photo

[sage reef]

yea but here is another problem... it's a denoising process, so even if you could stop at 70%, the result will be noisy, since it didnt reach the full denoising at 100%. then again, im no expert with how that works, so maybe there is a way to use the 70% output, but have no idea how 😦 i also had a similar idea to take intermediate results, but not sure if it's possible, and also the quality might take a hit like i said, since it's not fully denoised

[echo aurora]

I saw somewhere else from youtube, and it is possible, but i forgot the youtube😟

[sage reef]

huh, if you can find it and link it, i would be very interested in that

remember to check your browser's history, and filter by youtube, maybe you can find it

[echo aurora]

thanks, will look into it

[supple fulcrum]

hello

[trail lion]

Rather than stopping it, a better approach is probably to use the seed, maybe feed the depth map into cnet to ensure your get the same photo again, tweak the #steps, negatives, other params until you're satisfied

[amber cairn]

so why are you sharing pirated software in this discord?
@sudden ruin I guess this ping is valid?

[amber cairn]

@sudden ruin I guess I can ping for this as well, maybe you guys already knew about the whole spy.pet thing, but just in case (read the convo in the reply)

[karmic cedar]

Here’s a supreme controlnet challenge: img2kodachrome!

Can it be done? How would it be done? I wonder

[honest mica]

Thats a very expensive and time consuming "challenge" if done right.
Step 1: Buy around 223 rolls of kodachrome.
Step 2: Take around 8.000 unique pictures with a regular camera and with kodachrome from the exact same position with the exact same framing and focal length.
Step 3: Let the 223 Rolls develop with the dead K-14 process.
Step 4: Train the controllnet on your image pairs.

So i guess with "supreme" you are right

[fervent thunder]

I hear SD3 was all a big hoax

[sage reef]

you heard wrong

[fervent thunder]

That's what people are saying

[amber cairn]

yet #🆕｜sd3 exists

[honest mica]

What people?

[fervent thunder]

Everyone that isn't a part of this community lol

[amber cairn]

which is mainly reddit?

[fervent thunder]

Not all of them really care to come around a discord to find out info

Nah, hit other discords where people are doing Ai stuff but aren't on this server, there are plenty of them. For lack of better words, that is what is said

[stuck sun]

its true i asked my mom and she was like "stable confusion 3 is a conspiracy! It isnt real! Open your eyes sheeple!"

[fervent thunder]

nah its more like "sd3 is bs just ignore it, not comin. what idiots release papers that early"

blah blah

they go more in depth lol

[sage reef]

so basically some sweet summer children

[amber cairn]

I mean.... I guess the example of "rome wasn't made in a week" applies

[fervent thunder]

im sure people weren't announcing the build just before the release either

They just did it

[honest mica]

They didn't release a paper on building rome though

[amber cairn]

also, sd3 exists
#📣｜announcements message

just not as a local model AFAIK

[sage reef]

and it's coming May 10

[earnest anchor]

is there any channel I can use for help on how to use stable diffusion?

I just about managed to download it and now I have no idea how to even use it and YouTube isn't helpful

[sage reef]

there is tech support channel

but what videos were you looking at, i remember when i started i just saw maybe one or two vids from some of the bigger sd youtubers and i understood everything

[trail lion]

Getting it installed is way more difficult than using it, unless you're doing comfy... Start with auto1111

[honest mica]

Ouch. Isn't that a pretty bad recommendation nowadays? I mean, Auto1111 was good, but that was in 2022. Isn't forge and stableswarm the new player in this area?

[sage reef]

im personally not using any of those, but i think i heard that forge is not really for 1.5 tho, i think its mainly for sdxl

or maybe im confusing it with another UI

[earnest anchor]

can U send me the videos if possible

[earnest anchor]

no idea what those even are xD

all I know is I have version 1.9.0

[sage reef]

but wait im confused... you said you saw that video and you managed to download it and you dont know how to use it? so that video just told you how to download and not use? what the hell?

[earnest anchor]

yeah lol

[sage reef]

ok that is a terrible video

[earnest anchor]

I mean I can do basic stuff

but I can't make anything good

like I want a vid or guide on how to use loras properly how to make them which checkpoints are the best. the differences between the versions. if the version affects a checkpoint or lora.

and like what everything does

without using random abbreviations I have no idea the meaning of

[sage reef]

i mean everything you are asking can be found on youtube man, just search

like type: stable diffusion lora or stable diffusion controlnet, etc and you will find a tutorial on that

also add maybe a1111 or automatic1111 since you are using that

[karmic cedar]

Exactly. The amount of fidelity this would require translates to oodles of chemically processed materials

Imagine undertaking a grassroots initiative to shoot Kodachrome and pair it with a ground truth SLR capture

[honest mica]

First of all you would face the impossible task of finding unexpired film stock.

