💬｜general-chat

[vale steeple]

SDXL sure

[rugged mirage]

Im sure anime stuff will be great and will get the most finetunes, it's like the style you should be least worried about

[vale steeple]

I will say anime is not even a challenging part in generation stuff.

[sudden ruin]

I mean SD3 tho

[formal swan]

I need to make friends in this community, so I can bug them about issues I run into, that have probably already been solved by others.

°^°

[sudden ruin]

use the search function in discord

[formal swan]

:O You expect me to know how to do that?

[loud solar]

btw for A1111 and Forge there's a Civit.AI Extension I really like. Downloading Models,Picture, Model Information ... "CIVITAIHELPER"

[formal swan]

Oh? I have been just manually going to civiai

Can you dm that to me so when I sleeb, I can see it still.

[loud solar]

You search in your Forge or A1111 @ Extensions

[formal swan]

Ahh, like I had to do with controlnet

I still couldn't get the 3d model tab to appear.

Which irked me so much.

[loud solar]

ControlNet is already integrated in Forge ... but the same you did this in A1111, yes

[formal swan]

Oh, you've got me so excited for forge now.

Am I gonna have to transfer all my saved prompts over?

[loud solar]

And ... you can use your downloaded models in every UI don't need to download again. Youtube will help with that ...

https://www.youtube.com/watch?v=q5MgWzZdq9s&t=20s gave me some nice information

(Maybe not for that problem)

[sudden ruin]

Youll have to transfer everything if you plan to remove A1111 but its worth

Forge is king right now

At least compared to A1111

[loud solar]

Running A1111 and Forge here ...

no need to remove

[sudden ruin]

I dont see the need for A1111 when you have forge but you do you

[formal swan]

I don't have unlimited space, so if I don't need them both I'll just switch over.

As long as I'm not missing out on anything, then I'll be fine.

[loud solar]

Sometimes an update breaks a system, so I keep both 🙂

[pearl ocean]

Forge ftw

[formal swan]

Hearing that scares me.

[loud solar]

It's just that other parts need updates, too. So you have 1-3 days to play around with your other A.I. system 🙂

[formal swan]

I'll keep both 😭

I am terrible with writing prompts, so I need as much practice as possible. I wish I had a cheat sheet of useful prompts, and what order to place things for the best results, but I know it differs from model to model.

[loud solar]

It's like every other language. The more you practise, the more you learn. If you wanna try out things, it might be an idea to try with a fix seed to see the difference.

Fix seed ...

And try out things I think about small changes in one prompt ... I'm on the run to work, sorry 🙂

[pearl ocean]

no need to cry about it

🤔

[formal swan]

Have fun at work!

[pearl ocean]

how far is the run?

[loud solar]

Marathon ... 🙂

[true canopy]

anyone tried ella?

[rugged mirage]

there were examples from yday in #🏞｜general-with-images

I set it up but forgot I dont even have 1.5 models anymore when I tried to run it, so left it for today, will probably test in a bit

tho having to use 1.5 in 2024 kind of yuck

[pearl ocean]

No, don’t even know her

[true canopy]

ella somehow cant do anime for some reason, even tho they have anime examples in their paper

[outer warren]

:D

[narrow kernel]

What kind of "anime" are you trying to do?

[true canopy]

any checkpoint thats anime is not working

it completly butches the images

[warm junco]

some extensions just dont work on forge, so having both is better

[narrow kernel]

No I mean what are you trying to make it do with the anime

[true canopy]

i dont get ur question im sorry, just normal prompting, its the quality that is ruined

[pearl ocean]

I hear some people who have the Pro Max iPhone say that they don’t need an iPad coz the phone is big enough.

[narrow kernel]

Did you follow their directions exactly? Also, how exactly are you using it like are you using their GitHub stuff or some kind of implementation within another app like comfyui?

[true canopy]

im using in thru comfyui

and someone else on reddit are getting same ruined quality as i am getting

[drifting knot]

Sup guys! Maybe some of you know a good guide of how to use EMBEDDING for SD 1.5 model in code by using diffusers?

[narrow kernel]

What exact model are you using it with?

[oblique elk]

This is done with Ella #🏞｜general-with-images message

The quality is indeed not as good as the images generated without

[true canopy]

ive tried several, i even tried the exact one they use in their research papers, witch is counterfiet v3

here are the examples anotehr user posted

https://imgur.com/a/UMQhBhy

im not sure why ella wrecks the generation on anime

but its unusable

[narrow kernel]

what GPU are you using? it might be autocasting to some other precision level or something

[true canopy]

im using a 4060 ti 16gb

feel feel to try it out urself, if u have the time, id love to get it to work

[pearl ocean]

I’m just using a yi old 4080

[narrow kernel]

looking at their github code, looks like it has a default cfg of 10, uses stable-diffusion-v1-5 and DPMSolverMultistepScheduler for scheduling

[true canopy]

i tried 10 as well, no difference

some checkpoints was way worse then others as well

[narrow kernel]

give me a few, i'll test it out some. i don't really mess with 1.5 models anymore, so i hadn't bothered to install it yet

[bleak matrix]

Good morning, everyone! How are we all this morning?

[pearl ocean]

I like XL

[narrow kernel]

yeah i pretty much exclusively use sdxl based models or cascade

[pearl ocean]

Bass model keeps having deformed legs, arms, body, and multiple body@pars everywhere

[true canopy]

1.5 can be superior in some cases it seems, especially when it comes to anime

and even animating it seems

[narrow kernel]

there are a few decent overly trained 1.5 models that can put out decent stuff, but they are so damn biased toward pretty waifus its not even funny

[true canopy]

but its so much i have not yet had time to try out

[sudden ruin]

Problem with XL are the not so good controlNets

[narrow kernel]

the control nets work plenty fine and i use the hell out of them. you just need to learn how to make proper depth maps or canny maps

but yeah so far, this ella is god aweful for anime stuff lol

[sudden ruin]

Nah they degrade the image quality a lot

[narrow kernel]

i'll have to check the author's wrapper, he might have bad code in it or something

[sudden ruin]

Especially openpose

[pearl ocean]

I don’t use control nets, what do they do?

[narrow kernel]

they do not, you're just using the wrong. the strengths are misleading because it's actually based on steps under the hood. i don't really use openpose when it's easy enough to make a quality depth map with zbrush/blender/ue5/etc

a good depth map>>>>>>>>>>>openpose

[trim magnet]

when making videos u need to use openpose whos gonna waste time making a depth map for a video with hundreds of frames lol

plus controlnet is more than depth and canny,theres shuffle,softedge,instructpix2pix tile etc

most of these are not on XL and the cn tile sucks

[narrow kernel]

you can render out sequences with 3d packages, it's fast. stop using tiktok dance videos lol

[trim magnet]

why would i

i can just do that with openpose 1.5 faster then upscale frames with XL

[narrow kernel]

you do you

but for people to claim that controlnets are "bad" in SDXL is a lie

[trim magnet]

they are not bad they are just incomplete

[narrow kernel]

the main ones that the vast majority of people use work perfectly fine... depth and canny.

the vast majority of people aren't making videos

[trim magnet]

yea they are fine

the 1.5 are better but 1.5 models are less precise so nothing is perfect

[narrow kernel]

and another issue people have with controlnets and sdxl are that they carry over a bunch of bad 1.5 habits like having a million things in the negative prompt

or poorly wording the prompt

[rugged mirage]

it's not a lie, they are not even the same as 1.5's (and youd hope they'd be better) but worse

they are still usable but they are a step backwards when you expect a step forwards

[static cape]

Thank you. I always say that and also always get the "nah, you're just doing it wrong" response...

[verbal oriole]

Can someone recommend me a model best for generating Pokémon that fits in 8GB of VRAM?

[static cape]

Nintendo wants to know your location (also probably Cascade).

[verbal oriole]

I’m not using a Web UI as I’m building my own Pokémon generator thingy

shhh

It’s gonna be totally unique too lol

‘Totally unique’, to an extent lol

My current model gives me blurry and obviously AI generated stuff (runwayml/stable-diffusion-v1-5)

[nocturne wharf]

Good afternoon! i would like to make an ai female character that seems life real. I also want to get into pornography here. I'm still new to this and have no idea where to start. Who wants and can help me?

