💬｜general-chat

[steel fable]

I need a little help please, im trying to make some pic, and all the time the image have 2 char, how i can fix this?

[near marlin]

does anyone know how i can make realistic selfies?

[fervent thunder]

guys see this

https://youtu.be/SrcIC5RvKWs

ConsiStory: Training-Free Consistent Text-to-Image Generation | NVIDIA Research

gotta love nvidia with all these research

[fervent thunder]

No

[fervent thunder]

do you mean to put your face in another body ?

sure i made my brother look like a prince with long hair.

https://www.mage.space/build

i used this site, click on selfie

[fervent thunder]

hey all. does anbody knw if forgeui has their own discord?

[rain crest]

Anyone know how to train 2 Loras of tv characters then have both of them in the same image

[desert dagger]

you're going to want to use SD3 for something like this, other models lack the conceptual understanding to even create that sort of imge correctly.

[rain crest]

But this isn't a new concept right? Or is there a better way to do this?

[desert dagger]

SD3 is the first of the stable diffusion models that has the ability to actually understand something like 2 different characters in a prompt. the other models will make multiple people but usualy it'll just be the same person repeated.

[rain crest]

Even more so when the characters aren't well known. Do u have any idea where I should start?

[desert dagger]

i would personally just start with prompting - see what the core model will do without stuff like loras

[rain crest]

SD3 Medium from huggingface right?

[desert dagger]

sd3 2b medium from huggingface, yes

[rain crest]

I'm probably confusing something, but I get this error:
changing setting sd_model_checkpoint to sd3_medium_incl_clips.safetensors [3bb7f21bc5]: AttributeError
Traceback (most recent call last):
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\options.py", line 165, in set
option.onchange()
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\call_queue.py", line 13, in f
res = func(*args, **kwargs)
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\initialize_util.py", line 181, in <lambda>
shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: sd_models.reload_model_weights()), call=False)
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\sd_models.py", line 860, in reload_model_weights
sd_model = reuse_model_from_already_loaded(sd_model, checkpoint_info, timer)
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\sd_models.py", line 793, in reuse_model_from_already_loaded
send_model_to_cpu(sd_model)
File "C:\Users\aaaa\AppData\Roaming\StabilityMatrix\Packages\Stable Diffusion WebUI\modules\sd_models.py", line 662, in send_model_to_cpu
if m.lowvram:
AttributeError: 'NoneType' object has no attribute 'lowvram'

[desert dagger]

you can't just load the model. sd3 doesn't use unet. you need to download not just the models but also everything else, and then open the SD3 comfy workflow

[rain crest]

Is it Comfyui only?

[desert dagger]

auto1111 has a little support for it, mostly it's comfy

but you still can't just use it as an interchangable model with say, sdxl setup

it doesn't use U-net

[rain crest]

I'll look up a tutorial

[rain crest]

Basically it doesn't understand who the character is and justs generates a random person

[desert dagger]

okay. then i would probably make a lora that was half images of the first person, half images of the second person, with very specific lables

[rain crest]

It wouldn't just combine the two?

[pale latch]

regional prompter extension in a1111 i know for sure allows loading different loras into different regions. The regional prompting nodes in comfy surely will allow this too. swarm has regional prompting built in and could potentially allow loras per region. On the extension in a1, you have to set it to "latent" mode which is slower and more memory, but works

[rain crest]

Thank you!, I'll look into this tomorrow

[pale latch]

newest version has full support now. but it's grossly inefficient and comfy is still better for running it with t5

auto's is fine if you're just using the clip layers though

[pine fossil]

bro is trying to get custom char loras and you guide him to an untrainable base model

@rain crest your best bet is one of the realistic xl finetunes

[desert dagger]

if it's untrainable, how are there so many trained objects on civit for it, and how did @pale latch make a lora?

[pine fossil]

they're all garbage

[desert dagger]

everyone has an opinion, you're welcome to yours, however i wouldn't say that to @pale latch face

also, you might learn to read so you know what the guy was actually asking about to start with, unless that's too much to ask of your pea sized brain

[pine fossil]

i don't want to particularly insult anyone's training ability, it's not their fault the model is bad

[desert dagger]

no, but you did insult everyone - everyone that's been training with it and everyone that uses it. good job. now, how about you go find something productive to do with your time?

[pine fossil]

this phrases belies the level you are at, you are literally not qualified to type to me

with all due respect

[desert dagger]

reported for harassement

[pine fossil]

what more worthwhile use of time than to teach others

[hallow mortar]

hello everyone I'm trying to know which model and setting are the most optimized for snfw or carefully generating porn on stable diffusion

[pale latch]

sd3 works really well in many ways still. While there's not much for the gen pop yet, enthusiasts are making good fun out of it.

The tools to use it easier will iterate. The skill floor is slowly lowering.

Theres even naked girl models out for it now

[nova lily]

oh so which cards need no half vae exactly?