[earnest anchor]

what's that

[sage reef]

you mentioned v1.9, which i can only assume is the latest automatic1111 version

but wait, you dont even know what you have installed?

lol

[earnest anchor]

no I have 0 clue

I followed a yt guide

[sage reef]

but wait dude... how does that work... like you randomly click on some links, no idea what you are installing... you can then be tricked into doing anything like that, click this to send someone money...

i mean idk

[earnest anchor]

i didnt randomly click it

i clicked it with installing stable diffusion in mind

and went to all the pages

and did it

not a random link or video xD

[grizzled palm]

I have this awesome stable diffusion model for you. Install and then I'll tell you what to do next. 🙂
https://wallet.bitcoin.com/

[sage reef]

lol

[earnest anchor]

:000000

tysmm

!!!

[loud solar]

Thought the same 🙂

[earnest anchor]

i was 99% sure this wasnt a scam https://github.com/Stability-AI/stablediffusion

seemed pretty legit

[sage reef]

that video sent you to that link?

[earnest anchor]

no i went there myself

[grizzled palm]

Oh no LOL. Not a scam, but how on earth did you end up there?

This is like the worst possible outcome...

[sage reef]

wait im so curious now, can you link the video you saw

[earnest anchor]

by typing install stable diffusion github into google

[grizzled palm]

Not only do you have a environment polluted with weird python libraries, you can only run SD2, which is comletely worthless.

[earnest anchor]

there was like 10 of them

[sage reef]

ok but show me the one you saw

[earnest anchor]

oh

[earnest anchor]

there was like

10

of

them

[sage reef]

show me one

[grizzled palm]

Ok. Calm down. We can fix this.

The video doesn't matter. Please focus.

[earnest anchor]

yessir

🫡

[sage reef]

i was just curious what he saw and how he ended up in this mess

[grizzled palm]

Comfy has its own isolated environment. It shouldn't matter if you have pip and python elsewhere on your system... I hope.

[earnest anchor]

yeah (no idea what comfy is, no idea what pip is, i know what python is though)

👍

[sage reef]

i mean before we all waste our time here... do you even have the pc specs to even run this locally?

[earnest anchor]

uhh

my pc is pretty good

[grizzled palm]

My laptop with no GPU can run Comfy. (Just takes a few minutes per pic.)

[earnest anchor]

its not an nvidia gpu tho

[grizzled palm]

Here's the latest self-contained Comfy link:
https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z

[sage reef]

i dont think this dude is gonna go with comfy

[earnest anchor]

whats comfy :((

[grizzled palm]

Oh. Hmm.
It isn't user friendly...

[sage reef]

he seems to want everything simple to understand

[earnest anchor]

its my first time doing this stuff

[sage reef]

right

[grizzled palm]

Okay... Hmm.

[earnest anchor]

my bad i dont understand all the terms

[sage reef]

it's ok

[grizzled palm]

I just don't know what to do here... Is there some sort of custom node that gives Comfy an easy-to-use UI?

No way you can get A1111 installed on a system with global python stuff installed.

[earnest anchor]

cant i run it locally?

[grizzled palm]

Yeah it's just an endless ocean of errors if you're not installing on a "clean" system. (One without python and torch and cuda installed.)

[earnest anchor]

ah

[grizzled palm]

Comfy is easy to install and self-contained, but it's not user-friendly.
Want to try anyway?

[earnest anchor]

yeah

wdym by user friendly tho

oh easy to understand

nvm

[grizzled palm]

I mean it's not going to be, "type text and get image." You have to hook up nodes in a spaghetti-UI and copy-paste models into folders and stuff.

Anyway, I bet you can do it!

[earnest anchor]

:0

[sage reef]

i love the spaghet

[earnest anchor]

the thing i have right now is type text get image tho

cant i just download version 1.5?

[sage reef]

before you download models, lets focus on some interface

[grizzled palm]

This is the fastest way to get to all the awesome stuff.
(The absolute best way would have been to start by installing Automatic1111, and you can still try if you want.)