[verbal oriole]

bruh.jpg

[verbal oriole]

I’ve also tried https://huggingface.co/lambdalabs/sd-pokemon-diffusers and got the same poor result.

[nocturne wharf]

🤣

[sudden ruin]

They most definitely arent perfect, I had a lot more consistent looks with the openpose model from thibaud

As well as using it together with highresfix

That exterminates most issues for me

[tame willow]

Hi! Can you tell me how to set up the display of all information under the generated image? Where are these settings stored?

[thin oyster]

Hello, if you need someone for mixing, mastering or effects for your instrumental does not exist to contact me

[sudden ruin]

Usually they are stored in the meta data, you can drag and drop images into the PNG Info tab in A1111/Forge or get the full workflow from ComfyUI images by dragging and dropping them into your comfyUI

[lime sorrel]

Hi everyone, does anyone know how to load a local DreamBooth adapter (not from HF) to a pipeline? Not using A1111

[charred mesa]

finally, weekend is coming

almost 1 week closer to SD3

I wish the waitlist beta test would come next week or something

https://github.com/DataCTE/ELLA_Training

Dear... God...

[trail lion]

people keep talking about ella, why? Is it better than kohya or something?

[sudden ruin]

Ella isnt like kohya at all, kohya is for training. ella equipps SD models with LLMs for better prompt understanding and improved visuals

[trail lion]

ah ok

that github page mentioned training

[sudden ruin]

Yeah that git is about reverse engineering ELLAs code to make a SDXL version

[ivory shale]

[loud solar]

Reverse engineering ... they did with humans a long time ago ^^

[frozen helm]

[frozen helm]

Aboutta do it w u

imjk

[hybrid ivy]

What's the best UI these days? Let's say that I'm programming in python

I like the UI for rapid testing of prompts, and then I like to run a larger job through python

[loud solar]

Quick'n'easy sounds like SD Forge to me. Your abilities more sound like ComfyUI could be fun ...

[hybrid ivy]

Oh yeah, that's right I heard about ComfyUI

I took a break for a few months and things keep changing, so felt it was worth checking if there was new hotness

[loud solar]

Forge is the new A1111 now ... comfy still the same 🙂

[hybrid ivy]

Yeah A1111 had some annoying things Ididn't like (such as the weird weird notation with [1:2] and [1.0:2.0] having completely different behavior)

depending on whether it was decimals or integers, it completely changed the meaning of what you were doing lol

[loud solar]

Forge is a Fork of A1111, hnadling some stuff a bit cleverer, a bit faster, and in my eyes more complete, cause it comes with many useful extensions

[hybrid ivy]

class Character:
  def __init__(self, traits):
    self.traits = {}
    self.traits['age'] = Age(traits["age"])
    self.traits['gender'] = Gender(traits["gender"])

Yeah I was working on basically some code like this, so that you could select the kind of character you wanted, and it has predefined way to construct the prompt.

[loud solar]

Bad luck I'm not good in understanding code. I still have some basics cause of Borland Pascal ... but never used it ^^

[hybrid ivy]

Nah I don't need help with code, just show and tell

My idea was like: say you wanted to make a picture of a 16 year old, it has code that automatically adds the positive words Teen and negative words middle aged

so instead of choosing the words, you just pick the age you want and it makes the right prompt for you

[loud solar]

Well with SD XL many don't really work with neg any longer ... like like in SD1.5 times...

[sage reef]

i rarely if ever use negatives with finetuned sdxl models

[hybrid ivy]

Yeah XL just came out when I paused my work

[loud solar]

SD 3 could change even more ... so you may be a bit late for that idea

[hybrid ivy]

yeah next step for me is just getting XL to run in comfyUI I think. Not needing negative prompts would be pretty great

[sage reef]

comfyui is actually very easy, not sure why people think it's some black magic

[sudden ruin]

Yeah it barely needs negative prompts, I only ran into a few issues where I depended on them

[hybrid ivy]

I remember the other big problem was getting consistent colors (such as "red shirt, yellow pants" instead of getting the colors swapped, or ignoring one color etc)

I had ideas on how to get around that, but I think XL doesn't have that problem nearly as much

[loud solar]

That problem still exist sometimes ...

[sudden ruin]

There is an extension to solve the color issue tho

Whats the name uhm

[hybrid ivy]

Eye color was very annoying lol

[sudden ruin]

https://github.com/hnmr293/sd-webui-cutoff

This

[sage reef]

the color problem can be solved most of the time if you put more emphasis on the color and the object, like (red sneakers:1.2) and (orange shirt:1.2) or something, but yea it's not perfect, maybe SD3 will solve all that

[sudden ruin]

That extension pretty much solves it

[sage reef]

oh i never saw that one 😮

ah but that is not for comfy right?

[hybrid ivy]

I remember trying it for A1111 and it didn't work that great iirc

but I forget

[sage reef]

at the end of the day, we are all waiting for SD3 to solve all of our problems 🙂

[hybrid ivy]

yeah!

[loud solar]

🤣

[heavy lark]

I've been playing around with Ella a lot and if sd3 can do that, we should be in good shape.

[loud solar]

I hope it can wake me up with my good morning coffee 😄

[sage reef]

well i saw their technical report or paper and it does do that it seems, and a lot more, so i am excited

[loud solar]

SDXL wasn't really good first ...

[sage reef]

btw, what's currently the best approach to latent couple or regional prompting with sdxl in comfy? i saw this civitai page, it seems decent, but wondering if it's outdated or if there is some better methods: https://civitai.com/models/21100?modelVersionId=94130

[charred mesa]

yeah SD3 will also do a wider range of stuff like celebrities. ELLA refuses to generate some (if not most of them? haven't tested)

It knows Brad Pitt as a random black man

[sage reef]

hahah yea ella was giving me some weird stuff too

[charred mesa]

but now that ELLA training is possible, it's going to change the game

imagine uncensored trains of ELLA

do you guys this could make SD3 DOA (for AI corn addicts)

[heavy lark]

Isn't T5 a Google creation? They're into making black nazis

[sage reef]

i think by the time people train ella, sd3 will be released, so yea....

[charred mesa]

hehe I get you but I think it's just ELLA's dataset not being wide enough

or maybe its just censored, idk

[sage reef]

the ella we got is 1.5, not sdxl, so no wonder it's bad

[charred mesa]

idk SD3 is a superior pretrain with a superior VAE so I would personally try to train AI corn (if I were to be into that lol)

[heavy lark]

Indeed. I tried throwing prompts at the new cosxl. It was horrible. Shows how far the fine tunes have come

[charred mesa]

CosXL sucked for some reason

like there's a CFG threshold or something where above that everything becomse a weird abstract art

[sage reef]

i mean cosxl might be bad, but it is cool tech to edit images with instructions, the edit version of sd3 gonna be awesome

[charred mesa]

even with merges its bad for me

and it literally looks like someone took up the contrast effect

[loud solar]

Is it that blurry thing?

[charred mesa]

I bet there's some mathematical explanation behind how it's like superior but I don't know

it just makes stuff look worse for me

[heavy lark]

I wonder if that merged data is still bad though because it's based on the old model. It probably has to be retrained with the source imagery.

[charred mesa]

haven't tried it extensively but I suppose its a sneak peak at what 2B edit and other models could do

[sage reef]

there is a way to maybe fix cosxl outputs, you generate first with cosxl, then you apply a very light img2img with some sdxl model to fix the output along with upscalers, or whatever

[charred mesa]

idk I just hope SD3 will have bokeh optional somehow

[sage reef]

but yea sd3 hopefully will do better

[charred mesa]

like you can do a negative prompt or type these into the positive prompt: "deep depth of field" or "narrow aperture"

I hope that will work

[sage reef]

and yea i hope we can deal with bokeh, cause i hate it most of the time

[charred mesa]

I like it for portraits and cetain occasions but

if I want believable generic images it ruins it

boring reality Lora is so good for this

[sage reef]

we need anti bokeh lora

[charred mesa]

there is, but it makes everything daylight for me

[sage reef]

😦

it's hard to train a concept

[charred mesa]

I think its just the dataset not being diverse enough maybe

also the guy behind boring reality lora is making a new batch of loras If I recall correctly

I hope its gonna be even more amazing

it honestly shows the bigger parameter size superiority much more than a bunch of generic bokeh portraits

[rich kestrel]

or...30 seconds in photoshop

ppl get tunnel vision for some reason and forget that photoshop still exists

[loud solar]

Or they don't want to pay for Photoshop 🙂

[rich kestrel]

who tf pays for photoshop...

its like paying for windows. r0fl

[sage reef]

im not a huge fan of having to constantly switch tools to do a job, so i would prefer if sd3 or something can do all that by itself

[narrow kernel]

Enjoy your malware rootkits, they all have them

[pine jasper]

Use photo pea…..

