💬｜general-chat

[gilded lantern]

is matrix open source? or blockchain right

i dont have videos for that but have lots of links in my discord server and website https://autosynthetix.com/info but website is still a work in progress

[hallow sable]

coding is a huge pain, im at that stage where I can see what most of the code is doing but not good enough to write anything complicated

[gilded lantern]

cool ya that does look nice, looks like discord

[hallow sable]

matrix is good but no one is gonna use

all the gamers like the easy to use discorb

[gilded lantern]

these times are very interesting.. the AI coders will certianly get better rapidly but i still wouldnt undo any of the 30 yrs coding experience i have lol

[hallow sable]

I dont think AI will ever replace artist or coders

[gilded lantern]

ya

just only steps up the game

[hallow sable]

ai that replaceds lawyers would be pawgers

[gilded lantern]

lol

cool ya that's what i'm looking forward to also when AI can read full projects and unravel all my speggetti code! ha

oh wow cool

nice UI !

cool, wow i didnt even know you can run php desktop, i like php

cool ya that's exciting for me too finally fisnished old projects like that

finishing*

ya i have a 3060 (good price point) and a 3090 with 24gb.. cant afford the 4090

ya and python libraries have been a real pain getting all the different pipelines working together.. thats when i started coding it from scratch, all using the transformers library which is compatible with almost everything

yes with coding definitely

i only prompt it mostly 1 function at a time

definitely

im also making a task runner, that one involves a bit more

i think everyone would like to have an automated web bot post all our stuff but all the platforms dont allow a simple python web bot without lots of workarounds so not sure the best way companies will solve that

[fervent thunder]

:3

[gilded lantern]

dreamshaper

ya turbo still seemed to need like 10-15 steps and started to look ok

i run dreamshaper_8 at 20-25 steps usually

ya.. i'm sure it could be possible.. there's multiple factors i noticed.. if you only prompt 1 concept to the image it seems to 'require' less steps and works 'easier' on the model

similar to code AI

turbo could probably work for very simple image prompts and might be better for some realtime image animation/video project or something

cool might have to try that one, i dont think i tried the lightning

[pine fiber]

emad just said sd3 invites begin tomorrow or day after, and code is being sent to partners on friday

they will be launching with comfyui support with various controlnets "and other things" at launch

[vagrant fox]

not really, there is stuff called inflation, its not 2002 anymore, also noone forces you to look at top tier cards, if you want cheap hardware buy yourself a console or second hand previous gen card, with nvidia you are not paying for just hardware itself also but all technology on software side that's included

[autumn spire]

Anyone running sd on unraid?

[gilded lantern]

yes would really be nice to see AMD come back! always been for AMD

[vagrant fox]

and how much more tech and software upgrades you have since then? dlss, raytracing just to throw quick targets

nobody forces you to buy nvidia

well pay issue it's systemic issue and peoplke earn less and less for last 40 years or so, that's have nothing to do with private company trying to make profit

[crude bay]

I just saw the tweet, but I don’t know where you got tomorrow or the day after from. Got me real hyped! Wouldn’t surprise me if it takes a week still

[vagrant fox]

if they keept your oldschool prices there would be not much innovation

[gilded lantern]

they definitely do code libraries probably specifically to standarize nvidia only tho which isnt the issues iteself but then all AI used cuda which kinda sucks

[pine fiber]

in replies

[vagrant fox]

following that logic u never buy a card

if u stuck in minimum wage job then its not on nvidia, you looking for top tier tech being accessible to every pleb, it does not work that way

[gilded lantern]

lol

i think crypto started that though, from all the crazy demand

[vagrant fox]

but top tier cards are not made for all people tho

[pine fiber]

rent out ur gpu overnight and it pays for itself after like a year

[vagrant fox]

thats why u have 4060 if you on budget?

[gilded lantern]

software is still lacking.. im sure sooner or later there will be multi-gpu-stackable vram, models hopefully get smaller, etc

[cloud plume]

Idk I very much doubt that, also don't forget you also need to factor in electricity. Furthermore gpus go brrr... so good luck sleeping.

edits: better english

[vagrant fox]

also if you dont use gpu just for gaming its not hard to make your investement back, you only feel cost of gpu if you are stuck on generating useless waifus for personal use

[gilded lantern]

ya it is concerning there is a trend towards big models, but some demand for small models, hopefully it stays balanced

[gilded lantern]

i havent made a penny lol yet have all the AI features working. not sure what im doing wrong

[vagrant fox]

well you creating tool not providing paid service, i made cost of 3080 back (at scalpel prices) making ai music videos

[gilded lantern]

btc is going back up, porbably is worth it again

[vagrant fox]

not worth it unless u have free electricity and making 90 cent a day saves you

asic is more optimised for crypto alghoritms

[gilded lantern]

ya i agree , crypto didnt seem too great , that is unless you have cheap or solar

[vagrant fox]

still stressing out your gpu will pay off more when generating animations and selling them, before adobe stock banned me i used to make ez 50-80 usd a week from ai generated pictures for a while too

[gilded lantern]

music videos!? on youtube? i will hopefully have my tools production ready soon. i do have stripe and patreon setup

[gilded lantern]

nice

[vagrant fox]

but im not making money on views im making money on making animations for musicians, just get on fb groups, do small portfolio on google drive and get some clients

10 orders will pay off for your 4080/90

[gilded lantern]

cool, interesting.. so like freelance?

[vagrant fox]

yes

[gilded lantern]

i have seen soneone doing that on fb actually

[vagrant fox]

nothing comes to you, you have to go out your way and promote service you can provide, staying in one spot and complaining gets you nowhere

[gilded lantern]

ya

[vagrant fox]

be creative, there is still shit ton of people not knowing how to use those tools or not having hardware

[gilded lantern]

ya

lots of generic stuff i see .. definitely lots of room for more

[vagrant fox]

gpu pays for itself if you are willing to spam TF on groups and get clients, just make sure you always take 50% upfront and 50% halway through project

so u dont waste time like loser

[gilded lantern]

i'm apparently not 😀 but the tools im making are getting pretty cool

1 click text to slideshow makes the images into a slideshow with AI music, then 1 more click to upload to YT and they do get ~100view/day/video

[vagrant fox]

no need to be good at marketing just say shit like" Have you ever wanted have nice visuals for you music? Get some animation but it always feelt like out of your budged? I have today offer for you! Depending on animation loenght I am offering animations starting with 100usd" and you can go on and on, say what features you offer and link your portfolio

its not that hard, hardest part is consistent presence in high amount of musician groups

[gilded lantern]

oh man good advice, my first freelance the guy took all the advice i gave and ran lol. lesson learned

[vagrant fox]

how much it cost to register? in uk is only 12-18 gbp

[hard pawn]

very very shady

[gilded lantern]

cool cyber security will probably be even more of a thing in the near future

[vagrant fox]

llc seems like good starting point, i made ltd in uk

ltd is same as llc i believe

[gilded lantern]

all this AI redteaming is needed with all the companies, me included

[vagrant fox]

you have acess to biggest encyclopedy in world - ethernet

spend some time and do your research, nothing comes easy and by itself

[gilded lantern]

maybe soon AI will give instant mind training like the matrix though

joking..

[vagrant fox]

fuck it i wont become cyborg

and get chip in my brain

only slaves will do it

[gilded lantern]

same here i wont either

[vagrant fox]

concept is cool reality will be far worse than bad dreams

[gilded lantern]

we'd still be able to do practically the same thing i would think without having to go that far

[vagrant fox]

have you seen movie Gamer?

with people having chips in their head

[gilded lantern]

i would at least like something better than screens, 30yrs of LCDs is hard on the eyes

[gilded lantern]

havent seen it

[vagrant fox]

im looking for proper ar glasses that are small factor like normal glasses, something like apple vision pro but small

[vagrant fox]

recommending, its about enslaving society via chips

also have some elements of gaming

[gilded lantern]

glasses would be cool.. i hope theyll be more besides fb and apple though.. i like the alternative options

[vagrant fox]

poorest people are to rent like sims and people controll them via chips, convicts in jail play irl call of duty where winner gets freedom, and guy behing chip company wants to take over world

[gilded lantern]

i even tried ubuntu touch for phone os was really cool and almost doable but just didnt have certain important features

lol ya i do like scifi but so much scifi is dystopian.. sometimes can be ok if its a good story

[vagrant fox]

it is good story and its not far from close reality

[gilded lantern]

Dark Matter was my fav

tv show

damn, is it new?

[vagrant fox]

its made in 2009

[gilded lantern]

@vagrant fox you think open source will 'win' the 'ai race' ?

🤞

[vagrant fox]

even open ai shifted from being so open

but let's hope so

[gilded lantern]

lol ya

[fervent thunder]

do i need a very powerful graphics card to produce anything of any passable quality

[warm junco]

Nope, but it should have at least 4gb vram to be usable for image generation

[fervent thunder]

alright then thank you

[wet grotto]

hi guys

[dim minnow]

hi

Is stable diffusion web down?

[wet grotto]

I have a wron python installed, how can I delet than and install the correct?