[blissful edge]

hello i was using a website to use the ai it was the oficall one but i cant find it can anyone help

[grizzled palm]

Install Comfy:
https://github.com/comfyanonymous/ComfyUI

By downloading this file and extracting it somewhere:
https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z

[sage reef]

@earnest anchor i just want to verify, do you have currently this installed? https://github.com/AUTOMATIC1111/stable-diffusion-webui

[earnest anchor]

yeah im pretty sure

[sage reef]

right

[earnest anchor]

lemme send a pic in other channel

[sage reef]

i mean that matches when you said you had v1.9

so ok we know you have a1111

now you need some models

[trail lion]

Forge is auto1111, it's like a tweaked version, and yes it's better.... But for someone struggling to get it going that seems like an unnecessary point to cause confusion

[sage crag]

what is this game 👀

[sage reef]

what game?

[sage crag]

isn't this a game?

[honest mica]

No

[sage reef]

huh

[honest mica]

We are a self help group for generative AI addicts.

[sage crag]

oh sorry

okay

[sage reef]

in what way would this be a game anyway? like what.,,,

[stuck sun]

anyone here do block merging in supermerger/sdxl?

there's a great tutorial for 1.5

but the settings in XL are a little different

and its much slower to merge

so if i wanted to take lighting style/aesthetics and faces from from one model and poses/actions/etc from another its less clear how you weight it, which blocks tend to have which details etc

[trail lion]

🤣 I seriously wonder how some people end up here

[sage reef]

is that dude using the word game most people are using the word game? LOL

weird... or what?

[shell hazel]

LMAO

funniest thing to happen here in awhile

[sage reef]

must be some alien

[pearl ocean]

I hear SD3 can change a life!

[charred mesa]

it certain will change mine

I will be 1% happier

[tacit tiger]

man, in a few years there probably will be Ai game generator and this will become a gaming discord

[fervent thunder]

Hello new friends! 😁

I love AI art and I’ve been into it in some fashion since 2022, starting with Disco Diffusion (this was back before you could even make passable humans!) 🖼

I’ve previously been using online services to make SD images, but it makes me feel like a complete sucker. It’d be much better to get a proper PC and figure out how to run it locally I think 🤔

Anyways, my favorite model is “epiC Realism”, although I don’t believe it’s supported anymore, I’ll have to check.

[sage reef]

epic realism is still a thing

i think he focuses on xl versions now

[fervent thunder]

Yaaaay! I’ll need to try those

[sage reef]

yea very good stuff

[fervent thunder]

I’m particularly interested in photorealistic models, which are only trained on AI-generated images. I think epic realism qualifies for that, and I suspect there must be other ones as well 🤔

[sage reef]

just search civitai 🙂

[fervent thunder]

Ooooooh thanks! I’m still pretty new in some ways I guess 😅

[low moon]

Los Angeles looks like Detroit.

[karmic cedar]

hi

I saw a guy do an hour long video on how to do it but I don’t have time for that shit. Give me the colab??? Plz?

[low moon]

Inject SD3 in my veins.

[deep berry]

Hi I'm new. How do I create Images

[nova zodiac]

Hi new, I'm dad

[deep berry]

Hello Dad bod

[nova zodiac]

and proud of it - didn't eat these donuts for nothing!

[deep berry]

Lol

[nova zodiac]

to answer your question though - you either install a user interface locally (if you have an Nvidia graphics card), or you head over to one of the many websites that allow you to put in prompts and generate stuff

[low moon]

You meditate and focus intensely on your monitor.

[deep berry]

Lol you know I'm kidding right

[low moon]

We're all just kidding here.

[pearl ocean]

even the kids are kidding

[nova zodiac]

goats are kidding..

[nova zodiac]

no idea.. so many people come and go through here asking that question

[deep berry]

@nova zodiac i get it. I was just messing around. I posted alot already in the SD3 thread

[worn aspen]

I think I reached an overload with models and workflows, that SD has lost it's fun.

All the 1.5 stuff and playing with controlnet things. Then realizing XL is the thing, downloading, troublshooting, grabbing all the extra models to follow tutorials. Finding Cascade, liking it, only to realize that it's never getting adopted.

It feels like I download things, get it to work, be exhausted, and by the next day there is a new thing.

[nova zodiac]

Ah right - I don't explore this discord much, stick to general chat mostly

[nova zodiac]

welcome to the world of technology

tbh, new stuff still being released for 1.5 too - ELLA for example

[worn aspen]

I've been a tech guy for decades. I've experienced nothing as chaotic and disjointed as all of this. Lol.

[pearl ocean]

😮

[nova zodiac]

chaotic I disagree, but disjointed I'll agree with

[knotty rivet]

anonyone know a easy to use node in confyui to automatically calculate and resize to closest sdxl ratio/sizes

[worn aspen]

I downloaded all that; haven't really played much yet because of the burn out.