[warm junco]

GTX cards for example. But it also depends on the VAE file and model used.
Some 1.5 anime VAE and the sdxl VAE needs it too

When using kl-f8-anime-v2 as 1.5 VAE and sdxl fp16 VAE for sdxl, --no-half-vae isnt needed

[remote depot]

Is the creative upscaler available yet? I know it’s in nightcafe but I don’t know how to use/access it in automatic1111, I am new to this though so I may be missing something

[desert dagger]

not sure how what angus has created for nightcafe has anything to do with auto1111

[remote depot]

who is angus?
nightcafe says they're using "the new creative upscaler from stability Ai" im just wondering how i can access it running stable diffusion on my pc?

[desert dagger]

the guy that owns Nightcafe and the lead programmer - and you'll need to ask him what he's specifically calling that

[main snow]

hey guys, kinda new to this and was hoping someone could tell me what an "activation token" is

[remote depot]

ah ok. i mean on the website they say they're literally using "the new creative upscaler from stability Ai"
...is that not whats happening maybe?

[desert dagger]

and - again - he's calling something that, no idea what, there's a number of things he 'might' have decided to lable as that. there's nothing from stability specifically labeled that. you'll have to ask him specifically what he's calling that.

[remote depot]

right, thats what i was clarifying, i'll see if i can find a way to contact him

thanks for the info

[desert dagger]

log into the nightcafe discord and ping him

[fervent thunder]

the terminology gets confusing cos not everyone follows this rule
but generally "creative upscale" means diffusion like Magnific
and "upscale" or "conservative upscale" means something like ERSGAN or the transformers (DAT/HAT etc) which are almost all based at least somewhat on the ideas of the SWIN-IR paper

[desert dagger]

and none of that is something that stability has recently released, but angus is really good at making stuff up

no telling what he's doing. he might even be calling SD3 the 'creative upscaler'

[fervent thunder]

there was a random reddit comment or two
where Emad mentioned an internal upscaling GAN at Stability
but very sadly it has not been released

[remote depot]

ok that mostly makes sense, do you have any idea how i could emulate any of it in stable?

[remote depot]

dang haha, is it possible its just stability stuff that hasnt been released to the public?

[fervent thunder]

my main recommendation is to do 3 stages of resolution boost

1. generate with deepshrink/hidiffuson/resadapter 2. small upscale with ERSGAN or transformer 3. diffusion model like supir to finish

if you want to swap out any stage that's fine

e.g. magnific or others

[desert dagger]

nope. angus uses the same APIs everyone else does. though he is part of the commerical beta testers

[fervent thunder]

but the key point is
don't try to get like 16x out of a single upscaler
that's when things start to go a bit funny

[desert dagger]

are you using comfy or auto1111?

[remote depot]

thank you!

[remote depot]

i see haha.

i am using auto1111!

[fervent thunder]

I personally like the transformers best
but 99% of the community seems to prefer Magnific or tiled diffusion upscale
so that is definitely also an option

[desert dagger]

okay, first thing to do is get some examples of how it works together and then post in #🤝｜tech-support and see if anyon ethat uses auto1111 can tell you how to get the effect

[remote depot]

okay, ill give this a shot, thank you!

[fervent thunder]

luckily auto1111 accepts transformers
you can put this in the upscaling folder:
https://openmodeldb.info/models/4x-RealWebPhoto-v3-atd

if I remember rightly auto1111 also comes with Remarci and Siax

which are some nice ERSGAN models, those are still ok

[remote depot]

you the goat 🙏 thanks a bunch m8

[fervent thunder]

no problem

I actually got into upscaling before stable diffusion lol

weird order

in case you come back and find that transformer way too slow, here is one of the fastest ones:
https://openmodeldb.info/models/2x-NomosUni-span-multijpg-ldl

in my experience people find my default methods too slow

[desert dagger]

the annoyed feeling you get when you realize you just saved frames for you animation instead of saving an mp4

[remote depot]

so the previously mentioned stuff is slower but maybe more effective?

[fervent thunder]

oh yeah drastically slower

ultracompact:      153.56 fps (0.0065 seconds)
compact:            82.92 fps (0.0121 seconds)
span:               60.46 fps (0.0165 seconds)
realcugan:          34.58 fps (0.0289 seconds)
esrgan_lite:         6.54 fps (0.1530 seconds)
omnisr:              4.89 fps (0.2043 seconds)
plksr:               2.65 fps (0.3770 seconds)
realplksr:           2.21 fps (0.4522 seconds)
esrgan:              1.97 fps (0.5083 seconds)
swinir_s:            1.07 fps (0.9375 seconds)
atd_light:           1.05 fps (0.9536 seconds)
srformer_light:      1.05 fps (0.9545 seconds)
swinir_m:            0.69 fps (1.4509 seconds)
hat_s:               0.44 fps (2.2763 seconds)
swinir_l:            0.39 fps (2.5610 seconds)
srformer:            0.27 fps (3.6405 seconds)
atd:                 0.27 fps (3.7223 seconds)
dat_2:               0.27 fps (3.7284 seconds)
hat_m:               0.23 fps (4.3972 seconds)
hat_l:               0.23 fps (4.4004 seconds)```

look at that speed difference lol

[remote depot]

oh damn lol, ok i feel you

[fervent thunder]

but this is nothing compared to diffusion

you could set up a diffusion upscaler with a slow sampler e.g. implicit midpoint or 4th order implicit adams
with tight tolerances and small tiles
and it could take days

[nova lily]

oh im using a pony with a built in vae, I wonder if it's that one. I've had it off for the past week and have been going fine i believe

[rugged rune]

hI! Who knows why every time i Render anything on any model - on the last step at 95% The rendered image goes Red Af?

