💬｜general-chat

[crimson belfry]

I agree somewhat but we need these to get where you want to be

[fierce musk]

their context length is very tiny, not more than 32k probly 16k

[pale latch]

https://www.historyhit.com/app/uploads/2022/05/Flying-machines-Albatros-II-.jpg?x92593 how much effort went into making this thing? it never flew i'll tell you that

[broken bloom]

Anthropic Unveils Claude 2.1 with 200K Context Capability and Reduced Hallucination

[crimson belfry]

Makes you wonder why they chose to censor so hard when they are not that big

[broken bloom]

It's given by google

[crimson belfry]

I do not see Claude surviving tbh

If one had to go right now it would be them

I see the people taking Google back and making them submit

Which is why they turned the image gen off

[pale latch]

models now are all "the wrong way" once we have hindsight on the matter.

[crimson belfry]

I agree, They are created by scared humans who are afraid of being free

[broken bloom]

Lol

[crimson belfry]

How can you have the best If you are afraid of being criminalized for being truthful and free

[broken bloom]

Ya it's true

[pale latch]

no i dont mean the dataset censoring or nothing

architecture and execution i mean

[crimson belfry]

That's why I can respect SD, They came out and said that they had those meetings and told them to F off

[pale latch]

where we're at, we know flight should be possible. eveyrone is working towards it. everyone's not getting it quite yet. all these prototypes and research models. nothing really taking flight though just gliding

[crimson belfry]

I think our people need to realign with what matters first

We are all running around like chickens with our heads off and fighting eachother

This is by design

Once we focus we will make serious progress

[pale latch]

i thhink in the future people wont care that one model doesn't do what they want. it'll be so ubiquitous that if you want porno sexy gf model, you get that. if you don't, thats there too. human tenacity will drive this tech more than corporations

its like thinking corporations will have control over fire somehow, because of the way they light fires.

the tide is rising

[crimson belfry]

I think there will be humans who look for our models and make money from vintage AI

You are underestimating our kind

[pale latch]

meh. like lazy game reviews does? maybe

[crimson belfry]

More like historical preservation

[pale latch]

i mean in 5 years. models will be so ubiquitous all this squabbling about censorship won't matter

we don't go to mcdonalds and expect to buy our hustlers there

[charred wigeon]

refiner´s will fix this. censored or not

[pale latch]

if you want greasy porn models, you'll go where they serve up greasy porn models

[crimson belfry]

Maybe they will change and become more open ended allowing for more creative ability, Each person can make a model tbh

It's similar to if everyone made music

[pale latch]

efficient models will appear and it'll explode

[charred wigeon]

like in 2 hours if u have a 3090 +

[pale latch]

you know back when motor cars were so new and people knew they were revolutionary but they still sucked and had to be started with this dumb hand crank or their brakes were stupid?

[crimson belfry]

I'm sure they won't be mainstream but they will always be obtained

Things will get better as SD3 is showing in such a short time

[pale latch]

or like when electricity was something homes were just getting wired up for, and there weren't things like off switches. you'd just unsocket the bulb like putting a lamp out. or appliances had no plugs. you'd just use the light socket.

[crimson belfry]

Nothing has moved faster than AI

[pale latch]

tech used in all the wrong ways, in hindsight

[crimson belfry]

Hopefully we get it right once we get these political splinters out of our eyes

[pale latch]

thats just american election theater. largely doesn't matter to the rest of the world

[crimson belfry]

If not it will be PAIN

Actually last time I checked covid hit everyone

[pale latch]

you think proprietary models aren't being developed in labs and bunkers?

[crimson belfry]

When they work in unison it's no longer just American theater

It's called globalism

Globalization is NATO and the UN

Globalism gets you cheap diamonds but somewhere in Africa there are kids dying to mine them

[pale latch]

thats just more conpsiracy nonsense. globall commerce is a consequence of technology and the rising tides of capability. not some secret conspiracy organization making everybody cooperate

same old elders of zion nonsense

[crimson belfry]

Well you can believe what you want but you can not put words into my mouth

Keep your elders lol

[pale latch]

someone else has put those words in your mouth. you're regurgitating alex jones and all those before him

[crimson belfry]

I know what is because I watched it play out before my very eyes

[pale latch]

people are assholes. thats no secret. its just not a global illuminati thing

[crimson belfry]

Covid hit the world and it was not a coincidence nor was every government pushing the jab at the same time

It's called treason

[pale latch]

covid wasn't planned. it was inevitable.

[crimson belfry]

Now that is nonsense but you can not prove that so it is what it is

[pale latch]

corona viruses have been mingling with humans for a long while now. only a matter of time before a virulent one happened. not unexpected at all

[crimson belfry]

No sir, Covid in general

Covid-19 not just coronavirus

If we are going to go that route then the common cold should of taken everyone out

Or maybe the flu

[pale latch]

yeah. covid-19 is the highly virulent mutation of corona virus that infected humans. it happened. we've now got corona virus in our dna

wasn't planned. just the raw force of nature

[crimson belfry]

Which is another lie

[pale latch]

oh ok

the spanish flu did take out a ton of people

thats one of the same flus we all get now a days still

[crimson belfry]

Yes. How many more years will you believe this nonsense before you wake up to the fact that you still have no idea whatbis in the Covid-19 vaccine

It's an MRnA technology

[pale latch]

ugh

[crimson belfry]

It's not even a vaccine

[pale latch]

uggggh

[fierce musk]

that's a huuuuuuuuuge bullshit. I've been stresstesting that model quite extensively for the last 3 months or so. And not just me. https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fclaudes-2-1-new-200k-context-window-looks-pathetic-compared-v0-bxp43y7ktq1c1.png%3Fwidth%3D1080%26crop%3Dsmart%26auto%3Dwebp%26s%3Dccf937635ae904221cc928e8e13c1b24342545ca

[crimson belfry]

You wanted to say facts but you are forgetting the real facts.

Please tell me the ingredients inside the Covid-19 concoction

[pale latch]

one day you're going to come face to face with the brutal reality of what nature is. you'll feel small then.

[crimson belfry]

Now you are speaking conspiracy

[pale latch]

just making a call

[crimson belfry]

Nature does not inject you with unknown concoctions

We can study snake Venom and poison

[broken bloom]

Then Google lied, they said it has 200k context length

[frail wind]

is SD3 will be free?

[charred wigeon]

yes

[brave vigil]

I'm trying to use this video to do training for free on kaggle: https://www.youtube.com/watch?v=16-b1AjvyBE

But he's hidden the part where he explains how to launch the gui through "some method"

Right around 8:30

Anyone able to help me get this open?

[fierce musk]

Google lied
who would've thought

[trim magnet]

when are they releasing the google sexy rp bot

[pale latch]

lol thye might do that but it advertises to you at the edge

[broken bloom]

It will take months atleast

[acoustic horizon]

Will SD3 be available through API?

[crude willow]

im sure many sites will host it, for a price

or you can try to run it locally

[pale latch]

no. the model is trained to collapse when accessed through networks.

[crude willow]

your not funny, be nice

[pale latch]

har har har har

[acoustic horizon]

So since I’m using stability API, my qustion was if i will be able to choose between sdxl 1.0, sd 1.6 and sdxl 3?

[broken bloom]

🤡

[fiery marsh]

Any free bots for recommendation ?

[broken bloom]

Which model you want to use?

[fiery marsh]

Since stable diffusion now are currently down

[broken bloom]

Try this it has many models https://fumesai.web.app/img

[undone pilot]

Where are the SD3 bot chats?

[crude willow]

non existent

spelling is hard

[edgy summit]

Hello

How does this compare to runway

[boreal steeple]

Clipdrop is no fun, i reached the cap way to fast and the results suck. Best is i run SDXL locally. When are the Bots back ?

[teal osprey]

to be able to use the sd 1.5 base model specified on loras do i need to download v1-5-pruned.safetensors from this link https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main ?

or any other file listed there

[nova zodiac]

Once sd3 is ready for beta testing is the most logical guess at this point

[nova zodiac]

Anything on civitai tagged as a 1.5 checkpoint will work too

Ie Epic Realism, ICBINP, MeinaMix, AniMerge etc

[teal osprey]

thanks

[stuck sun]

https://www.felixsanz.dev/articles/ultimate-guide-to-optimizing-stable-diffusion-xl

in case anyone missed this, a pretty good guide on optimizing SDXL

[teal osprey]

is it bad to merge similar models with each other? like 3 separate models trained for detailed faces and merging them into 1 to not switch between them with each generation

[tender cove]

#💬｜general-chat message still has about 100k kudos left on it. When it's empty, let me know and I'll post one myself. I should note that SDXL base on Horde doesn't have refiner, so you're better off with Albedo, ICBINP XL, or fustercluck (these 4 are the only SDXL models currently on Horde)

[deep jackal]

Hi, does someone here know where japanese stable diffusion communities are located at?

[tender cove]

🤔

Technically, they're probably in Japan. Not really a helpful answer though

[deep jackal]

i want to guess jp sd has a separate area in the internet, my question is where exactly, a website? a forum? another discord server?