[fervent thunder]

I'm thinking about upgrading my computer to get faster gens. Does the more vram you have speed up the process?

[hasty hornet]

more vram might let you generate more images at the same time

[trail lion]

so you've never heard of linux? seems like you're overdramatizing the experience. There are many front-ends to experience stable diffusion, much like there are many flavors of linux distributions, but in the end it's all a means to an end, and the journey is what matters

[shell tendon]

yeah. it's in such a "breaking news" research phase that everyone's scrambling to chase the next big hit, publish it, dump the bare minimum code needed for it to funciton (if even) and then race off to the next thing

hence, barely anything is documented, the github nodes on comfyui are chaos, etc

no one has stopped to consolidate anything really

i imagine this will continue for a while

[narrow kernel]

There's no need to because the other researchers have what they need to work on the next big thing that takes some of the things learned from the prior projects. The whole AI field is moving exponentially faster than almost any industry in human history.

Not every pitstop along the way needs to be made user-friendly

[trail lion]

hopefully it wont, imo

like forget about 1-click installs, it becomes a drag on the community

[fervent thunder]

just ran an upscale that took 1.5 hours. some of the stuff i had edited, the upscaler decided to add hair to it

[trail lion]

not sure I would have the patience for that

[fervent thunder]

doesnt take any pateience, just set and forget

[narrow kernel]

Means you're probably running out of vram and it's bottlenecking due to using system ram, which slows things down by magnitudes

oh and you can maybe use a tiled upscale instead and get roughly identical results if you know what you're doing. would likely bring that 1.5h down to a handful of minutes

[shell estuary]

Does anyone know of any AI tools that allow 3d mesh generation from a set of photos of an object from different angles rather than just image to 3d?

[verbal osprey]

Why is adetailer not listed in the extensions list when I search for it in auto1111?

[hot kettle]

I'd say papers from stability ai and from others neatly document new architectures and training methods

[shell tendon]

yes, and then there's implementation

[hot kettle]

Yeah those aren't documentent as much

[shell tendon]

yeah, generally not at all

i have countless nodes where i have no idea what anything does unless i dig into the code

[hot kettle]

I prefer to just read the paper about something instead of checking the code when it comes to ai

[shell tendon]

doesn't work when you're actually using the extensions or nodes

who knows what actually went into those

or what paper they referenced

rare to see a referenced work at all

[hot kettle]

Well it's easy to find sources for basic nodes I think

Even without references in comfyui documentation

[craggy geyser]

Can someone provide me with a good negative prompt or tips for T-ponyai3 3.5?

Please

[loud solar]

Have a look at the model page on civit.ai

[warm junco]

Its called !After Detailer in the list

[pearl ocean]

imagine using SD3

[sage reef]

imagine using a1111 in 2024

[narrow kernel]

Imagine all the CO2 created by weirdos redlining their PCs 24/7 making lewd waifus. Probably gotta be close to the output of small country or more at this point

[honest mica]

No, thanks.

[white python]

Hello

[pearl ocean]

CO3 when

[pine helm]

Hi, does it say on CivitAI in which order he used Detail Tweaker XL, Juggernaut XL and SDXL Lighning LoRAs?
https://civitai.com/images/10123212
I tried Juggernaut XL (same promt, seed, etc.) and I get something completely different :/

[warm junco]

In the prompt you see where he added the loras

[pine helm]

Sorry again, but which ones?

Ophelia's dream, Low saturation color photography, vintage, grunge, top light, masterful painting in the style of Anders Zorn | Marco Mazzoni | Yuri Ivanovich, Todd McFarlane, Aleksi Briclot, oil on canvas

[warm junco]

Oh I checked the link.
Its not in the prompt in that case.
So we can't know. He may used an older extension where you could load the Lora's in without using the keywords.
He could have also used comfyui maybe.

[pearl ocean]

Will SD3 ever come?

[warm junco]

I think so

[bleak matrix]

Good morning, everyone! I hope you are well today!

[sharp marlin]

#dream logo name, vk living, design, high-end, home living,

[worn aspen]

I have, and I've pretty much run Debian for years and years. True there are is chaos and cutting edge distributions, but there are distributions that support getting things done.

It sure would be interesting to see what is under the hood. I've messed around with the introductory challenge at Kaggle (I know it's not the same), and that was super time consuming just to dip the toes in. In this realm, it seems people who aren't even really bent towards development are writing nodes and doing some advanced things.

[compact panther]

If I want to build a specific service utilizing Stability.ai's APIs, do I need to purchase a membership subscription in addition to the API usage fee?