[warm junco]

Uninstall the wrong one via system/apps or system control panel

[lapis urchin]

civitAI down?

[still glacier]

in maintenance

[still glacier]

which one ?

Cause the only "stablediffusionweb" website I know of, is not an official one but a very fishy one.

[dim minnow]

Oh that's not offical?

they had stuff linking here, mb

[still glacier]

that's why I'm asking which one

[dim minnow]

stablediffusionweb . com

[still glacier]

yup that's not an official one

[dim minnow]

oops

[waxen geyser]

is there a way to make adetailer less sensitive for small faces

and only focus on the main faces?

[lapis urchin]

yes

adetailer>detection

[waxen geyser]

i know kinda, but which value i have to change there ? 😄

[lapis urchin]

max area ratio probably

[lapis urchin]

actually it would be min

[coral lark]

When will SD3 be released? And any info on what it takes to finetune?

[gaunt pulsar]

Most likely

[vagrant fox]

i heard people that keep asking for release are being put at the end of waitlist for access to sd3 😄

[astral goblet]

https://www.reddit.com/r/StableDiffusion/comments/1b8hgbq/stability_ai_accused_by_midjourney_of_causing_a/ heh

[vagrant fox]

someone said about it while it was happening on office hours

why those MFs dont do voice recordings of this gathering like twitter spces do idk

yikes

scared of what they saying usually or what?

would be nice to have voice of this to use in reporting of that story

[stable geyser]

hello i am ai

[oblique willow]

We got like 9999999.. money of image Game, that all images feel worthless now which once amazed like 'wow its impossble for me to draw something like this" and now there's no value..

This make me sad and happy at same time.

[fervent thunder]

sounds fishy, why would stability scrape midjourney, their prompt adherence is terrible

way too kitsch images every damn prompt

[astral goblet]

the funny part is that mj, a saas, can't keep it up.

[fervent thunder]

besides they admitted it wasnt against their policy until they just changed it. so they made up new rules to justify bans

which is extremely unprofessional behavior

among other things like having 2 accounts able to turn off your app

[astral goblet]

effective immediately we're banning ALL stability accounts. "ohnooo"

[fervent thunder]

seems like they're looking for a scapegoat for their own fuckup

if a user can break your app, usually thats your own fault and banning a group of users by association is just a diversion. trying to hide just how bad the app is

and acting now like scraping is such a bad thing... their entire business model is built on scraping other peoples images

😂

same as everyone else, so going on a high horse is stupid

[astral goblet]

yeah its a circus

[frank kernel]

should i get stable diffusion frist before downloading comfy ui?

[shell tendon]

way better

[vagrant fox]

https://youtu.be/I1y1TP96N2g

my only concern is, if new models are trained from ai generated images would that mean we should prepare for regress in quality of images?

[slate dust]

The SD models go inside the Comfy models folder so install that first

[frank kernel]

is comfy recommended?

[slate dust]

It has a learning curve but is very powerful. Worth a shot and lots of people here can help you. or else try a simpler one later

[frank kernel]

Gotcha gotcha

[shell tendon]

Fully updated forge, fully updated a1111: over double the it/s with forge

[frank kernel]

what is 1111 in all of this is that just a model?

[slate dust]

Like Comfy, it's another UI to run SD models

[frank kernel]

ohhh

[shell tendon]

14 seconds vs 30 seconds in a1111

50 steps sdxl dpmpp_a

[slate dust]

Nice! I might try Forge in a bit. I'm super happy with Swarm atm

[shell tendon]

Both are great

I use forge cuz it's usable on my phone

[vagrant fox]

i need to get used to comfy this shit looks like one big spaghetti monster for me lol

[shell tendon]

Comfy on my desktop

Comfy rules

[frank kernel]

loool

[vagrant fox]

used to auto1111 too much

[frank kernel]

anyone know VAE?

im not too sure if thats importnat

[slate dust]

Ya I can't ever go back to the slowness of A1111

[vagrant fox]

you make long animations in comfy without any issues? stuff like 5-10k frames long

[slate dust]

I haven't done any animations

[slate dust]

I think you just need this one in the VAE folder https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

I'm not sure anymore, most models seem to have it baked in

[vagrant fox]

speed with single img / batch of 100 generation don't bother me, i would only swap if animating long stuff is noticable

and works flawlessly (just works i don;t expect animation to be perfect by visual standards, just keep generating without any issues)

[frank kernel]

gotcha thanks

[slate dust]

ya you definitely want to use what works best for that. All the animation stuff is wild. But i've always bene a still image guy haha

[vagrant fox]

i need to give comfy ui a go i had it once but tried for day or two then i went a bit out of loop away from ai having some personal stuff going on and i am back again

im fucking overloaded by all new shit

try to catch up its like 30 years of shit released in year

hahahaha

[slate dust]

haha ya you can't take any breaks on this stuff

[shell tendon]

Yeah no kidding

A week is a long time

A month is like years

[pearl ocean]

then what would be year be like?

[edgy scarab]

decades?

[shell tendon]

Sd15 feels ancient

It was the state of the art for stable diffusion less than a year ago

9 months

[pearl ocean]

makes GTA5 look young

[shell tendon]

Hell, controlnet is barely a year old

[edgy scarab]

I just got into this stuff, geez, sounds like things are advancing at pretty fast pace

[nova zodiac]

Yep, between this space and the large language model stuff

[pearl ocean]

I like the Euler sampling method, as it doesn't change the image every time 🙂

[edgy scarab]

Ohhh, euler doesn't change much? I have been euler a since people doing tutorials say it makes generation faster.

Huh, maybe I should try euler

[shell tendon]

Euler is deterministic

Others are too

Highly rec dpmpp Karras or exponential over Euler

You'll need half the steps and you'll get the same result

[pearl ocean]

is Karras faster?

[shell tendon]

Not really faster but different

It's more aggressive

Tends to create more change which usually means more interesting

Exponential I tend to bust out more when I'm inpainting and want to refine without restructuring the image

Sgm uniform is a lil like Karras but kinda in between those two

[pearl ocean]

the problem with Karras is, if you take the seed, and put it in the Hires Fix, it will be a complety different image

[shell tendon]

Yeah use exponential or dial back the denoise

Vague rule of thumb... 0.5 denoise with an ancestral sampler will act like 0.6 denoise for non ancestral

Same for Karras vs exponential

[pearl ocean]

I hear 0.3 is the best

[shell tendon]

Also if you're upscaling and dNt want the image fucked with I suggest dpmpp 3m sde with exponential 0.45-0.55 denoise

Min 30 steps

That's pretty light imo

[pearl ocean]

ohh, I normally does hires steps 20, and also sampling 20

[shell tendon]

I also suggest doing the upscale with something like ldsr or ldsr plus

So 4x then scale by 0.5 or even 0.375 to do 1.5 overall

That preserves a lot more detail than a latent upscale even with the vae encode/decode step

So you can get away with less denoise

Yeah 20 is way low imo

I never do under 25 anymoreo usually 30 is my floor

You have a 4080 so you won't need to resort to desperate measures lol

[pearl ocean]

How did you know!

[shell tendon]

Why wouldn't I lol

[pearl ocean]

1080 images generate in like 40secs which is nice.
2048 images take awhile

[shell tendon]

40 sec? What step count

Do you mean 1024x1024

[shell tendon]

Lol at unstable diffusion discord

I keep getting warned for shit that's fine here

Last one I don't even know what it was but apparently a character in one of my goofy images looked too young or something

Zero nudes or sexual content from me

[opal hedge]

SD3 waiting room

[topaz mirage]

Hi everyone! Noob here... My outpainting misteriously stopped overnight (from the logs it seems an hard crash). Anyway, is there a way to restore the work done?

[sudden ruin]

If it did not finish, probably not

[topaz mirage]

guessed so, thanks

[faint plume]

Hello, is the. bot up yet

[oblique willow]

#1047610792226340935

[faint plume]

Ok

When it is back up?

[oblique willow]

no idea.

[sour hollow]

hey. ive found a lora which used a model i like, but the model isn't mentioned in description. where can i send the images for people to tell what model looks similar?

[sudden ruin]

I found this kind of useful https://www.findsd.art/find-sd
No official website or anything tho, so take it with a grain of salt

[sour hollow]

thanks

[merry kestrel]

Guys what are you using for image caption/interrogating an image ?

[oblique willow]

Hi, If you are using Topaz video ai, Please answer a queation of mine. I do not have unlimited internet (country issue). And I see when I select a model and run for first time, It will download models from server. So there is no info on file size of those models. And there are 8 of those there. Assuming I want to use them all. Can you pls pls check your topaz file to answer what size they are. Someone told me they are 12 GB in total of all. And on forums i found link of all 66GB models. and 1274 models.... While the UI only shows 8 models. I searched on google a lot. Found nothing related.

[dry trellis]

Are they going to do the rating of the SD3 models here on the bots like with SDXL? Emad said invites going out today or tomorrow

[hidden hollow]

Hi everyone. Hope you are doing well.
Do you have mind to update your business website with latest technology and design or do you have any plan to make new business online websites or mobile Apps?
As a full stack developer, I am available to design and build the entire website and app that is related online business like AI, B2B, Web3, Blockchain from scratch.
If someone is interested, kindly DM

[ornate wagon]

bot doesn't exist anymore?