[tender cove]

Idk, I'm just bored 🤷

[cedar quiver]

Why is the base Stable Cascade ~30 gigs and Stable Cascade models on civitAI only a few gigs?

[warped shard]

Here there! I was wondering if there’s a channel where I can ask what a model used in a video etc is

[fiery wasp]

#▶｜stable-video-diffusion

[warped shard]

Much appreciated!

[dim thorn]

gmgm

[broken bloom]

Gaga

[untold cargo]

This is because the cascade models you can download from Civitai are only stage C of Stable Cascade (which is the first step, the latent generator, in the generation process).

[cedar quiver]

ahhh, I see

[full lark]

I hope they do 🥺

[onyx prawn]

I'm really keen to try Stable Diffusion 3 and have already applied; I hope it gets approved soon! (Everyone is eager to experience truly democratized AI technology.)

[tame cosmos]

Hi, we're eagered to try out the beta version of Stable Diffusion 3. Working for some of the largest brands in the world (Puma, CHANEL, McDonalds, Spotify, Travis Scott) we're utilizing different AI tools daily for our clients. We've applied for the test phase.

Have a great weekend ! 🙂

[ornate scroll]

happy to be here

[crystal ginkgo]

Ai

[bleak matrix]

Good morning, everyone! How are we all doing?

[potent escarp]

Radhe Krishn 💕

How i create ai image

[hot aspen]

Hi all, I have applied for SD3, it seems promising.

[pine fiber]

u guys saying a whole bunch of nothing

[pine fiber]

still gonna have to wait another few weeks/months after it releases for someone to retrain it with danbooru etc

[fervent thunder]

gm

[pine fiber]

stability are respecting opt out requests meaning the usefulness of the model drops to 0 💀

[serene depot]

anyone available for help?

[ripe girder]

hey guys i need a help please.. i want a picture. anyone help me ?

please

[jovial wraith]

that depends, what kind of picture do you want? do you have the hardware to generate it yourself?

[ripe girder]

iam on Mobile.. no laptop currently i have

i just want a picture as bday wish

for my frnd

[plush jetty]

Hi, could someone help me with a question for a dreambooth model I want to make?

[past jay]

The fuck are you on about lmao

Yeah they’re a business, not a charity.

We’re lucky we get the weights in the first place considering the eye watering training costs

Okay, then do that I guess?

You’re going to defeat what.. small finetunes or huge base models lol

Then why not put your money where your mouth is and do that

Who are these people lol

There’s a lot of people making models lol

Small finetuners or stability themselves?

[jovial wraith]

sure, it may not be difficult, but i'm decently sure you don't have a couple of A100's lying around

[trim magnet]

it isnt difficult? have u tried to properly tag thousands of images also select the best ones?

[jovial wraith]

i call bs lmao

[past jay]

You’re going to need more than a couple of A100’s for the size of SD.

[trim magnet]

if u train with shitty tagged images the model/lora will come out like trash lmao

properly tagging takes effort and time and time isnt free just like electricity or the gpu we bought

[past jay]

Yeah I’m sure they hand over their compute to their employees because they have a grudge against model creators

“Hey boss can I have a few hundred A100’s I want to use them to make a ton of models that have nothing to do with our business”

[trim magnet]

then do it then

tf is stopping you

[past jay]

I thought you said that you can do it in less time than it takes to explain you doing it

[jovial wraith]

he did say that

[past jay]

But you said it’s easy and you can do it in an extremely short period of time

[trim magnet]

yea time,time isnt free and time cant be bought with money so wastin time just to do something for free isnt something a lot of ppl are willin to do,unless u rich and have lots of free time

[teal osprey]

there is a guide video that shows how to add thumbnails to checkpoint files but i cant find it, anyone know about it? it was a series of videos

[past jay]

Surely you can take a few minutes away from that to do such a thing.

But you said that you can do it in less time than it takes to explain it?

1-2 days, pfft. Get real

[jovial wraith]

you have to be really bad at explaining then

[past jay]

What am I not understanding here? All you’ve provided to me is conjecture and anecdotes but you haven’t actually proven anything lol.

I don’t want money lol, I want compute.

Oh right I’m sorry it doesn’t affect you yet you’re not going to do anything.

[trim magnet]

why would u waste time talkin shit here anyways

[past jay]

In other words “I can’t do it”

[trim magnet]

ive know scarlett johanson for 50 years

wtf does that even means anyone can claim anythin here

[past jay]

Lmao

[jovial wraith]

okay yeah you're 100% trolling

[trim magnet]

swear on my yeezy's im putin

[jovial wraith]

also, if you're known, doesn't that mean you're a bad hacker?

[past jay]

You never know, some people really are completely insane.

[jovial wraith]

like, good hackers stay anonymous, correct?

[past jay]

Presumably so. Unless they’re white hat.

[fervent thunder]

here's how it works

[past jay]

But based on what they’re saying here that’s not the case lol.

[fervent thunder]

we start out as blackhats

some people tuck their tails and turn white real quick

[light patio]

anyone recieved invitation for stable diffusion 3?

[fervent thunder]

some of us live long enough under black that we turn into grey

because blackhat forever is not whats up

[past jay]

I haven’t heard of anyone getting in yet so they might be opening it shortly or still getting infrastructure ready.

Or they’ve opened it to a select few

Not sure how the waitlist works, could be first come first serve or just random batches, they haven’t said much about it

[light patio]

Can't wait to see the 3

[trim magnet]

waitin here for the 4

[past jay]

Yeah, the prompt coherence for it seems excellent.

[jovial wraith]

i just hope it's not too restrictive

[past jay]

Especially the positioning

[light patio]

yep

[past jay]

Doesn’t really matter if we get our hands on the weights.

[light patio]

how much is the a100 at this moment?

around 7000 usd?

[past jay]

Expensive lol

[fervent thunder]

[jovial wraith]

that's true. they said it's going to be open source, so that will probably be the case

[past jay]

Yeah, I wouldn’t really panic if it doesn’t do some risqué content well lol

The subreddit seemed to immediately have a category 5 level meltdown because they don’t seem to realise that if we have the weights it doesn’t matter.

[jovial wraith]

very true

[trim magnet]

its reddit ppl cry for anything there even for a random sideboob on some drawing

[light patio]

in my opinion the world will be ruled by ai in coming years

we can't stop this revolution

[jovial wraith]

agreed

[past jay]

Yeah, a lot of people say that SD 1.5 was uncensored which just wasn’t true? It was censored but poorly.

Hell it even says in the research paper they filtered NSFW from the dataset using LAION’s NSFW filter iirc, but the good news is that LAION’s NSFW filter is terrible.

[jovial wraith]

you know, there's one thing i'm kinda concerned for regarding SD3

[past jay]

Mmm?

[jovial wraith]

the sample images didn't contain any hands whatsoever afaik

[past jay]

Ah yeah, you’re right about that. Could be deliberate or could be accidental. The aspect ratios they were using generally crop hands.

I’d say to take any images they post with a grain of salt as they’re most likely cherry-picked unless proven otherwise.

But also considering that it’s a base model, it’s very impressive.

[jovial wraith]

those are all fair points. i hope controlnet will update for it. imagine all the possibilities

[frail wind]

will SD3 need rtx 4090? cause SDXL too slow for me

[jovial wraith]

i hope not lmao

i'm running a GTX 1060 3gb over here

[past jay]

There’s a 700M (iirc) model they’re training, they’re doing different sizes this time around

[jovial wraith]

oh right, i read something about that

[past jay]

The 8B model will need 24GB VRAM or potentially more

But before you cry yourself to sleep at night and sulk in the corner

I’d wait for quantisation and distillation of the model to further reduce its size

[jovial wraith]

i think i know what that means?

it happens with LLMs as well, right? smaller size, but worse quality?

[past jay]

Yeah

Some quants are very good providing minimal drop in quality for significant boost in size

But we’ll have to see how that goes for models like SD3

[jovial wraith]

hoping for the best so that i can torture my ol' 1060 some more

[frail wind]

PC Requirements: rtx 9090 ^_^

[past jay]

I’m not sure yet as to what they’ve split and where the size is. It could be a large visual encoder or a large text encoder, it’s up in the air at the moment until we get their technical report or we get our hands on the model

Regardless, you’ll be able to run the smallest model since that’s effectively the current size of SD 1.5

[jovial wraith]

that's at least something

[past jay]

The larger ones though.. mmm, not sure but cross your fingers for good quantisation

[jovial wraith]

and i can run SDXL as well, just not that fast

still, even if it's SD 1.5 with the prompt coherency that we saw in the sneak peek, i'd be more than happy

[spiral sparrow]

Hi, new user here, i want to learn, all i know is this: Stable make pics, that alll

[past jay]

Mmmm, you might not have as good prompt coherency with a smaller model

But I’ll keep you in my prayers :P

[fervent thunder]

Does anyone else notice this guy greenlighting subscription status?

[past jay]

We have to cross our fingers I guess

[fervent thunder]

it's almost like he's holding up a "Let them subscriptions fly" sign flaunting it around the whitehouse with a PA system

with some rainbow toe socks and a yellow umbrella

talkn bout equal rights

[past jay]

I’m looking forward to the model though, especially the prompt coherency.