Photopea

[formal swan]

I'm baaaaack, good morning people c:

I am going to attempt to switch to Forge now O^O

[loud solar]

YaY! Good luck! 🥳

[gentle plume]

Hello, community! Do you know if an image generated via stable diffusion several months ago for ourselves or someone is findable or not?

[ruby lake]

Y'all checked out that new audio model

https://www.udio.com/songs/gnFD9LGCUZx7NvECfpKdRy

You can make some hype stuff with it

I almost have a hard time believing it's AI

[tame willow]

Thank you!

[true canopy]

kinda looks like an exact copy of suno ai

or the other way around, who knows

[ruby lake]

It's very high quality, highest I've heard

[sage reef]

im personally more interested in models that i can locally install and use :3

hopefully some version of stable audio comes out

@woven panther haha you are a legend man... brushnet now too? you are implementing all the cool stuff for comfy ❤️

[woven panther]

Pretty quick to wrap that stuff now that I've learned a bit, I was surprised how good brushnet seems to be for outpainting especially

[sage reef]

i also saw powerpaint, but not sure if that is older or newer than brushnet

yea it's some cool stuff

[woven panther]

Just looked at that, they adapted to use brushnet now with some changes, I'll look into including that

[sage reef]

😮

nice

btw i saw the marigold thing, is that like a new depth map thingy?

il try it out

[woven panther]

Marigold is oldest of them now, but still one of the best

There's also depth-fm and geowizard available now (wrapped both)

Geowizard creates absolutely astounding normal map predictions for a lot of stuff, like architecture

[sage reef]

wow i didnt know about those, im gonna spend a lot of time on your repos

probably missed a lot of cool new stuff

[woven panther]

Yeah haven't had time to actually use anything much... every day something new comes out 😅

[sage reef]

i know right :3

SD3 just around the corner too

[woven panther]

Hope so

[charred mesa]

Just around the corner -> 2-3 weeks left

[sage reef]

gib mii

[charred mesa]

fr

[static cape]

Well... if it turns out true. Another week went by without any news at all. They promised API access like a month ago.

I hope they don't spend all that time lobotomizing it further...

[smoky grotto]

Armed Jewish colonialists attack a village and burn homes and vehicles

[sage reef]

insert that Godfather scene :3

[broken cave]

the boring reality lora was made from just a few images. i think you can achieve the same thing with vanilla unclip

[charred mesa]

they cant

they could restart the entire training process and prune the dataset with 99% NSFW checking strength like with SD2.0 and Cascade

then it would be lobotomized for sure

[static cape]

Cascade can learn and trains fast... but it's a Diva to use. The strange cousin of the SD family.

[pearl ocean]

Did your look around the corner?

[loud solar]

Hard for me not to make a joke 🙂

[sage reef]

Im too scared to look around the corner :3

[mortal delta]

any of you have any plans for when sd3 comes out, and what if you are let down?

[sage reef]

i expect to be let down as any base model of sd so far, and i expect the finetunes to be better 🙂

[mortal delta]

Finetunes are usually better for some odd reason.

[fervent thunder]

i'm honestly afraid that they'll maxlobo SD3 and basically make it PG6

[sage reef]

lol

hopefully not

[mortal delta]

that would be very bad

[sage reef]

just give us something we can work with and we are on our way to greatness

[mortal delta]

i hope it will run on my low end device because i can run sd1.5 but not sdxl.

[sage reef]

if you cant run sdxl, it might be hard for you... but will see

[trail lion]

ouch, you havent run sdxl this whole time?

[sage reef]

people will maybe come up with pruned versions and whatnot

[mortal delta]

yeah, and im scared of google colab for some reason.

[loud solar]

Just download on Torrent (Joke off)

[trail lion]

well, there are other options, but they cost money

[mortal delta]

infact im scared of most stuff.

so i might have to man up and try google colab.

[sage reef]

you are scared of your own shadow too :3

[trail lion]

the free colab technically it's against the EULA

[mortal delta]

dang

[trail lion]

people still use it, but you risk google banning your account

[mortal delta]

is there a free alternative by chance?

[trail lion]

there's something called kaggle, but I havent tried it

[mortal delta]

I might look into that.

[nova zodiac]

kaggle is pretty good - you get 30 hours of access on their t4 machine per week, but they may block a1111 installs IIRC

aws sagemaker has a free tier

[loud solar]

IIRC? Maybe better don't use shortcuts ...

[trail lion]

you can do cheap, you can do good, you cant do both

[karmic cedar]

Suno getting generous with their free credits, I wonder why

[latent prism]

Quick question fellas, what's the easiest prompt to generate a desired text on a t-shirt?
example: "white t-shirt with LOL word on it"
?

[trail lion]

woman wearing a t-shirt with the word "LOL" try that, but dont expect it to work every time

text is kind of tricky

[karmic cedar]

OMG, BBQ, and LOL are considered the Holy Trinity of Internet Abbreviations

Also WTF is too, but more like in a fourth band member sort of way

[trail lion]

yah, dont expect to be able to generate a paragraph

[karmic cedar]

I suppose once SD3 drops you could try generating something like “I live in the age of nerfed datasets and quantified love”

[latent prism]

Hm aight, thank you

[karmic cedar]

https://www.reddit.com/r/LocalLLaMA/s/ZlpSbo7f8Q

Whoa!!

So cool 🤯 🌈

But then comes the kid that wants a new and unique story every day, and then suddenly they want a new and unique story every day as an adult and then suddenly they are some kind of sociopath

So… 🤷

[pearl ocean]

If it ever drops

[charred mesa]

it will

[placid mango]

hi discover that AI from a youtube vid, is it still free to use ?

[charred mesa]

I hope multi subject Loras will be trainable for SD3 8B on 24GB, but I bet that's a stretch

[pearl ocean]

I only got 16gb

[charred mesa]

yoinked said that training an 8B with 24GB of vram is barely possible

2B Loras can probably be trained with as low as 12GB 🤷‍♂️

since 3.5B SDXL requires around slightly more than 12GB

[heavy lark]

You guys see the sd3 thread on Reddit? He does lots of peoples prompts. Better than sdxl in a lot of ways. Lots of the sdxl problems like intersubject interactions are still there. Hands are still rather messed up. It'll be a nice movement forward but it's no Dalle or ideogram. Not by a long shot.

[languid shale]

yes

[placid mango]

thank you, installed the thing but can't figure out how to make it work XD, trying to recreate images using other image i have for pixel art

[languid shale]

you might want to search for a model that is good for pixel art

although im not an expert at this stuff.

🤷‍♀️

good luck!

[pearl ocean]

Here’s a question, how come there’s always UFO sightings in the US, but not Australia??!! Are we not good enough for the aliens???

[shell hazel]

that's because the kangaroos are the aliens

[nova zodiac]

Have you seen Australia - there could be an alien race having a giant rave party in the middle of the outback and noone would ever know!!!

[placid mango]

yeah that's ok, thanks for the answer haha

[nova zodiac]

Yamer has a decent pixel art model - and searching pixel on civitai should net you 5-10 models that will all do a reasonable job

[heavy lark]

https://civitai.com/models/381734/hysteria he has a new horror model coming out in a day

[placid mango]

Honestly i have no idea how to use the tool like at all, i don't even know what to do with that info XD, my first goal is to recreate a really similar art style for a indie developper, he lost his art guy and we want to add more sprite in his game, the style is pretty specific, so yeah. But it's noted, i'll look into it probably tomorrow, going to stream a bit on youtube for now lol

[nova zodiac]

Ive seen the previews, its pretty sweet

[pearl ocean]

[proper grove]

hello

[pearl ocean]

Hello God Like SDXL user

[karmic cedar]

sightings tend to be correlated with two things: the amount of military bases nearby, and also the population density

[pearl ocean]

Ah, so we screwed:(

[karmic cedar]

but with statistics, outliers are predictable

and those reports are simply outliers

and these days, if they aren’t actual antigravity devices then they’re one of four things: russian hypersonic technology, chinese hypersonic technology, united states hypersonic technology, or nation ___

or, honorable mention: a made-up news blurb designed to distract from someone’s affair or murder

or war

[trim magnet]

its def the obama reptilian,he can fly

[pearl ocean]

😮

[karmic cedar]

yes i am told you can hear his “whoofle-whill” sound if you leave a bag of kumquats and cola nuts in front of the savannah at dawn

[pearl ocean]

We don’t even get these ghosts everyone is raving on about in the US lol

[karmic cedar]

key thing to remember is that it isn’t everyone

it’s just one person saying it’s everyone

sometimes two

[formal swan]

I believe in the reptilian armada hiding within our ranks.