[trail lion]

it exists, it's just not accessible

[pine fiber]

you have to apply to the early access

[dry trellis]

I did but wasn't sure how it worked if you got access...like if it will be a different discord server or whatever

[pine fiber]

not sure

maybe this server XD but they said it was secret

[dry trellis]

Me too.

[crimson belfry]

I hope Joe Biden, the CIA or FBI or DoD are not influential in the decision to release this technology

Backroom meetings with government officials needs to be made public

Humans better see the wood through the trees 🌳

We do not scale humanity or decentralize AI by working in the dark

[full lark]

Nah its not illegal,also MJ didn´t ask permisions to scrape images from the internet

[tame thorn]

how to generate image on stable diffusion discord server ?

or bot is still down?

[sudden ruin]

Bot is still down

[pulsar lintel]

Theory I bet the bot will be back when SD3 comes out

[thorn ridge]

when will the image generation bots here be back?

[dry trellis]

Me too.

[fervent thunder]

i wish the smaller sd3 models will allow us to generate quick images to get basic structure and composition then use the 8b for a final render

[vagrant fox]

K, woken up, 9k frames generated, mid way through upscaling, let's soon see what random prompt for animation created in gpt4 made while i was asleep

[vagrant fox]

well smallest model will be 800m

[shell tendon]

here's something interesting i just noticed...

[fervent thunder]

ik but i don’t know if it will be possible to recreate a prompt from one model to the larger 8b model with keeping the image 1:1 and adding more details

[vagrant fox]

not everyone has 4090 and 8b model works so far only on 24gb vram on 4090 as stated in papers

it need optimalisation

[fervent thunder]

smaller model go brrr

[dry trellis]

Emad stated he thinks SD 3 and it's fine tunings will be last models because they will be perfect. If that's true I'd definitely invest in whatever GPU it takes to run it properly lol

[fervent thunder]

Will be interesting to see how cherry picked the results are and if it's worth the time trade off vs SDXL or 1.5. Hopefully we'd get an SD3 Lightning soon after. I don't follow twitter anymore, are they still posting examples?

[trim nymph]

hopefully! finetuning would be so much fun 🙂

[charred mesa]

without T5 and probably at fp16 + xformers and other optimizations it will hopefully fit on 8-12GB of vram 🙏

if the 2B model is fine then people might just finetune that and it might become the most popular model

[dry trellis]

I havnt seen any new ones in a week or so.

[dry trellis]

Definitely

[pulsar lintel]

Ull just need the latest GPU any gaming pc is fine

[pulsar lintel]

It’s a very strong theory esspecially sense they wanna compete with midjourney

[dry trellis]

I have a good gpu that runs everything so far really well...hopefully it will do SD3

[bleak drift]

How does this sdxl prompt look?

(masterpiece, ultra detailed, 4k, photo-realistic, highres:1.5), A dark seductress succubus (long and curly black hair:1.3) (Cat-like eyes glowing red:1.2), full body image, she is standing in the middle of a bloody battlefield, dark clouds adorn the sky behind her as her black dress blows in the wind, (dark, gothic, menacing atmosphere:1.3), cinematic lighting, ashes in the air

Any tips for improving it?

Or organizing it better?

[jovial wraith]

yo guys, you may want to check this out: https://www.youtube.com/watch?v=nVSpIhNY2IQ

MattVidPro is giving away a 4080 Super

[dry trellis]

Are you using negative prompt too?

[astral goblet]

anyone getting invites yet? https://twitter.com/EMostaque/status/1765498520235131149

[jovial wraith]

i haven't

[astral goblet]

youtube giveaways... meh. no regulations on those contests. they're all rigged. Just view count hypes.

the guy who wins it is often just a friend of the fam

[jovial wraith]

i mean, you miss all the shots you don't take

not like i have anything to lose

[astral goblet]

its just click bait marketing

FOMO

[jovial wraith]

again, not like i have anything to lose

[astral goblet]

go for it . it's not a real contest though. it's just a sham for view counts.

the practice is 100% normalized for youtube since nobody regulates those contests

[elfin condor]

Does anyone have experience using Flowframes? When I try to interpolate my photos to make a video, not only does it play the image out of order, but some of them don't smoothly change to the next from and it's cutting out half of my photos and not including them at all...

[astral goblet]

i rmeember one contest, the guy was giving away a pc they built ons tream. but the girl who got drawn live wasn't a name he liked so he tossed it and drew another, which was someone he knew

conveniently

[jovial wraith]

oh damn

[astral goblet]

https://youtu.be/vnrZq3nKcaw this one was a black swan event but it shows just how normalized the practice is

he was emboldened not because he was pioneering new frontiers. that's just how it low key works all the time

[static cape]

He just loves to exaggerate and generate Hype. The "let's just use even more Clip Encoders" doesn't feel like anywhere near a "perfect" solution to me. And what we have seen so far is far from perfect as well... but I also hope it will be an excellent base for Fine-Tuning.

[astral goblet]

t5 isn't a clip encoder

base models shouldn't be "perfect" because what's perfect to you, is not for other purposes.

photographic sdxl models and sd15 models are over fit and have destroyed a ton of the latent space.

open weight models are different from SAAS models. The base should be a base.

[static cape]

True. It's still a strange exaggeration again.

That being said - is there a way to keep up with his tweets without registering a Twitter account? I don't want to ever go back there...

Oh and Stable-Audio isn't Open Source as of it, right? I thought I had only seen it as a web-service so far.

[nova zodiac]

web service only - audiocraft is the open source version of it from Meta

[nova zodiac]

nitter?

rembg - or you can use LayerDiffusion to create an image with a transparent background

[astral goblet]

layer diffusion sucks for faces, especially trained faces. background removal is best

lots of other purposes for layer diffusion though

LOLOLOLOLOLOL

[nova zodiac]

https://huggingface.co/spaces/KenjieDec/RemBG

Meet ya halfway 😛

[astral goblet]

https://www.remove.bg/

[static cape]

Ah, I wanted to take a look at that one for a while now. Do you know if there is a GUI for it by now?

[nova zodiac]

https://github.com/gitmylo/audio-webui

[static cape]

I didn't even know that exists.

[nova zodiac]

I'm on a roll this morning, AMA lol

[astral goblet]

Are you the gate keeper?

https://youtu.be/lVAiBP8GeSc

[nova zodiac]

Goal keeper yes.. gate keeper no

[static cape]

Amazing, thanks. 🙂

[nova zodiac]

And no, I've never actually watched Ghostbusters

[astral goblet]

😮

👻

but.. bustin makes you feel good

[nova zodiac]

I mean I do know who to call if there is something strange in the neighbourhood...

[astral goblet]

imo, they still hold up. Jim hensons did the SFX. Pros of the day

[static cape]

Today has been so great so far... now if we could just start with the SD3 testing... hoping

[astral goblet]

to your credit, you did recognize the key master line. exceptional referencing skills. haven't even seen it

[bleak drift]

I haven't used that yet

[astral goblet]

less emphasis. 1.5 is gonna blow out your latent space in many cases. there shouldn't be a need for it and the tokens you're cranking up don't really do too much. what lead you to these values?

i like to understand somene's reasoning for crafting the prompt and then we improve the reasoning

[nova zodiac]

I've seen some models handle :4 weights, but that is sdxl and rare and depends on the token and the model

[static cape]

Why would you even try that? I feel bad using :1.5 in very rare cases... I usually feel like :1.2 is enough to get my point across (in SDXL).

[nova zodiac]

I wanted to try force a cartoon model to make a colouring book line art 😛

It did actually do it at 4

[static cape]

That sounds pretty cute actually.

Care to show the result?

[nova zodiac]

That was a while ago so havent got it.. but the model was fustercluck if you want to try it yourself

[brittle maple]

neither sd nor mj could draw a character from the game

[nova zodiac]

which game?

[brittle maple]

residentevil)

[nova zodiac]

you may need to find a LoRA for the character on Civitai.com to make SD be able to draw it

[brittle maple]

i found but it bad

[static cape]

Funny name.

[nova zodiac]

Thanks, I try make the names of my models interesting 😛

[lost harness]

Hey everyone! Has anyone received access to Stable Diffusion 3 yet? Id love to see some images with prompts if you could point me in the right direction. 🙂

[brittle maple]

i want sora

[static cape]

Haha, got me good there. 😄
I hope you had some of that gingerbread recently?

What are you using to Fine-tune XL? I personally found it to be a pain to fine-tune locally... With 1-5 it was as easy as making a LoRA now.

(curse 4 times the resolution 😛 )

[nova zodiac]

Had some earlier this week, it's fantastic 🙂

[nova zodiac]

I make LoRAs using the trainer on Civitai, then use supermerger in a1111 to merge models and loras til I get a mix I like

[nova zodiac]

I do and I don't, the world is not ready for a society where you can't trust what you are seeing

[brittle maple]

need ai generation detector

special digital stamp

[nova zodiac]

ai forensic expert going to be a very important role

one thing to note is that some of these models are built using methods where they are trying to beat a detector (GAN)

[static cape]

Interesting. So merging LoRA's into Models really makes them perform better? I tried that once or twice, but never got results that made me consider it further.

