💬｜general-chat

[shell tendon]

ha zf0

[light lily]

b r u h

[trim magnet]

dont project yourself here pls

[shell tendon]

real hackers worth fn jack shit don't waste their time on pissant BS like discord fights

[hallow sable]

discord sucks honestly

[shell tendon]

they're too busy doing important, consquential shit

[hallow sable]

imagine paying for nitros :S

[trim magnet]

hes like snowden but he can bench 500pounds each leg

[shell tendon]

yeah, i know about that stuff

[light lily]

nice man

[hallow sable]

no void pointering!

[trim magnet]

we in presence of a very dangerous individual

[forest trout]

What you guys listening to rn? I'm listening to Oneohtrix Point Never.

[light lily]

I noticed as well everyone hold on to your lan cables
he might go hack you so you can unplug lmao

[trim magnet]

only danger you pose is to the taco bell toilets

[mortal delta]

i dont really listen to music...

[hallow sable]

ha im on wifi on my xbox nice try

[trim magnet]

you do have experience digging

[light lily]

bro u done no one gives a rat's a$$ about ur threads

[hallow sable]

silk threads?

[light lily]

my name is julien actually

[shell tendon]

there's a lesson i learned a long time ago

the most dangerous people don't talk much

because they don't want you to get your guard up before they strike

[hallow sable]

I was molded by the black hats

[trim magnet]

what a badass

[light lily]

nice bro you propably have a 4min ping lmao

[hallow sable]

im pogging right now

[trim magnet]

he did say he was about 40, probably 40 iq

[light lily]

ok bro

[hallow sable]

did you know comfy ui was made using punch cards?

on basic

[shell tendon]

lol 15 languages

once you learn like... 3 the rest are just semantics

[forest trout]

If you don't program in Turbo Pascal, can you really call yourself a programmer?

[mortal delta]

i feel like we should move the conversation to something else? like a different topic...

[shell tendon]

ah here we go, thank you mods

oh, self-moderation, like a good corporate boy

[mortal delta]

im no mode btw.

[trim magnet]

im actually enjoying the show rn

[light lily]

fkn same

lmao

[hallow sable]

how do you do a inverted mask with comfy ui? I want to change everything not masked

[trim magnet]

yes put it all over my face cowboy

[light lily]

you finally done with your intellectual posturing

[clear oyster]

what happened here lol

[light lily]

you must be the most insecure person irl

[mortal delta]

where can i ask about tech help and such, in this server?

[forest trout]

#🤝｜tech-support

[shell tendon]

grew up in the hacking community myself... the talkers never did fn shit

[mortal delta]

thank you.

[shell tendon]

my friends were all serial killers

[vague pond]

but he's behind 7 proxies!

[shell tendon]

they hacked up nerds in the woods

[light lily]

bro is this the SD discord or flarkin cartel wher am I here

[forest trout]

https://tenor.com/view/hacker-hackerman-kung-fury-gif-7953536

[hallow sable]

EMBEDED FAIL LAUGH AT THIS USER!

[forest trout]

Aww. Forgot this was the general chat channel.

[vague pond]

The last three times I have randomly checked this channel there have been bizarre trolls. Insanity.

[clear oyster]

wtf is up with this chat today

[clear oyster]

same question

[clear oyster]

lmfao that too he has a black role hahahaha

[clear oyster]

can someone moderate this server lol

[trim magnet]

dont worry is just unglued who was been sniffin glue for the past hour

[mortal delta]

i just posted there and i hope things go well.

[clear oyster]

hopefully it passes fast lol

[shell tendon]

and one even bounces through vietnam! and there's a shell on his buddy's server in canada and i heard their police and very uncooperative when the FBI calls

[forest trout]

I'll give you some more freebies: I used to moderate this server too.

[mortal delta]

so... ummm how is everyone today?

[forest trout]

Apparently everyone in this channel is off their meds at the moment.

[shell tendon]

nah, just one guy

[mortal delta]

thats crazy.

[clear oyster]

um ok lol?

[clear oyster]

that makes eevryone else a little off too lol

[winter pike]

almost 29/02

[shell tendon]

so does emad believe in midnight drops?

i would strip naked and run down the street screaming im jesus! im jesus! if the models dropped at midnight

[winter pike]

he dont

hahaha

[mortal delta]

i just got buried in the tech support channel, dang

[pale latch]

looks like i picked a bad day to stop sniffing glue

[trim magnet]

u are suppose to eat it

[forest trout]

Or a good day to pick it back up...

[trim magnet]

you are hilarious already

[steel mural]

Hi guys, could someone help me answering my question in #🤝｜tech-support? Pleaaase

[mortal delta]

was me getting an amd card a mistake if i love ai? im trying arch linux but this os is so hard and nothing seems to work the way i want it to.

[rich kestrel]

yes

especially with 50 series on the way

[mortal delta]

dang

[rich kestrel]

theyre gonna change the AI landscape forever

[wispy oasis]

when

[rich kestrel]

u can still use it

[rich kestrel]

Q4 this year or at most, Q1 next year

[wispy oasis]

oh wow

[mortal delta]

but for what?

[wispy oasis]

long time

[rich kestrel]

linux with some patches

for rocm compat

but honestly, I wish everyone was team cuda

[mortal delta]

im too cheap for cuda.

[rich kestrel]

good to anticipate to save

that shit aint gon be cheap

[wispy oasis]

how much you think

[rich kestrel]

if we're lucky, 1800msrp

if they go nuts, for 32gb version maybe 2500msrp

maybe 2600

its a crapshoot but prices will get leaked eventually

I think it'll be somewhere in the middle. I'd say 2000

[wispy oasis]

damn

[rich kestrel]

bot hasnt worked for a long time

[forest trout]

#1047610792226340935 message

Looks like it's been down since late January.

[fervent thunder]

sadly im about to buy a 4070 cause I cant wait xD

[trail lion]

it only looks like that because it has

[fervent thunder]

Hopefully i can return it when 50 series comes out

[pliant tiger]

Can somebody please tell me if there is no function for generating on board now?

[nova zodiac]

#1047610792226340935 is search broken again??

[inland swan]

can someone give me an image and caption with their lora training so I can see how detailed people are

[pale latch]

no you dont return a year old card. sell it duh

[teal yacht]

hi, so, most ai generation online isnt free or limited, do if i use my own 4060 laptop i can use as much as i can and with good quality?, which one is worth

[radiant tusk]

Do people still use dreambooth? or is it all training loras now

[hallow sable]

I use dreambooth

[inland swan]

Thats what im uysing

[hallow sable]

I think kohya ss is better

but ive never used it

[inland swan]

Thats what I thought KohyaSs is

Oh well Im uysing kohya ss

now answer me a few questions

can I train mutiple characters on the same lora?

[hallow sable]

no

i wont

[inland swan]

ANSWER THE QUESTION

PLEASE

i BEG

[hallow sable]

😠

hmmm I dont see people doing that with people. but with buildings and other objects yeah

[inland swan]

well thre reason is

I wanna use lora to be able to make pictures with both OC characters in it

[hallow sable]

OC ?

[inland swan]

original

character

My own creations

[worn radish]

yo guys
every time i create a image (using fooocus),
the character is looking at me/ the camera
i always write stuff like "looking at laptop, staring at laptop"
but it doesnt work..
anyone know why?

[hallow sable]

lora is made to make images with strong resemblances to the dataset it was trained on, you can use lora with a low value

[inland swan]

Right ideally so I can use two characters
now one more questions what are Regularisation images meant to be

Its meant to be like close to but not exact images to train the model on those images compared to one another?

[hallow sable]

if you had a lora based of bin laden it would change everyone into bin laden, you can work around that with a mask

[inland swan]

Whats that website you use to grab a whole bunch of images at once based on what you enter?

[hallow sable]

gooogle?

[shell tendon]

You can do regional Lora

[inland swan]

Whats that mean captain?