[tough canopy]

gm

gm

[fervent thunder]

VAE

[remote depot]

hey! i thought i could google the stuff you mentioned but it turns out i have no idea what im doing lol, where do i put the transformer? i cant find the "upscale" folder anywhere

[zealous dew]

What with Scale for AMD for windows?

anyone has news?

[oak latch]

any good adetailers models to use?

[abstract linden]

Holy F*nk Comfy barely uses any recources at all

That thing is a NIGHT and day difference

[desert dagger]

yup 🙂 oh, also https://comfyworkflows.com/ something for you to explore

[low moon]

Everyone is happy when they give into the noodles.

The only question you will ask after you adopt ComfyUI is "why didn't I do this earlier?"

[abstract linden]

OOOOH THIS LOOKS SUPER FUNNNN! Thank you, will 100% take a look!

[abstract linden]

Oh gosh for reaaaalll!
This is honestly the good stuff. Wish I started here! \XD

[desert dagger]

welcome home?

[nocturne jackal]

guys whats the prompt to generate images here?? and from which channel?

[desert dagger]

read the information here #artisan-faq

[nocturne jackal]

Thanks dude!!

[primal void]

如何下载

安装

有大佬带一下吗

[low moon]

😆

[tribal fractal]

Hi, trying to download some stuff from stable diffusion off huggingface. I made an account but it asks me to create an organization for certain repos.

[tribal fractal]

So has everyone who has downloaded the latest models needed to be part of an organization?

[sudden ruin]

Shouldnt be a requirement afaik

[tribal fractal]

Ok. I'm trying to use GIT after downloading lfs. Since I'm new I'm not sure if I have to use a certain version of GIT or something. When I type in the clone command there, it gives me this:

Cloning into 'stable-diffusion-3-medium'...
sq remote: Access to model stabilityai/stable-diffusion-3-medium is restricted and you are not in the authorized list. Visit https://huggingface.co/stabilityai/stable-diffusion-3-medium to ask for access.
fatal: unable to access 'https://huggingface.co/stabilityai/stable-diffusion-3-medium/': The requested URL returned error: 403

[sudden ruin]

Open the page in the browser and ask for access

[tribal fractal]

It says that having an organization or affiliation is required.

[sudden ruin]

Just type in "none"?

[frigid star]

Hi guys. If I change the name of a saved lora in Auto1111 will it break anything?

[sudden ruin]

It shouldnt, but old prompts that use it might not work if you plan to reuse them

[frigid star]

Ah, makes sense. Perfect, ty

[tribal fractal]

That fixed it. Thank you.

[old lichen]

I've forgotten how to promt conditions or write
(blonde or brunette or redhead)
there was some syntax

[tribal fractal]

Ok, sorry, that helped me with the first repository. I want to download two more and when I type the clone command for them into GIT I don't get anything. It just doesn't say anything.

[sudden ruin]

Why are you cloning them in the first place btw?

[tribal fractal]

Ah I thought that's how you download lol. Is there another way to do it?

Lmao I legit thought that was the only way

[sudden ruin]

https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main
There is a files and versions tab

Then you just download the model you need

[tribal fractal]

Thanks again

I just trolled myself for more than an hour over that lmao

Was telling myself "There's gotta be a way to just download all the necessary files at once" didn't even bother trying to read the readme file until I could figure that out smh

[cosmic mural]

画一个屈家岭考古遗址内部照片，要求遗址内部的文物要能够体现出立春节气

[lusty beacon]

https://blog.fal.ai/flux-the-largest-open-sourced-text2img-model-now-available-on-fal/

[lyric owl]

Hello! What do I need to do in order to use Artisan?

[tribal fractal]

Do you need the source repository to use a model? I am trying to download the 4d video one and the source for that is a StabilityAI>generative models github, link here: https://github.com/Stability-AI/generative-models

On that page, at the bottom, it gives installation instructions. Long story short, it wouldn't download because I need to integrate python into git. Do I need to learn python just to download the source repository so that I can load the model?

[pale latch]

if you wanna play around in this field, you should probably have some cursory knowledge of python. Not enough to code in it, but you should have the basic familiarity with python environments and execution

you can avoid learning it in some cases, like many UIs hide the python and you dont have to learn. But at the bleeding edge, like fresh models that haven't been integrated into any UI's yet, you'll need the tools

[tribal fractal]

Like to use stable diffusion at all or to download the source repository?

I know it's probably pretty basic to create a directory to have a virtual server, I just would rather get into the video creation. I'm barely downloading my first stable diffusion models rn.

[tribal fractal]

Hmm. I was planning on using comfyui. I don't know if I can use that for the 4d model though?

[pale latch]

not a virtual server. virtual environment. hmm, the distinction seems obvious to me, but i am thinking you don't know what a system environment refers to.

[tribal fractal]

Ah. Yeah I have no clue, this would be my first experience with coding. I tried an html course once in middle school iirc through khan academy, but I don't think either of us would count that LOL.