[static cape]

But we are already in a society where we cannot trust what we are seeing... even though the part about most people not being ready is true.

[crude bay]

I thought about that and there should be some kind of digital markings inside the pics and metadata for every digital camera

[static cape]

I wonder how many people didn't realize there is a new superior 3d Model because of the SD3 announcement on the same day.

[static cape]

But just like Metadata of Cameras you can scrub those. As long as there is reverse engineering, nothing like that will hold people with ill intent back forever.

[crude bay]

Im no expert but silicon has flaws so maybe every sensor has already some kind of unique identifier? On the other hand we might only trust pics without scribed meta data in the future

[brittle maple]

I think we need some method of including this in generation

but you can delete this like in crack game

if you have skill rev eng

[astral goblet]

since the beginning people have been encoding meta data into their generations, a1111 or comfy or whatever. generation data in the file has been here. But people have also been mad about the watermark code that stability puts in their implementation everytime a model releases. As if they forget they're putting meta data in anyways

[crude bay]

It’s already there but platforms need to be forced to display those and at least declare ai content as such

[astral goblet]

"forced" meh. nah.

people will use the platform that provides the best services. no reason to force anyone to make a website a certain way. that's how we end up with stupid eu regulations on every single website, as if asking permissions for cookies matters at all

it just uglied up the internet

regulations should prevent abuse, not dumb shit like telling them how to design their interfaces

[crude bay]

Just inform people ai is ai. Protect the consumer

[astral goblet]

Often, "protecting the consumer" is just a guise for "protecting corporate interests"

regulations prevent new services from competing

regulations should be there but in the right way. enforcement at the top. not at the end user

[crude bay]

Huh? I can see the need to protect the consumer. How would that protect corporations?

[astral goblet]

imagine if AOL was succesful in convincing the US government to regulate media on the internet? They were trying to get netscape shut down because it was dangerous to provide such wide access to audiences

This is the precise reason that Sir Tim Berners-Lee was knighted

[crude bay]

We are talking about some specific thing relating to misinformation and declaring something that is a fake photography. Yes the government shouldn’t overstep boundaries but that isnt a problem to my idea but a problem of specific governments

[astral goblet]

oh ookya. the laws will only regulate the bad guys not the good guys okay. that makes sense

corporations wont use those to plow over competition at all

[crude bay]

The good guys would care at all. I post so on insta and always mention ai generated on my posts because it matters

Wouldn’t *

[astral goblet]

Free market. Let the end user figure it out. It'll be fine. Old laws still apply

All regulations should be doing is forcing services to release their weights publically

[crude bay]

Yeah you don’t make sense or counter my points at all

[astral goblet]

your points are ideal thinking. "good guys woudn't care!" sure.

[crude bay]

Then make me understand why disclosing you used ai harms you

[astral goblet]

real world historical evidence shows regulations get demented and perverted in most cases.

A good one i like is "No insider trading" i bet there was a lot of hub bub about that too. Something as drastic like that is needed today. "No private model weights"

if your model operates publically, weights are public

[static cape]

"Oh, honey. Don't be so indecent! Has your Mom not taught you not to share your weights?"

[crude bay]

And? There is also a law of not murdering people. Does that also only harm the good guys?

[astral goblet]

okay so you want ALL material an end user publishes to require a document of source, because only bad people murder?

Strong points. Good case

if you're going to be absurd of course i'm not going to counter your points directly

[crude bay]

Im being intentionally absurd, because you can’t stay on point

[astral goblet]

oh sorry for not operating in your scope of assumptions. mbad. I shouldn't have gone and applied all previous experience on the matter to this topic of regulation

[rich kestrel]

or.... just hear me out. ppl who spread fakes as real pics get punished. the other 99% of the population doesnt

its really not that complicated. so ppl will deepfake videos. so what. don't ban the tools, punish the bad guys instead

[crude bay]

That’s what I mean basically. If you make life like pictures you should declare that’s ai

[astral goblet]

Guys. Some trolls will call 911 and swat streamers. I think we need govenrment to regulate all phone companies and screen all calls to make sure that 911 isn't being dialed as a prank

its the only way to be sure

[rich kestrel]

911 calls are already screened....

[crude bay]

Haha

[astral goblet]

its not enough! it must be prevented

[nova zodiac]

you should, but say you just dumped me in an ugly relationship.. now I don't care about the consequences of making deepfakes of your images and plastering them irrecoverably across the internet

havent got an alibi, corrupt policeman deepfakes your image onto cctv footage putting you at the scene of the crime....

[loud needle]

How long left till the bots come back

[astral goblet]

i dont think a teddy bear and an alien would be an ugly relationship. yoiu guys should work it out.

[nova zodiac]

How long is a piece of string

[crude bay]

And that’s what regulations and laws are exactly there for. You get punished by doing that

[astral goblet]

maybe we need government regulation to come in and prevent people from breaking up beforesomeone gets hurt

[loud needle]

Idk what string

[nova zodiac]

but the reputational damage is already done...

I'm not saying ban everyone from the tech, as that ship sailed long ago...

[loud needle]

Bluds yapping

[astral goblet]

old laws all still apply. we don't need new laws for ai. libel, slander, hate speech, all of it.

[nova zodiac]

The point is that we don't know....

[crude bay]

Yeah but that is not a counter to my argument? Should those people not be punished after the fact?

[astral goblet]

no way i know the guy an he's not that bad. he's a chinese man an makes great noodles

[nova zodiac]

oh agreed.. but if the evidence led to me getting the death penalty, the cop getting punished later is a fuckload of good to me

[nova zodiac]

I wasn't going to make the racist joke, but was thinking it 😛

[crude bay]

That’s why I’m against capital punishment and lucky to live where that’s not being used

[nova zodiac]

agreed, but this is the world we live in

[astral goblet]

i dont know if just mentioning races makes something racist. plenty of those priest pastor an rabbi jokes that AREN't racist (tons are though)

maybe the govnerment should just make it all illegal to be sure

[nova zodiac]

they should also take all the cars off the road as that kills a few thousand people every year too.. 😛

it's coming whether we like it or not

the ai, not the removal of it

[astral goblet]

government regulations so strong that even thinking of driving a car is punishable. its the only way to prevent deaths

[nova zodiac]

technically the only way to prevent deaths is to prevent births

[astral goblet]

naw the universe still dies. we need more regulation

[nova zodiac]

#banthebigbang!

Douglas Adams was so right with "In the beginning the Universe was created.
This has made a lot of people very angry and been widely regarded as a bad move"

[astral goblet]

He's my personal Jesus

preach

Adopting leaves as a currency, realize there's kabillions of leaves, so you burn down the forests

ez money

[nova zodiac]

I also like his theorem that "if someone discovers the meaning of the universe, then the universe will instantly implode and be replaced with something more absurd"

[astral goblet]

some theorize that this has already happened

[nova zodiac]

This must be a Thursday. I never could get the hang of Thursdays

(well it's Friday in my timezone lol...)

[slate dust]

All you figments of my imagination are pretty entertaining

[nova zodiac]

Ford! There’s an infinite number of monkeys outside who want to talk to us about this script for Hamlet they’ve worked out!

[astral goblet]

love the h2g2 vibes

shoutout to dirk gently. i actually loved the netflix adaptation but oh well.

[nova zodiac]

can't say I saw that one

[sullen sleet]

is the bot channels permanently down?

[nova zodiac]

"Listen, three eyes," he said, "don't you try to outweird me, I get stranger things than you free with my breakfast cereal."

[nova zodiac]

#1047610792226340935 - the theory is that the gpus powering the bot are currently doing the final finetunes of SD3, and once that is done they will come back online. No-one has disproven this theory, just as noone has disproven the theory that I am a sentient robot, but it's what I'm going with

and given Emad said access will be going out to the waitlisters in the next few days, full public release won't be far away

[nova zodiac]

We are all just part of the rats' simulation anyways...

[slate dust]

Despite all my rage, Im still just a rat in a cage

[nova zodiac]

between that line and the song title, there are some great prompt ideas there!

For instance, on the planet Earth, man had always assumed that he was more intelligent than dolphins because he had achieved so much—the wheel, New York, wars and so on—whilst all the dolphins had ever done was muck about in the water having a good time. But conversely, the dolphins had always believed that they were far more intelligent than man—for precisely the same reasons.

It is known that there are an infinite number of worlds, simply because there is an infinite amount of space for them to be in. However, not every one of them is inhabited. Therefore, there must be a finite number of inhabited worlds. Any finite number divided by infinity is as near to nothing as makes no odds, so the average population of all the planets in the Universe can be said to be zero. From this it follows that the population of the whole Universe is also zero, and that any people you may meet from time to time are merely the products of a deranged imagination.

[modern halo]

anyone know the best free sd browser bot?