[shell tendon]

In comfyui you can do work flows that allow you to only use a Lora in one place

Another approach is called attention couple

If you don't care about perfect likeness the easiest approach is a pair of ipadapter with attention masks

And start with a controlnet for the first 0.1 to 0.2

That sets the characters in the right spot forh the ipadapter to then take over

[pearl ocean]

I wanna make a Disney film with Sora

[inland swan]

Dont have that or I would use it

I only use easydiffusion

[shell tendon]

Comfyui is easy to install

[inland swan]

I know but I dont like using it

its wacky

[glossy drum]

execution.py how to enable

[pale quarry]

future boom😅

[pearl ocean]

😮

[wet roost]

Im looking to buy an online service

whats the best for stable diffusion

[fervent thunder]

no, like if I buy one tomorrow could I before Q4

i have a 5700xt but im gonna upgrade to 4070ti

Then a 50

[pale latch]

return policies are typically 30 days

tops

[fervent thunder]

smh

ill return one and buy a new one every 30 days

[pale latch]

wtf kind of entitlement comes up with returning it a year later

[karmic radish]

Does anyone know where we go to get help with the official stability AI API? It looks like the "platform" channel no longer exists?

[lone hawk]

it's been 84 years... SD3 still hasn't released

[gaunt pulsar]

I just need SD3 in my life and I'll never be bored again

[nova zodiac]

it's been up and down for about 3 weeks, i'd email them

[nova zodiac]

It's been less than a week, and the reddit comment from one of the stability team said 2-4 weeks...

[inland swan]

SD3 being released soon?

[nova zodiac]

literally look up 1 comment chief

[trim magnet]

SD3 when and where

[trim nymph]

hope that user gens gathering starts soon 🙂

[static cape]

Came here to ask this. 😛

[proper wharf]

hi new here

is it possible to run lora on collab?

[nova zodiac]

run as in generate with one, or run as in create a lora?

[nocturne pewter]

Been watchin this ’movie magic’ docu series from youtube. Guessing it's from early 00's somewhere. It’s portraing matte painters, cinematographs, cgi ppl and other envolved with movies making the visual possible. Super interesting, highly recommend! A bit sad to watch it though knowing all that is history within 5 years or less. Surely new kind of ways to work arise at the same but still.

[sudden ruin]

When SD3?

[proper wharf]

i tot lora is another form of stable difussion

[nova zodiac]

no.. lora is "LOw Rank Adaptation" - the adaptation being it adapts the model that it applies to

[nova zodiac]

https://tenor.com/view/beating-baseball-bat-harley-quinn-noxville-gif-21079527

the lack of gif on this page ruins that effect...

[hushed quarry]

When SD3

[dry stratus]

When SD3 😛

[worthy bone]

Yes, when SD3?

[pale quarry]

Mid April

[lone hawk]

stop torturing us, give SD3 😄

[nova zodiac]

What's gonna come first, sd3 or Half Life 3....

[lone hawk]

https://th.bing.com/th/id/OIG3.DLm.071icHXENnqBLS_w?w=1024&h=1024&rs=1&pid=ImgDetMain

[nocturne pewter]

Hey, has anyone already asked when sd3 is coming? 🤭

[trim magnet]

3 questions partna's
when the heck is SD3 releasin?
where the heck is SD3?
what the frick happened to SD3? ⁉️

[shell hazel]

SD3? Meh

Give me SD Sora thanks

[upper topaz]

heyhey. uuuh. i dunno if this is the place to ask. but would anyone be willing ot help me get an good setup?
i know theres like extentions and ive seen some before, but its been a while
and frankly, im a little stupid. so can anyone here help me get set up? already got the SD set up an such. so i can like make good pictures and within my vram(i have six dedicated ram. and 22485 "total available"

[limpid tapir]

SD3 is the first step to that hehd

[warm junco]

Hey, do you already have the right commandline Args in your webui-user.bat for Max performance?
Also, do you need help with setting up extensions like Controlnet?

[upper topaz]

oh im getting help. but more is alwaysgood
they reccomended Forge for me.
also. never heard of controlnet. whats it?

[jovial wraith]

It's an extension that allows you to get more control over an image in several ways. For example, you can add poses to an image, declare the general layout of the image, that kind of stuff

[inland swan]

Im having every issue in the world

just trying to make a lora

[desert nebula]

I m trying make pictures for a friend that look like her, I use IP-adapter in Forge. What is the best workflow? Do I take the seed of the result with the most similar face and change the prompt?

[karmic hamlet]

Hey, i need some recommendations on decent laptops to run SD smoothly without any hiccups

[pine fiber]

any nvidia gpu with 16gb vram and 32g ram

something like 3090

if you want to run sdxl

but sd3 will require like 24gb lol I would wait a bit

[jovial wraith]

Is that just speculation or do you have a source?

[pine fiber]

im just guessing based on the fact the largest model was said to have 8b parameters

[jovial wraith]

Ah fair

[shell tendon]

https://www.reddit.com/r/StableDiffusion/comments/1az3jwe/comment/ks0imjk/

Emads comments on that

[jovial wraith]

i think it's pretty safe to say i have no hope of running this on my machine lmao

[naive belfry]

Hey guys! What's been your go-to model as of late? Really been curious about what everyone's been using. Getting back into ai art after a bit and wondering. I found a new(ish?) model i'm really enjoying, astica.ai. Anyone else tried it out?

[pine fiber]

thanks, didn’t see that!

[karmic hamlet]

Thanks for the response. I'm learning with the web UI right now and my current laptop runs on 8gigs ram but a 4GB GTX from earlier 2021. Tried launching the UI and using models but the whole laptop freezes and then crashes. Even tried modifying the bat file to reduce the load but no luck as yet.

[pine fiber]

mhm on 4gb vram you can probably barely load 1.5

it’s really the limited factor in open source

[winter pike]

are you using sdxl? i also have a 4gb GTX (1050) but sdxl takes several to finish one pic. Sd1.5 is good, less than a minute

[shell tendon]

Yeah def a good idea to sit tight a bit if you're an inference only person

For training 24gb vram is a no brainer but inference... 12 might be fine, might need (or really want) 16 who knows

Kinda like 10, vs 12... I was so glad I got the 12gb 3080 long before I discovered SD and not the 10

Most of what I was doing with SDXL got me to 10.7 or 11.5gb vram used

[naive belfry]

Sorry if not allowed, jst wondering if there's anyone here with experience using astica.ai. It's open source and seems great so far but i would love some pointers. nobody in my inner circle seems to have heard of or used it

[pine fiber]

never heard oif it

[shrewd linden]

Anyone care to explain why this prompt "three characters, a boy and a girl and a dog, looking at a spooky american mansion, (Makoto Shinkai, Greg Rutkowski), (anime:2) japanese ink painting " usually only gives me a girl, a girl and a dog, or a boy etc... but always never the three subjects together like prompted?

[haughty brook]

https://chat.openai.com/g/g-0pA1w4IJm-the-memory Play The Memory

[cobalt turret]

with this one you can run sdxl on your 4gb gpu: https://github.com/lllyasviel/Fooocus (but it takes 2 minutes per image)

[charred mesa]

yeah this is a limitation with current open source models. Stable Diffusion 3 will solve this problem

its not out yet sadly

[cobalt turret]

if you have a rtx... gpu, gtx still needs 8gb for some reason

[cobalt turret]

has anyone got private access through the waitlist yet?

[charred mesa]

nope

I heard from someone that someone heard from another guy that heard from a random person that they heard that today would be when people get access through the waitlist

guess not

or maybe in the evening in the american timezone 🤷‍♂️

THIS IS JUST GUESSING BTW

like I have no idea they tell us nothing

its been radio silence for like 2-3 days

[cobalt turret]

isnt the goal of a waitlist to give people access gradually over time? I find it weird that literally no one got access. Normally at the start of a waitlist some people get access and then more over time. Its pointless to have a waitlist if you give access to people at a specific time at once

but I guess thats how they want to do it

[charred mesa]

not even the subreddit talks about SD3 rn

I thought people would be hyping but it just died like it wasn't even announced lmao

[worn trail]

because there's no updates on it

[charred mesa]

true

[cloud plume]

idk i think they might do it gradually. With the last bot they also gave access first to only a "few" people, but idk.