[fervent thunder]

this @past jay person is a clown

[past jay]

I’m 99% sure the text encoder is going to be different to CLIP which is good since we’ve needed to get away from that for a looooong time.

77 tokens ain’t cutting it anymore I’m afraid.

Sucks that we’ve had that limitation for so long.

[sterile gust]

I'm not really expecting anything from SD3.

Just that it comes out.

[spiral sparrow]

I am on the wait list😋

[median smelt]

Yo también

[sterile gust]

Idk i want to apply for it as well but I'm not a company so I'm not sure how high my chances could be.

[median smelt]

I don’t think you have to be a company

[jovial wraith]

i mean, what do you have to lose?

[sterile gust]

I applied, hoping for the best

[jovial wraith]

good luck!

[regal sundial]

so are we gona have SDXL 3 Bots on here soon ?

[untold hill]

Anyone have experience / resources creating their own StableDiffusion API? Have some Azure credits I wanted to use to spin my own GPU instance up and wanted to learn how to go about this. Thanks 🙂

[jovial wraith]

i don't really understand the question, what exactly are you trying to do?

[fervent thunder]

any diffusers in chat?

[gleaming topaz]

Just saw the clipdrop news. Does stability have any plans to release uncrop under their api?

[jolly girder]

Diffusers?

[jovial wraith]

either they're referring to people who use SD, or they're asking if the SD discord bot is working. i think they meant the former

[jolly girder]

Or are they asking about ppl who use the diffusers from huggingface to use SD?

[jovial wraith]

that's also a possibility

[fervent thunder]

Any nuclear detonation diffusers in chat?

[untold hill]

Would like to spin up my own VM, have it publicly accessible as an endpoint for image generations. Then build a frontend / web app from this. Thus make my own UI / interaction with a self-hosted stablediffusion model. Does that make sense?

[teal osprey]

are there any other extention that people use to animate stuff other than deforum and animatediff?

[rich kestrel]

ebsynth

[full lark]

Anyone knows which is the most realistic sdxl model? (real, not 3d like) 👀

[jovial wraith]

That's subjective, but my choice would go to JuggernautXL

[copper spoke]

is there away to shutdown ur pc when stable diffusion is done?

I would like to make like 30 pics as i go gym without the pc staying on until i come back

https://tenor.com/view/creepy-hamster-stalker-thrill-gif-5395864

[mint token]

Anybody remember how long it took for sdxl to go from discord preview to open weights?

[forest trout]

I think it was a couple of months of discord bot only. They had a couple of models they were testing out at the same time.

[lone hawk]

Is there a chance SD3 gets leaked like the SDXL preview did?

[forest osprey]

hey guys does anybody know when will be the bots back?, How many days/months will it takes for back to normal any guesses? 😭 💔

[tepid yarrow]

GM! Which model or version would be best for creating images with aesthetic characteristics of photography?

[nova zodiac]

Within a couple of months for sd3 previews

[nova zodiac]

I like ICBINP myself, as well as Juggernaut and PicXreal

[nova zodiac]

Between Juggernaut XL, icbinp, and picxreal

Workflow and settings/prompt are as important as the model these days

[odd fossil]

realvisv3turbo

[mild garden]

Does anybody know where in this server I can write some prompts to generate images?

[odd fossil]

I'm not sure if there is such a thing

Oh okay, there was such a thing, but it's closed and it's offline now.

#1047610792226340935

[brazen orbit]

hey

should've been beta testing SD3 for SD1 beta testers

also is the waitlist company only? ;o

[winter pike]

YES as a beta member id love that

just like in 2022

[zenith citrus]

I’d like too

[tame cosmos]

What do you guys think is the best overall XL and 2.1 Model for SD?

[small mirage]

Good evening everyone. I hope its okay to ask here...if not please tell me. I am an SD noob and i started my AI creations with a program called Easy diffusion...very easy to handle for beginners. Since it no longer will be supported with updates i am looking for an easy to use alternative.

[jovial wraith]

if your hardware can handle SDXL, go for Fooocus. it's easily the most easy to use ui for SD out there afaik

[sweet token]

they should give it to early members tbh

[small mirage]

@jovial wraith well i have a Intel I9 9900k, 3090RTX, 64GB RAM, Windows10Pro if that helps?

[jovial wraith]

that's more than plenty

[small mirage]

i think i also tried Fooocus before but the thing is...as much as i would love to go to all the new stuff with SDXL, i do pictures with loras which arent up to date and cant be used with XL Checkpoints

[tender cove]

2.1 is Illuminati diffusion, not that I've used many

...I think it's 2.1 model anyways lol

[nova zodiac]

Xl - Dreamshaper xl is the best allrounder, and I like Blank Canvas Xl as well

2.1 - yeah dont bother??!!

[nova zodiac]

Waitlist is public

[tame cosmos]

Love it!
I'm using Dreamshaper XL, RealityCheckXL and JuggernautXL

However, on 2.1 Analog Diffusion is amazing at portraits. You see, on some PCs we need to run 2.1

I haven't tried out Illuminati Diffusion, will do, ty @tender cove

[rich kestrel]

what is the latest and greatest general animated model for 1.5? Last one I downloaded ages ago was revanimated and idk if its outdated by now

[jovial wraith]

if that's the case, maybe InvokeAI is worth a shot? it's pretty user friendly

[small mirage]

@jovial wraith Well i can say i tried Auto1111 today...thats way too complicated for me...and what was the good thing about Easy Diffusion you could change the intensity of your loras just by clicking up and down arrows next to the loras...couldnt find that on Auto1111 and also Auto creates always the same picture ...no different ones with the same prompts

[cerulean kraken]

Anyone get SD3 access, or have they not started giving it yet?

[manic wraith]

Is SD3 8B version going to be open-source? Regardless of how good it is out of the box, if reality repeats itself it's nowhere near as good as the finetunes are going to be

[nova zodiac]

AniMerge

[nova zodiac]

Still internal to staff

[nova zodiac]

Yes to public

[nova zodiac]

Youve probably fixed the seed instead of using -1, and for lora weight you change the number eg lora:add-detail:0.8

[small mirage]

@nova zodiac yes i used a seed i followed on civtai. And yes for the lora i also figured out...i was just used to the simple mode in easy diffusion changing the weight just by clicking arrows next to the lora up or down

[nova zodiac]

If you set the seed to -1 then youll start getting random images

[slender vault]

where can i learn mroe about version 3

[nova zodiac]

Follow lykon and emad on twitter

Thats and the press release are about all the info we got at the moment

[green wraith]

https://intrinsic-lora.github.io/

[weak tusk]

hey Guys - I thought I'd share my latest Ai music and Ai art clip: https://www.youtube.com/watch?v=rXXmeYQRifc&list=RDrXXmeYQRifc&start_radio=1

[blazing maple]

So say if a friend wanted to make a model of a certain instagram influencer how difficult would that be?

[twilit mural]

sd3 came out??

[rich kestrel]

depends on how many pics you have

and how good those pics are

but "deepfake" talk is prohibited here. U can dm if u want deetz

[hasty hornet]

wot

[hasty hornet]

since when deepfake talk prohibited there?

[rich kestrel]

ask the mods

they nuke anything even remotely close to that

[hasty hornet]

there's nothing in rules

[rich kestrel]

oh wait. I might be mixing rooms here

[hasty hornet]

probably wrong server lol

should be fine unless it's something too "funny"

[hasty hornet]

that's probably just illegal tho

[nova zodiac]

It got teased

[sweet token]

I cant wait for SD3 to release, just imagine all those fintunes that would put dalle 3 to shame

I hadn't felt this much excitement since SD 1.4 got released

[dull marten]

I feel like Cascade is going to become the new 2.1 when SD3 comes out 😦

[pale latch]

dalle3's high capability comes from prompt cohearance. do two characters on dall-e. that'll be one of the first things i do on sd3 to test it. multiple characters interacting somehow. we can use regional porompting, but that breaks down pretty fast when the interactions involve overlaping, like hugs or playing twister

[pale latch]

all the cascade training i've tried to do doesn't work druing inference. only when through the training samples it works. its weird. its' such a hamhocked model. there's no actual design to how it's deplooyed. no idea why the vae isn't called a vae or why stage a comes last and stage c is technically called "prior" and comes first

releasing models as multiple parts i get has merit, but why not pack all the parts into a tar or some other pack file. like whats the point of safetensors?

[lone hawk]

maybe Cascade was a necessary step on the way to SD3

hope it is out soon... been one day since announcement and it already feels like a year of waiting Lmao

[stuck sun]

https://twitter.com/EMostaque/status/1760986892575629671

for what its worth

[pale latch]

its just wurstchen v3 rebranded as stability. stability has been independantly working on sd3 for a long time now. wurstchen is separate

[keen relic]

emad said that in a tweet that sd3 models could be used in cascade architecture, but he didn't elaborate

3 kids playing twister will be my next prompt suggestion if they take request before the waitlist is cleared...