[elder lily]

hello

[fiery roost]

Hi

My AnimateDiff does not show

Webui v1.9.2

And I has check the version of AnimateDiff that shows lastest version

[broken cave]

what was the wisdom of retiring the stable cascade channel?

[storm shard]

Lack of, I would say

wisdom

[fossil karma]

hey guys, weird personal question,
so basically i am working on my thesis to finish university,
i decided to discuss about ai, generative algorithms and all the things that we have around nowadays,
do any of you perhaps know someone that could help me gather texts, books, anything really, about chatgpt, dall-e, dlss and fsr?
i tried googling them but i get a bunch of not in-depth articles, any help would be much appreciated

[loud solar]

Yes ... a lot of A.I. generated sh*t ...

[chilly pike]

Since this discord no longer works, what can you use to generate unlimited images for free?

[weary socket]

Happy Earth Day!

[fossil karma]

yeah i am aware its a lot, but i really need documentations and not random articles to work on

[broken cave]

bing image creator

what is the relevance of / audience for generating anime? what am i not clued into here?

i know there's no right answer but is there an intellectually honest answer that is true most of the time? is it for porn?

[karmic cedar]

there is an app that will take a bunch of your photos and then make you a bunch of AI portraits

But I forget the name 😢 😞

[verbal osprey]

I ain't NEVA GONNA STOP using a1111!!!

[sharp fulcrum]

Do the 1070 and 3070 use the same graphics driver? I'm upgrading and I'm not sure what to do.

Just swap the card out?

[amber bloom]

yeah, swap it out and then re install the driver.

get the latest driver for your card from nvidia.

[marble portal]

Where the channels for generating stuff

[harsh skiff]

please guys help

I don't have a permission to chat in finetune chat

[hot vortex]

hello

[harsh skiff]

how can I pause my lora training state in kohya ss??

[hot vortex]

I'm new to this, so I haven't tried any extensions or plugins yet.

[red lynx]

Ah man of culture, William Montgomery reference

[harsh skiff]

hello, welcome

[hot vortex]

I wrote something into tech-support, but it's bugging me. It seems people use different programs for stable diffusion other than AUTOMATIC111. I started having a problem with RAM and VRAM overflow evne though I have everything setup for low performance.

[trail lion]

well, there is a way to save the state at an interval you specify, but it does consume a decent amount of disk space. I find that unnecessary, as you can also resume from any saved checkpoint. So if you save every epoch or every X # of steps, you can simply resume from one of those...just know that it will start over in terms of dataset, but all weights will continue training from that point

[harsh skiff]

you mean loading last lora from LoRA network weights right?

[trail lion]

yes

so just put your last saved checkpoint there and you're good to go

[harsh skiff]

I heard that doesn't make it like when one run go.

so I want to use (Resume from saved training state) option

but it says path to last-state folder to resume and I don't know where is last-state folder

[trail lion]

right so if you ran 12 epochs, it's not the same exact thing as running 6, and the resuming from that 6th checkpoint for 6 more, but it's pretty close

[harsh skiff]

I can't find it in model or log or image

[harsh skiff]

do you know how to use the (Resume from saved training state) option?

[trail lion]

yah, it works similarly to the load network weights, you have to tell it to save them and you'll see them in the output folder

[harsh skiff]

which output folder?

[trail lion]

wherever you tell it to save your checkpoints to

[harsh skiff]

I can't find it

[trail lion]

it would go to the 'model' one

so if it isnt, something is wrong

[harsh skiff]

always there is something wrong

is it a folder? or like models?

in the model folder there is just one folder named sample and it's empty, and .json files and .safetensors models

another question, what are the folders in the log folder?

[hot vortex]

try using AI Copilot or you.com or anything similar, personally, I've found these things usefull, it can at least give you examples of such books, give it clear and specific prompt

[mint wolf]

Hey Y'all, quick question.. is there a way to add a dreambot to my private discord server?

[hot kettle]

Papers with code maybe?

Or just arxiv

[broken smelt]

yep arxiv is where you want to go for academic work

[hot vortex]

yes, there is many of them, e.g.: arxiv, google scholar, elsevier

[hot vortex]

Are there any trending models that can be useful in image generation? mostly for fun and some high quality, I used a basic upscaler and so far so good

[loud solar]

Just look at civit.ai 🙂

[hot vortex]

well, I see sd3, cosxl and similar one, laso people use different programs for these diffusion models

such as comfyui and so on, so I just wanted to know briefly how good they are

[loud solar]

SD3 not available lokal at the moment. If you want the most flexible UI Comfy sounds like the best decision.