Also most people likely aren't aware of the waitlist. Idk how many are actually on the waitlist. (the announcement didn't ping iirc)

(also for all we know the bot might already be online but hidden to people without access 🙃)

[forest trout]

They started with early users, then beta users, etc.

[pine fiber]

https://vxtwitter.com/EMostaque/status/1762187622326186078?s=20

[charred mesa]

oh my bad

thanks

[pine fiber]

but yeah they havent said a lot

for an open source model they are awfully quiet

[charred mesa]

I signed the waitlist the day it came out and only like 4 days later I noticed that I didn't receive the email

I bet I'll be put so goddamn back

unless its completely random

[pine fiber]

lol?

I got the email

[charred mesa]

and I didn't

[pine fiber]

did u just put your email wrong or what

[charred mesa]

yeah that's probably it

its just so goddamn unlucky

maybe I used up all my luck by getting the Tekken 8 Closed Network Test and Closed Beta Test

[worthy bone]

I don't understand, Did you receive the email when you registered on the waitlist?

[pine fiber]

I just got one that said I was in the waitlist

[worthy bone]

I signed up the day they opened the waiting list but I still haven't received any emails

[pine fiber]

u sure

[worthy bone]

Yes absolutely, I followed all the steps to the letter

[pine fiber]

weird

[worthy bone]

But did you receive this email after a few days? Just to understand

[charred mesa]

Oh no, I probably typed in my email incorrectly the first time and therefore I didn't receive and email at all. Around 4-5 days later I noticed and signed it again and THEN I got the email

[charred mesa]

though this worries me

maybe it was just a bug or something lol, I don't understand

yeah I signed up 22th, nothing, tried to sign up again on 25th and THEN I got an email instantly

[pine fiber]

no it was immediate

u both probably just put email wrong

[charred mesa]

yeah that's a high possibility

[fervent thunder]

same i signed up again and got the email now

[charred mesa]

BRUH

[pine fiber]

dub looks like im getting into early testing

[charred mesa]

I'm starting to doubt myself both ways

[pine fiber]

I hope its today bc

[charred mesa]

yeah today but probably in american timezone + evening or something

that's my guess

[pine fiber]

im hoping emad didnt just hype up that stability ai crossover thing

[charred mesa]

what is your timezone

[pine fiber]

gmt

[charred mesa]

okay you should probably be fine

[charred mesa]

he was posting on his twitter about whatever crossover it was

[pine fiber]

yeah

morph x stability

[charred mesa]

yes morphaistudio

don't think that's it

unless they are trying to mislead and disappoint us lmao

yeah they have a waitlist too, this is worrying

then again, the images were generated with SD3, so it must mean it's FOR SD3

[pine fiber]

I hope it can do res higher than 1mp

[charred mesa]

ehhhh

I wonder if it will be even around 2048px lmao

[pine fiber]

it should be able to technically

[charred mesa]

aren't they staying with 1024px, at least when training

still no technical paper actually

damn

[pine fiber]

probably same dataset as sdxl if i had to guess but idk

[charred mesa]

even the tech paper release would satisfy me for now

I just want to see the more technical improvements, like how flow matching EXACTLY helps and stuff

[pine fiber]

the new architecture allows for lots of resolutions so i dont see why not

true

[charred mesa]

you sure it's not just aspect ratios you mean?

[pine fiber]

some of the image quality emad posted werent exactly miles ahead of sdxl

let me check

"Sampling flexibility
Sora can sample widescreen 1920x1080p videos, vertical 1080x1920 videos and everything inbetween. This lets Sora create content for different devices directly at their native aspect ratios. It also lets us quickly prototype content at lower sizes before generating at full resolution—all with the same model."

and they can make images up to 2048x

so 8mp

or 4mp sorry

[charred mesa]

idk even 2048px from a single pass sounds good

[pine fiber]

I dont think sd3 model is large enough for that tho

idk enough about it

[charred mesa]

well, if highresfix will work fine then it will be okay

especially now that it's smart, it will probably KNOW what is being upscaled

img2img will be so good with this

[pine fiber]

we will see

[charred mesa]

you know what would be surprising

its if they were switching between the model sizes for all the images they have been posting

[brave matrix]

does anyone have sd3 yet?

[shell tendon]

No

[trim nymph]

oh man, ive signed up 10 minutes after announcement only to get an email now after trying again.

[grizzled palm]

Checking Discord every thirty minutes and waiting for the first SD3 images to start dropping...⏳

[charred mesa]

What the fuck so its more common than I thought then

[worthy bone]

Luckily I tried again and the email arrived

[charred mesa]

yeah I tried like 3 days later

so it may not be that terrible

[worthy bone]

Let's hope so

[trim nymph]

maybe we got on the list & its just confirmations that started coming later/confirmation emails were setup later

[storm shard]

Are we supposed to have received an email after signing up as confirmation?

[shell tendon]

Congratulations you've been added to the Stable Diffusion 3 early preview waitlist!

I got that when I signed up

A while ago

[storm shard]

😦

I signed up approximately within an hour after the message on X, haven't received anything

[shell tendon]

Feb 22 1pm minutes after I signed up

Did you use Gmail? I did

[storm shard]

No, I don't use gmail

[wispy oasis]

can i sign up if i dont have twitter

[storm shard]

"You'll notified by email with an invite to our Discord server when you've been granted access to the preview"
This sounds like you'll receive an email at the moment access is granted

[cloud plume]

yes (idk if it's still possible though) #📣｜announcements

[wispy oasis]

cool

i think we need faster billing support before sd3. my card just work on stripe 2 days ago but wont on stability website

well he message i replied to is gone now but ok

[storm shard]

I made a mistake with my discord-id, signed up again and now I got the email

[real kernel]

Mee too! 😢

[fervent thunder]

it's fine we'll all get access anyways lol

[real kernel]

but wen

[mortal delta]

what can be done with ai for fun, because i find myself to be bored alot is there a way i can use ai to cure my boredom?

[grizzled palm]

Yes! (At work sorry later.)

[grizzled palm]

Is no one posting here because SD3 dropped somewhere and I'm the only one who doesn't know?

[jovial wraith]

i don't think there would be silence if SD3 dropped lmao

[charred mesa]

its just that the server is dead rn

we are just waiting for SD3 access from the waitlist tbh

[jovial wraith]

this is probably a bad idea, but i'm bored outta my mind. so if anyone who doesn't have access to SD wants an image made, hmu ig

[shell tendon]

I keep checking wondering.,, damn all the silence

[grizzled palm]

I want an image of an easter egg with a weird design dyed on it.

[shell tendon]

Are they going nuts over there at stability drinking gallons of spiked coffee frantically trying to fix some bug that causes "high quality photo" to spit out porn every time lol

[trim nymph]

something is releasing/started today, its being worked on i read just now

[shell tendon]

Got the impression something was coming out today, thought it'd be a lil earlier in the day

[past rock]

Do you have access to SD3?
Can you please generate:

biped cute shorthaired caracal walking from a local grocery shop holding two white plastic bags with food, in a cloudy day, russian winter town streets, February, typical soviet panel buildings at background, dramatic cinematic film still

[karmic brook]

[jovial wraith]

not sd3, just what everyone has access to

[grizzled palm]

The community would love that bug. Really... I wonder what. Maybe lots of articles, contacts, getting the bots ready for the eval comparison, checking repo dependencies on something other than Ubuntu. (The last one was a joke. They're not doing that.)

[outer crescent]

weird notion of what "the community" is. 😛

[charred mesa]

SD 1.5 community

[jovial wraith]

i don't know what it is, but i cannot get that thing to be bipedal. and with controlnet it turns into a robot of some sorts?

[past rock]

"biped" should put the caracal on its two hind legs. It works great in DALL-E

[shell tendon]

Use the word anthropomorphic too

[jovial wraith]

i even put the weight up, but it did not work

[jovial wraith]

did that too

[shell tendon]

FenrisXL as a checkpoint

[fervent thunder]

guys i'm a time traveler, sd3 comes out tmr

[jovial wraith]

i don't think i can run that model lmao. it's twice the size of my vram

[hallow sable]

what makes SD3 better than the rest just the dataset it was trained on????

or is there more behind it

[jovial wraith]

prompt understanding, text, hands, just to name a few

[fervent thunder]

the rest? as in dalle 3 or sdxl? if you compare it to dalle 3 it's promising to be dalle3 that can run on consumer hardware

[hallow sable]

ive only used SD, midjourney, bing dalle

[charred mesa]

wdym better, in what way

just prompt understanding or aesthetics too?