[sudden ruin]

#artisan-faq

[pale latch]

rough explanation and don't treat this like gosple. ok so your system has a global enviornment. that's your OS where all the programs run. your browser is a virtual environment. Javascript an junk can run in your browser, and it wont' do anything to your system. other programs won't conflict with every other webpage you visit. Lots of javascript code going on on every website, but they're all contained to a virtual environment.

Python is similar. You create a virtual environment for python project so that it doesn't affect the global system or any other python projects you have. contained in it's own environment

[pale latch]

i count it. it's experience. don't throw it out.

virtual in IT usually means "exists only in software" . virtual vs physical.

[tribal fractal]

Hmm. Ok. I'll try to learn how to do it eventually if I need to I guess. It's not like it's actually developing something with it, I'm hoping it won't get too complicated? My main question is just if I need to to use the model I'm looking at. Do I need the source repository to use the model? Because up until a few hours ago I'd never even heard the word repository. If I need to, I can try to learn what I need tomorrow or something, I still need to learn the image modeling so I can throw that into the 3d render that I'll have to learn to eventually put that in the one we're talking about.

All this just to make a custom dynamic wallpaper lol.

[pale latch]

At first it'll seem complicated , but then over time, itll feel like it's a part of you

[pale latch]

I'm not sure there are many tools for SV4D yet. All of this stuff is pretty bleeding edge still.

When new tech hits the market and you see the Apple Vision in the stores or a new PS5, that's the cutting edge. It's a term that migrated over from manufacturing factories. Imagine you print a big sign. So you leave whats called a "bleed". Your sign's image file should be bigger than the final planned printed sign. So when you cut it, the print goes right out to the cutting edge and looks super duper clean. The bleed is the colors that print onto the paper outside of the cutting edge. Those are often messy and will have a less crisp edge for many technical reasons.

That's where this stuff is right now. It's out on the bleeding edge. Hasn't been tidied up to be a final consumer product yet. This is the domain of tooling and experience. Welcome to the club.

also i love saying bleeding edge because it sounds badass and like, i'm all bloody from battle

[desert dagger]

so create an organization

[tribal fractal]

I wasn't sure if putting down a false or unregistered logo would have gotten me in trouble. Someone figured it out for me though, even though it says required you can put none.

[frigid star]

Hi all. Am I able to create "groups" when prompting? Say a car, a tree, and a person - can I organize the prompt to include all things about the car, then the tree, etc? For visual organization I guess.

[fervent thunder]

guys

will the rtx 4070 12gb run sdxl faster then 4060 ti 16gb ?

More cuda VS 4gb vram more

[cedar salmon]

my guess probably faster at very basic levels till you add like a Cnet or IPAdapter then the ram would be more useful

[pale latch]

4070 would probably be barely faster. not enough to justify trading 4gb.

[bleak matrix]

Good afternoon! How are we all doing?

[clear oyster]

when 800m sd3??

[grave veldt]

if i upload a image can someone tell me what is the model or style?

[full lark]

We can try 😁

[grave veldt]

i cant upload image

[fervent thunder]

is sd3 based on sdxl or its new version of sd 1.5 2.0 ?

[full lark]

You can on #🏞｜general-with-images

[fervent thunder]

https://www.reddit.com/r/StableDiffusion/comments/1ehiz51/flux_image_examples/#lightbox

this is insane

looks like i need to save for 24gb gpu

i was going for 12-16gb XD

[sage reef]

5090 waiting room

[desert dagger]

you realy want 32 if you can swing it

[fervent thunder]

Hopefully nvidia will release a 5060 ti sd edition with 32gb of vram

I am really suprised why they havent done it. but we all know why.

[frail iris]

Hey there, is there any channel specific for stability API usage?

[spark tinsel]

you know everyone who plays vrchat, cvr, neosvr, second life, former playstation home etc is thinking: Stable Fast 3D Launch?! I'm gonna generate blender-compatible models of my waifu from 2d images and upload them!

[drowsy robin]

sand

flux can run on 12gb of vram

they modified the "load diffusion" node in comfyui to allow it to run as fp8 weight type

[pliant jasper]

schnell

ITS TIME FOR SCHNELL

MAC GANG RISE UP

[fervent thunder]

what even are .sft files

sussy

[full lark]

vae and checkpoints, you can rename them to .safetensor (only the extension)

[fervent thunder]

flux really worth checking out?

ty btw

[full lark]

It depends on what gpu you have, In my opinion is one of the best models

but its extremely slow on my PC

[fervent thunder]

got a 4090

[full lark]

Damn then yeah you gotta download it

I´ve seen examples of it on par with dalle3 or even better

[quartz siren]

yes, and they also provide a 4 step version that is faster then sd3 or sdxl. I would recommend using 8-10 step on the 4 step one since that provides similar quality to the flux-dev. Even with 10 step, I believe it should be faster then sdxl or sd3(i think).

[fervent thunder]

idk i'm kind of burned out of all hype

too many crazy releases in AI lately, it turned into a constant flux of releases

[quartz siren]

one thing i love about flux are the fingers and hands, they are usually very good. humans overall are good too.

[brave matrix]

Are people able to train their own models with flux?

Also will they be on citivai?