[sage reef]

So about a week ago, there was a thread on reddit and Emad replied:
https://new.reddit.com/r/StableDiffusion/comments/1bvt1yb/emad_is_no_longer_the_main_shareholder_of/ky3cf09/?context=3

He says at the end "SD3 will release in a few weeks probably API much sooner worry not."

The API part seems true since there is a dude on reddit with API access it seems, telling people to post prompts so he can generate them. So, Emad was right on the API access part when he said "much sooner", so we can perhaps trust he will be right for "few weeks" as well? which would put it still around April 26 idk? which i guess is approx the same range as the new lead dude said at Stability. 😮

so if we are lucky, maybe like 2 weeks from now

[pearl ocean]

You, change your mind, like a girl, changes cloths

[crimson belfry]

I signed up and never got anything back, I would like to see the bot return

Is everyone still on the wait list?

[pearl ocean]

It will be out before you get the invite

[amber bloom]

It could be that stability has decided that having a massive amount of beta testers via the API/discord bots isn't particularly useful at the moment, so they're not going to send out any more invites.

The whole wait list may have been an Emad stunt more than anything they really really needed to develop SD3

[earnest warren]

Wooooowwww wow wow wow this is huge why didnt i hear of this

👀 👀 👀 👀

[stuck sun]

is it just me or has progress on realistic SDXL model training really petered out, civitAI seems to be overrun now with the most repulsive and depraved Pony variants imaginable

[trim magnet]

pony popular,realistic models are good enough already

[stuck sun]

good enough for what

[narrow kernel]

All of the realistic models tend to be extremely overtrained. You tend to run into a diminishing returns effect

You'll see the same handful of facial structures and looks

It's easy to test just by setting the cfg to 1

[stuck sun]

yeah, and flexibility for doing action poses is very limited

[narrow kernel]

So it's kind of pointless for a lot of them to continue training more into the models because it will break their vertical slice presentations of the models

[stuck sun]

hello world was very promising, then they just stopped.

[narrow kernel]

By adding more variation to the model, you'll make sacrifices in quality elsewhere

There's only so much information you can pack into X number of gigabytes

If you want an ultra quality model with a ton of variety in say juggernaut, you'd probably have to double or triple the model size, which would knock it out of usability for 95% of local users

[stuck sun]

still thats fine if you are seeing fine tuning to customize it for specific things, but thats not really happening as much any more, except for pony models that specialize in various juices overflowing from engorged anuses

[trim magnet]

you can always do what pony did,get a few a100's and train them yourself

[stuck sun]

they spent like 50k to train pony

[narrow kernel]

I only recently learned about the pony stuff and was absolutely disgusted, but not surprised by it

[stuck sun]

someone said the midjourney model is 80GB in size, not sure if true, but interesting if it is the case

[trim magnet]

no one really knows since its closed source

[stuck sun]

since its so much more flexible than any individual SD model

[narrow kernel]

MJ probably has a dozen different models that get used based on the prompt. They use LLMs to prompt expand and probably pick the correct model to use with it

[stuck sun]

yeah hence the disclaimer

[pearl ocean]

image that much VRAM

[stuck sun]

i mean that would mean its running on A100s or whatever, again i have no idea if thats true, its just something someone told me, and i forget who, also your fingernails don't keep growing after you die, that was a lie too.

[narrow kernel]

i mean you can simulate having that much vram pretty easily, it will just be extremely slow. you can just do CPU renders and use regular ram. 128gb isn't really that expensive. but you'd probably have 5 minute render times lol

[stuck sun]

or Nvidia could put some more fuckin vram into their goddamn cards

[pearl ocean]

😮

[narrow kernel]

no need for consumer cards. the vram gimmick is a lie. even the most demanding 4k games don't really need more than 12(even though clickbait youtubers think they do). pretty much every major game engine uses pooling and just dedicates some chunk of vram. the actual usage of the allocated space might only be a small fraction of it. also, asset streaming and virtual textures. no point in keep a 4k texture for a rock that's 5% of your screen space when the texel resolution is based on your screen pixel count. so it will drop it down to like 512 without any loss in visual quality. but there are also so garbage game engines out there that don't manage pooling well, so they use the extra vram as a crutch

only 27.72% of GPUs have more than 8gb of vram btw, according to steam hardware survey

and only 4.8% have more than 12gb

[stuck sun]

talking about using it more for 3d rendering, compositing, visual effects, editing, professional work

[narrow kernel]

you dont use consumer level cards for that

you use quadros

[stuck sun]

lol

what do you think the advantage of quadros over consumer cards is

[trim magnet]

just the vram and nvlink/ai support

[stuck sun]

yep

[trim magnet]

but the chip is the same

[stuck sun]

yep, well almost, but yep

quadros use less power

and they still let you use blowers to run 4 cards in parallel on the one machine

and pool the vram

but all that stuff they used to let you do on consumer cards

until they took it away

so they could charge more

and if you render in redshift you dont even need nvlink any more tbh

nvlink is shit anyway

[narrow kernel]

point is, the vram is there for those cards since they actually need it. gaming/run of the mill cards do not. nvidia knows this and anyone else in the gaming industry knows this as well. but hey, there's an 8khz polling rate keyboard out there that will give you a 300% advantage in fortnite (sic)

so they aren't "holding" the vram out from rtx cards, they just don't need it right now and it's a waste to put it in there

[pearl ocean]

They say more VRAM = happy life

[stuck sun]

its not a waste to 3d artists, they just dont give a fuck about them as a customer base since they are a relative minority

[trim magnet]

there was a card that needed good vram and nvidia killed it with a shitty bandwidth the 4060ti with its 128bit bus

[narrow kernel]

its just a dumb gamer fixation for now, like idiots that think they need more than 120hz monitors

[stuck sun]

but anyone who does 3d has been screaming about getting more vram for years

except for those kinda people who say "why wld u need that" on every forum when you ask for anything to get better ever

[narrow kernel]

not really, you can do quite a bit with even 8gb cards. if you find yourself needing more than that, you need a render farm anyways lol

[pearl ocean]

Imagine having 4gb vram

[narrow kernel]

oh and i should say that i've been doing vfx and game development for over 15 years now

[trim magnet]

they know they can just gatekeep the 48gb cards and sell them for 10k usd instead of putting them in the x090 ti's and selling them for 5k usd

[stuck sun]

yep

[stuck sun]

8GB vram is useless for 3d rendering anythign but really basic shit, i presume you are being hyperbolic here

[pearl ocean]

Imagine SLI

[narrow kernel]

if you need more than 8gb to render out 1080p videos (you dont waste time rendering 4k content on a single machine, it would be assinine), then you need to learn how to optimize the scene and settings

and that's counting gpu accelerated rendering

[trim magnet]

they didnt offer good suppor for SLI either,they just abandoned it,instead they chose to support NVlink cuz they know only the $10k+cards support that

[pearl ocean]

Tbh, they should allow SLI on modern cards, even if it isn’t efficient. Good to have options

[stuck sun]

well i have a small render farm of rtx6000s and 8000s and 4090s at home, i used it to render a bunch of 4k shit recently for films i was making in redshift and vray gpu, i can tell you none of that would have rendered in 8gb vram, big fluid dynamics sims with millions of foam particles blah blah,

[narrow kernel]

sli was just carryover from 3dfx iirc

[narrow kernel]

then youre not optimizing the scenes out and not precaching the fluid sims correctly

[pearl ocean]

Ocean sims are amazing

[narrow kernel]

a long time ago, i learned the importance of proper optimizing. you can get within margin of error on images but the poorly optimized one might take 5x longer to render and use far more ram/vram

[stuck sun]

uh huh

[narrow kernel]

oh and again, rendering at 4k on consumer level hardware is pointless. it would be 10x faster to render at 1080p and AI upscale to 4k, with an absolutely minimal loss in quality

[pearl ocean]

But did you learn coding?