[jovial wraith]

i've seen some impressive images from it posted on twitter, so i'd say the quality improved significantly as well

[charred mesa]

I think the quality is around SDXL base tbh

so its aesthetically decent

just not EYE POPPINGLY BEAUTIFUL like midjourney

its not like it won't be possible, finetunes are gonna bring it up to ideogram quality if not further

[bleak rampart]

Sorry if this is a obvious question but can people contribute to the code base of SD?

[charred mesa]

I guess if you make a pull request on the repository then sure? 🤷

[bleak rampart]

I'm a data scientist who works w/ torch often and would love to help out

[charred mesa]

damn nice

but which Stable Diffusion

[jovial wraith]

https://twitter.com/Lykon4072/status/1762126186581242333 i find the dragon in particular very good

[charred mesa]

2.1 repository or SDXL repository

[shell tendon]

Oh shit sd15 only then?

[bleak rampart]

I guess I need study the code base first

[charred mesa]

oh crap I forgot about that, yeah it looks pretty good

still, not desaturated enough 😉

[bleak rampart]

Is there a "start here" in terms of knowing the technical structure ?

[charred mesa]

uhh

I think there's documentation on huggingface

[jovial wraith]

i can run any sd1.5 model, and only some sdxl models, albeit very slow

[charred mesa]

I think it's diffusers and you might find code examples there for multiple models or something

I don't know myself

[bleak rampart]

Hmmmm ok I'll google it then

Thank you

[charred mesa]

You're welcome, good luck with that

[charred mesa]

wdym some?

like fp16 models only?

or just like stuff like lightning and turbo

[jovial wraith]

that, yeah

[charred mesa]

ahhh

[jovial wraith]

i mean, i can generate on regular sdxl, but my memory nearly runs out when it's going through the vae

[charred mesa]

no refiner?

[wispy oasis]

how do i get stable if billing support wont respond to me

[jovial wraith]

haven't tried yet, but i don't think so

[charred mesa]

oh so you only use base

hmm

how much vram

[jovial wraith]

3 gigs

[charred mesa]

huh

I think someone got stable cascade running on that

probably dog slow

[jovial wraith]

really? my memory ran out completely when i ran it

[charred mesa]

they probably offloaded a lot and applied 50 optimizations

just hold on

[jovial wraith]

like, it used all my vram, and 8 gigs of shared vram as well

[grizzled palm]

Any minute now guys...

[charred mesa]

yeah

[jovial wraith]

do we have confirmation that it's coming or are we just speculating?

[charred mesa]

I hope it's not like 2 am in our timezones

if not worse

[lone hawk]

any minute SD3 dropping? Oh finally, good to know!

[charred mesa]

well, the online early test version that is...

also what about the techinical paper?

[lone hawk]

what a perfect timing to release SD3 on april fools day Lol

[charred mesa]

nahhh wait

https://twitter.com/EMostaque/status/1763286314160460224

"not quite there"

on a post where he posted an image of a robot with 29th on it

[shrewd linden]

I can wait but as a relative noob it really surprises me how ‘bad’ SD is in understanding prompts compared to midj/dalle

[charred mesa]

Shit I wonder if that means they are suddnely delaying it last second

[charred mesa]

the dataset was literally a tagging system lmao

"black dog", "in park"

[shrewd linden]

what does that mean?

[charred mesa]

lol

like

[lone hawk]

Nooooo...

[jovial wraith]

to be completely honest, i'd rather wait a bit longer than have an unfinished model

[charred mesa]

but we're testing it early, that's the point

we're testing an early model through a service to give them feedback before it's fully trained

they'll release the weights when it's finished

[lone hawk]

yeah that's the point, why the stalling... please just release it! Go!!!

[jovial wraith]

yeah, fair point

[fervent thunder]

hiii guys who knows how to hack fortnite ? i would love to make a present for my little cousin

[pale oyster]

most likely the delay is due to issue in the integration / infra, not related to the model

[fervent thunder]

tf?? hahahahaha

[fervent thunder]

told y'all

[grizzled palm]

No SD3 today guys. 😦

[fervent thunder]

so where is it..

[fervent thunder]

i said tmr

watch it not release tmrw😁

[lone hawk]

it is Emad, get him guys! 😄

[fervent thunder]

yes!!!

[lone hawk]

ofc not, it's obvious 😛

we are prob two weeks away....

[trim nymph]

https://x.com/emostaque/status/1763270146938786142?s=46

[charred mesa]

AHHGHGHGH

"no just waiting for that to be out, long day"

[trim nymph]

„waiting for that to be out, long day“

[fervent thunder]

so tmrw?

or the end of this day idk

[trim nymph]

so maybe something happens in the next 10 hours

[fervent thunder]

hmm i hope so

[trim nymph]

would be fitting that user generating starts a week after announcement

[charred mesa]

23:59 in the most western American timezone

[trim nymph]

just hope they think of us europeans

[lone hawk]

don't give me hope

[charred mesa]

nahh

[fervent thunder]

guys!! omg i just got the sd3 discord server invite🥺

i didn't say the time zone 🤥

[charred mesa]

and don't even mention the aisans bro its like almost morning for them when this will get released

if at all released

[fervent thunder]

liar 😢

[fervent thunder]

nuu >w<

[honest spear]

gonna stay up and wait to see SD3 releasing at probably 5 a.m. just as preview for selected users.-.

[wispy oasis]

hi

all people get invites on list or its only some

[charred mesa]

no idea

[worn trail]

SD3 has at least 3 parameters

[trail lion]

impressive

[charred mesa]

?

[fervent thunder]

we need more info about it pls🙏

[pine fiber]

let me guess 0.2342634 0.6239587 0.892374

the usual suspects

[charred mesa]

exactly

correct

[static cape]

Maybe it's not pruned / lobotomized enough yet. 🤫

[shell tendon]

Id like the unfinished model now plz

[cloud plume]

Tbo I mostly can't wait for people to (complain about that)/(be confused by) the prompts they optimised for an older model no longer producing the exact same results 🙃.

Explaining that people "overfit" prompts for specific models and also things like regression to the mean never gets old 🥰

[charred mesa]

SD2.x moment lmao

not the same clip model guys aaah!!!!1!

also I wonder if the more literal and sensitive SD3 prompt structure would incentivize the usage of llms

[radiant tusk]

What's the best models/loras for text in pictures?

[jovial wraith]

you're best off just waiting for SD3

[charred mesa]

https://civitai.com/models/146548/bettertextredmond-improving-text-on-sdxl-concept-lora I think this is alright for SDXL, I didn' try that much

but yeah, just wait for SD3 at this point

its not gonna be slightly better, it's gonna be WAAAY better for text

like near paragraph length good

[honest spear]

Mmh I'd just use controlnet, like a canny or depth and a model with the style that I need

[lone hawk]

Stability staff posting on twitter more boring SD3 pictures, honestly what's the point? It doesn't showcase the new model capabilities...

[charred mesa]

yeah they sometimes show regular images that are already possible with current models

we want more complex prompts

unless of course we get access from that waitlist thing, then we can post our own stuff

I'll watermark them with like "SD3 EARLY" of course

or just simply say it beforehand

[lone hawk]

what is this guy doing, like what's his job? https://twitter.com/andrekerygma

[pearl ocean]

X

[pearl ocean]

Do you use SD or SDXL?

[alpine hull]

Anyone know where SD3 is available once you're past the wait-list?

Wanna know if it's in dreamstudio

[charred mesa]

SDXL right now

[charred mesa]

a discord server

[pearl ocean]

What’s the difference?

[charred mesa]

SD3 will be smarter

[pearl ocean]

😮

[charred mesa]

it will know how to generate text and it will treat multiple subjects better

[alpine hull]

You have to use it through a discord bot? :(

[charred mesa]

yup...

at least, that's what I think

[alpine hull]

Shoot

[charred mesa]

since you will get a discord server

why what's the problem

[pearl ocean]

I have noticed that when I tell SD to do coloured eyebrows, it won’t do it. Would SDXL do it?