[low moon]

thats an understatement. 😄

[pale latch]

the license allows it. and distribution is allowed

[pale latch]

this was no hype. they never said anything before today. black forrest just showed up and gave us new base weights like how beyonce puts out new albums

[low moon]

yeah thats amazing about these models they come out of nowhere

[fervent thunder]

yeah but lots of DiT models have been coming out

[low moon]

flux is like edogram at home

[pale latch]

from teams who aren't the og

imagine wutang just put out an album today

[fervent thunder]

whos wutang

[pale latch]

just some future hw for you. no big deal

[low moon]

of course the side effect of this progres is that the pc i got 6 months ago now is barely keeping up

:))))

[pale latch]

truth

[low moon]

if anything much beyond flux come si wont be able to run it

[fervent thunder]

how long do gens take?

[low moon]

around 90 secs rtx 3060 12 gb vram 32 ram

[real pewter]

hi

I'm new to this AI thing with stable diffusion, my PC isn't one that has the most advanced resources in the world but I'd like to know if there's a way to run stable diffusion on google colab without requiring an nvidia card, since I was only able to use the stable diffusion interface in the public URL once, but when I want to start it again it asks me for an nvidia video card, which I don't understand because the first time I used it it didn't ask me for anything. Could you help me, if it's not too much trouble of course?

[verbal osprey]

Just seeing the flux news. This is great! Does anyone know if it can be run through automatic1111?

[fathom marten]

Not yet

[verbal osprey]

I downloaded the models but I'm not familiar with the format which is .sft

[low moon]

its safetensors

i renamed them safetensors just in case and it works

[low moon]

Only on comfy. noodle flex moment.

[verbal osprey]

Ah ok thanks. I don't like comfy. Using comfy is like eating a bowl of spaghetti. It's just strings everywhere. I'll wait until it works in automatic.

[shell tendon]

with a1111's issues with vram efficiency you'll prolly need an a100 when it does lol

[desert dagger]

just hide the strings

[abstract linden]

Isn't there also a node that can make them straight lines too?

[desert dagger]

there's a setting for that, sure

[desert dagger]

click the little gear icon on the Queue Prompt toolbar, then scroll down and find where it says "Link Render Mode". by default it's spline. change it to what you want those links to look like. pick hidden if you don't think you need to see how things connect

[regal fulcrum]

Hi, I'm wondering what AMD/Intel GPU users use for cloud GPU/SD solutions. Rundiffusion, Thinkdiffusion, Paperspace, RunPod, Vast, Salad...?

There are so many platforms and so many GPUs offered that it makes my head spin.

[fervent thunder]

have used Rundiffusion RunPod Vast

[regal fulcrum]

How did you like/dislike each?

[fervent thunder]

runpod and vast are same
rundiffusion is doing something a bit different, you pay a lot more but the server is set up for you and is more reliable

[low moon]

and now stability comes out with the 3D stuff.

so this 3D stuff is interesting as it shows models are starting to understand things in 3D so thats the key to consistency later on

[drowsy robin]

This depends

Schnell, yes. Commercial or non-commercial. Schnell's got the Apache license.
Dev? No. Non-commercial distribution of fine-tuned or trained model derivatives based on the dev model is only allowed.

[rugged vale]

anything happening with SD3?

[lone hawk]

Black Forest just destroyed Stability, Flux is incredible

[lost yarrow]

I found a platform, 1 image generation costs 1 buzz, and you get 300 buzzes per day. This is great, and it seems cheaper than renting a server

[fervent thunder]

hi

[fervent thunder]

300 images per day is actually an incredibly small amount
it might sound like 300 is a lot
but $0.20 per hour on Vast.ai can get you 24GB, or sometimes even 40GB servers
which can put out hundreds of images in that hour

[urban cliff]

hello

[nocturne pewter]

Currently testing it and responds nicely to prompts. Qaulity is also great. I just have to figure out how to add negative prompts.

[low moon]

nah

it cant handle negatives 😦

[nocturne pewter]

oh, ok. Pity though, I could have used some negatives in my promting

[low moon]

someone tried and only got blurry artifatcs

like VHS tape effect

[nocturne pewter]

I got something similar, static noise haha

[low moon]

"Our model is so good you don't need negative prompts."

[nocturne pewter]

Yeah, I'm sure it's very very capable with all sorts of images. With negative prompt I could exclude something, negatives are not just for fixing things.

[low moon]

yes

I think the same that's why I wanted some negative prompt ability...

I am 97% sure it will come tho...

[nocturne pewter]

Agreed, most likely will come soon

[wet grotto]

flux run on SD?

[errant yacht]

Have you been comparing it side by side?

[pale latch]

Most distribution of models is non commercial. Everything on hugging face and civit.

What did you even correct?

[ornate folio]

Guys, I need some help with the Text Encoder learning rate... If I have my "Learning Rate" and "Unet Learning Rate" set to 0.00005, what should I set the Text Encoder to? I don't fully understand the relationship between these settings, what the Text Encoder does and how they affect each other, so I just need some general guideline

[pine fiber]

did u see sd3.1?

also flux is huge so of course it would beat sd3 (which im not claiming is any good in the first place), and people saying its prompt following is not as good as something like aura flow

[river sparrow]

Flux still has errors, for example missing limbs or body horror when women try to lay on the grass

[drowsy robin]

What I was going to add to that before I imploded

Was if you actually look at the licensing for the flux dev model

it very clearly states the outputs of the model are usable for commercial purposes

[pale latch]

1/10 times instead of 9/10 times

[drowsy robin]

What isn't allowed however

is making finetunes of the dev model and selling those

[pale latch]

It's better in most cases

[desert dagger]

a buzz? does your platform have a bee theme?