[hot vortex]

thank you, but I'll stick to AUTOMATIC111 for now, I need to at least make it work again without eating much RAM/VRAM

[loud solar]

I totally understand that ... you could try Stable Diffusion Forge ... it's a fork of A1111 but works faster and handles (V)RAM usage better ...

[hot vortex]

Does ComfyUI work with Python 3.10?

[loud solar]

As far as I know, yes. But it's a complete new System based on Workflows and Nodes ... easy to update and play with the lastest tools. But needs to learn a bit and it's not so comfortable as A1111 or Forge if you are used to them.

[hot vortex]

if it doesn't have issues such as VRAM/RAM overflow issues then I'm all for it

so ComfyUI should be the best one

[loud solar]

It handles (V)RAM better than A1111 ... but that can change every day ... but I would think Comfy is one or two days ahead 🙂

[hot vortex]

so it make tables and then you can just match them together? interesting

[loud solar]

A1111 has prepared workflows you don't change that much. You can add ControlNet and so on but it's all shown in the UI.

Comfy you download workflows that are build a bit like Lego with different bricks (Nodes) ...

Or you can completly build your own thing. That's why it's so flexible and can react to new stuff pretty fast ...

[hot vortex]

ok, thanks, it looks like a lot of stuff to do

[loud solar]

Take your time. It's fun,too! One thing you don't wanna miss using Comfy is: https://github.com/ltdrdata/ComfyUI-Manager

It can download missing nodes and keep everything uptodate

[hot vortex]

thank you

[loud solar]

And I am using Python 3.10

Oh and I use Forge and Comfy. Forge for easy picture creation cause I am so used to that easy workflow and Comfy for more complex things of if there's no solution for forge yet

[hot vortex]

OK, I'll definitely try to use both somehow

if it was Linux, would the files work with links?

because idk how I would use these large models, I downloaded a hefty amount of them

[loud solar]

The link I gave you is just the manager add on for comfy. But on huggingface they tell you how to install ... it's pretty easy. and there are ways to download models only one time but use with all UIs

[hollow gorge]

Would anyone like to try out a new bot I have been working on for stable diffusion? I would really appreciate the help in testing it. Just lmk or dm me.

[hot vortex]

Is it normal for AUTOMATIC111 to use max VRAM and almost half of RAM? It lags my notebook.

[sage reef]

i think automatic1111 is like the worst option these days

[honest mica]

Sounds like A1111 to me, yes.

[hot vortex]

yes

I was just wondering if it is normal.

I literally only loaded it and it started using almost my entire VRAM.

[sage reef]

do you have amd gpu?

[hot vortex]

no, NVIDIA and Intel CPU

[sage reef]

so you are using by just loading it, it uses all your resources? is it automatically loading a checkpoint? it's been a long while since i used a1111, so i dont remember how it works LOL, i just know that i hate it

[hot vortex]

I just loaded it and it is loading checkpoint automatically

[loud solar]

I think it's just loading checkpoint at start while others do later

[hot vortex]

I just read stuff that the certain things don't get freed from memory

resulting in crashes etc.

[sage reef]

it's terrible

[hot vortex]

the thing is I had to click on cmd or else it'd be stuck at a certain point, but I had python 3.11

[sage reef]

i would personally recommend comfyui, but that's just me :3

[hot vortex]

now I'll try with python 3.10.6

I alrady have that installed. I'll just need to set it up.

[shell tendon]

yep, also i hear stableswarm is coming along nicely too

[sage reef]

yea

[pearl ocean]

Until you used Forge

[shell tendon]

as @neon oriole said recently, "a1111 is deprecated" 🤣

[pearl ocean]

[sage reef]

it really is deprecated

[loud solar]

Stableswarm has still a pretty uncomfortable and bad UI ...

[shell tendon]

ultimately stableswarm and comfy are going to be the way to go

once all the kinks are worked out

cuz then you get the best of both worlds

[sage reef]

literally ANYTHING is better than a1111

[charred mesa]

lol

[honest mica]

I prefer painting in oil on canvas over A1111

[sage reef]

hahaha

[nova zodiac]

Yeah but after the comfy vs illyasviel argument about forge, I dont believe comfy is all that

[honest mica]

It's faster

[shell tendon]

last time i fired up a1111, i hit 20+gb vram using juggernaut and a single lora

[nova zodiac]

Forge is a1111 done right

[sage reef]

@honest mica btw, i discovered today about a lora family called boring reality for sdxl, i never knew about it, gives some cool stuff 🙂

[shell tendon]

illyasviel also went awol after pushing a couple commits that broke stuff

no commits since march 8 i think

though there were 74 private commits in april lol

[sage reef]

i mean i respect illyaviel, he did so much for the community, so he gets a pass ok 🙂

[loud solar]

Respect to all people programming for us ...