[alpine hull]

Oh you're speculating

Hopefully it'll be available in dreamstudio

[charred mesa]

possibly, no idea

SDXL is slightly smarter than SD1.5

[pearl ocean]

😮

[pine fiber]

not really way better but

[charred mesa]

might as well wait for SD3 and the finetunes of SD3

[pine fiber]

idk

sd3 image quality kinda looked like ass

[charred mesa]

have you seen the shirt text?

[pine fiber]

yes

[charred mesa]

so...?

[pine fiber]

its alright

not perfect

[honest spear]

Give sdxl a chance, it has 2 text encoders, some models follow the prompt better than others

[charred mesa]

yeah its a base model + it's half baked/not finished training

[pearl ocean]

I wonder if SD can make a good Sonic

[pine fiber]

1mp is kinda bad

details are blurry

[pearl ocean]

Why is SD max at 2048?

[charred mesa]

yeah the model is probably still 1024px

I always use highresfix anyway so I don't mind

and considering the VRAM issues that might arise thanks to the HUGE 8B model, it's understandable

[pearl ocean]

Hopefully a 4080 is enough

[pine fiber]

they should have done a 13b but I can understand 8b for training purposees

[charred mesa]

the 1.58 Bit paper might help with that, or for even larger models, but they'd have to start training all over again, even if it was faster 😬

[fervent thunder]

are you talking about the twitter pics? i think that's cause twitter compression is bad

[fervent thunder]

can't diffusion transformers be quantized similar to llms?

[pine fiber]

perhaps

[charred mesa]

I think so yeah https://www.reddit.com/r/StableDiffusion/comments/1b2fxeg/this_revolutionary_llm_paper_could_be_applied_for/

its interesting

[charred mesa]

yeah the compression is terrible, but I have noticed that images are sometimes rather blurry in a way

but, again, highresfix usually solves these problems

[pine fiber]

its not been tested with higher parameter models yet

but theyre in training

[fervent thunder]

well that paper says you need to retrain from scratch, quantize methods are done post training

[charred mesa]

oh you mean that

oh yeah

I don't know about stuff like if int8 would work

or GPTQ 4-bit and stuff

I don't understand what would work

[pine fiber]

u get speedups because the multiplications turn into additions

[charred mesa]

int8 could be applied to the text encoder of pixart-a and deepfloyd to heavily decrease their vram requirements

to like ~8-12GB vram

we don't even know the text encoder or clip model for SD3

[fervent thunder]

ik pixart-a has a LCM version

[charred mesa]

no technical paper STILL

[cosmic herald]

Is sd3 out yet

[lone hawk]

no, but you may enjoy the most boring and random pictures of SD3 posted by stability stuff on shitter 😄

[radiant tusk]

like if I wanted to generate a sign saying "McDonalds" I can use controlnet to make it so the text isn't all fucky?

[honest spear]

Yes, you can use controlnet or even ip adapter if you have a reference image, just don't set the strength too high or the model will struggle

[jagged flower]

Hi at all 🙂 in which cannel i can create ai images? 😁

[jovial wraith]

i'm afraid you can't currently, as the bot is down.

[jagged flower]

Ah ok :/ but which channel is it when the bot works 😊

[jovial wraith]

i'm not sure. i've only been active here after the bot got shut down temporarily

[winter pike]

#1047610792226340935 will be green when the bot is available again

[trail lion]

update, still red

[summer shell]

Imagine/

[crimson belfry]

imagine/ photorealistic discord channel ((pitchforks)) demanding SD3.

[nova zodiac]

I don't need to imagine, I just scroll up in here 😛

[buoyant moss]

I thought they were going to release something today

Like Bot channels again 😄

[nova zodiac]

that much hope can kill a man...

[buoyant moss]

I'm only interested in bot cuz of the prompts

I've scraped 300,000 prompts

But MOAR is better

[trail lion]

think of the possibilities, or dont

[pine fiber]

they are soon

[fervent thunder]

😆

[pale latch]

i wonder if sd3 will benefit a lot on Ada cards because it has the transformer engine and it's a transformer model. not sure how that works

[pearl ocean]

SD4 when

[inner lintel]

SD4 will reproduce the real world more accurate than the actual world is looking. 🚀

SD3 soon though

[bright marsh]

max steps by default is 150, what if I put in 10,000 steps? kinda wanna try it lol

[nova zodiac]

9875 wasted steps lol

[bright marsh]

true

can confirm

[nova zodiac]

probably 9960 wasted, depending on sampler.. hell with the new lightning models, that is up to 9996 wasted steps!

[pine fiber]

I know youre right but why is it bad to have that many steps?

just wondering the reason

I never really tried

[nova zodiac]

It's just wasted effort.. the algorithm is just adding a small amount of noise, then removing it over and over again at that point

[alpine hull]

Just wanna know if SD3 beats ideogram 1.0 tbh

[wispy oasis]

yes

[fervent thunder]

probably will

yes since its open source and we can finetune it

[wise bear]

Waiting for SD3 is excruciating, even when it comes out, not sure if everyone on the waiting list would get access to the models immediately, or if the models would be up for download any time soon.

[shell tendon]

Lol so SD3 not happening today obv

[rich kestrel]

im not in a rush to get 1.5 replaced anytime soon

i can wait.

[alpine hull]

I'm not fine tuning it, just hoping the base model beats it

I mean, ideogram 1.0 is really good

It's main problems are hands, faces, and artifacting. Everything else is awesome

[wise bear]

I'm sure in the short term, there may be the requirement of Lora's or new training to undo some of the fine-tuning that SD3 may apply. I'm worried they'll butcher the hell out of it, leaving it down to a mixup or trade off between 1.5 and it's generative and adherence to prompts.

[pale latch]

as far as weights, they're moot. ideogram only sells a service and doesn't provide weights

stability is in the weights game

[alpine hull]

Idm open source or not, I use either way

[pale latch]

far as we know, ideogram is a room full of a million monkeys. we can't be sure

[alpine hull]

Also, I'm hoping SD3 early access is through dreamstudio anyways

[pale latch]

what diff would that make? they'll almost certainly do the initial preview through a discord bot

[alpine hull]

I like web UIs over discord bots

[pale latch]

not really consequential to the model weights thoughaint it

the preview is just going to be the preview. not the actual release. we might be a couple months out from that

[alpine hull]

Just my personal preference

[alpine hull]

I mean... It makes testing easier if I can just switch between tabs and access a web UI without signing into discord on everything I got

[pale latch]

the testing is more for their purposes too. i think they'll be employing rlhf again.

[fallow wren]

how to use sd ? i'm a newbea

[wise bear]

Be interesting to know if the bot would allow private channels and sessions for creation. If we're testing it, the implication is to abuse it and spot for discrepancy. Unless they mean looking for errors in generation, in that case, I'd assume the testing phase would be much quicker.

[wispy oasis]

well do you haev a good enough computer or you want to run it as a service

[fallow wren]

i'm just download some ckpt files , install comfy-ui locally and run default prompt successfully but i'm very confused about the image

[fallow wren]

yeah , my computer configration can run most llms.

[wise bear]

What models are you using, 1.5, XL, 2.1?

[wispy oasis]

how much vram for sdxl is needed again

[fallow wren]

bmp 10Core 30Gpu 36G Ram

[fallow wren]

sorry , i don't know what model i'm using now,the model name is v2-1_768-ema-pruned.safetensors

[wise bear]

VRAM ≠ RAM

[wise bear]

Only generate 512x512, not higher using that model

Or at least that's what someone online said as the output got weird at higher resolutions using that model

[fallow wren]

OK，ehh , I will choose another model best suitable for me but first i should learn how to begin so where can i learn from

i have learned some prompt skills but the sd prompt is so different

[wise bear]

Different to what?

[fallow wren]

in normal case ,we will using Role + detailed requirements + few case + some others to construct a prompt , but in sd i just see much more adjective

[pine fiber]

emad im waiting

[fervent thunder]

[sterile gust]

SD3 will probably be the most ethical AI model ever made

so take a guess

[fallow wren]

Is there any introductory post?