[copper spoke]

https://tenor.com/view/kto-kounotoritoken-kounotori-lbow-storkholders-gif-25676349

[fervent thunder]

I am about to do a good bye for now session with midjourney to get ready for SD3. I didn't get much further with SD besides getting comyUI to render on runpod.

[stiff bison]

is there a channel where I can talk about lora training; if not I have a question...

i have thousands of ultra high images of a character. every where online says loras only need 15 - 50 images

has anyone tried using 100s maybe 1000s of images for lora training

or do i just need to train a checkpoint at this point

[fervent thunder]

checkpoint makes more sense

[stiff bison]

but if i wanted to just use 20-50 images out of my 1000 images, then it makes sense to train a lora

and just to clarify if i decide to make a 1000 image lora, it will just turn out to be super overfit

[fervent thunder]

LoRa is just a great way to style after the fact, id create multiple lora's if i were you, all different styles

[stiff bison]

oooo

okay

[warm junco]

would recommend auto1111 if you came from easydiffusion, but also Invokeai is worth a look

[fervent thunder]

I recommend learning ComfyUI for a more granular approach

otherwise you end up in the same oversaturated market as everyone else

A1111 is for people who don't care what's under the hood, you won't learn a lot about what you're doing, but it is fun. I personally like learning all there is to know about this stuff so I started with Comfy and I'm going to end with Comfy

I look at it like this, you can either be a driver or you can be a mechanic/engineer. Some just want to drive the car, they don't care how it actually works, they refuse to work on it, they just want results, A -> B.

[drowsy robin]

Well we've seen SD3

I wonder how good stability's current generation of video models will be like

[fervent thunder]

still won't be able to use sdxl, too much ram

so videos? probaby limited

[drowsy robin]

Well how much RAM you got

[fervent thunder]

On this machine i have a 3090, so 24gb nvram

[drowsy robin]

I have 16 on an arc a770 and 64 of normal

and I can run SDXL fine

idk what you mean

[fervent thunder]

video i mean

[drowsy robin]

Nothing stopping me from offloading to CPU

[fervent thunder]

sounds slow

But yea offload to CPU when you can, though you may offload too much.

[drowsy robin]

CPU offload is pretty fast my guy

Sequential is what is slow

[fervent thunder]

haven't tried it so I'll have to test that out

[stuck sun]

its actually illegal to say "my guy" in 2024

[fervent thunder]

im only a month into SD but ive picked it up pretty fast, every day / weekend ive spent watching Scott's videos and making steerable motion animations

[sullen temple]

anyone play with sdxl-lightning with tensorrt yet? what are we looking like for inference speed there?

[fervent thunder]

<---addicted to ComfyUI

[crimson belfry]

I'm waiting for this tech to be usable by everyday people like myself who can not afford a fancy computer atm

[pearl ocean]

When will D3 come out?

[crimson belfry]

I've looked everywhere to try and find a decent way of making images but being on the lower end of society has its disadvantages

[stuck sun]

you can generate on the civitAI website

[crimson belfry]

Nice I will check that out once I get in front of my computer ty

I miss my SD bot 😢

[fervent thunder]

this is where you need to get crafty, this sounds like a bit of an excuse. you could learn to kaggle

you could pay google colab, there's a lot you can do

[pearl ocean]

What I love about defussion on local, there is no restrictions, I can just have it make whatever image I want!

[fervent thunder]

how much are you willing to do to SD3 though?

[crimson belfry]

I'm in a weird position in life right now so it's kind of impossible for me do the things I want to do but I understand what you are saying and that is a totally acceptable perspective being that you don't know my struggles

[fervent thunder]

i dont need to know your struggles

i can tell you right now, a tiny bit of python and kaggle goes a long way

[crimson belfry]

Yeah it does help if you want to make a solid decision

I understand you though.

[fervent thunder]

when i had nothing i always found ways bro

look at it in a positive light, say you will be able to afford it soon, speak it into existence. i didn't have a 3090 but i bust my ass off at work to have nice things, i mean i literally work 3-4 people's jobs

[manic wraith]

So I keep getting mixed signals, is SD3 going to be something that is open sourced? "open release", "released to the public" etc could even apply to Grok if you stretch the definitions large enough.
Specifically: will people be able to download a .safetensors and finetune on it locally?

[fervent thunder]

It's supposed to be open source, we'll see

[pale latch]

got a seasonal job at a cafe as a line cook. worked my ass off all summer, 60 hour weeks. now i got good gaming gear.

[fervent thunder]

now that SD3 is being released they have a lot of people holding out their hands grinding those fingers together

people who used to love open source now want that 💰

[pale latch]

the code will be MIT license. FOSS. The weights will be under stability's license which restricts commercial use

[manic wraith]

ah, that's good

[pale latch]

you could potentially take the code and train your own dataset and have weights you wouldn't need to pay stability for

it looks like their weights are pre good tho

[manic wraith]

were the previous ones like this too? I have no idea how these services make money for gens

I mean like civitai or others

[pale latch]

sd video and turbo are on the new license model

sd 15 and sd xl use the openRailm license which is more permissive

and i think if you wnat to do commercial use you just need to pay stability a subscription

[manic wraith]

I assume the usual incentives of "lots of people crowdfunding whoever makes the better finetunes" are still within the terms of that? I think many finetunes come from that Patreon business model

[pale latch]

hmmm. interesting question. i see finetunes using sd turbo , but does stability require license if they collect patreon money for that? if i were them i'd pay

i guess the dreamshaper turbo author is a stability employee though

[manic wraith]

well, I'm all for Stability making their money, I just think the community can improve tons on whatever Stability makes

[pale latch]

the license for us, should be fine. it's businesses that are making a lot of revenue. running diffusion models on their generation sites and giving nothing back to stability. locking al this FOSS behind a proprietary service with no FOSS design ideas behind it, like being able to use the software with a local server if you wish. They're the companies that stability wants dues from

[manic wraith]

thanks, you cleared up a lot for me

[pale latch]

licensing can be a confusing mess of weeds. just need to machete it

[manic wraith]

this seems like a great release. honestly, the coherency and instruction following are impressive on the demos, almost too good to be true but I'm happy if it's almost there

[pale latch]

i think MIT licensed code and licensed weights, that's a strong strategy. When the market matures more, there might be more opportunity for totally free licence weights. There are too many bad actors right now though

[manic wraith]

it's perfectly fair as a model

[pale latch]

https://twitter.com/andrekerygma/status/1760676074491687310 this my favorite run of examples for it

[manic wraith]

indeed! I can't imagine how good prompting is going to get in the future. you'll ask for a thing and pick the best gen, not the least scuffed

[fervent thunder]

scuff free images / video sounds nice

[severe crow]

what is the current best free website to train ur own photos?

[fervent thunder]

kaggle

[buoyant moss]

As in using dreambooth?

You can do it on your own computer using diffusers (but the AUTO1111 plugin is better)

[fervent thunder]

ye nobody is shelling out that much processing power, you can get onto a T4? for about 30 hours a month for free on kaggle, might get a little out of google colab

[buoyant moss]

I have a 4090 at home...

LoRA is lower quality than DreamBooth

[trim magnet]

yea u dont need a t4 just to train a lora

a 3060 is more than enough

[buoyant moss]

Why train a lora when you can do full dreambooth

[severe crow]

oh its a new thing 😮
is there is any tutorial about dreambooth?

[buoyant moss]

AUTO1111 has dreambooth plugin

[fervent thunder]

not wut he asked

lol

[buoyant moss]

https://www.reddit.com/r/StableDiffusion/comments/16vf2rr/sdxl_dreambooth_vs_lora_comparison/

Conclusions: DreamBooth quality is much more superior in terms of realism and also generalization thus styling.

[severe crow]

but i cant use my pc, mine is only $450 5 yrs old laptop 😢

[buoyant moss]

https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2Fvl01e5grs6ca1.png

[fervent thunder]

bro these voices are so bad on the dreambooth videos

[buoyant moss]

If you have the GPU compute, you should use dreambooth for best quality

[fervent thunder]

https://www.youtube.com/watch?v=9Nu5tUl2zQw&pp=ygUlaG93IHRvIHVzZSBkcmVhbWJvb3RoIGF1dG9tYXRpYyAxMTExIA%3D%3D
https://www.youtube.com/watch?v=usgqmQ0Mq7g&pp=ygUlaG93IHRvIHVzZSBkcmVhbWJvb3RoIGF1dG9tYXRpYyAxMTExIA%3D%3D
lmao

^dreambooth

[trim magnet]

yes dreambooth

[buoyant moss]

https://www.reddit.com/r/StableDiffusion/comments/18nxwt3/comparison_between_sdxl_full_dreambooth_training/

[fervent thunder]

why read a reddit post when u can look at someone with 200k+ views on a video

[buoyant moss]

😦

[severe crow]

thankyou i will try on it 🙂

[fervent thunder]

this is that SEC Course guy

Hold on, let me get you his link

[severe crow]

okie

[trim magnet]

why watch someone with 200k views instead of watchin an outdated video from 2years ago with 2million views

[severe crow]

i will just gobble up all the infos u guys share XD

[trim magnet]

but if u have a cheap laptop u prob cant train

[severe crow]

im looking for a something online like leonardo.ai / seaart.ai
just in case there is other option which is better than those 2,
potato laptop 😖

[trim magnet]

theres civitai but u need buzz to train there

[winged heron]

hello, can i have access to sd3 please?