[narrow kernel]

yes

[pearl ocean]

Nice

[narrow kernel]

im an engineer by formal education

[pearl ocean]

You know, I never felt like an engineer as they call it

[sonic tundra]

[narrow kernel]

oh and another big hit in render times fabricatedgirls: your sampling settings. you can also allow for far noisier renders and just use a good denoiser. gotta find the right balance though, if you go too low with it, the denoiser will smooth the scene out too much

[pearl ocean]

Me when I solve a problem

[sonic tundra]

lol

[stuck sun]

why are you trying to give me a remedial beginners lesson in optimizing 3d scenes

[sonic tundra]

because internet

[stuck sun]

lol

[pearl ocean]

Back in my day

[narrow kernel]

the whole point of remedial is to relearn something you either didn't learn or didn't learn correctly the first time

[sonic tundra]

back in my day we waited 35 mins for a single frame

[stuck sun]

appreciate the condescension, i think most people who do 3d professionally would not make the claim that you can output any render of any scale or complexity using 8gb vram if you just work smartly to optiimize the ram usage, since some things tend to have more detail than other things and to do the things with more detail it uses more vram no matter how much you dial in your render settings, and even IF you use a "good" denoiser.

anyway

[pearl ocean]

back in my day, a frame was mailed to each other 1 at a time

[loud solar]

Back in my day we used oil paint to create a frame ^^

Just wanted to be part of the "Back in my day" thing ^^

[stuck sun]

Oil paint is cheating, back in my day we used manganese dioxide and charcoal on cave walls

[true canopy]

lol i cant figure out a good way to make characters grab stuff like sword = /

[loud solar]

Repeat, repeat, repeat ...

[true canopy]

not sure what im doing wrong, ive tried inpainting, img to img

[stuck sun]

controlnet + openpose

[true canopy]

well, i figured i could generate an image, add a sword over the hand in editor like photoshop, then regenerate it and have her grab the sword

im not trying to make it in one go

maybe its just too few pixels to work with

[stuck sun]

can you try openpose to help guide the inpainting?

idk never had to try this

specifically

[true canopy]

posted image in #🏞｜general-with-images

[stuck sun]

ah i see

if you placed the sword in photoshop when not just take a photo of a hand gripping a sword from somewhere and past that roughly in, then im2img will have more info to work with

[pearl ocean]

hmm thats a good idea, we might need some cool lookinh swords in #🏞｜general-with-images

[fervent thunder]

Hey everyone! I'm offering $100 to anyone who can help me install two local instances of voice cloning software like TortoiseTTS, X-TTS, etc. I'll pay $50 after each successful install, which means it should consistently clone the voice of a chosen person.

I've run into some snags trying to do it myself, so be prepared for a few challenges along the way.

Specs wise, I'm working with an NVIDIA GeForce RTX 4070 Ti, 13th Gen Intel(R) Core(TM) i5-13400F, and 32 GB of RAM, so hardware shouldn’t be an issue.

I’ll be checking my DMs tomorrow at 18:00 BST and will give everyone a fair shot—first come, first serve. Looking forward to your messages!

[true canopy]

i have to pace a whole hand in to just make it understand? T_T

[stuck sun]

well you can do it really quick and rough, and it'll be a lot faster to get what you want done than the way you are going about it now imo

anyway thats the first thing i'd try, since you are using photoshop anyway, its a trivial amount of work to test it out

oh i see you got it working anyway, nice one

[narrow kernel]

kind of odd because i just made a vdb in embergen with a 512x512x128 resolution(~33m voxels) and two data channels. rendering it with two lights and at 4k, i didn't even go over 3gb of vram in the task manager(i forgot to check the peak within the render window, but it would have been much lower). don't have maya on this pc, so so i'm doing it with blender 4.1 with experimental feature set. i'll try running some more sanity checks with actual scenes+vdb fluid cache to see how high it goes

[low moon]

SD3 out yet? haha just kidding! -_-

[loud solar]

Download on Torrent ...

[pearl ocean]

April fools was awhile ago

[humble ravine]

hello

[narrow kernel]

Had some time to rig up a vdb from embergen into the demo junk shop scene. the smoke effect was around 50 million voxels and the majority of them are in view of the camera. I set blender to render at 4k and cranked up a few light bounce settings. Peak memory usage was 5043mb and my gpu never wen't above 7.7gb(didn't go into shared memory) and that's with the viewport scene assets still cached in the vram(it would have evicted them if it needed the space). Kinda a slow render, like 2m30s per frame, but this PC only has a 2080 in it. But yeah, that's a fully fleshed out scene, with a very large amount of voxels and rendered at 4k, all fitting within 8gb of vram.

oh and render tiles set to 2048 and the vdb step size set to auto(it will determine based on voxel size). I don't really use blender much, so I can't think of any other settings to really mess with to make it more or less difficult to render.

[sweet chasm]

Me when I solve a problem

[past charm]

hello

[molten quest]

I'm looking for some help. I have some very stylized text from a game logo and I would like to try and use comfy UI to generate different words/letters in the same style. Anyone know how I might be able to achieve this workflow?

[frozen flower]

How to create 360 panorama? (ComfyUI)

[pearl ocean]

Dang it! I forget to make Sword images

[fervent thunder]

my god supir is painfully slow. the electric bill for using it on an entire video must be insane...

[molten quest]

What?

[fervent thunder]

I said "MY GOD SUPIR IS PAINFULLY SLOW" 😄

its used for image restoration / upscale. typically need around 30 gb vram ( you can use some methods to decrease that a bit) to run it

[pearl ocean]

I only got 16gb 😦

[fervent thunder]

you could run it with that im pretty sure. ive got 20. a 1.5x upscale took just over 30 minutes though, 2560x2560 to 3840x3840. Its good for sharpening up an image as like a final upscale pass without changing the actual details. just for putting that final polish on something you really like

ok... i just zoomed in to compare side by side. im actually blown away. well worth a 30 minute render

[low moon]

Supir is the best upscaler in the known multiverse.

[trim nymph]

what paramters you used?

[fervent thunder]

oh man. which paramaters lol

more or less default on everything, was just a test run since i just installed it

[thick stream]

Is there anyone here that has been able to deploy stable diffusion serverless to the cloud?

[fervent thunder]

did enable background restore, ran it at fp16 , and used tilevae

[thick stream]

I'm kind of lost and was wondering if anyone could give me some tips

[sand flax]

BRO a grandma will get done walking from state to state before this company be ready to finish what their doing with SD3

Like Tacobell will already have came out with a brand new menu before this AI company gets ready

[fervent thunder]

my guess is that its been ready but "safe guards" are being put in place. same as with sora. been ready, they just want to ensure it cant be misused 😐

[sage reef]

i personally never understood the point of "safe guards". to me this would be the equivalent of like manufacturing a weapon and then nerfing it, which doesnt make sense. the problem is the "person" using said weapon, said software, etc.. go after those people, but dont nerf the product, i mean idk.. might as well then not make anything at all if we are this worried about misuse or whatnot..

[formal swan]

I got a question. Do you need to download all the controlnet models if you are using forge? I installed forge that has controlnet baked in, but I haven't installed the models for controlnet, but they have still worked, all the ones I have tried at least.

[sand flax]

Like the SD3 has the capability to create something of a celebrity, a photo that never existed. So this SD version to celebrities could be threatening to them

Or making some form of fabrication, or worse or unspeakable things

[sand flax]

Nah, if that were to be the case we all shouldn't exist if mother nature didn't knows we would misuse the purposes of our living, like hurting or unaliving another person.

[crude notch]

would you like a half baked model?

would you like something barely better than sdxl with higher vram reqs?

or would you like something that is Basically Sota On Your PC?

(also unfortunately they will lobotomize it, not as much as 2.x lol)

[charred mesa]

so they CAN lobotomize it?

how so, especially somewhat late in the training?