[wise bear]

"Ethical" according to whose subjective morality? >.<

A lot of the times, people assert "ethical" in their "AI safety" papers but have no real argument for particular content they wish to butcher in their models.

Like, names of people, personal identities and various particulars of the sort is certainly one thing, but purely fictional content that can be abstracted from these models have no "ethical" boundaries.

[rich kestrel]

only n00bz care about base models

make it as censored as you want. community will expand its knowledge on anatomy farther than you could ever imagine

[sterile gust]

when a model is heavily censored there's hardly a way to fix it.

just look at 2.1

[rich kestrel]

well if it trains like ass cough cough XL

2.1 had a crapload of issues with text encoder

we need a model that is as smooth to train as 1.5

[sterile gust]

and i have a feeling SD3 won't be any different or maybe even worst

[rich kestrel]

until then, its all just recycled XL

yup. wasnt born yesterday. thts why I smh when I see ppl hyped

cant see how this will be any different

thankfully 1.5 kicks ass. I can be as patient as they want

[hallow sable]

what people still make nudes in sdxl ...

[sterile gust]

we won't have another great influx of users or hit of relevance unless a propietary model leaks again.

there's no way an open model does that on it's own

[fervent thunder]

i have hope, stability learned from the previous releases

[cedar salmon]

open or closed model has zero bearing on its performance

[sterile gust]

no they haven't, why can midjourney and other propietary model makers can train on anything but it's just SD the one that has to remove artist data and have a way to opt out?

[trim magnet]

depends,if its a closed wallet model vs an open wallet model

[sterile gust]

doesn it make sense that the corporate models are the ones that have the liberty of manupulating any data while the open source one is the one that has to comply?

[cedar salmon]

i prefer SAI than the models that while damn good seem to produce their distinctive look that you are stuck with

[winter pike]

it will prob be as censored as SDXL and its okay, those custom sd1.5 model will not go away

[fervent thunder]

by performance you mean aesthetics? closed models will probably always be better than open cause closed models are backed by capital

[sterile gust]

resignation is not a solution

[cedar salmon]

time will tell, money shifting, some of the winners now will be the first to fail

[karmic cedar]

an overly censored model is a lobotomized model

When it comes to model training and construction—unless they are re-training the model without the questioned data, then trying to reduce its influence based on an arbitrary statistical threshold is only going to weaken the overall performance little by little

IMO

[winter pike]

ill judge when its out, and well see if those removed artists really make a difference or not

[karmic cedar]

for sure

[pine fiber]

Definitely will

it looked half baked

[radiant tusk]

does animatediff always generate super blurry gifs?

[sterile gust]

Why do you need to wait to know it's censored? We already know it is.

[trim magnet]

if u use a low resolution,yes

[radiant tusk]

what resolution do you recommend, i've been using 512x512, does hi-res fix work?

[trim magnet]

yes u have to use high res fix, 1.5x or 2x depending if you have the vram for it

[radiant tusk]

Would you use hi-res 2x for 1024x1024 or no hi res and generate at 1024x1024?

[trim magnet]

512x512 then hi-res 2x or 512x768 then hi-res 2x

[winter pike]

"it will prob be as censored as SDXL and its okay"

the thing is that a lot of yall think this will be another sd2.0, judge when its out

[pale latch]

people still to this day say that sdxl is censored. the unstable diffusion community is filled to the tits with misinformation. anything to keep the donations flowing in

even stable diffusion 2.1 the censorship was fully overblown. people just didn't know how to prompt for it

[trail lion]

well there are things the base model cant produce sure, but where I'm struggling to understand is the notion that 1.5 offers something that sdxl doesnt, that makes no sense. what exactly cant be trained into sdxl?

[trim magnet]

if u have money u can train anything to it

[trail lion]

yes, this

money, or will, or time

or the desire to learn how to train in the first place, not just think that prompt enigneering is the beginning and the end

[pale latch]

2's biggest problem was that the architecture was locked in and difficult to train. i don't think that was intentional

[radiant tusk]

What's the quality differences of sdxl vs hi-res sd1.5?

haven't really seen big noticeable diffs

[fervent thunder]

only reason i like sd1.5 is the speed, helps me visualize a prompt way quicker. I wish sd3 supports smaller resolutions for faster inference

[trim magnet]

for animatediff, 1.5 works better the sdxl motion model is not very good

[pale latch]

afaik, stable diffusion 3 will have different versions.. one with an 800m parameter count

same as 1.5

[trail lion]

ok, maybe, but after a couple 5 sec clips of flickering video, that gets old and back to generating something worthwhile

[fervent thunder]

do we know if the 800m model supports 1mp? or it's only 512x512?

[shell tendon]

Advantages in coherent composition

[pale latch]

dont know. i imagine a transformers based model is going to have a wider resolution operating space too

[trail lion]

will can only eat so much spaghetti

[pale latch]

on 1.5? hard press on x to doubt that one

[fervent thunder]

hi res sd1.5 is slower than sdxl normal res

[trim magnet]

no

[shell tendon]

sdxl has advantages in composition, yes

[pale latch]

a lot of people sleep on prompting the seperate clip layers of xl. the openclip model is exceptional at catching prompts

[pale latch]

ah misunderstood you thinking you meant the reverse

[trim magnet]

doesnt really matter when only thing u create are portraits or asian girls

[pale latch]

heh

[fervent thunder]

i'm creating artistic porn so for me it does matter

[trail lion]

I'm not sure those 2 words were ever meant to be used together

[trim magnet]

they were if your name is Sarah Lucas

[wise bear]

Well, it is significantly more coherent than older diffusive keyword based models, so any ability to train easily makes for a significant hype around it. I do also hope it can train easily, but how butchered the model becomes before we get it is definitely a worry. The hype is more so the hope that we'll be able to easily train with it.

[sterile gust]

i don't expect anything, in fact im now hoping the opposite, maybe the model is bad and is the worst of them all, maybe that'll be the only way i could ever find SD3 as an advancement.

[trim magnet]

well if u create the hard type of stuff its better for u to stay with 1.5

[radiant tusk]

Ahh, is there a dreambooth for sdxl? Or would you recommend just training a lora

[trail lion]

lora is dreambooth

but yes you can do full checkpoint dreambooth training or lora training via kohya scripts

[rich kestrel]

lora =/= dreambooth

[shell tendon]

Depends on your patience and resources

Lora is a lot faster and a lot smaller

If you want lots of models the disk space can add up

Granted that's cheau as shit now compared to a 4090 so that shouldn't matter too much if you got one

[hallow sable]

wtf is a cheau

[trail lion]

it's literally called dreambooth lora, it's more or less a tweak to reduce the trainable paramaters and the resulting file size

whatever, it's not dreambooth 😄

[rich kestrel]

thats because engrish from japanese transration

dreambooth on kohya is a lousy choice anyways, at least for 1.5

[radiant tusk]

which one is higher quality and closer to the trained faces?

[trail lion]

most say the full checkpoint training is, and you can also extract the lora from a full checkpoint, if you are ok with generating a bunch of 6G files until you are satisfied

personally, I'm always content with a lora, but seeing is believing

[shell tendon]

Celestial Hovering Entertainment and Utility device

[shell tendon]

Finetine yeah

I've gotten really good results training a Lora and a lycoris-ia3 on the same dataset

Then combining them

[hallow sable]

hopeful SD3 region locks china...

I mean western taiwain

[trail lion]

btw, the fact that you can extract the lora from a full dreambooth checkpoint is more evidence that it's in fact dreambooth, but I digress

[trim magnet]

hope SD3 includes a lactose intolerant mode

[radiant tusk]

finetine?

[pale latch]

It's not. It's just creating a lora from the difference of two models. Turns out that's really effective at separating what dream booth changed. Thing is, they're all just doing math to floating point weights

[radiant tusk]

So training both a lora and a lycoris on the same dataset then combining them outdos the quality of a dreambooth checkpoint in your experience?