[sullen sleet]

If you are a sentient robot, you would have traveled through wires to SD3 testing server, and found out what the server is cooking. You would not be pasting a theory baked by other mortal creatures. Does it automatically disprove the theory that you are a sentient robot?

[nova zodiac]

best is relative.. but do start with the one on civitai.com or the one on happyaccidents.ai and work from there

[nova zodiac]

Who said the theory was baked by a mortal...

Who said I didn't go through the wires and am just refusing to share due to my stupid NDA chip actually working

It’s a well known economic phenomenon but tragic to see it in operation, for the more shoe shops there were, the more shoes they had to make and the worse and more unwearable they became. And the worse they were to wear, the more people had to buy to keep themselves shod, and the more the shops proliferated until the whole economy of the place passed what I believe is the termed the Shoe Event Horizon, and it became no longer economically possible to build anything other than shoe shops.

I feel like this is coming but with AI

[opal hedge]

Do we have any confirmed vram/compute reqs for the highest param SD3 model

[astral goblet]

4090 will run the 8b model and thats without optimization

if you run it in fp8 without the t5 encoder, should be pretty versatile

[nova zodiac]

yeah.. 24gb for fp32 should make it possible in fp16 on a 3060 12gb

fp8 and xformers and I'd be surprised if it couldnt fit on a 8gb card

[cedar salmon]

2.6B

[nova zodiac]

the base bit is 3.6B and the refiner is another 2.6B to give roughly 6.5B - it only runs on 8gb as you don't run the refiner at the same time

[astral goblet]

yeh i was wrong

[nova zodiac]

dw I did the same thing a couple of weeks back too

the 24gb thing in fp32 - 19gb of that was t5, so it could well be that it'll run fine on 6gb

assuming you can circumvent the t5 encoder somehow

or run that on cpu

[astral goblet]

that i didn't know. t5 is huge eh? i remember i ran a model with it on my pc. pixart?

must've been autocasted t5 caue i have 16gb

[nova zodiac]

t5 is massive (11B params for the big one IIRC)

[pearl ocean]

How Stable can the Diffusion possibly get?

[high portal]

How to keep my characters consistent when using SD draw the pictures?

[dry trellis]

A lot of people use various controlnet options

[sour needle]

Can we use stable diffusion 3 locally on mobile?

[vagrant fox]

currently you can't use it at all

[sour needle]

Oh

[astral goblet]

apple phones can run 1.5 i think

[sour needle]

Too bad, android user is here

[sonic anvil]

how to run on apple phones?

[astral goblet]

uses the m1 chip. i dont know how. there must be tons of apps by now

wait, m1 is the macbooks. i dont know. it does stuff on the chip. uses core ml

https://github.com/apple/ml-stable-diffusion i dont have apple anything. just trust these guys

[shell tendon]

We need a CheckpointsAnonymous channel

Got another 57gb downloading... Just hopeless lol

[karmic cedar]

If you’re going to locally diffuse on fruit hardware you’ll want Draw Things. Otherwise, go have some fun in the cloud perhaps?

[vale steeple]

Actually

[nova zodiac]

I think draw things would work theoretically on the new ipads, but not on iphone

[vale steeple]

CES 2023 already have some couple mobile phone running lightweight version of Stable Diffusion

using I think Snapdragon 8 Gen 3?

[pearl ocean]

Running on Apple Watch when

[vale steeple]

Alright, it is Snapdragon Summit, not CES.

I think it could be doable for phone as newer as Snapdragon 8 Gen 2 ( who had a NPU upgrade from Gen 1+ )

Mediatek Dimensity 9300 could run SDXL Turbo

[halcyon crescent]

Sorry newbie question here
Is stable diffusion free to use?

[lone hawk]

why tho? I stopped using sdxl, cascade and all that crap, this is so inferior now Lol

[nova zodiac]

the model itself is free to use.. you either have to run it yourself, or pay for someones gpu time to run it for you

[shell tendon]

Inferior to what?

[pearl ocean]

Just use a 4080 🙂

[shell tendon]

Feel bad for ya if you gave up lol

[opal hedge]

Dunno mate, I've been generating some quality images with sdxl

The kind people would have paid 300 dollars for not even a year back lol

[shell tendon]

Yeah I'm churning out incredible shit

[inner canyon]

Hello, where can I found the expected sate of the bot return ?

[shell tendon]

Wow the unstable diffusion discord kinda sucks

I posted one of those crazy images I've posted today on the sdxl showcase and I guess there was a tiny character somewhere that maybe didn't look adult or maybe did? I don't even know

Anyway, suspended lmao

Think I'm done with that server

[trim magnet]

prob one of those twitter/reddit mods who care more about a drawin than whats happening at the world right now

[shell tendon]

Three times now

First was because one of my shark clowns had school stuff in the background

No kids, just a weird ass shark clown character in what looked like a classroom

Just too jumpy

[pulsar lintel]

It runs on the new iPad but make sure u have the cloud storage… it takes up a lot of ram space when downloading and using civati

[shell tendon]

I don't spend all my time when generating images obsessing over whether there's a kid somewhere in the background or the body proportions of an anime character and what they could suggest about age

[trim magnet]

imagine those mfs going to the Sistine Chapel they would probably get an aneurysm and try to spray paint on all of the naked angels painted there

[shell tendon]

Lol yeah

Posted the offending piece of the image in general images

Apparently that's a kid? Lol

That cutout occupied about 0.3% of the image

[errant pulsar]

Ботовете все още ли ги няма?

[bleak matrix]

Good morning, everyone!

How are we all this gorgeous day,

[uncut haven]

I feel a passion for ai bubbling up within me

yes, I'm a total newbie

[bleak matrix]

I can approve this message!

[cloud plume]

Btw does reacting with ⚠️ to messages still report them? (did that with above link/message)

edit: it has since been deleted :)

[sudden ruin]

It should do that yes

[loud quest]

Newbie question: on Automatic1111, when I generate an image, there's a starry button to apply hires fix to that image

How do I go about doing that manually? Let's say if I want to do that to an image that's not on the last batch I made

I tried looking at the img2img menu but it seems to have different parameters entirely

[zenith lance]

Hello everyone!

I'm new to
StableDiffusion and I need some help importing a character and converting it to an anime version (comic book, manga).

Where can I make my request?

[bleak matrix]

You can learn more about SD in #1080946152318443610 There's also #🍥｜anime

[zenith lance]

ty 👍

[fervent thunder]

Hello

I have an question

Whats the difference between: https://github.com/AUTOMATIC1111/stable-diffusion-webui and https://stability.ai/

Stability.ai was the first thing i seen when searching it (Stable Diffusion

[karmic cedar]

Welcome!

[sudden ruin]

Stability provides the base models, with the ui you can use them

[fervent thunder]

Aint there loads of Free models tho?

[sudden ruin]

Yeah those are all custom trained based on the official release

[fervent thunder]

So whats so special about the stability.ai ones?

[sudden ruin]

Without them there wouldn't be any custom models

[fervent thunder]

I dont really understand now

Is that stability.ai something i need to create things?

[sudden ruin]

You don't "need" stability.ai, you can download any model from civitai for example and use that model in the ui of your choice

But stability releases the base models for people to work with to release their custom tunings of that model

[fervent thunder]

Whats your favourite UI

[sudden ruin]

I'm using forge and comfyui

[fervent thunder]

Ohhh, i understand. Was Just kind of odd that that was the first thing that shows up

[sudden ruin]

Forge is a fork of automatic1111

[fervent thunder]

Why both and not Just one?

[sudden ruin]

They serve different purposes and work differently as well. Comfy is node based

[fervent thunder]

Just installed ComfyUi but this is confusing

I wanna be able to use the upscaler etc

Whats a good upscaler btw?

I didnt do this for a long time

[sudden ruin]

Yeah if youre not familiar with nodes it can be very difficult

Depends on what you need

SUPIR is the newest tech

For upscaling

[fervent thunder]

Should i try installer Forge then?

Need something with an UI then probally

[fervent thunder]

Oh sick, il try that out

[sudden ruin]

Yeah forge is prob your way to go for an easier interface

SUPIR for example only works in comfy right now

That's why I use both

[fervent thunder]

Is there some kind of run down on how to work Comfy

Or do u have an upscaler thats good for Forge

[sudden ruin]

Ultimate SD upscaler does well with tiled diffusion

But tiled diffusion doesn't work with forge yet, only the og automatic1111

[fervent thunder]

Oh yeah i seen u can create videos these days

How does that work

And can my 2080 even handle that

LMAO

[sudden ruin]

Animatediff, stable Video or external tools like pikalabs

[trim magnet]

the vid2vid extension doesnt work on forge

[sudden ruin]

I like comfy more for Video stuff

[fervent thunder]

Tbh im too new to know whats going on

[trim magnet]

yea animatediff works better on comfy

[sudden ruin]

Way more customization options

[fervent thunder]

Im in comfy rn and i see Just nothing basicly

[trim magnet]

on auto1111 and forge u can only do limited text2video

[sudden ruin]

Double click to create a node

[fervent thunder]

Just letters and numbers

[sudden ruin]

Watch some tutorials tbh

[fervent thunder]

yeah i probally wil

Forge isnt in the #1080946152318443610 ?