[plucky musk]

is there a colab notebook for a finetune of sdxl working? kohya crashes when starting the epochs on the t4 gpu free. loras work but I think the sdxl model finetune is not working

also does anyone know if its possible to "--medvram" for kohyaa notebooks in colab, or configure the gpu so it uses system ram fall back instead of oom?

[hardy nexus]

where is it ?

[pine fiber]

ill get it one sec

[pine fiber]

thread https://x.com/virushuo/status/1819097766255079734

[hardy nexus]

i saw this, but where is the anatomy displayed ? Nowhere

[pine fiber]

beats me

[hardy nexus]

it was the reason why they are doing a 3.1

which tells me it is not ready yet

[pine fiber]

they probably didnt fix it yet.. if they did fix anatomy they would be showing it off

it is the biggest issue because it doesnt matter so much if the image quality isnt perfect that can be fixed with some finetunes

but the base model has to be good at generalising things like anatomy

[hardy nexus]

datavoid fixed a lot of it in one epoch. Why don't they get him on the team to help

[pine fiber]

lol no clue probably something to do with their safety team

[river sparrow]

Yeah they said two weeks, when was that...?

[hardy nexus]

more than 2 weeks ago

[river sparrow]

[hardy nexus]

look at the data of the new licence than you would know

i'm guessing datavoid will release his sd3 finetune before them and it would be bonkers if it's better

[river sparrow]

Wut? Is another party now trying to finetune sd3, not waiting X amount of weeks??

Seems inpatient

Are the usual request finetuned, like boobs and anatomy?

[hardy nexus]

a single person that got good results overnight

makes me wonder wtf is stability doing

[river sparrow]

Lol

Im yet to find an easy way to make sd3 loras, not even thinking about finetuning the model itself.

Civitau seems to accept sd3, but no online lora training

[hardy nexus]

https://x.com/DataPlusEngine/status/1816252032929435850

Just achieved full rank finetuning of Stable Diffusion 3 Medium 2B on a SINGLE consumer GPU!

Only months ago, this was thought possible only on 80GB VRAM cards.

He said on his tweet. He is not using a lot of ressources for this since he trains a lot of models

[sage reef]

isnt that the same group that said they will release ella for sdxl a while ago, have they?

[hardy nexus]

yeah but it was hard to do, they trained other things since then.

[pale latch]

From a week ago. He should give up and fine-tune the true mmdit model

[sage reef]

I tried Flux and indeed it is awesome!

[hardy nexus]

he is just one person who is an electrician. Not even a ML engineer

[vale steeple]

If so you will never knew the potential of SD3 lol

[pale latch]

2b is garbage next to flux and not just due to parameter counts

[vale steeple]

... And not just its hardware requirement

[hardy nexus]

i still hope sd3 get fixed though. Flux is way to large

[pale latch]

Flux is the potential. Original team that did the original research

[vale steeple]

It is like, super great. But remember that Flux-dev is distilled version

The real Flux is even larger from what I can see

[pale latch]

So is sd3. Dpo safety training was a distillation process

[hardy nexus]

same number of parameters, same vram size, just faster generation

[sage reef]

pro version prob needs around 32gb vram or something lol

[hardy nexus]

no pro is the same size

[sage reef]

ah

[hardy nexus]

just longer inference time

[vale steeple]

Are you talking about the schnell version

It got Pro, dev and schnell

[pale latch]

Dev isn't a turbo distillation

[hardy nexus]

both are distilled. Schnell is more distilled

[pale latch]

They're different distillation process

[hardy nexus]

dev is guidance distillation, schell is guidance and step distillation

[pale latch]

Not more than one another

[sage reef]

i tried both schnell and dev, and i personally like dev more, schnell seems to produce a certain aesthetic or kinda bit burned images for my taste, idk.. maybe il try it again

[river sparrow]

If size does not matter, just throw away ML. Just put all existing images into 1 database, then do some interpolation 😅

[pale latch]

Only need 4 steps for Schell or it looks burned

[hardy nexus]

tried schnell on hugging face space but it's very different from my comfyui one. Maybe because i'm using the fp8 version and not the fp16

[vale steeple]

That will just caused a catastrophic amount of storage taken lol

[pale latch]

I didn't say size didn't matter. I said it's not the single reason why flux is the true implementation of mmdit

[vale steeple]

Plus it is better to just improve its training method

Rather than revamping the whole dataset

[sage reef]

flux finetunes 😮

[river sparrow]

I just wish they would make miniFlux for 12 GB vram

[pale latch]

Sd3 is what happens when inexperienced executives make decisions that force the researchers to handicap their work

[sage reef]

i tried flux on my 12gb vram laptop, it works

[river sparrow]

Huh, how?