[pale latch]

So btw, your biases are showing

[trail lion]

we all have biases

[shell tendon]

Yes, like a sharp antler

[pale latch]

Full checkpoints are typically the domain of finetunes on large datasets these days. I don't think many dreamboothed checkpoints are doing too well. Loras make more sense now a days

Extract a lora from juggernaut XL and it won't be nearly as good, since it changed a lot

[acoustic solstice]

hey guys i am trying to get stable diffusion running on a local VM. my problem is that i can't seem to figure out how to get a pass through GPU working on vmware workstation 17. is virtual box better?

[true ermine]

VirtualBox does not support GPU passthrough

[acoustic solstice]

what are my options for VM software then?

[true ermine]

Unfortunately that's all I know

[trim magnet]

im gonna train a lora to region lock ppl

[shell tendon]

Are you just trying to run it on Linux on a windows box? If so. Wsl2

[acoustic solstice]

yes, just trying to run some ML stuff on linux VM

on windows host

[young hatch]

@acoustic solstice Nvidia Toolkit with WSL2 running on an nvidia GPU. If you do "WSL--install" and let that complete, it will install the latest Ubuntu env on your macine... THEN install/reinstall Docker Desktop. THen you will have Nvidia Toolkit for a Passthrough to containers... --- you can run BASH from Powershell, or be in the Ubuntu environment. You can then run "wsl --install any time to get into ubuntu... run nvidia-smi and it shows your you GPU passthrough system : https://i.imgur.com/UXXt3lg.png -- Watch this vid: https://www.youtube.com/watch?v=PB7zM3JrgkI

[radiant tusk]

If you use animatediff on forge, does the length of the gif determine how much vram it uses? or is it resolution

Im using hi-res fix on a 512x768 video with 400 frames , says I have 14gb vram left but it just hard lags my PC and it becomes unusable

[acoustic solstice]

@young hatchawesome, thank you!!

[trim magnet]

too many frames

[radiant tusk]

Ahh, I will try 300frames to see if it keeps lagging

[trim magnet]

still too many,u can probably push it to like 120 frames at like 8fps or 12fps

depends on your gpu

[radiant tusk]

without hi res

I was able to do 500 512x768

[pearl ocean]

lol

[radiant tusk]

at 30fps

Does hi-res add that much vram usage?

[trim magnet]

no its just that animatediff wasnt meant to do that many fps default fps is 6 and u can prob push it to 12

[radiant tusk]

ahh

Im using controlnet as well

does that make a difference?

[trim magnet]

yes,heavier with controlnet

[radiant tusk]

ahh

[pearl ocean]

Sora when

[radiant tusk]

How do you make the generations less snappy? Like random things disappearing and appearingin the background. Are there any settings that fix that

[trim magnet]

thats just normal u either remove bg or you edit each frame with something like photoshop

[trim magnet]

but if u are doing a vid2vid workflow u can do it at like 15fps + controlnet,it depends on your gpu,whats yours?

[radiant tusk]

4090

[trim magnet]

then yes u can do it

try with 16fps

[radiant tusk]

for 60 fps videos

Do I just do 15 fps but 4 interp frames? to make it smoother

[trim magnet]

yea play with interpolation or just gen a 15fps vid then do interpolation in another program

[radiant tusk]

Okay I will try that

If I want to generate images with text in them, are there specific loras that can do it better?

A lot of my text end up wonky or with missing/extra letters

[trim magnet]

yea harrlogos lora for XL

[radiant tusk]

Can I mix that with anything else?

to be able to get the text first try

or is there a way to prompt it

[trim magnet]

yea u can prompt for whatever text u want

[pearl ocean]

I got the 4080

[radiant tusk]

When using controlnet on a batch images, how comes it preproceses both before and after the video?

[trim magnet]

sorry idk i dont use animatediff vid2vid

[fervent thunder]

cowroll

[pale latch]

i thought the fps setting was just the playback speed of the file. if i put 15 fps with interpolate of 4, it plays ultra slowmo

[radiant tusk]

I get that issue too

interpolate is supposed to raise the fps so 15fps with inter 4 supposed to be 60fps

but I also get just 15fps but .25 speed

[pale latch]

it just adds frames

[radiant tusk]

yeah on the animatediff doc

it says it adjusts your fps

so added frames don't slow down your video

[pale latch]

its the context size that the model cares about. 16 context batch. the fps i'm certain is just for the post process when stitching the file

[radiant tusk]

if I change context batch does it do anything

[pale latch]

thats the batch size i'm pretty sure

[radiant tusk]

hmm so if I did 32 it would need more vram but would generate faster?

[pearl ocean]

I made some really nice looking dragons in #🌠｜show-and-tell

[mighty cipher]

For those interested in the academic discussion on generative AI in general, and in this case particularly in visual generative, Alan Warburton (https://alanwarburton.co.uk/) is hosting a webinar in a couple of hours as part of an AI webinar series at Umea university (https://www.umu.se/en/events/fraiday-the-wizard-of-ai_11886311/)

[nova zodiac]

is it being recorded, thatll be 1am here 😦

[swift juniper]

does anyone know any method or extension that would allow me to inpaint in high resolution (>1024x1024) without leading to distortions? like something similar to hi res fix but for inpainting

[warm junco]

It depends on the fps at which you decode the images back with ffmpeg

[warm junco]

For that you need to use an sdxl model

Its trained on 1024

[swift juniper]

right

[swift juniper]

is it possible to merge 1.5 models with sdxl models and get decent results?

[nova zodiac]

not a thing

you can use the result of a 1.5 model, and then inpaint that with a sdxl model sure, but mixing the models themselves is not a thing

[swift juniper]

rip

my sister in law has asked me to edit some of her wedding photos to hide her baby bump but theyre all in 4k

any good sdxl models you can recommend to do that lol

[pulsar marsh]

ay what WebUI are we all using?

[still glacier]

the one that suits your needs. You re welcome

[pulsar marsh]

[still glacier]

mostly it s a1111 (or forge), or comfyUI or fooocus.

Depends of what you want to achieve

[nova zodiac]

Is she pregnant of her own free will?

[swift juniper]

lol ofc its my brothers kid

[pulsar marsh]

I was fishing for new ones to try because my A1111 bricked and I can't make it work anymore

[nova zodiac]

Then why try hide the bump, pregnancy is a beautiful part of the human experience..

[still glacier]

might want to check with #🤝｜tech-support to fix it then

[nova zodiac]

people are weird...

me especially lol

[swift juniper]

ig she wants some wedding photos where she looks slim lol

[pulsar marsh]

Nah I tried that already, and we got it working briefly but it went back to not working

Online generators are fine, but it's limiting after using a local SD Webui for so long

[nova zodiac]

what a1111 extensions did you have?

[pulsar marsh]

None

Pre-installed ones

[nova zodiac]

in that case try installing forge from scratch in a different folder (its a1111 but actually fast)

[pulsar marsh]

tyty

[nova zodiac]

once forge is installed, then copy your model files over

[pulsar marsh]

I'll have a look

[jovial wraith]

do note that not all extensions work in forge

[pulsar marsh]

That's alright, I don't really use any

The CivitAI one for uploading images and stuff

And that's about it

[still glacier]

Personally I also don't notice any significant speed improvement when switching to forge :p. Not sure why so far.

[jovial wraith]

what gpu are you using?

[still glacier]

2070s, so I should get "30~45%" inference speed improvement according to them

when testing my speed when from 8.9 to maybe 9.2 it/s on some simple 1.5 stuff, no lora, no lcm, no controlnet, etc

[jovial wraith]

yeah, i think they padded the numbers a bit, since i'm getting about that much improved speed on my 1060 3gb

[still glacier]

I did notice it used about 500mb less vram tho

checking back my notes from a few days ago

=================================================================
sdwebui 1.7 --xformers
Total progress: 100%|██████████████████████████████████████████████████████████████| 60/60 [00:06<00:00,  8.91it/s] 
=================================================================
sdwebui 1.7 (dev branch) --xformers
Total progress: 100%|████████████████████████████████████████████████████████████| 240/240 [00:27<00:00,  9.14it/s]
=================================================================
sdwebui forge
Total progress: 100%|████████████████████████████████████████████████████████████| 240/240 [00:26<00:00,  9.36it/s]
=================================================================
sdwebui forge --xformers
Total progress: 100%|████████████████████████████████████████████████████████████| 240/240 [00:26<00:00,  9.74it/s]

But many people do report speed improvements. So it s worth a try and it s probably a "me" problem.