[sudden ruin]

It's rather new and these channels aren't kept up to date I feel

Also as I said it's a fork of automatic1111

So it's very similar

[fervent thunder]

Oh i see

I think i found the github

Bro theres like a million models

I havent done this in months lmao

[fervent thunder]

Do u have an favourite model for just general ai content?

[sudden ruin]

Juggernaut is really nice

Or dreamshaper

[fervent thunder]

Yeah it was the first one i seen on the civitai website

Figured it would be good

LMAO

Im installing forge rn

I know a little bit of how to work this still

The last time i used this was a year ago wtf

Time goes by quick man

[sudden ruin]

Yeah and time moves fast with AI especially

[fervent thunder]

Yeahhh man stuff goes quickly

I seen ai videos on reddit

They look insane bro

[last cape]

👋

[dense wave]

hi there!

I wanted to know who do i speak to for a potential partnership? Thank you!

[pastel turtle]

So, T5 encoder will not be optional. Too bad for my 20GB VRAM GPU.

[sudden ruin]

Partnership with who? If you mean Stability.ai please contact their support

[rain osprey]

anyone know a way to use/run Relight locally?

[still glacier]

relight is a tool specific to clipdrop, you can emulate it using controlnet depth map + some light / god ray / lensflare image + img2img

[stone stag]

Yoo when are the bots returning

[rain prairie]

Presumably the GPUs are being used for SD3 testing.

[pine fiber]

it is optional

[charred mesa]

^

absolutely it is

[karmic cedar]

Is today SD3 Day? 👀

[charred mesa]

and it won't decrease visual quality, only prompt adherence a little

honestly, its mostly the TEXT capabilities that get hurt

which people aren't as excited about comparted to PROMPT ADHERENCE

which is something I have personally been anticipating all this time

text is easier to figure out anyway

[placid wasp]

is there a place i can compare the itterations per second with hardware that generated it? I have a local ionstall of SD on my windows 11 box, Nvidia 4090 and a i9-14900 with 192GB RAM. I am getting an it/s of 18.x on single image generation, and an average of about 13 when generating 3 images in a workflow. just be good to compare to see where I am. I too m new to this

[charred mesa]

apparently? I bet I hope they will get people from the waitlist into that server where they host some SD3 bots or something

they promised it on feb 29th but it got delayed

now emad claims that their base model is a better candidate

[karmic cedar]

I wonder if they’ll go in a similar direction as SVD and work up an online platform of sorts…

[boreal mauve]

does SUPIR work on mac? I saw a video last week, that got me interested. You know where I can find instructions?

[severe zealot]

i just openned and updated Stable diffusion automatic 1111 and it made a lot of changes, now its saying this error when i try to do a basic generation "NotImplementedError: No operator found for memory_efficient_attention_forward with inputs: query : shape=(2, 4096, 8, 40) (torch.float16) key : shape=(2, 4096, 8, 40) (torch.float16) value : shape=(2, 4096, 8, 40) (torch.float16) attn_bias : <class 'NoneType'> p : 0.0 cutlassF is not supported because: xFormers wasn't build with CUDA support flshattF is not supported because: xFormers wasn't build with CUDA support tritonflashattF is not supported because: xFormers wasn't build with CUDA support triton is not available requires A100 GPU smallkF is not supported because: xFormers wasn't build with CUDA support dtype=torch.float16 (supported: {torch.float32}) max(query.shape[-1] != value.shape[-1]) > 32 unsupported embed per head: 40"
[11:35 PM]
the update seemed to install and uninstall torchvision but im not good with this stuff

[sudden ruin]

No clue tbh, not a mac aficionado

[sudden ruin]

#🤝｜tech-support

[severe zealot]

its on windows 10

[boreal mauve]

no, I meant instructions for SUPIR in general how to get it up in comfy?

[boreal mauve]

website to track?

[severe zealot]

oh sorry that wasnt for me

[sudden ruin]

Oh my bad, check the github https://github.com/kijai/ComfyUI-SUPIR

[boreal mauve]

I just posted a related question in comfy before about upscaling, hires fixing and tiling

[boreal mauve]

dang ... xformers ... So I have to wait until someone adapts

[sudden ruin]

xformers is automatically detected and enabled if found, but it's not necessary, in some cases it can be a bit faster though
From the Git

Should work without

[boreal mauve]

oh, okay. I'll just give it a shot. Thanks!

[charred mesa]

The subreddit will explode after the SD3 server starts getting invites, but I'm worried if it will get a mixed reception or mostly praise

[sudden ruin]

As long as people give valuable feedback its fine I think

[charred mesa]

I'd be right happy to

[pulsar lintel]

consider the following steps:

•    Ensure CUDA Support: Make sure you have CUDA installed and configured properly on your system. CUDA is essential for running deep learning models on NVIDIA GPUs efficiently. You might need to reinstall PyTorch with CUDA support explicitly.
•    xFormers and Triton Installation: Check if xFormers and Triton (if needed for your use case) are installed correctly and built with CUDA support. This might involve compiling these libraries from source with CUDA flags enabled, depending on your setup and the instructions provided by these libraries.
•    Adjust dtype and Model Configuration: If your GPU does not support float16 operations efficiently, consider using float32 for your model operations. Additionally, check if there’s a configuration option or model variant designed to work without requiring float16 or specific hardware features like those available on A100 GPUs.
•    Dependencies and Environment: Reinstall torchvision if it’s required for your project. Ensure all other dependencies are compatible with each other and with the version of Stable Diffusion you’re using. Using virtual environments (like conda or venv) can help manage dependencies more effectively.

Given the complexity of configuring deep learning models and their dependencies, especially with GPU acceleration, it might be helpful to consult the documentation of Stable Diffusion, xFormers, and other related libraries or seek support from their communities for specific installation and configuration instructions.

That’s what chat gpt said

[severe zealot]

oh gosh its been so long since i set up sdauto1111 and did all his i woouldnt know where to start

[pulsar lintel]

1. Update CUDA and NVIDIA Drivers

   • Ensure your CUDA toolkit and NVIDIA drivers are up to date. The latest versions can often resolve compatibility issues. You can download them from the official NVIDIA website.

2. Reinstall PyTorch with CUDA Support

   • It’s crucial that PyTorch is installed with the version of CUDA that matches your setup. You can reinstall PyTorch by following the commands on the PyTorch website, ensuring you select the correct CUDA version during the installation process.

pip uninstall torch torchvision torchaudio

Visit https://pytorch.org/get-started/locally/ to get the correct installation command for your setup

3. Check xFormers and Triton Installation

   • You might need to reinstall or compile xFormers and Triton with CUDA support if they are necessary for your specific use of SDAuto1111. This step can be more involved and might require following the specific installation guides from their respective GitHub repositories or documentation.

4. Switch to float32 if Necessary

   • If float16 is causing issues and your GPU does not efficiently support it, consider modifying the configuration to use float32. This change may impact performance and memory usage but can improve stability.

5. Use Pre-configured Environments or Docker Containers

   • If available, using a pre-configured environment or a Docker container can significantly simplify the setup process. These environments can package all the necessary dependencies, libraries, and configurations in a way that is easier to deploy and use.

6. Consult Community Resources

   • The community around SDAuto1111 and similar projects is often very active and helpful. Forums like Reddit, GitHub issues, and specific community discords can be invaluable resources for troubleshooting advice and updated installation guides.

7. Start Fresh

   • If you’re still stuck, sometimes starting from a fresh environment can be easier than trying to fix a complex issue. Uninstalling and then reinstalling your libraries and dependencies from scratch, while time-consuming, allows you to follow updated guides and documentation that may have changed since your last setup.

Remember, taking it one step at a time and leveraging community resources and documentation can help navigate through these complex setups.

[fervent thunder]

What is a "Sampling method" @sudden ruin and why do u use Forge over the other one

[pulsar lintel]

I put ur response into chat gpt that’s what chat gpt said back

[fervent thunder]

The stable difussion UI

[sudden ruin]

Forge tends to be faster these days, especially on lower end hardware

[fervent thunder]

Do we still need to use the "best quality" prompts etc?

Or is that all automated these days

[sudden ruin]

Samplers are a little difficult to explain, for further information I suggest reading this https://stable-diffusion-art.com/samplers/

[sudden ruin]

Prompts are model dependent

[karmic cedar]

I think we’ll likely see a lot of success right out of the gate. Folks will find the vanilla in it right away, but that’s where time, community and fine-tuning come in.