[sage reef]

default comfy workflow, make sure you run in fp8 for both unet and t5xxl and it should work

[hardy nexus]

mine works on 6gb vram hurray

[vale steeple]

Say what you want, you can simply use your other 100% open source model in your disk happily

[river sparrow]

Better situation compared to rocket launch and ignoring engineering advise

[pale latch]

14 researchers left stability to build a proper mmdit model and effectively utilize their research. There is a ton of subtext in this event.

[sage reef]

[hardy nexus]

sd3 8b is good no ?

[vale steeple]

It take one minute to generate with 8GB card

So... not too bad

[sage reef]

but that minute is worth it 🙂

im personally waiting for 5090 to drop so i can finally upgrade and be free lol

i mean upgrade my whole pc

[vale steeple]

Can you tell where those researchers went to

[pale latch]

8b is good. 2b is good too. just not compared to flux.

Black Forest team built the foundation of sd3 too. They would've pretrained 8b

[pale latch]

Founded black Forest labs

[sage reef]

imagine flux 2 😮

[desert dagger]

@pale latch do you know of, or know any documentation of, what sort of videos specifically that SVD was trained on?

[hardy nexus]

have you compared 8b with flux with the same prompts ?

[pale latch]

They're being sued by shutter stock because of the watermarks so probably scrapped data

[sage reef]

woman lying on grass? lol

[hardy nexus]

yes

[desert dagger]

more specific - like how much of the dataset was vehicles, how much was animals moving, etc

[vale steeple]

Yeah I will just take it with a grain of salt

[hardy nexus]

does that work well on sd3 8b ?

[vale steeple]

Or 2B API?

[pale latch]

I did hand stand and composition tests on both on different days.

[sage reef]

i hope they release svd2 or something

[desert dagger]

agreed, and sometime soon would be nice

[pale latch]

Haven't directly tested but 8b isn't better. There are a few fundamentals that are extraordinary

No more pixelation blowouts past 1024 resolutions is big

[pale latch]

Confirmed by Emad plus all these guy's names on published papers

[hardy nexus]

8b should be enough for anatomy right ? And i don't think size has to do with anatomy since pixart sigma does better anatomy than sd3 and it's onmy 0.6B

[sage reef]

flux is very good at generating what you want first try, compared to so many other models, but of course it's not perfect, but that is ok, cause as a base model, that is crazy good as is

[hardy nexus]

than the 2b model anyways

[hardy nexus]

it is extremely good, way better than midjourney

[pale latch]

Flux is all the research that built sd3, and more, and no corporate meddling in the training process

[sage reef]

man, the good ol days with sd 1.4 LOL

we have come a long way

[vale steeple]

Generic NovelAI art, yeah

To nowadays AI porn infestation

[pale latch]

Black Forest labs also has 33m in funding while stability has 3

[sage reef]

wait i didnt even check, does flux have a research paper or technical paper? i want to read

[pale latch]

It's coming. They showed up and gave us weights with no hype

The blog post announcing it has some info

[sage reef]

i wonder how long until controlnets, loras, finetunes 😮

[vale steeple]

Pretty sure it is something coming soon

But I dont think it is any different from other MMDiT

[pale latch]

They're the guys who invented mmdit

[vale steeple]

Also pretty certain that they gonna training a what, v0.2? If they have the power to do it

[hardy nexus]

that was bonkers. I like that style very much. Like "here friend, take SOTA model, bye". Thank you ?

[vale steeple]

Or just went on doing optimization work

[pale latch]

Beyonce style

Auraflow 0.2 is out already. They is flux.1

[sage reef]

i tried also i2i with flux, does it very nice too

[vale steeple]

I mean Flux ofc

[hardy nexus]

auraflow is not that good. The dev said he had made so many mistakes. And it does not have 16ch vae, which makes it unhypable

[vale steeple]

Auraflow's aethestics is shit as fuck as for current

[pale latch]

They're making a sota video model next

[vale steeple]

Currently it is still about prompt coherency or adherence

[sage reef]

wait, what is the flux vae? is it 16ch?

[vale steeple]

Wait

[pale latch]

Must be

[hardy nexus]

yep

[sage reef]

😮

[vale steeple]

There have no attention for Flux's VAE at all I realized

[sage reef]

i mean the results do look good

[pale latch]

Flux is the sota currently

True SOTA open weights

[sage reef]

it's gonna be the thing to beat

[hardy nexus]

100% is 16 ch, it renders far away faces and text with high accuracy. THat's not sdxl vae. idk if it's 16ch but it's for sure a better vae than sdxl

[pale latch]

I'm gonna coom again brb

[sage reef]

lol

and this is just August 2024, year isnt over yet 😮

[hardy nexus]

i hope pixart devs don't dissapoint, they said they are aiming for something this year not the next

[sage reef]

but come on Nvidia... release 5090 already... i cri

yea i heard the next pixart model is coming

it's always nice to have competition

and SAI has yet to fix 2B version and release 8B

[hardy nexus]

i like what they have done with sigma. Such a small model, but it is so good. They aim for effieciency so their next model should aim 8gb vram cards

[sage reef]

yea

[vale steeple]

Guess it take time for them to recoil the entire management and training

[hardy nexus]

or amd, be cuda compatible with more vram. If only they were smart

[vale steeple]

You know after all these mess lefted

[pale latch]

Stability is still radio silent as far as I'm concerned. Not as cool of a stealth mode from what bfl did

[hardy nexus]

they only need one competent guy on it.