[wispy oasis]

Hey everyone

does anyone have a 4070 graphics card who uses sdxl

[nova zodiac]

ive had good luck with it then.. had a 1070 that gets 1.3it/s go to 2.2it/s with forge, the rtx5000 on paperspace went from 7.7 to 13, and the a4000 went from 9 to 14.8

[swift juniper]

i also find forge seems to load models faster

[nova zodiac]

all tests with a 1.5 checkpoint, 50 step euler a at 512x512 with no prompt and 7 cfg

[swift juniper]

although i havent tested

[still glacier]

I d recommend setting higher steps value. It gives more time to SD to stretch its inference legs properly for this test.

[wispy oasis]

i would like to know aprox how many images per minute or seconds per image it will be for sdxl @ 1024x1024 with a 4070. benchamrk i found only shows old version of stable

can i also use sdxl to generate 512x512 at faster speeds? ty

[still glacier]

you can check the sdnext benchmark database for that https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

[nova zodiac]

i run batch count 4 and take the speed of the 4th image

[still glacier]

it s not perfect but it should give you an idea of what to expect

[nova zodiac]

that way it has "warmed up" so to speak

waits for mod....

[wispy oasis]

dont see any using sdxl model

[still glacier]

hmmm yeah but you re still gonna hit the VAE stage pretty early. Which could mess the numbers. It s like you re giving SD an hiit workout instead of a long run. Not sure if I m clear on this one.

[nova zodiac]

speed difference between 1024x1024 on a 1.5 model and 1024x1024 on a sdxl should be neglegible

[wispy oasis]

ty

[nova zodiac]

I did try 100 steps initially, but found 50 gave pretty much the same result plus or minus .1 it/s

[wispy oasis]

can sdxl generate 512z512 or only resize after

[still glacier]

Ok then.

You did your homework

[nova zodiac]

sdxl turbo is designed for 512x512 - full sdxl models are designed for 1024x1024

[wispy oasis]

ok ty

[nova zodiac]

copied off toms hardware, but without access to any cool gpus 😛

[swift juniper]

figured out i can just inpaint in 512 then go over the inpainted bits in segments with a low denoising strength to get better quality

[static cape]

I hate it when industry people post dates and then can't meet their deadlines. Thanks for generating hype and making us wait, Emad!

(to clarify - no date and then have it drop in our laps 1-2 weeks from now would've been preferable... afterwards words gets around fast)

[lone hawk]

we are in a such a bad situation, there's no point using SDXL anymore and we don't have SD3 yet 😄

Good god i'm so tired fighting that censor dog in Dall-E, just... just kill me please =((

[static cape]

We have Cascade... but I don't feel like trying to wrangle that strange beast when we have a potentially way better Model with different architecture around the corner.

[lone hawk]

Yeah, Cascade is dead on arrival since SD3 "soon" 🙂

[cloud plume]

Wdym? It's been less than a week, and iirc they said at least 2-4 weeks...

[trim nymph]

think they refer to user generating, inside discord

[honest spear]

I just hope SD3 will have some controlnet or ip adapter weights released together with it, I don't care if they take one more month to release it

[static cape]

It was a really strange way to go about things... I mean the SD3 announcement came like 1-2 weeks after Cascade? (and tbh Cascade is like a talented, but strange alien cousin of SDXL)

[static cape]

There was an Emad Tweet hinting at yesterday... and then yesterday he was like "oops, lol, not there yet."

[static cape]

That would be amazing, but I don't think so.
I'm absolutely looking forward to see if we will be able to properly fine-tune it on consumer hardware though.

[charred mesa]

im interested in what the multi-modal aspect is for

cause with LLMs that usually means it is able to use images

maybe the technical paper could explain a lot more to us

[static cape]

Probably language/image/ maybe motion

[unkempt glacier]

https://tenor.com/bGoN5.gif

[gaunt pulsar]

Did anyone get called from the wait-list?

[swift egret]

Still waiting for XL based inpainting models. I hope SD3 will have one.

[shell tendon]

There's a bunch now

[swift egret]

I only found controlnet based or similar hacks that only works is you set inpainting resolution to same as the image. Well I do inpainting in 6k renders professionally and need a proper inpainting safetensor model for a1111.

[shell tendon]

There's a sdxl base one, a juggernaut one, a realvis one, I can't believe it's not photography one etc

[swift egret]

Link?

[shell tendon]

https://civitai.com/models/245423?modelVersionId=288402

Look around on civitai the rest are typically on the same page as the base model

[pine fiber]

where mah sd3

[swift egret]

"not yet fully implemented"

[shell tendon]

https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1

[swift egret]

I mean, read the description. It's not a real inpainting model.

[trail lion]

Only for automatic1111, sounds like it's potentially working on a number of others such as comfy,foocus. It also says it's achieving results similar to 1.5 inpaint. Worth checking out imo if you really need it

[trim magnet]

no one inpaints XL with comfy

[cedar quiver]

this is a bit of a dumb question but does anyone know of an a1111 extension to make notes somewhere in the UI
even if it's just the hackiest thing possible

i keep forgetting the valid resolutions for this SDXL model because they're specific strings of numbers i literally never use otherwise and i don't like having a separate notepad/obsidian file for it

[fervent thunder]

How much do negative prompts like disfigured body, two heads, extra fingers etc. help?

I am pretty much getting the exact same number of weird proportion images as without them

[trim magnet]

i think theres an extension for a111 that has a bunch of default resolutions

[fervent thunder]

I tried copying the list from pins in #📝｜prompting-help

is there such a thing as too many negative prompts that result in worse results

[cedar quiver]

perhaps aspect ratio helper will help

[trail lion]

I think if you use the tabs vs the drop down at the top to select the model, you can input some metadata, like preview images etc, can't remember all of the choices

[trim magnet]

yea i think that was the name,there were 2, one was abandoned and the other one

[cedar quiver]

yeah like the specific resolutions are
896 x 1152
832 x 1216
768 x 1344
640 x 1536

i usually shoot for the second one (1024x1024 is also an option so 832 is technically 'middle' option)

but there's nothing like, cementing those numbers in my head so i have to check my notes

i'll probably get used to it
i should probably just set my default res to be 832 x 1216 since i almost exclusively run this sdxl model

[alpine hull]

Surprised that people here really do think SD3 will surpass Ideogram 1.0

You'd think the open source model wouldn't surpass closed source because of it being artificially limited

[pine fiber]

you mean the other way around ?

How is an open source model artificially limited

[alpine hull]

Because

I mean, look at Gemma compared to Gemini

Or well, literally all of open source LLMs

Open source LLMs barely surpass gpt 3.5, let alone even get get NEAR to gpt 4

[pine fiber]

you’re assuming ideogram has the same amount of compute and funding than openai and google

what a stupid argument

[alpine hull]

Ideogram 1.0 is really good tho

I tried it, holds itself up well to the SD3 demo images

[pine fiber]

yeah and you can compare it to the early preview pics of sd3 and conclude sd3 is slightly better

even before rlhf

[alpine hull]

??? I tried it myself, I prefer Ideogram's images to the SD3 demo pics with the same prompts more than half of the time

Won't say it's objectively better

But damn, it's not bad

[pine fiber]

they are equal at best but I wouldn’t say it’s better

it’s also crippled by being closed source

censored, will be slow to implement in painting controlnet etc

open source is just much faster at this stuff

see what they did with 1.5

[alpine hull]

Well, I don't really care about half of that stuff

Sure, more control is better

But all I need is text to image, inpainting, and image/image and text to image

BUT

[pine fiber]

I don’t think they have all of that yet

[alpine hull]

my biggest hope is

SD3 is super good with text

[shell tendon]

Ideogram does a lot of composition that looks like a collage approach

[alpine hull]

My biggest problems with Ideogram 1.0 are problems they themselves solved before...

Hands, faces, and artifacting

So weird