[fervent thunder]

Ohhh im using Juggernaut XL rn

[sudden ruin]

Check what people use on civitai to get a feeling

XL tends to work without those weird best quality, 4k, etc. prompts imo

[fervent thunder]

Ohh okay

SD VAE

Whats that btw

[sudden ruin]

VAE is the image encoder/decoder between the latent and the image space

[charred mesa]

for a base model it already looks somewhat aesthetic. My only worry is that the bokeh might be unremovable yet again (Like SDXL and Cascade)

[karmic cedar]

oooo does Stability (or anyone) have a fine-tuned GPT agent for diffusion-related questions? Seems like that would be an incredible way to get people up to speed with the dynamics of these architectures

[charred mesa]

I want narrow aperture images

while still being aesthetic

[sudden ruin]

You could probably train one on SD papers and similar stuff

[charred mesa]

when I used a Lora to remove the bokeh for SDXL, it made everything brighter (the dataset was probably not diverse lol)

[fervent thunder]

What is CFG scale and why is mine on 7 and some people use a lower one

[charred mesa]

the higher it is, the higher the contrast, but eventually it will become "deepfried" and too distorted

lower CFG is safer, smoother, usually is better for faces and portraits

if its too low then it becomes liquid goop

it also depends if you use a special model (like SDXL Turbo, SDXL lighting and other stuff like LCM, which require a lower CFG in order to look good)

[fervent thunder]

what an good middlepoint

[charred mesa]

for normal models, use about 5 CFG

[karmic cedar]

Fooocus uses that great CFG hack to allow for less cruddy images at higher scales.

[charred mesa]

and just go up and down where you feel like it looks sharp/defined and contrasty enough

[fervent thunder]

How much do i put Hires steps to

[charred mesa]

comfyui also has cfg rescaling and tonemapping which does something similar (or the same thing?)

[karmic cedar]

Yeah I haven’t explored it as much in the nodes

But that is a neat trick for sure

[charred mesa]

are you using focus?

[karmic cedar]

I don’t think they are

[fervent thunder]

whats that

Im using Forge

With Juggernaut model

[karmic cedar]

https://github.com/lllyasviel/Fooocus

I recommend that distro along with the Juggernaut V9 model

Still waiting for a decent pixel art style, though…

[fervent thunder]

I Just put that into my Forge folder?

[karmic cedar]

Nah you’d have to run that outside of Forge

But it’s an alternative to it

[charred mesa]

ok wait

so if you are using forge

use like 10-20 steps at best

in my experience that's enough

if it has a *denoise* function in it then don't go above 0.6

[jovial wraith]

have you checked my guide to pixel art yet? https://discord.com/channels/1002292111942635562/1212330984192614440 i'm personally quite pleased with the results

[karmic cedar]

Hey, this looks nice. I’ll check it out!

[jovial wraith]

:D

[pulsar lintel]

Yes my personal GPT does

[karmic cedar]

Smart.

[pulsar lintel]

Yes he is a beast

[karmic cedar]

He got a dick eh

J/k

[reef shard]

I was hoping to get early access to SD3 for research purposes, including determining the efficacy of various watermarking algorithms on the new model.
It's been two weeks since I signed up for the waitlist and I've heard nothing. Has anyone here gotten into SD3, and if so how long did it take you?

[jovial wraith]

nope, none of us have access yet afaik. Emad said i think yesterday that it would come soon tho

[karmic cedar]

I did the thing, just waiting for the thing

[pine fiber]

could be today or tomorrow

but they don’t like giving exact dates

they are secretive for some reason

as in, don’t interact with the community a lot

I mean, we don’t even have an sd3 channel yet

[pulsar lintel]

Well who knows I named him after Dreamshaper hahaha

I am a girl so everything is a he 😂

[karmic cedar]

That does sound better than Dreamshapess

[pulsar lintel]

Rite

[charred mesa]

Yeah hopefully tonight we'll get access to that secret server where we get to try out and rate the model

I don't think any downloadable models were promised

the 2B model would be fire though

everyone could try that

[reef shard]

aw :( that makes most of the research impossible

[charred mesa]

yeahhhhh 😬

you can't test the censorship for example

[reef shard]

I guess I could still try the prompt engineering

[charred mesa]

do you mean stuff like emphasising tokens?

or just simply testing a lot

seeing how much the CLIP (and if T5) could differ

[reef shard]

things like trying to break safety filters with adversarial prompts

[charred mesa]

there won't be any safety filters (at least I really hope so 💀), it is probably mostly the dataset that's censored

I don't remember anything about harcoded safety filters or stuff like that

but if you do then please tell me

[pulsar lintel]

I think u can keep eyes open on civiti for the downloadable models keep in mind they are in comp with midjourney

[charred mesa]

I want to know about these "safety guardrails" and stuff

[reef shard]

I just want to see how the MMDiT architecture breaks any of the existing diffusion-model research/knowledge

[pulsar lintel]

I am dying to see what SD3 is gonna be like

[vagrant fox]

a bit better than sd 2.1 that's all

[pulsar lintel]

Or if it was only for hype

Agreed

[charred mesa]

💀

if the censorship is only as bad as SDXL then we're fine boys

if it's Cascade or 2.0 we are cooked as f*ck

[pulsar lintel]

Agreed

[real kernel]

they removed all nsfw-related images in the dataset as stated in the paper

[charred mesa]

so far the images don't look botched and the anatomy seems fine (especially for a half-cooked model)

[vagrant fox]

don;t worry if you into fake porn someone will train models on sd3

[charred mesa]

maybe we're getting the ken barbie doll treatment

and thanks to the better dataset, introducing new concepts might be way more effective (Thats my guess)

[reef shard]

...especially watermarks. there's some great watermarking code that I maintain, but it integrates directly into the diffusion pipeline and I want to make sure it'll still work with SD3

[pulsar lintel]

I don’t care for NSFW art lol I’m a girl

[charred mesa]

that could include corn 😬

[real kernel]

no I'm not into fake porn lol I just gave some info about what I already have read in the paper in case you guys didn't read that part of the paper

[pulsar lintel]

Watermarks are annoying

[charred mesa]

idk if stuff like CLIP would drastically change, but the diffusion process probably will

watermarking will probably be a built in feature anyway

[pulsar lintel]

I don’t read I just generate lol jk

[vagrant fox]

in every model they removing nsfw

*every official model

[pulsar lintel]

Pretty sure it’s only for three

[real kernel]

but they haven't for 1.5 or have they? just not much?

[charred mesa]

1.5 was a finetune of 1.4

so most of the NSFW stuff stayed

[real kernel]

anyways the model looks amazing and prompt understanding is on a whole new level so I don't really care that much about nsfw tbh

[charred mesa]

same

Neither did I, until it's about Gore (I want to make zombies) or botched anatomy (NSFW detection being too aggresssssivvee)

[pulsar lintel]

Same like I said NSFW is not my type of art style

[vagrant fox]

as long as you can fit t5 thats 19 gb itself on your gpu

[pulsar lintel]

But sucks u have to have it turned on for uncensored gore and horror art

[real kernel]

I think there is enough space on 2x3090's

[vagrant fox]

nsfw is not art style is type of content tho

[real kernel]

but T5 can be quantized with GPT-Q or similar

[charred mesa]

I hope without T5 and loading the model in fp16 we can fit the 8B model onto 8-12GB 🙏🙏🙏🙏🙏

[charred mesa]

or just launched at int8 with no conversion

[pulsar lintel]

It’s a whole sanctioned of various content just not soft core pron

[vagrant fox]

yeah i hope my 16gb of vram do it's job too

[real kernel]

I think we can quantize T5 to a good amount so the whole 8b model fits in 12gb vram

[charred mesa]

oh that will probably work

[vagrant fox]

yes but it's still type of content not type of art style

[charred mesa]

without T5 it should be working on 16GB

maybe even T5 running at int8

[pulsar lintel]

💯 same here not mine either

[real kernel]

T5 Q4, INT8 and keep in mind that it's a diffusion transformer - transformers can be quantized quite well.

I wouldn't worry too much about vram, in the paper even the 2B DPO version looks quite good

I'm wondering if it's possible to integrate diffusion transformer into GGML

[pine fiber]

uh huh

[languid helm]

is it better than sdxl?

[vagrant fox]

should be

[pine fiber]

it beats dalle 3

[languid helm]

can i run it in my rtx 3060 tho 😭

[pine fiber]

probably not

[languid helm]

damn

i was able to run sdxl iirc

[pine fiber]

it’s slower than sdxl too

until we get turbo you’ll be looking at 1 minute gens

[still glacier]

you have lots of information for something that isn't even released and optimized yet :p

[radiant tusk]

Anyone here experienced with using animatediff past 16 frames?

[pine fiber]

these are all things from the paper and things emad have said

[still glacier]

sure but it s all still beta stuff

[vagrant fox]

50 step 1024x1024 takes 34 sec with 8b model

says right in paper

so idk where u get 1 min plus from

[trim nymph]

3090 24 gb ram will be fine to run it 100%?

[vagrant fox]

i know its fine on 4090 since u have same amount of vram should do

[pine fiber]

with 4090

[vagrant fox]

ofc they dont test it on 2080

[pine fiber]

on a 3060 idk why you think it would be the same speed?

[vagrant fox]

thats obvious

[languid helm]

bruhhh whats the best model currently for the 3060 one?

[vagrant fox]

best in terms of what?