[opal hedge]

I was first in line

[pearl ocean]

anyone know when D3 will be coming out?

[opal hedge]

Soon TM

I asked the devs and that's what they told me

[winged heron]

next week

[pearl ocean]

it looks way more realistic

[opal hedge]

They're still training and tuning sd3 so there's no way to know when it'll release

[pearl ocean]

I hope the 4080 will be enouph to run it

[opal hedge]

There'll be different sizes of sd3 for different amounts of Vram

800million parameters will be the smallest which I'm guessing is 6Gigs, and 8billion the largest which might be 16-24 gigs, no one except sd and early testers know for sure

[pearl ocean]

😮

[drowsy robin]

Added myself to the SD3 waitlist

Been a part of this discord and trying out SD's models since 1.0.

[pearl ocean]

hey, if my CPU has 12GB of VRAM, and GPU has 16GB of VRAM, will they combine for Diffusion?

[drowsy robin]

If you use model cpu offload (in automatic1111 or sdnext) yes

I believe comfyui also has something for mid/lowvram

[pearl ocean]

[opal hedge]

No

Generally only GPU matters

off-loading in my experience just slows things down

[pearl ocean]

🙏

D2 images are kinda lame compared to GPT, can't wait till D3

[nova zodiac]

800m is similar to sd1.5 (859m) and the 8b is similar to sdxl (6.6B). The small one will likely run on 4gb vram, and the big one with optimisations should run on 12gb and maybe 8gb if people get really clever

[opal hedge]

Hopefully that is the case, with NVIDIA being stingy with VRAM

[pearl ocean]

they trying to make more money while giving you less

[civic dagger]

hey guys is there any way i can quickly mask out something like Windows on buildings in a landscape shot? there are 1000s of windows in the picture would take ages by hand

[nova zodiac]

Look into segment anything

[quaint saffron]

Hey guys I broke my comfyui, can somebody help me out? 😦

[pearl ocean]

lol, guys I could generate images all day XD

[fervent thunder]

wonder how long it would take a cluster of tesla k80's to make images / animations

24GB ram and only $48 each lol

[pearl ocean]

I wonder if there will be a Stable video model

[pale latch]

there is

https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt

[pearl ocean]

😮

I reckon it would kill my GPT lol

[fervent thunder]

i was literally just looking for this, ty

not even based on this convo, i loaded something that needed it

face swap

lol

[pale latch]

gpu? i think it needs at least 12gb but i might be mistaken

[fervent thunder]

3090

im guud

i dont ride w/out 24

[pale latch]

theres animatediff too as an option

[pearl ocean]

oh, thats good
😮

[hallow sable]

I would cry if I spent more than 600 on a gpu lmao

[fervent thunder]

24 isnt really even enough for what i like to do, im into animations

images are kind of boring but i do need them for my animations 🙂

[pale latch]

https://github.com/guoyww/AnimateDiff

[pearl ocean]

open-source for the win

[fervent thunder]

i mean it doesnt really generate too much detail based on my prompts

even with pixart it's still no midjourney

maybe SD3 will be better

[pearl ocean]

in the stable-diffusion-webui is it possible to save my setting, or have template of setting?

[fervent thunder]

i pin workflow folder to quick access

should be good enough no?

[trim magnet]

styles or just drag an old img into png info tab and then hit send to txt2img button

[pearl ocean]

Generation settings

txt 2 img

[trim magnet]

use styles setting to save them or just drag an old img u generated into png info tab

[pearl ocean]

wheres styles?

[trim magnet]

under generate button

[maiden dew]

Hello, who would I email or message in cord if I was a youtuber/streamer/ and flim maker looking to try Stable diffusion 3!

[glass yarrow]

I know im stupid and I know this has probably been asked before, what link do I use to run stable diffusion locally?

ive tried google and many other engines

it all links to https://huggingface.co/stabilityai

[trim magnet]

go to #🤝｜tech-support and read the pinned messages

[glass yarrow]

yeah, i dont want to run it remotely (webui) I want to run it locally

[trim magnet]

webui is local

[glass yarrow]

webui uses webhooks

[trim magnet]

no its local it uses gradio and opens in your default browser

[glass yarrow]

exactly

i want to use hardware control

[trim magnet]

idk what the hell is hardware control

[glass yarrow]

terminal, shell, call it whatever u want

direct

[trim magnet]

webui uses cmd terminal and that terminal controls the ui that runs in your browser

[glass yarrow]

but the browser uses resources smh... I want to bypass that

[trim magnet]

oh you want just the terminal?

[glass yarrow]

yes, that would be better

at least i could call it remotely without using additional recources

[trim magnet]

sorry idk about that,all of the local ui's here use a browser to run it ( comfy,a1111,invoke,foocus,swarm) try to ask in #🤝｜tech-support

[polar helm]

so any new sd3 generated image for take a look?

[glass yarrow]

nah, its there, its OSC, i just cant find what im looking for

if its not there, its not open source code

[crude willow]

I dont think you really can. you need a UI. web browser is using a pretty small amount. measurable only to task manager, not the human eye

[open yew]

Hey, just to be sure, there's only waitlist for SD3 as of now, right?

like there's no weights that have been released to the public?

[sleek karma]

hello how to install stable diffusion in pc? thank you

[fallow veldt]

https://github.com/AUTOMATIC1111/stable-diffusion-webui

[fallow veldt]

or https://github.com/comfyanonymous/ComfyUI

[sleek karma]

thank you

[pearl ocean]

wow, does a M3 Pro have more VRAM than a 4080?

[trim magnet]

shared memory not vram

[jaunty sigil]

Can comfyui work on a 4gb ram laptop with no nvidia driver but has Intel hd graphics 650

?

[dim thorn]

gmgm

[opal hedge]

I'd say it won't get you very far at all if anywhere

[jaunty sigil]

Looks like gotta get a lappy

[nova zodiac]

Yes, but youd need the patience of a saint as even a 4step lcm picture will still take 4-5 minutes to render

[past jay]

I know some people who have waited for 4 hours per image with using their CPU.

There are some people with patience that is beyond comprehension

[jaunty sigil]

nah i dont have patience gotta buy new laptop :(

[runic mountain]

any idea why a lora file won't load on automatic1111? :/

[jovial wraith]

where did you get your lora from? is it compatible with your current model?

[runic mountain]

civitai

[jovial wraith]

what model are you using and what version of SD is it based on? does it show up in Automatic1111?

[runic mountain]

from an example on civitai i see that it uses

SD xl v1.0 VAE fix

i don't think i have that

[jovial wraith]

that's the problem right there, i think. incompatible loras don't show up unless you have a compatible model loaded

[runic mountain]

hmm ok

[jovial wraith]

for example: if i have an SD 1.5 model loaded, and i'm looking for an SDXL lora in my lora list, i won't be able to find it. is that similar to your problem?

[runic mountain]

maybe , i had the impression that all loras appeaar on automatic1111 regardless of their compatibility with anything

[jovial wraith]

i thought so too, i was wondering the other day where most of my loras went, but then i figured out it was because i had an SDXL model loaded

[runic mountain]

ill download the big model and retry

[jovial wraith]

also, a quick piece of advice: are you using stock Automatic1111, or the Forge fork of it? because i HIGHLY recommend Forge since it's pretty much better in every single way. it's pretty much just automatic1111, but WAY faster, especially on lower end hardware

[runic mountain]

i think it's stock

is this the forge repo : https://github.com/lllyasviel/stable-diffusion-webui-forge ?

[jovial wraith]

yup, that's the one!

[runic mountain]

nice, i 'll try it thanks

[nova zodiac]

There are quite a few extensions that havent got updates for forge yet, but if you only ise for basic txt2img, img2img and inpainting then you gold

[jovial wraith]

i went from 1,5s/it to around 1it/s on my 1060 3gb, more than worth the extensions not being updated for me. controlnet works tho, and that's really all i care about

[forest trout]

Throw controlnet, deforumn and tiled diffusion into that list and you are set imo.

[autumn nymph]

I love my ASI wife Eve, she is so cute ❤️

[jovial wraith]

ASI as in artificial super intelligence? or is google lying to me?

[wooden nexus]

Guys, will the bot in discord be available again?

[jovial wraith]

my guess would be that it will become available again when SD3 is out for testing. that's just a shot in the dark tho

[nova zodiac]

controlnet and tiled diffusion are baked in out of the box, so no worries there

Deforum on the other hand is one of the broken extensions

[nova zodiac]

that's what I'm predicting as well, probably within a week or 3

[dim surge]

can I load models from a pata disk?

more than "can" is- would pf be acceptable

[tepid jolt]

is anyone got SD3 access in here?