[crude notch]

LECO or something like that

more likely will just be training on pure sfw for a while

[charred mesa]

so erasing concepts

damn I hope that doesn't hurt the model's quality

if not then idc, cause guns and blood are still available in SD3

or maybe not blood, I don't remember having seen blood

[crude notch]

its mainly just to remove pure nsfw, e.g. what most anime models excel at

guns n [relativly low] blood, while not pure, are in the data afaik

[charred mesa]

I mean I thought about this too, removing concepts from the finished model somehow, but I was always worried about it accidentally hurting other parts of the model

[charred mesa]

I mean blood will probably look like thick ketchup anyway

just like with current models

I have seen someone on reddit try TF2 for me and it didn't know about TF2 still, it just made "generic cartoony hero shooter"

idk if the stable assistant was using 8B though

[crude notch]

yeah i doubt it has enough tf2 data

[charred mesa]

I thought it might have a bit of TF2 like with other dataset, but with a higher parameter size it would perform better with more obscure stuff

but I suppose that's not how it works

I dont even know the current dataset

[crude notch]

the way i think it works is that while it might know what tf2 is, it probably only has like 4-5 images, since tf2 is really obscure in the grand scheme of things

re-ran it, yeah still with "Generic Hero Based Shooter"

[charred mesa]

aw you have access

good for you hehe

well yeah

I'm trying to search it in Clip Front (laion clip retrieval) but the backend url it uses is literally 502 bad gateway rn

so I cannot even see if Laion-5B had it

[crude notch]

reading people hate on sd3 is amazing lol

[charred mesa]

Laion-5B had almost everything, including angry video game nerd, which SDXL and SD1.5 failed to catch of course

so I thought maybe an 8B model with like a higher portion of Laion would perform better in more obscure territories that previous models would have missed upon

[charred mesa]

fr

on reddit they keep comparing it to Ideogram and DALLE3 and Midjourney

[crude notch]

it is not dall3, it is probably better than mj (i have not used mj)

[charred mesa]

and they're like "eh not following the prompt 10/10 and hands aren't perfect so this model fckin sucks 👎"

lol

[crude notch]

the person who got kicked literally gave it a 500 word prompt and complained when it didnt follow it perfectly

[charred mesa]

oddly better and worse than MJ, from the Stable Assistant results it seems to perform worse in 2 subjects, yet better in every other territory

it's weird

[crude notch]

i think it more or less decided that it shouldnt waste params for an obscure image and use them for something else

[charred mesa]

uhhmm!!!!! 500 tokens are within the 512 context length!!!1!

[charred mesa]

hmm

maybe they also use aesthetic filtering that might accidentally prune these as well

🤷‍♂️

awww I wish I could use Clip Front man

I mean its not like DALLE3 was miles better

it did give the characters gloves and ammo belts so they all looked like heavy weapons guy

so it was somewhat like TF2, but it's still not it

[crude notch]

sort of the same for todays model

[charred mesa]

if we can train multi concept loras on 24GB on 8B I'd be super happy and I would not even mind missing info, but you said that training simple loras on 8B with 24GB is already barely possible

[charred mesa]

wait whaaat

today's SD3 version?

[crude notch]

lora vram reqs are gonna suck for sd3 (not as much though)

[charred mesa]

well 2B Loras are obviously going to be trainable on 24GB (right? 😬)

[crude notch]

oh yeah the bot's model updates every now and then

[charred mesa]

damn

[charred mesa]

so there is some hope wow

like I wont expect PERFECT TF2

[crude notch]

yep, if you can train xl you can train 2b (unet)

[charred mesa]

but if it can get to whatever style DALLE was trying to depict then that's awesome already

[charred mesa]

ahh so it does somewhat work like that

point is, SD3 is already amazing for a base model

[charred mesa]

https://www.reddit.com/r/StableDiffusion/comments/1c35onn/remember_when_emad_said_sd3_looks_better_than/

these are Stable Assistant SD3 images

[crude notch]

afaik stable assistant isnt using the 8b model

idk

[charred mesa]

WHAAAT

so thats why people say it looks like 512px upscaled

if it's 2B then I'm amazed

[woven kindle]

hi guys, anyone can help me? I'm trying to do face swap video, I use roop unleashed, it works +- when the person is facing forward but when the person turns their head it completely bugs. Is there any special configuration that needs to be done? Or another software?

[crude notch]

the only example of 2b ive seen is a picture of yellow elmo (its better than xl)

[charred mesa]

and alex (mcmonkey) made some cat images that compare 2B-512 with SDXL

and it looks as good as SDXL (if not better), and prompt adherence wise it obviously won

the 16 channel VAE is really showing it's worth with those 512px images

[crude notch]

the new vae is amazing

[charred mesa]

I thought it was only going to be for better colour accuracy

[crude notch]

and it doesnt look fingerprinted

[charred mesa]

wdym

[crude notch]

remember xl1.0vae?

[charred mesa]

ohhh

the weird artifacts yes

we use 0.9 vae anyway

[crude notch]

yep

[charred mesa]

also its funny that people complained about an SD3 result, because the prompt had "very low resolution" and "pixelated background" and it make the whole image pixelated, yet when people generated DALLE3 results they were all 2D pixelart games, whilst SD3 was 3D geometry with big pixels

but I mean DALLE3 modified the prompt anyway so we don't know

@crude notch https://www.reddit.com/r/StableDiffusion/comments/1c2je28/comment/kzc0ltv/

Amazing painter strikes again

wtf he complained about all those stuff but could not take 5 seconds to read about the dataset captioning from the paper

its like a 15 second job with the search function

also what's funny is that the TF2 prompt I had where a blue soldier is holding a rocket was terrible with DALLE3, it kept holding some weird rifle or grenade launcher

[crude notch]

oh would you look at that, someone posted a comment explaning each point and why its inaccurate

[charred mesa]

LMAO

lemme refresh

[charred mesa]

who could that be

[crude notch]

i have no idea

[charred mesa]

also wasn't that painter guy using an older version too since he got banned like weeks ago

[crude notch]

yep

even then, the model still could do that

[charred mesa]

the funniest thing would be if he prompted it like SD1.5, I would have laughed my ass off

[crude notch]

i wish i could post the prompts.

i dont want to lose sd3 access, but it would be really funny.

[charred mesa]

makes sense

ah

[crude notch]

its more like if someone treated sd3 as a local dall3 clone

[charred mesa]

hmm I was thinking of Superprompt v1, but that would need more training to get it to work better with SD3 tbh

it would help a lot for the SD1.5 community though

make them learn natural prompting and etc

[crude notch]

which, as i have to say, it is not

yeah, tag based prompting (at least for launch sd3) is dead.

[charred mesa]

wooowwwww a model that's less than half the size as DALLE3 is not as good wow its trash it's literally DOA on god fr fr

[crude notch]

"i cant believe this 8b+12b model is sllightly worse than a 20b+gpt4 model!!!!"

[charred mesa]

I wonder what would happen if we omitted or left the prompt for the clip model empty which performs better on tags

clip_l or clip_g

I don't know how the two clip conditioning models work alonside eachother

[crude notch]

they dont want to get sued

[charred mesa]

^^^^

and they need good PR, eg: "we worked a lot on SAFETY so that the prompt wont destroy our lives and take over the entire world"

they always target the open models

[sand flax]

Thing is...people are going to misuse their product anyway, and it's funny because they know this

[charred mesa]

sure, people will train it back in

but as a company, you need to censor it.
if you don't, then people will know Stability as a company which gives out models that can generate nsfw images of celebrities

[crude notch]

they have to cover their asses, legally

[charred mesa]

if they censor the model and people train it back in, it's less of a hassle for stability, since they promised a safe model which would prevent bad actors

[crude notch]

if you make a model that can do nsfw of any celeb, then those celebs will get mad, and will sue you (most likely)

[crude notch]

pushes the blame onto those other people

[charred mesa]

Yessss

epic win!!!

[low moon]

Sigh why does anyone care anyway what Harry Potter looks naked...