[languid helm]

best in terms of overall prompt reproduction

i would be using it in a pipeline

not for generating arts

kinda creating a pipeline for generating comics from stories

[unique quiver]

can someone please send me a link to stable diffusion so I can begin generating my muscular Waifu with a dodge ram

[languid helm]

bruhhh

start with a1111

[unique quiver]

what

I'm on a phone

[charred mesa]

maybe without T5 and at fp16 it has to possibility to run on 12GB, but I'm not so sure 😬

but that's without mentioning optimizations

and stuff like xformers

[languid helm]

when its even releasing 😭

[pseudo lark]

anyone who knows the best and fastest interference speed LLM that I can finetune on a 3090 (lora finetuning maybe)?
I'm trying to make a successor to magicprompt

[charred mesa]

and if all else fails, there's gonna be a 2B and 800M model

[languid helm]

yeah 2b can def run

[charred mesa]

the 2B probably being the best option for us

I only care about 8B if it has way way more superior prompt adherence or more "knowledge"

otherwise peopel gonna use the 2B for sure lmao

[languid helm]

instead of 800mb, 12b models they should keep the release conventions according to the gpu memory spec

4gb model, 8gb model, 12gb model

lol

[charred mesa]

yeah 8B unoptimized is 24GB

[vagrant fox]

they do top tier model for top tier card

whats so hard to understand

[charred mesa]

yeah

[languid helm]

I see

so mid tier for the mid

[vagrant fox]

community will optimise it anyway just wait

[charred mesa]

I just wonder if with optimizations and omitting the T5 we could get it on 12GB 🙏

I don't care about Text, without T5, prompt adherence doesn't get hurt that much, its mostly text

[vagrant fox]

t5 itself eats 19 so it will be interesting to see what it will evolve into

[charred mesa]

exactly

[languid helm]

i have to finish my project within 1 week

[charred mesa]

and of course you can run T5 at int8

[languid helm]

i think ill go with sdxl

[keen rose]

Hey guys, I just started trying out this AI generation thing with CivitAI's on-site generator. I'm seeing that there's a whole bunch of generators online though, what's supposed to be the difference between those and StableDiffusion? and is CivitAI's generator the same as StableDiffusion or independent of it?

[vagrant fox]

they mostly use different models but most of those are based on SD still

[keen rose]

models as in like checkpoints and loras and stuff?

[vagrant fox]

yes

[charred mesa]

https://pixart-alpha.github.io/PixArt-sigma-project/

PixArt-Σ achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters).

This means that the 800M and 2B SD3 models might still have oustanding prompt adherence

[vagrant fox]

idk who choosen to put that Eiffel tower as example but looks like garbage lol

[charred mesa]

honestly all of the examples look painterly and blobby

im more interested in the prompt adherence

[keen rose]

if prompt adherence is what I think it is that sounds really useful cause when I'm using civitai's on-site generator often it just...straight up ignores them, no matter how early in the prompt I put something

[charred mesa]

Our model is trained on a mix of long (Share-Captioner) and short (raw) captions with a ratio of 60% and 40%, respectively.

Similar to SD3! (But SD3's Dataset had CogVLM captions and it was 50/50 with raw)

[charred mesa]

yeah

what are you prompting exactly?

[keen rose]

is this PixArt thing also something you use LoRAs for and stuff?

[charred mesa]

are you using a complex prompt with multiple people or descriptions?

[charred mesa]

god, no idea

it never really caught on so probably nobody even made loras for it (if it is even possible to make for it)

[keen rose]

oh I'm basically just trying to make solo woman porn. which is annoyingly more difficult than I thought it would be considering how common it is

at least, to make anything seem consistent anyways (face/proportions with different character details)

[charred mesa]

what model are you using to make that

[vagrant fox]

get porn trained model for it

[charred mesa]

there are bunch of 1.5 based models that apparently excel at corn

(not like we should be talking about this in here though )

[keen rose]

I've tried using (on civitai) Hassaku, Pornmaster, MeinaHentai, and sometimes AnyLora (solely because of the name assuming it has better compatibility) as the checkpoint

[charred mesa]

damn

well no idea man

try asking in #📝｜prompting-help

[vagrant fox]

positions are hard to replicate you better off training model yourself if you not satisfied with outcomes

[rich kestrel]

it is not difficult

you must be doing something wrong

[keen rose]

as a total noob, is there supposed to be some kind of way to find checkpoint compatibility with loras?

I've noticed at least some spit out decent looking results while some...are...abstract to say the least.

[rich kestrel]

all checkpoints are compatible with all loras as long as theyre for the same architecture

ie. 1.5, 2.1. xl

[keen rose]

I've managed to make the porn itself but trying to achieve any kind of consistency with artstyle is my main issue (and trying to get the characters to wear revealing clothes instead of full-cloth or full-nude.)

[vagrant fox]

every lora on civitay says for what model it is

then get lora for underwear

[keen rose]

ah ok that sounds like an idea

is there a way to maintain pose consistency if I get an image that does it right? cause it feels pretty random, I read generating with seed helps but it still makes a bunch of overly different variants in pose

[vagrant fox]

controlnet and pose

[rich kestrel]

controlnet definitely is a must

I have this insane pose for spread open legs which is 🧑‍🍳 🤌

fits perfectly on a 536x672 frame

[keen rose]

I don't seem to be able to find this controlnet thing on the civitai onsite generator so I'm guessing I'll have to locally run SD for that (I'll have to wait until I can use my main computer with a graphics card)

[rich kestrel]

yup. running local is also a must

I don't regret buying my grafx card a year ago since I've used it nonstop, both in gaming and SD. Still addicted a whole year later.

[pine fiber]

19gb is so much just for text

[upper pelican]

Any predictions for SD3? I was hoping 2/29/24 because emad made some teasers with that date but turned out to be nothing

[buoyant moss]

So it seems like SD3 8B is around 5 times slower than SDXL (base model only, no-refiner)

[pine fiber]

maybe mid may?

[vagrant fox]

no point in predicting first release will be gatekeeped for online generation only we will still have to wait for models to be released into wild

[upper pelican]

I just wanna play around with it a bit even if its not open yet

[buoyant moss]

They claim 34 seconds on a 4090 for SD3 8B, and I get 7 to 8 seconds with SDXL base model on a 4090. Both 50 steps

[pine fiber]

i assumed they would release weights on release

[upper pelican]

Testing base model capabilities are important and can be expanded later

[vagrant fox]

and how many parameters model u use have? 2b? less?

[buoyant moss]

SDXL?

[shell tendon]

what sampler?

[dry trellis]

He teased limited access starting today but who knows

[upper pelican]

I believe they said something about having a demo first, unsure if may be mistaken

[buoyant moss]

50 steps each, so why does sampler matter?

[vagrant fox]

sampler change also generationspeed

[shell tendon]

some samplers are faster than others

yeah

scheduler doesn't but sampler does

[pine fiber]

just try ideogram

[buoyant moss]

Most samplers have very similar cost, just from my benchmarks

[upper pelican]

Ideogram is pretty good on tests

[vagrant fox]

and some of them take twice as long

[shell tendon]

they can be very different

[buoyant moss]

DDIM vs DPM plusplus 2nd order are within a second of each other

[shell tendon]

dpmpp is twice as fast as the ancestral version

[keen rose]

fits perfectly on a 536x672 frameI've been assuming that the pictures on civitai's on-site generator are lowres just to reduce costs, I've seen 4K AI images before (which aren't blurry) what makes the difference / why choose to go low?

[rich kestrel]

bc upscaling is easy

you always start with a low res and upscale as wanted

[vagrant fox]

to generate faster, also you wont be able to generate 4k image even with 20 gb of vram i believe u need upscaling and it also saves you time

[buoyant moss]

Regardless. SDXL base is like 3.5 billion parameters. SD3 8 billion being double that, and a transformer architechture, being 5 times slower is not surprising to me

[pine fiber]

is that including t5?

[vagrant fox]

but you compare sdxl to sd3 with t5

[keen rose]

is upscaling something more sophisticated than say photoshop's image size changer? cause if it's like that I wouldn't expect much of a quality difference

I'm assuming the upscaling is done natively in SD and not some other program

[shell tendon]

oh yeah way more sophisticated

[buoyant moss]

Ah lol, funiest story, I just met one of the first authors of T5 flan in an airport waiting room in November

I asked him what it was like to train T5 at Google

[shell tendon]

what the hell are the odds of that lol

[buoyant moss]

He said it was mostly debugging and restarting training when the model blew up

Very high

[rich kestrel]

yah, it is nothing like traditional upscaling. new upscaling method basically re-renders the image in tiles

[buoyant moss]

I mean... if you go to a ML conference, lots of authors of various models are there

[shell tendon]

ohhhh

makes sense

[buoyant moss]

But anyways, it sounded training T5 was less like science

And more like being someone who had to debug and fix things when it fails

[shell tendon]

fun

[vagrant fox]

isn;t that most of new tech? method of trying and failing untill u succeed?

[buoyant moss]

Maybe

There's a joke for that

[shell tendon]

(so what sampler were you using? and was it comfy or something else? curious because i have a 4090 and have been doing some half ass benchmarking)

[vagrant fox]

nothing works first try

[fervent thunder]

What type of performance is a 4070ti supposed to get?

I saw someone's post saying they get over 30

[vagrant fox]

how many cuda cores does it have

[fervent thunder]

Meanwhile I am getting max 14 it/s