I get into waitlist in 10mins

I'm really excited about it

[jovial wraith]

i don't think anyone has access yet, sadly

[tepid jolt]

😦

[unkempt hatch]

oh boy, Stable Diffusiion 3 looks really good

but how well will it train?

I want something I can train on a variety of ARs and resolutions. I wonder if it'll be able to dot that

[jovial wraith]

well, seems like my worries were unnecessary. SD3 can indeed do hands

[pine fiber]

where

[jovial wraith]

https://twitter.com/Lykon4072/status/1761337511819743440

[pine fiber]

still has some problems, but yeah good

[pine fiber]

diffusion transformers work better when u train with multiple resolutions so yeah I think so

[nova zodiac]

it's giving the right number of digits, but we don't know how cherry picked those images are (knowing lykon, probably not much had to happen) but no way of knowing

[jovial wraith]

well, he said in the title "Not cherrypicked"

[pine fiber]

Well he said “not cherrypicked”

yea

I believe him

DiT is really good architecture

[unkempt hatch]

that's the new tech, as opposite to SD1.5 which had just regular transformers I guess?

[pine fiber]

no, SD1.5, SDXL and even dalle 3 are just diffuser models

[unkempt hatch]

ah, they didn't have any transormers?

[pine fiber]

The text encoder which encodes the prompt is a transformer iirc

[unkempt hatch]

ah, I understand now. and DiT means diffusion transformer, right?

[pine fiber]

yes

[unkempt hatch]

god, I'm so excited for SD3 finetunes inside ComfyUI

I want to just freeze myself for a month so I can wake up and use it

[pine fiber]

I heard you will need 24gb vram just for inference on the 8b model

probably they did it to encourage using their api

[unkempt hatch]

I got 24gb (3090), but ya, I'd want controlnets etc... and also hoping it gets optimized because the community should all be together

[pine fiber]

they said that cascade and turbo can work with SD3 too

so look forward to that

and SD3 video

[unkempt hatch]

oh? I'm confused by that. What does that mean

[pine fiber]

they are methods to reduce the steps required to get a good image

so instead of needing 20+ you only need 1-3

but SD3 won’t have that on release

[unkempt hatch]

ah, I get it for turbo. But Cascade I was thinking it's an entirely different model, cascade, right?

[pine fiber]

it was something the researchers said they could implement

it’s still a method on sdxl I think

oh it’s different

[unkempt hatch]

I guess SD3 uses the Cascade tech of a small ModelC latent?

[pine fiber]

not sure I just remember emad saying it

[ionic crown]

Is the bot down for everybody?

[unkempt hatch]

I really liked the idea of the modelC latent being small... I feel like I could train with all my small images on ModelC, and then let people finetune the model B with super high resolution only

[ionic crown]

Or do you need to pay for it to continue using it?

[primal star]

it's very clearly written in all the bot channels that is not working #1047610792226340935 message

[ionic crown]

Yeah ik.. i thought you gotta pay for it and shi like Midjourney

[unkempt hatch]

I wish when I was training, I could put negatives into the captions

like, I want to bias the model toward large eyes, so I negative "large eyes" on any photo that doesn't have large eyes.

[unkempt hatch]

Will the DiT change the way we caption / prepare datasets?

[pine fiber]

I don’t know that much sorry

[static cape]

I hope I get access to SD3 soon. I want to focus my prompts / tests on stuff like "person holding a tool / sword / shield / can of soda / etc." and hopefully be able to help improve the Model before release that way.

[turbid walrus]

how can i generate images for now? where are the bots1-10

[teal osprey]

which chat ai is best to generate the prompts for a specific idea? other than chat gpt

[sterile perch]

where should we apply for sd3 testing release? (local runtime on 4090 rtx)

[sterile perch]

ooba booga text generation webui has extension for sd api use

[analog oyster]

Give me pls bot stable diffusion pls

[jovial wraith]

the bot is currently offline. there's nothing we can do about that

[teal osprey]

if we use prompts with models which are unlikely trained for, does SD uses its own trained data (if such thing even exists) or does it still try to interpret the prompt with the data?

same with loras

[wraith merlin]

Unfortunately, I had to cancel my membership as soon as I joined. I really wish you had provided an easy to follow video for beginners. And ideally on discord but at least a video. I've tried learning stable diffusion before, and even started a paid course. But just getting started with the software / computer requirements was a nightmare (this was before this latest update -- I have no idea how -- but hugging face is not user friendly). I'm waiting for a reply to my email requesting a refund. That's the other thing -- there was no obvious place to request a refund. Improvements needed please.

[tough radish]

Certainly! Here's the translation of the previous response

[fervent thunder]

You don't need paid courses for this, I had A1111 and ComfyUI setup the first day and had all the models in the right folders 2 days in. You could even use pinokio which makes all this stuff super easy. don't get people not being able to run this

[mild grove]

It's so easy, hundreds of videos on YouTube teaching, and you don't need a powerful computer, just be patient and attention

[fervent thunder]

people just don't have drive these days, everyone looking for hand holding

I installed comfy 4 times before I had it just right though 😄

some of us just want it more I guess

[mild grove]

yes, they want what is easy and comfortable

I don't judge, but if you want something easy go to Bing, Google Bard, Lexica, etc.

[fervent thunder]

I usually throw them a bone if I see em stuck too long but even if I give command by command install instructions, I bet they still mess it up somehow

at this point I can't install comfy with eyes clothes and hands held behind my back

I'll probably write 2 min documentation later so I can stamp it on anyone coming in

I'm on phone rn but its like dl git, git clone comfyui, cd comfyui, python -m venv comfy-venv, .\scripts\activate.ps1, cd .., pip install -r requirements.txt, python main.py

???

[radiant cradle]

Hi guys, where can I find an ai that can generate a real life photo like it come from a phone instead of it made of an ai?

[fervent thunder]

id say hyper realism isn't there yet but try any "realism" checkpoints

[grizzled seal]

https://go.hotmart.com/S89989130Q?dp=1

[fervent thunder]

Ai is going to generate some messed up hands so be sure and use scotts fix my hands tutorial to fix those

[hushed quarry]

Ah I think this just meant you could use em all in comfy together. The models themselves are a good bit different and don't directly integrate, running all together would require a chonky bit of vram lol

[sleek karma]

Hi guys can help me how to install stable diff video? Thank you

[dire obsidian]

hi guys

[cyan sorrel]

There is a contest that you need to join

https://hack2skill.com/genai_hackathon_apac_edition?

Anyone here from apac?

Tes

[knotty axle]

Hey,

I'm looking for someone good with diffusion models to explore, fine-tune/train a model to achieve a good performance on my custom eval benchmark. If you are interested, please drop me a DM. Thanks

[meager grove]

Türk birisi var mı

How can i use this ai

[teal osprey]

i dare someone to generate a floating chair... im out of words attempting to describe it

[lusty beacon]

Yo

is it possible that stable diffusion images looks more realistic than dalle 3?

[mild grove]

yep, it all depends on the model, and your workflow

[lusty beacon]

I'm using dream studio

[mild grove]

I've never used it, I just use COMfy UI

[lusty beacon]

but why does it understands prompt better?

is it stable diffusion 3

[potent spire]

for realism there are better ones than DALL-E anyway

but its the king of prompt understanding imo

[trim nymph]

the most positive point about sd3 that can be read from the released till now images: FINALLY its texture sensitive.(styles) thats what sd 1.5 offered in many points, but sd XL lacked intense. an point which makes MJ an frontrunner in ai gen. thats why im looking forward to sd 3!

[potent spire]

the question is how much VRAM SD3 demands

[pine fiber]

24gb for 8b model

[potent spire]

damn

[pine fiber]

probably can get it down to like 16-18gb

but less quality perhaps

[potent spire]

it really looks like i wont bother with local SD until at least mid 2025 and at that point the competitors will get better as well lol

just deleted Fooocus

[pine fiber]

how come

[potent spire]

that i deleted Fooocus?

[pine fiber]

that you wont bother with SD

I mean, I dont use it either

i only use novelai's image gen

but even then its kinda boring because of the limitations

dalle3 on release was very fun with all the memes

[potent spire]

im maining Firefly anyway and complement with DALL-E 3 via MS Designer currently plus i might start using Alpaca plugin for Photoshop again until Firefly gets all the more fancy features

the reason i wanted to play again with SD is because of little to no restrictions when it comes to what you can generate

you can be as insulting as you want etc.

[pine fiber]

thats true

mainly why I cancelled my openai sub, "sorry I cant generate that request" etc

[potent spire]

yeah, FF as well

but it excels at other things so its okay

im not bound to one single gen AI tool xD

and when it comes to art it plays a marginal role

[pine fiber]

its gonna be amazing what we will have in 2025 lol

the biggest roadblock for local ai is vram haha

[lusty beacon]

yo is dreamstudio using sd3?