[loud solar]

Sounds like prompt material ... thanks! 😁

This stupid bad guy might be more interesting ^^

[fervent thunder]

SD3 almost seems like a myth with how early the release papers were out and still no actual release

[charred mesa]

yeah idk why they announced it so early

[fervent thunder]

marketing stunts

[charred mesa]

they built hype too early

[fervent thunder]

wayyyyyyyyyy too early

this is subpac vibes (scammy)

[crude notch]

mainly because this is a whole new arch

[charred mesa]

lol

absolutely

[crude notch]

and they want to make sure that It Works Well

[fervent thunder]

https://subpac.com/ they've waited 2 years for their product

never came out

Looks great, but in reality it's not real

[charred mesa]

okay but SD3 will come out

[fervent thunder]

yea ok

I'll believe that when I see it

[charred mesa]

CTO stated multiple times that it WILL have an OPEN release

[fervent thunder]

The rest is just fluffy bs 🙂

[fervent thunder]

I could state this 100 times for 2 years

[mental crow]

guess we will have to wait for another two months

[fervent thunder]

does that make you feel better on signing up for something i haven't released yet?

I just put out some papers?

[crude notch]

its been like 2 months, cant you wait 2 more months/weeks?

[fervent thunder]

that's pretty ridiculous

to announce features 4 months ahead

anyway, g'day lads, just thought i'd chime this in

announcement: WAY too early

[mental crow]

Elder Scrolls VI

[crude notch]

yes, it was early, will it release? of course

[charred mesa]

yessir

[mental crow]

2 months till it releases then another month or two for the fine tuning

[charred mesa]

epic win

[fervent thunder]

that's a loss to me

and in the eyes of every consumer

[mental crow]

And only then we will be able to prompt shrek sitting on a chair made out of spongebob

[crude notch]

i doubt that wd will take that long lol

[fervent thunder]

I'm still patient, but here's what this did to my mind: I quit using SD all together until SD3 was released, yall keep pushing it back

[mental crow]

well wd is kinda dead

[fervent thunder]

I have not done any Ai because of the SD3 wait

[mental crow]

i'm more waiting for animagine sd3

[crude notch]

sort of, wdvxl is soon

[charred mesa]

tried deepfloyd, Pixart-Sigma and ELLA SD1.5

lol

pixart-sigma is the closest since its a DiT with a captioned dataset, but it's tiny (600M, smaller than the smallest SD3 model)

[mental crow]

last time i tried a wd model it gave me errors or was very low quality

[charred mesa]

but the quality and prompt adherence isn't even close to SD3

[crude notch]

1.5e3?

[mental crow]

probably yeah

[crude notch]

yeah its mainly since auto webui doesnt like 2.x

[charred mesa]

interesting

[mental crow]

hopefully they make it work this time around

[charred mesa]

I was using auto1111 in the 2.X times as well

and it worked fine

[crude notch]

its gonna be cursed.

thats all i can say

[mental crow]

lol probably

those guys commited to agi anime waifu anyway

or some llm thing i don't remember the announcemnt

[crude notch]

agi tomorrow (probably)

[mental crow]

lol

[crude notch]

oh kino's announcement yeah

[fervent thunder]

anime waifu? lol

[trim nymph]

[charred mesa]

emad

[crude notch]

yep lol

[trim nymph]

its really time for sd3

[mental crow]

I keep checking the stability.ai twitter feed every couple days like an addict waiting for sd3 this is kinda sad

[charred mesa]

I would try it out on the demo or even 2B locally so bad lmao

but it would mess things up sadly

[trim nymph]

for weeks/months ive read „next few days“

[fervent thunder]

I check each weekend

[charred mesa]

stop trusting those lol

[fervent thunder]

but it's gone so long that now im llke ??? i need to go to the discord and ask

[charred mesa]

CTO's ETA isn't even close to "few days"

[trim nymph]

i gave up

[charred mesa]

its still 2-4 weeks

it's probably 4 weeks tbh

[mental crow]

Did the beta access even go out ?

[charred mesa]

or 3 if we're lucky

[crude notch]

1. yes its happening
2. 3-4 weeks

[crude notch]

yep!

[charred mesa]

not for the waitlist people iirc

[mental crow]

I swear lykon is the only dude that got access

[charred mesa]

for the waitlist people too?

[crude notch]

i have it, just that cant share without permision

[trim nymph]

some even mentioned, beta finished two weeks ago

[charred mesa]

bruh

[crude notch]

oh that dude.

[charred mesa]

LMAO

FOR THAT PERSON YES IT ENDED

LMFAO

[mental crow]

wha'ts your twitter yoinked?

[crude notch]

i dont use twitter

[mental crow]

you posting stuff like lykon

oh

[charred mesa]

based.

[crude notch]

i only post on discord

[mental crow]

less brainrot

[trim nymph]

you have access to sd 3?

[fervent thunder]

he uses tiktok like a real man

[trim nymph]

[crude notch]

yep

[crude notch]

lol

[trim nymph]

how well it works with artists?

[charred mesa]

where 👀

[mental crow]

what's the wildest thing you prompted on it

and it got it right

[crude notch]

here or tohoai

[crude notch]

wildest? obscure video game character (didnt get every detail but it got shockingly close)

[trim nymph]

how good is the artist knowledge?

of the model

[mental crow]

I wonder are people already mixing models with it?

[crude notch]

not yet*

[charred mesa]

uhh I cant find it lol

[mental crow]

So you don't have access to weights just like an api thing?

[charred mesa]

can you DM link or something

[crude notch]

oh, should just be .gg/touhouai

[crude notch]

yeah discord bot

[charred mesa]

oh so it IS touhou

lmao thanks

[crude notch]

same for every non-sai employee

[trim nymph]

💋

[mental crow]

Oh got it yeah makes sense

[crude notch]

they really dont want it to leak

[native lion]

guys

i want to turn my pic to a cartoon and put a pikatchu hat on it

is it possible to do in here eh

no answer eh !

[true canopy]

not with any bot no

[low moon]

Try with R2 D2.

[forest trout]

I've been using Auto1111 for a long time now but I'm wondering if it's time to pivot. Are there any guis that have surpassed a1111 I should try out?

I've tried comfy but I'm not sure it's my thing.

[loud solar]

Stable Diffusion Forge is the new A1111

[forest trout]

I thought the repo owner abandoned it?

[crude notch]

its abandoned

lol

[loud solar]

Damn ...

[stuck sun]

is it?

thats a shame

it was better than A1111

except for the bugs

lots and lots of bugs

[crude notch]

oh if you like the performance then just use comfyui

[stuck sun]

if you like spending your time zooming in and out of things

[crude notch]

swarm then

[charred mesa]

ngl I tried swarm and I was more overwhelmed than just comfyui

I got into this custom workflow thing and I just cannot use an auto1111 type interface ever again

[crude notch]

auto is limiting

[charred mesa]

but swarm is really cool though

[stuck sun]

anyone care about foooooooocus any more

[charred mesa]

basically comfyui but you dont have to do the hard bits

you get fast updates and everything

[sage reef]

i would recommend comfy for three main reasons:

1. usually is the first to have support for the latest new stuff
2. has better memory management and requires less vram for stuff
3. there are tons of modules that are implemented almost on the daily, based on new papers released, etc,
   and are usually available way faster than A1111 or similar where you have to wait who knows how long till someone
   implements it, if at all.

Also, I think technically comfy has way more modules developed for it than any other framework anyway at this point,
just lot more things to connect and try and have fun experimenting with, gives you a lot of flexibility where as the others
you cant connect them in a way you want, you have to follow the way it works in the background and that's the only way.

[charred mesa]

stable cascade support, and now SD3 support WHEN IT COMES OUT

[low moon]

Comfyui was forged in Hell.

[charred mesa]

but I mean technically comfyui DOES run SD3 right now, we just don't have the code and weights

it literally already supports it lol

[stuck sun]

i take it there is still no cascade support in A1111, then, other than that useless stable cascade extension/tab thing

[charred mesa]

comfyui might take over, you guys... 😔

stableswarm will be your only option for now if you want to try SD3 on day 1

with very good VRAM management btw

[crude notch]

comfy supports sd3 day -1

[charred mesa]

^

literally right now

[stuck sun]

i bet 80% of ppl still use A1111 over comy tho (wonder what the actual numbers are)

[charred mesa]

yeah cause its simpler

[sage reef]

i switched from a1111 a long time ago

[charred mesa]

and many people just want to type prompt and click generate

[crude notch]

a1111 will support sd3in like 3-4 months after release

[charred mesa]

but I have been using stable diffusion since 2022 august so uhhh

[charred mesa]

lmao real

plugins for a1111 supporting SD3 will come out before A1111 implements it lmfao

[low moon]

With 64GB Vram needed.

[stuck sun]

why so slow, is the coder an alchoholic?