[sage reef]

yes

[nova zodiac]

He may be busy with personal stuff, he doesnt have the same flexibility as comfy does given comfy works for stability

[sage reef]

imagine working for stability, sounds like a cool job tho :3

[loud solar]

I have read somewhere the forge guy had an examn ... but who knows?

[nova zodiac]

Thatd do it

[loud solar]

I asked Stability for a job as panikbutten engineer ...

[shell tendon]

been six weeks though

[honest mica]

Unless you are one of the 10% who got laid off lately 👀

[sage reef]

im an actual programmer tho, so can they hire me? even for cheap salary? :3

[nova zodiac]

Would be, I did apply but no dice, but if @karmic brook et al are listening Im still available 😉

[sage reef]

im always available

but i have so many personal projects im working on

[nova zodiac]

Likewise - although probably more classified as a technical BA/functional analyst

[sage reef]

currently working on a 3d game engine, some neural network project from scratch, 2d games, digital signal processing stuff, and other smaller projects

im versatile

[nova zodiac]

Hi versatile, Im dad

[sage reef]

[nova zodiac]

He not happy, emad 😉

😛

[sage reef]

[sage reef]

panikbutten ?

[hot vortex]

I can't make A111 work anymore.

It simply loads model and overflows RAM. so I'll go with ComfyUI

[sage reef]

yay

welcome aboard

[shell tendon]

youo'll find we're pretty helpful here with getting you started

most of us leave our workflows embedded in our images here so you can borrow em

[loud solar]

The red button you push in case of emergency to shut everything down ...

[sage reef]

ah haha

time to do some spaghet 🙂

but it's very ez

and you can save workflows

[nova zodiac]

ComfyUI is the antethesis of easy unless you understand intricately how diffusion models work and have used node based UIs before

[loud solar]

Workflows should be saved in every picture ...

[honest mica]

Don't open Clownsharks workflows!!! Displaying them alone uses 12gb of vram.

[sage reef]

yea clownshark has some crazy stuff

[nova zodiac]

Once you have some preconfigured workflows, youre ok but it can be a steep learning curve

[shell tendon]

ehhhhh

comfyui is how i learned both

i'd never used a node UI, never read a single thing about diffusion

[nova zodiac]

It is a learning curve

[shell tendon]

i first found out stable diffusion existed dec 7

[hot vortex]

never used 🤡 I'll try

[shell tendon]

had no idea wtf was going on with a1111

[hot vortex]

I only worked with A111 until it stopped working outa blue

[shell tendon]

used comfy a week later

[honest mica]

I would say it's a pretty linear learning experience.

[sage reef]

i recommend for you maybe scott detweiler comfy tutorials to begin with, start his playlist from the start, and learn along @hot vortex

[pearl ocean]

Me either, I moved to forage ages ago

[hot vortex]

that's me, the last wee I was trying to run A111 on my own, then it worked and today it doesn't anymore, it uses too much RAM and VRAM

[nova zodiac]

It does resonate with some people more than others, but for beginners Id recommend fooocus or easy diffusion before picking up comfy once you understand whats going on

[hot vortex]

and there is absolutely no command, global variable that can save it

[alpine hull]

Bro

[pearl ocean]

And VRAM is at a premium

[alpine hull]

I expected SD3 to be really good at complex & long typography...

[nova zodiac]

I switch between forge and comfy, but prefer forge

[alpine hull]

Why does it suck at that 🙄

[shell tendon]

i don't think i ever would've figured out what was going on if i don't move over to comfyui

[sage reef]

@hot vortex he works at SAI too, highly recommended, very awesome dude https://www.youtube.com/playlist?list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x

[alpine hull]

Also, does Stable Assistant generate better images than the API?

[hot vortex]

whoa, that's a lot of tutorials, thanks for inputs

[nova zodiac]

Nope - the assistant just calls the api under the hood

[shell tendon]

comfyui really isn't hard at all to use, the main problem is the documentation for most nodes is shit (if it exists at all)

[sage reef]

you dont need to watch all of it, but at least couple first ones

[nova zodiac]

Its not hard once you have the hang of it, no

[hot vortex]

I started with this because usually these genrators are paid monthly or cannot generate what I want (such as real person, because of copyright and so on)

[pearl ocean]

I could never figure out how to get comfyUI to work properly

[shell tendon]

watch some tutorial vids, play around wit hworkflows here, ask us some questions with screencaps of your workflows or dropped as an image in here... you'll get going fast

[nova zodiac]

And things are getting better as more of the content creators have started making stuff for it

[loud solar]

No image here 😄

[karmic cedar]

What is SD3’s token limit? Anyone know offhand?