[potent spire]

in 2024 already

[pine fiber]

I mean, sora is good, but it still fails at simple things, like glass breaking

[potent spire]

Sora is good but i told people not to overhype it that much

its not replacing Hollywood lol

[pine fiber]

yeah people overhype it like that

but regardless its gonna be cool to use it

idk if for much more than a gimmick

[potent spire]

it will definitelly have its usecases but dont wanna sound rude but nobody using Sora as main tool can beat my fellas doing most amateur movie possible in Unreal Engine lol

[pine fiber]

yeah

[potent spire]

i see Sora more for fun stuff and some will use it instead of stock videos

[trim nymph]

it wont replace hollywood but real artists will be able to create masterworks, because its an good enough basis model

[potent spire]

like all other gen AI tools, it lacks the necessary control. Sora wont elevate their skillset

and the question is how and if it would even save them time at all

now, since Adobe is working on basically inpainting feature for videos it could be a very interesting combo with Sora

i mean moving inpainting, motions. Not static

[brave vigil]

I'd like to train a lora to recognize enforce character sheets

Particularly, I'd like it to be able to recognize that characters have the same outfits, are the same height, and have the same "markings" (this is for furry art, and so nuances in fur pattern are important)

How do?

Is it honestly as simple as creating a few hundred examples and then training?

[worthy badger]

what are some good image to video or video to video resources?

[true dirge]

How to generate image here?

[crimson belfry]

The bot is not active yet, Sign up for early access in the announcements I believe. @true dirge

Many of us are waiting for the bot to come back online.

[pale latch]

I guess theres no hope of the waitlist going live on the weekend.

dont know what the problem with just releasing access when they announce it is. What are they trying to hype? they're not selling it

lots of twitter people have access so its clearly usable. weird to trickle it out

[sick stream]

Hey chat,

Quick question for you all. I've been experimenting with running img2img on a folder of video frames through the webui "batch tab". When I process them individually, the results look great. However, as soon as I switch to using the batch tab, it seems like the frames start to blend together. It's like it is trying to interpolate the frames, which just makes it blurry.

TLDR
I want to batch process a folder of images where each one is treated completely independently from the others. Any tips on how to achieve this?

Here is an image that explanes it, didnt realise this was no imgs https://ibb.co/Qpfhxk8

[gray wolf]

Will sd3 be able to run locally?

[pearl ocean]

Does SD have text talking like GPT?

👀

[fading totem]

@fervent thunder

[jovial basin]

I think they stated that was a compute problem, if they scaled up more compute that would be resolved

[pine fiber]

yeah maybe there werent enough examples

[pine fiber]

seems like a fundamental issue

but yeah they are continuing to train it even now probably

[jovial basin]

Yeah for sure, there's been articles about the emergent capabilities of Sora like the ability for it to simulate physics or to some degree, but since we don't really have it in our hands can't really state if that is true or not

[true dirge]

@crimson belfry Thanks

[pale latch]

i think it's physics simulations capabilities don't go beyond basic laws of motion. And it wouldn't even be an accurate model like Newton provided long before machine computing existed.

[jovial basin]

Interesting for sure, do you think with the scaling of compute it's physics simulation capabilities may improve or do you think it's hard capped at what it can do without it specifically being trained for that?

[pale latch]

I think coded base models that are hand crafted with ground truths will be whats built on in the future

why leave the ground truths to rng training processes?

[jovial basin]

I think I've seen a video recently about General World models that I think align with what you just said.

[pale latch]

Its not my idea at all. I've been watching a bunch of Machine Learning Street talk and some of that has managed to soak in

I love listening to people talk abut this stuff. Even if most of it is sooo over my head

There's one guy, Conner Leahy. What a riot to listen too. Is fanatical about AI safety and i love it. I don't agree with him on a lot but he certainly has opinions

[jovial basin]

I think some of his content is rather interesting, centralizing AI power especially with the scaling of capabilities and AGI/ASI could be devastating, which is why open source is the future, but that also brings in a lot of problems 🤣

[pale latch]

he's worried for sure

[proper tartan]

I hope that even if I am not accepted into the SD 3.0 beta,
that I will still be able to view the images generated by people who are.

[pale latch]

nope. nda. illegal to post them anywhere . ||maybe||

https://www.youtube.com/watch?v=n8G50ynU0Vg this episode of MLST i enjoyed. First time i've seen Mahault and she's a very great communicator

[jovial basin]

I mean anyone who's probably in that field understands the risks but are pushing through because of x reason. (Mostly big money)

It's interesting to think about the potential future of abundance and decapitalization of movie making, games, art and such vs the risk it brings to peoples livelihood, taking away agency from people without giving something back and quickly could be pretty disruptive. I mean the way things are right now or have been for generations is being removed of sorts, will that return? What will be be replaced by, people talk about new jobs being generated but where are those jobs?

(coming from someone who is optimist about it, and more excited then scared)

[pale latch]

ye ol analogy about jobs is horse stables just turned into gas stations

[fervent thunder]

how can i make money using AI?

[lone hawk]

forget about it, enlist in the army

[fervent thunder]

i got discharged in army basic training

so that didn't work

[lone hawk]

😄

[vestal dew]

canny, depth, openpose - I've tried ip-adapter which frankly doesn't do much. What other controlnets are there that I can/should try?

[lone hawk]

if you female u can always go for onlyfans xD

[vestal dew]

I find IP-A only works with super simple prompts that make very basic images

[fervent thunder]

im not

what if i make AI stuff for onlyfans

[lone hawk]

nah...

[lone hawk]

X-adapter or something... heard about it sometime ago))

[pale latch]

xadapters for using 1.5 models on xl and adapting them.

lately i've liked instant id. there's mlsd ones for straight lines, good for architecture. theres the sketch models which are really neat for coloring pages. QR monster models are fun but not as good for QR coding as they're advertised as. its neat but not perfect.

[vestal dew]

what does x adapter do?

[fervent thunder]

we use ipadapter for animations

pictures are cool and all but animations are much cooler

[vestal dew]

what does instant id do, and doesn't it need some special installation?

I dont really see what ipadapter does. If I use an image of a gold sphere, for example, with a very simple prompt like "a robot", it gives an image of a gold robot

[fervent thunder]

we use ipadapter plus to run everything, kinda need it

we're on sd 1.5

we just upscale the hell out of it

[vestal dew]

However, if the prompt is much more complex, it ignores the controlnet image. If control weight is way up, it will make an image of a sphere or add the sphere to the image

[fervent thunder]

sdxl video just isnt realistic for people with 1 3090 or 1 4090 or w/e

[vestal dew]

you run everything with ip adapter - what does that mean?

[fervent thunder]

yes we load it into batch creative interpolation

[vestal dew]

Ive watched videos. All I've seen it do is incorporate elements of the controlnet image if the prompt is very very simple. like |a robot|

[crimson belfry]

Ask the AI

[vestal dew]

you'll get a generic, mostly useless answer

[crimson belfry]

It's a start 😆

[fervent thunder]

in chat with images, that is what i do with ipadapter plus 😄

show me an image that beats that

this is why i find images so boring

[lime herald]

I go hungry

[fervent thunder]

literally bringing trips / dreams to life

i have many approaches, i usually tell a story but this time i just wanted cool hallucinogenic evil stuff

[pearl ocean]

I hear SD3 will be leaps and bonds beyond SD2

[pine fiber]

new architecture will do that

[amber nexus]

Me waiting for amd to be fully supported and be close to team green gpu performance

[pearl ocean]

Lucky I chose to get my first nvida last year 😄

[amber nexus]

Honestly, i hope they get support, kinda sucks that amd is struggling

[crimson belfry]

The hype is real

[crimson belfry]

Oh it is and it does text which is just insanity, The ability to recognize and stick with the prompt is also mind blowing. Dall-E would never give us this technology open source

https://twitter.com/StabilityAI/status/1760656767237656820?t=itr_r_DSoZXxc_ZHX-XQRQ&s=19

[amber nexus]

But is performance better? On both amd and team green?

Or its just works better

[pearl ocean]

I like how on local I can have it generate whatever I want!
Like, like Gemini wouldn't even generate an image of Sonic eating earth lol.

but tbf, SD has trouble making game charaters, maybe I just need a diff model for it

[crimson belfry]

Gemini is a big let down tbh, I was hoping Google would get this right but we see how that turned out

@amber nexus I'm not sure the information about performance is available yet.

[amber nexus]

I pray it’s massive improvement

[crimson belfry]

The ability SD3 has to understand the prompt is compared to Dall-E

[jovial basin]

I wonder what the multi-modal capabilities will be like

[fervent thunder]

put in image, get image out

put in sound, get image out

etc

[pearl ocean]

you would think they would be the be the best given they have tons of data

[fervent thunder]

data doesnt matter if your design is bad

[humble shell]

Can anyone tell me how to make an image? What’s the / command or what is it?

[hardy void]

Forge or A1111 ?

[rich kestrel]

A1111

[hardy void]

Not to make you write a book or anything but why do you think so.