[charred mesa]

WHAAAAAAAAAAAAAAAAAAAAAAT

ik its a joke but damn

[low moon]

lol

alcoholic

[charred mesa]

hello

[low moon]

😄

[meager sage]

hai

[crude notch]

less than you expect

[crude notch]

codebase is a fuck

[charred mesa]

it is simply put... a fuck...

[charred mesa]

12GB VRAM for 8B fp16 and T5 at 4-bit?

[crude notch]

yea

[charred mesa]

and that's the biggest model

[crude notch]

correct

[charred mesa]

and if stable assistant is using 2B or 800M, then people don't even have to worry

2B beats SDXL

[stuck sun]

why was forge abandoned then, seems like it solved a lot of the problems.

[charred mesa]

alex (mcmonkey) showed an example of 2B-512 with equal quality + way better text and prompt adherence

[crude notch]

they have to keep up to date on both a1111 and comfy

[stuck sun]

forge is same coders as comfy?

[charred mesa]

no

if I recall correctly the guy who made forge has to do with controlnet?

[stuck sun]

yeah thats what i thought

maybe i'm just not following, whois "they"

[charred mesa]

Lvmin Zhang (Lyumin Zhang)

[crude notch]

yeah lllya

[charred mesa]

https://lllyasviel.github.io/Style2PaintsResearch/lvmin dayum

face reveal

[loud solar]

Didn't he write an examn?

[tepid jolt]

Hello guys 😄 do anyone have SUPIR?

[stuck sun]

yeah i have supir

[warm junco]

Any source?

[stuck sun]

might not be latest version tho, havent updated in a bit

[tepid jolt]

How long is the average processing time on your current GPU?

[low moon]

Even NASA can't run Supir.

[tepid jolt]

I own a 3090, I don't know if I can install it

[stuck sun]

3090 will be fine

[tepid jolt]

I heard it uses 32gigs of vram

[stuck sun]

nonono

works in 12

[low moon]

They tried it but some satellites started crashing so they had to stop.

They were trying to upscale naked pics of alien girls.

[tepid jolt]

How many GB does the download SUPIR? I'm currently on an internet package 😦

[charred mesa]

3090 will be more than enough for SD3

with vram left for controlnets and loras as well

[stuck sun]

my supir install folder is 68GB

[charred mesa]

jfc

[low moon]

i has 3060 12 gig vram

am i screwed?

😄

[tepid jolt]

https://tenor.com/view/nervous-courage-courage-the-cowardly-dog-tense-anxious-gif-17225992103088597327

[low moon]

wtf

[tepid jolt]

thank you for information man 😅

[low moon]

68 gigs?

are we serious?

and i thought 6 gigs for sdxl mdoels is a lot

[stuck sun]

yeah but that includes several SDXL checkpoints (juggernaut/helloworld and at least one more)

pluss it downloads llava

[low moon]

😮

[stuck sun]

we're talking about Dr. Furkans supir frontend here btw

and autoinstaller

[trail lion]

when there's a cloud front for it, it'll be cool, but I gave up after getting frustrated with OOMs

[stuck sun]

its good tho! Best upscaler i've ever seen, i paid for a magnific sub and cancelled it when this came out lol

[low moon]

I know its the best

I am convinced of that

[stuck sun]

only downside is it couldnt uprez to super high resolutions due to ram limits but maybe they've addressed that since, not touched it in about a month

[low moon]

It's incredible.

[stuck sun]

yeah its really impressive

[low moon]

Almost as good as those Blade Runner "zoom in" things.

[stuck sun]

I did all the magnific tests for furkan in his video about it, it pissed all over magnific in every single test without exception

[low moon]

🙂 yes.

There is always a better free alternative out there for all the AI services.

You just pay with time and your own resources.

[stuck sun]

actually did some upscale tests from tiny low rez blade runner images lol, and the results were amazing hahahaha

it preserves the subtleties of peoples facial expressions better than anything else

[loud solar]

Where to get it?

[tepid jolt]

Yes, I took a look at dr.Furkan's video. He is truly a competent person

[karmic cedar]

What version we up to on that?

Yeah i provided a few upscales to him as well

[charred mesa]

@crude notch my god I'm starting to give up on this community

[manic mantle]

hello

[karmic cedar]

More like goodbye apparently

[charred mesa]

oh im not leaving this server

and you guys are cool

[karmic cedar]

woohooooo

[charred mesa]

woohooo indeed

[stuck sun]

i'm subscribed to his patreon

dont think he made it public?

its 5 bucks a month

[loud solar]

Ahhh... thanks! I found that ...

[stuck sun]

went to sleep so missed this. I have no clue why you are trying to argue that because you can load a very low resolution voxel grid into a basic demo scene that ships with the software, that this somehow means people using gpus to render and comp in production dont need more than 8GB vram to render production quality 3d renders. I don't understand what you are attempting to achieve by arguing this? Literally what is the point? If this was the case why do they even sell quadros with 48gb vram that you can pool with nvlink? is it to render bigger scenes or not? People here will argue literally anything.

[karmic cedar]

fuuuuuuuuuuuuu

https://github.com/PKU-YuanGroup/MagicTime

metamorphic generation = more contextual fidelity

and they’re training an even stronger model with the help of that Open Sora project

so we got zords merging into an open source m e g a z o r d

[low moon]

We are cool as a polar bear's nipple.

[loud solar]

Ohhh... never touched that kinda nipple ^^

[low moon]

When SD3 drops we'll be as cool as liquid helium.

Midjourney and Dalle will melt in our bright light.

[pearl ocean]

Imagine how much VRAM you’d need

[foggy halo]

sd3 release when

[charred mesa]

sometime in the future

[low moon]

In the future.

[foggy halo]

as a real power rangers fan i can confirm

[stuck sun]

you guys excited for WW3?

[low moon]

Sort of

[charred mesa]

ww3 release when

[foggy halo]

i am excited for SD3

[low moon]

I try not to

I was hyped for the Phantom Mencace

[stuck sun]

they are beta testing WW3 right now

[low moon]

so I have learned to tmeper my expectations.

[foggy halo]

yalll watching coachella livestream??!

i am only for the edm

[low moon]

But WW3 is more likely to come out before SD3 right now.

[foggy halo]

no pop

[foggy halo]

😔

real

[low moon]

:/

I thought my game is a dystopia. But with the way things are going ym game is actually a utopia.

😄

[pearl ocean]

We need to generate some cool Swords in images

[nova zodiac]

I mean two fronts have opened up and we are just waiting for the Franz Ferdinand moment for full fun and games…

[low moon]

I am totally excited for WW3 yes. 😄

Pick your sides

Geriatric Biden or obese Kim

Lets go

!

[nova zodiac]

#🌶｜off-topic 😛

[low moon]

Right.

=p

[nova zodiac]

How is everyone doing today??

[low moon]

Prety good

[nova zodiac]

Cruisy Sunday morning here

Just smashing back marmite on toast atm

[low moon]

mmm toast.

[warped shard]

hmm does anyone know how i can install embeddings etc?

[pearl ocean]

wait till SD3 comes out

[warped shard]

wait its coming out soon?

[pearl ocean]

we told soon, but who knows when

[low moon]

It should

[warped shard]

ahh ok

[low moon]

like this month?

[stuck sun]

after WW3

[low moon]

idk

[warped shard]

also does anyone know how to install embeddings?

the pt files

and how to use them etc

cant really find tutorials online

[stuck sun]

put them in embeddings folder

thats it

[warped shard]

oh, like you dont have to type any promopt to activate them or anything?

[stuck sun]

are you using A1111

[warped shard]

no, im using forge

[stuck sun]

ok same thing, theres a tab with loras etc etc

[pearl ocean]

Forge can chnage a life!

[stuck sun]

there's one for these

[warped shard]

ohh ok

[stuck sun]

you just click on 'em in the tab and it ads them to the prompt

[warped shard]

ahh ok!

tysm!

[stuck sun]

np

[mortal delta]

Is it possible to earn money with free ai tools/software, or do you all think its impossible?
i really want to start working towards a better computer/device but im not allowed to get a job yet, and my autism limits me greatly.

[stuck sun]

i did concept art for an upcoming movie using SD and some shots in a TV commercial in dec, in combo with 3d rendering, etc. But people are a bit antsy now about using generative AI, due to how enraged it makes people, and copyright concerns etc