[nova zodiac]

154 iirc

[karmic cedar]

Wow

That be gud

[nova zodiac]

77 pos, 77 neg

[sage reef]

i mean they had to increase it because of T5 anyway

[karmic cedar]

Well yeah

[nova zodiac]

I dont know how the api is dealing with long prompts - t5 can do 4096 I think, but there is still a clip layer stuck at 77

Looking at the initial diagrams

[sage reef]

yea idk, i guess we will figure all that when everything is released

[nova zodiac]

Yeah - no idea if they are doing prompt chunking then passing the entire through

[charred mesa]

guys guess what

[nova zodiac]

??

[charred mesa]

nothing

[sage reef]

i facepalmed irl

[charred mesa]

[sage reef]

if we are lucky, sd3 in about 18 days from now 🙂

[charred mesa]

if only

[sage reef]

haha

[loud solar]

Just download from torrent ^^

[surreal jay]

is Forge missing anything that A1111 has? I just started last week and got a workflow going in A1111 but may switch after hearing Forge's better

[loud solar]

The other way round ... it comes with what you have to add on to A1111

[earnest lichen]

77 & 512

[surreal jay]

so there's no reason not to Forge?

[earnest lichen]

depends on how you define "limit" 🤣

[rugged mirage]

a111 got an update last week, forge hasnt gotten that one, and has existed for shorter

[astral goblet]

automatic1111 is upstream from forge, so changes happen there first. If you're riding the dev branch, this will have some stuff that forge doesn't. Most recently was scheduler options being seperate but that's caught up now.

animatediff extension is the one that has me staying on webui base for now. There is a forge specific version but it's behind development last i was playing with it

[earnest lichen]

a1111 pains me so much due to how weight normalization is applied to words/sentences x_x

[trail lion]

even when I was running auto1111 directly, I was only pulling like every 3 months

I prefer stability first

[earnest lichen]

switching between comfy & a1111 on a regular basis shows how A1111 is amazing for newcomers, since it auto-corrects typical mistakes in weighting, and literally turns it into a feature. but then the moment you switch to any other UI, you feel like the gens are worse, because no behind the scenes weight normalization is happening

[sage reef]

arent there comfy nodes that can mimic a1111 functionality?

[astral goblet]

i got different installs for stable and feature testing. lately too, these projects have matured to the point where venvs don't freak out so much and core libraries don't need to be 8 different versions depending on what extension you're using. i remember when i was starting out with a1 ui, "git pull" would mean disaster at times

[earnest lichen]

yeah, but I need the reverse XD I need A1111 to not normalize things. let me over-emphasize some word!

[rugged mirage]

you can still over-emphasize by doing brackets and weight to specific words

cant you

[astral goblet]

is there somewhere i can erad about the technical side of weight normalization and whats happening ? am curious

[hot vortex]

Any ComfyUI experts who knows which version of torch I should use? I'm still getting CUDA error when I start main.py of ComfyUI

version 2.2.2 for Nvidia doesn't seem to work (torch-2.2.2-cp310-cp310-win_amd64)

[earnest lichen]

in a nutshell:

Input: a (blue haired:1.4) woman is walking (down the street:0.5)

Output:
COMFY: a (blue haired:1.4) woman is walking (down the street:0.5)
A1111:(a:0.8 (blue haired:1.1) (woman is walking:0.8) (down the street:0.4)

math isnt correct, but theory is the same

[rugged mirage]

I just let it choose it's own, not sure why you are micromanaging it. if it doesnt work just use the portable version

[honest mica]

Use the portable version

[astral goblet]

your bracketing isn't correct either. broke my bracket parser

[hot vortex]

I used the command in the guide: pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121

and it isn't the correct version it seems

I'm using Python 3.10.6

[rugged mirage]

just use the portable version and let it figure out what to install

[sage reef]

i think using command prompt, use the command nvcc --version to check what cuda you currently have

[hot vortex]

Do you perhaps know the version?

nvcc --version doens't work in cmd

[sage reef]

so you dont even have cuda

[hot vortex]

wait, it's for linux

I have

[astral goblet]

Ahhh so you are talking about the way a1111 implements weight emphasis. I get it. Yeah it's a very opininionated implementation fo it isn't it? I tend to not rely on weights and just write longer prompts or go regional prompting. @earnest lichen

[sage reef]

oh you have linux?