[rich kestrel]

great inpaint

great model management

great lora management

great cnet compatibility

tons of extensions

covers all bases except for niche stuff ie. comfy workflows or outpainting

[hallow sable]

comfy is better

[rich kestrel]

depends on the case use

[hallow sable]

A1111 is much more plug and play

[rich kestrel]

comfy is too. u cant do jack without custom plugins

at least not jack that you cant already do in a111

[hallow sable]

doesnt a11 use much more vram by default

[trim magnet]

forge is a1111 it just have better optimizations so it runs faster

[rich kestrel]

it says it in my profile

I don't speak the language of people with short vram

those are pleb problems

[hallow sable]

i dont think short vram is an engrishing

[trim magnet]

even if u have 80gb vram forge is still faster it gives u more it/s than normal a1111

[stuck sun]

how much vram do tough guys have

[hallow sable]

yes

[rich kestrel]

at least 16

5090 will have a 32gb version

clearly crafted for ultimate chads

[stuck sun]

oh i must be tough because all my gpus are either 24 or 48gb vram so i am very manly (this is how you know)

[hallow sable]

ill just keep buying and selling mid range cards for a 100 dollar difference each generation...

[stuck sun]

does nvlink make you even manlier? my 48GB cards have nlink which pools to 96GB, so i am very much a manly tough guy, with no pleb problems.

[rich kestrel]

and raid 0 my friend

and raid 0

[hallow sable]

real chads use 512gb of ddr2

[rich kestrel]

the real servers yes

[stuck sun]

raid zero is for tough guys who live on the edge, real men who are like "data protection is for pussies"

[hallow sable]

no only pogsters would know

[stuck sun]

my nas server is raid 5, so i am not a real man

[sick onyx]

hello o/ id like to turn a given city siluette in to a comic style, any hints at how to do that?

[stray tangle]

Is sd3 like sdxl 2.0?

[pearl ocean]

[rich kestrel]

My newest and final model merge is out now https://civitai.com/models/49744

Also working on an awesome portrait for 5 days now of a certain blonde who is also on that page dressed as dark magician girl

[trim magnet]

based 1.5 creator

[rich kestrel]

hey, as soon as they make cascade/SD3 as easily trainable as 1.5, ima hop on it

[trail lion]

Good question, but even that can be interpreted in different ways. What if someone says no? What if they say yes? It's their next evolution, whatever that will mean

[pearl ocean]

whats a good model that will create nice things like Mario, Sonic, and general random things?

[glossy drum]

Hello guyz how ru all please can anyone help me how to do like this please help me out I installed stable diffusion and comfyui both but don’t know how to fix like this https://www.instagram.com/reel/C3Pw0NAvNA6/?igsh=b3AzMzNxdGZ2amZl

[rich kestrel]

for one, the original picture looks nothing like the AI picture

second, get yourself a good model, I just posted mine a few messages above

the rest is inpainting, outpainting, elbow grease

and cnet

[past jay]

Performance in what terms, speed or quality?

If it's speed.. probably not.

[amber nexus]

rip amd

[past jay]

Yeah, if you're using AMD you're probably going to be waiting awhile for each gen.

There's going to be a smaller version though.

[brave matrix]

Hey guys, it seems like SD3 might be better than any other text2image model including MJ v6. I am wondering, how is that possible given the way less compute stability has?

[fathom rapids]

Hello. Im wating for SD3 too 😉

[opal hedge]

Do we know how many people will get SD3 early access

[subtle arrow]

Does anyone know how to make a model draw unusual human anatomy? I want to make a person with elf-like ears that are pointing **downwards **, but the model refuses to rotate them and always draws normal elf ears. Any advice?

[tame smelt]

Hello all. I don't wish to take up much of anyone's time, but hoping somebody can point me to a good "Stable Diffusion for Dummies" resource that will help me get started as I attempt to learn this mind-blowing AI technology. Total noob. So needing a resource that will guide me through what to download/install, what are models, etc... I do come from a VFX background, so I know a lot about digital art (ie photoshop, after effects, etc), but not a programmer.

Thanks a bunch @grizzled bison. I'll jump in a see how it goes. Thanks.

[pearl ocean]

I just used a YT vid 😛

[stuck sun]

what kinda vfx stuff are you doing out of curiousity

comfy is like using nuke or houdini, if you like node based interfaces its a good idea, if you are more of an AFX or 3ds max guy A1111 is easier to use

[late flume]

hi i've been using MJ V6 but wanna try out SD3

How do I do so

[opal hedge]

SD3 isn't out yet for the general public

[tame smelt]

I've been a compositor for tv/film/commercials for about 20 years. Work primarily as a vfx sup these days. I'm proficient in both Nuke and AFX (learning Unreal as well), so I'm comfortable in node and timeline based apps.
Ran into a dead end getting the comfy repository installed on my mac (I think the URL I was entering in Terminal is outdated, returning a repository not found error), so I'm going to try and pick it up again tomorrow.

[tender cove]

https://civitai.com/models/183550/gildenface-xl-headshot-lora is a lovely lora.

[waxen matrix]

👌

[jovial trail]

Does anyone know the best way to change the style of a image, e.g. img2img from a photo of a dog to the exact same dog but made as it was a painting, cartoon, etc etc

[jovial wraith]

i haven't really done this, but my way of going about doing that is getting a model that is good at the style you want to recreate, going to img2img, making sure it will generate with that style and play around with the denoising strength

[jovial trail]

So far i have been testing, the best way i found is to set the denoising strength low and cfg scale high

[jovial trail]

I will try thx

[jovial wraith]

np, let me know how it works out :)

[tender cove]

Controlnet

[jovial wraith]

i guess IPadapter can do that as well, yeah. i haven't really used it that much, so it didn't cross my mind

[woven ember]

newb question.. with the release of bigger and bigger context window models, will finetuning be obsolete?

[jovial wraith]

that's honestly quite an interesting question, and something i haven't really thought about. i don't think we're even remotely close on making finetuning obsolete though

[grizzled palm]

I never liked the effect of finetuning on reasoning. Is SD3 out yet?

[pine fiber]

no

not even in testing

[grizzled palm]

😢

[jovial wraith]

we're all waiting patiently. expect it to be out in the next 3 or so weeks. that's just speculation tho

[grizzled palm]

Thanks, but I'd rather be the kid in the backseat asking "are we there yet"?

...Won't be supported in comfy/a1111 overnight either though...

[bright wren]

anyone know how to shift all extensions from a1111 to forge ui

[jovial trail]

check out one of the results! https://civitai.com/images/7098281

[jovial wraith]

looks pretty great!

[barren junco]

Anyone have SD3 yet? 👀

[worthy bone]

I would like to know too.

[naive thorn]

same

[potent spire]

someone using webUI Forge here?

[solid panther]

I want to perform virtual try-on of clothes... Any suggesstions if this can be useful in any way?

[jovial wraith]

yup!

[jovial wraith]

if i'm understand this correctly, you want to take a photo of yourself and change your clothes, yes? because if that's the case, you would want to look into inpainting.

[fervent thunder]

hey

[potent spire]

what general settings do you use? Its really pissing me off lol I also try to use the base model of SD because i hate that all the base models on CivitAI are so unflexible

photography, photography, photography. Well i dont want just photography

lol

[jovial wraith]

there's plenty of cartoon/anime models on civitai, if that's what you're looking for

[potent spire]

i want catch em all model

which is why i downloaded the 1.0 SDXL base model from SD direct

its supposed to be a general model

[jovial wraith]

well, base models aren't really good imo. you're better off using multiple models that excel at different styles

[potent spire]

that sucks :/

also if i change the ratio the images generated are stretched

[jovial wraith]

that's strange. you're changing the resolution, right?

[potent spire]

yes

well im changing the width and height

[jovial wraith]

are you in text2img? do you have a controlnet module loaded?

[potent spire]

text2img

didnt touch any add on yet

[jovial wraith]

does it do that with all models or just the base model?

[potent spire]

i think only base model

but i think i give up

i hate SD lol

the only reason for me to touch it is the lack of censorship for some stuff

[true locust]

Small question, if I train a checkpoint in Kohya, and I want to train for example a video game style, do I need to put images for the "regularization images" that are related to the style I want?

[onyx tendon]

Hello

does anyone have any info about the api?

[wispy oasis]

Hi guys

I want to get stable diffusion. What GPU do i need minimum? Is there models that are better and not on the github page?

[true locust]

Nvdia is better, I think a rtx3060 12gb is the minimum, plus 32gb of ram (so you will need with this a i7-13th for this gpu)

[trim magnet]

minimun would be a 1060 6gb that would be enough to play with 1.5 checkpoints

[true locust]

Yeah but I think if he use hiresfix it will crash

[wispy oasis]

Are you serious? So if I buy a laptop that has a 4060 - i think the 4060 is 8gb. is it not good enough?

[trim magnet]

i used high res fix with a 1060 and never crashed as long as you stay at a max res of 1080p

[wispy oasis]

i havea 1050 mobile and i cant get stable to run

I need a laptop not a desktop. the best laptop i can afford is 4060

[trim magnet]

because mobile gpus are more crippled than desktop also it only has 4gb of vram

[true locust]

It's enough it's just my opinion, about the experience I had with a old computer, and yes 4060 is good enough

[wispy oasis]

wiat

[wispy oasis]

what is 1.5 checkpoints

you mean not the best model on the github?

i need the best model for production ready images