[vale steeple]

Obviously it's for good or bad we dont know a thing for sure

[hardy nexus]

yeah Lykon has not come here at all lol

[pale latch]

We know they've been up staged and have still said nadda

[vale steeple]

Probably got hit by new management after his controversial guidance lol

[hardy nexus]

maybe, what controversial guidance ?

[vale steeple]

When someone complained about SD3, he simply said "it's your skill issues, you should have learn how to prompt properly first before anything"

[hardy nexus]

btw pixart discord was very active. It died as soon flux came out

[vale steeple]

And it got blown off

[pale latch]

Still can't help but laugh at how a community of PC users with gaming hardware acted so fragile about "skill issue" comments

[hardy nexus]

yeah wtf. Now we can say, look at flux. Even shitty prompts work very well with 1 less clip lol

[sage reef]

i mean it's like a ferrari being released next to a toyota

[vale steeple]

He also have multiple... opinions with AstraliteHeart and his training method

[pale latch]

Ferraris can only be bought by a private club of people

[vale steeple]

But anyway he just got muted from ever talking about it.

[hardy nexus]

trying to shup up the mob with lies. When has that worked lol

[sage reef]

i personally dont have anything against lykon, to me, i still look at him as one of the top dudes for finetunes, so im happy with that, i can forgive some drama or whatever

[pale latch]

Astralight is a bit of a sociopath imo

[hardy nexus]

yeah i think he was just playing the role of the employe and could not talk bad about his company

[vale steeple]

RunDiffusion ( creator of Juggernaut ) unsurprisingly stand for neutrality

[sage reef]

Juggernaut Flux 😮

[vale steeple]

I gonna off and go back to sleep

[hardy nexus]

flux does not seem to need a finetune, only maybe one for further aesthetics to completely shit on midjouney's neck

and for some boobas

[pale latch]

Ponyxl has shota and loli content dpo trained into it. So Yeh. He pretends to be about unity but also hard refined CSAM into his "base" model. It's unfortunate to see such success in his efforts

[sage reef]

it's kinda funny really, sd 1.5, you generally need like tons of negatives for it to give you something nice, sdxl way less so, like very small amount (at least personally), and then flux you really dont need negatives at all LOL

then again, it still would be nice to have negatives i guess, just to have it, idk...

[hardy nexus]

not necessary when the model is good at prompt following

[sage reef]

i dont really follow the pony stuff (not my thing), but what is the current status, is the creator working on the next version?

[hardy nexus]

but sure it would have been nice if you only gen nude but don't want any. has happened sometimes with schnell. That model needs negatives

[hardy nexus]

so he says but not an actual update with which base model

[sage reef]

isnt schnell the german word for fast?

[vale steeple]

He is working for v6.9 based of SDXL

After he openly despise SAI for closing SD3

[sage reef]

6.9 of course it's that version... haha

[hardy nexus]

yes

[pale latch]

He's a business operator after all. Makes money by selling sex to children and grooming them. So disappointing

[sage reef]

i had no idea about that part

[pale latch]

Pony is embarrassing to acknowledge is such a large part of the community

It's a cartoon porn model. Who do you think it's primary audience is

Rule 34 joke grew out of trends like furniture porn . Then it just became groomer art

[sweet token]

is there evidence of him grooming?, or are you talking without anything to backup your claims

[pale latch]

Releasing a model that is primarily purposed to depict child pop culture in hardcore penetrative porn qualifies as grooming in my book

[sweet token]

"child pop culture" define that, what counts as child pop culture, if you say anime, then your plain ignorant

[pale latch]

You may disagree but I view you in a whole new light now

[sweet token]

what did i say?, i asked you to define child pop culture, i am from iraq i dont understand western culture much

[pale latch]

Sure

[sweet token]

can you define it then?

[pale latch]

It's self descriptive

You should know I don't argue this topic. I just judge people

[sweet token]

if you mean by child pop culture, children cartoons like my little pony, if so i guess i understand your argument there

[pale latch]

Because lolicon anime didn't qualify. /s

[fervent thunder]

jesus flux is so slow even on a 4090

[pseudo coyote]

Hi all

[river sparrow]

ok, im gonna try flux schnell on comfyui using 12gb vram and 16 gb ram ... first need to make some room on my harddisk

[sage reef]

what speed you getting?

[fervent thunder]

dev model with fp8 clip takes like 8-14s/it

using --highvram loading the full thing

in lowvram mode on the fast model it was like 1-1.3 it/s

only got 32 gig ram so i dont think i can go full model

1.4 it/s on dev model with lowvram option and fp8 clip but its laggy asf

[sage reef]

on my laptop, 12gb vram 32gb ram, im not using any of the vram arguments, just default, and fp8 unet + t5xxl,
i didnt check the it/s, but for schnell 4 steps, it took like 20 seconds i think, and dev 20 steps was around minute
or something like that, forgot exact numbers

il do more tests tonight

[fervent thunder]

Prompt executed in 82.07 seconds on 4090 dev model in lowvram mode fp8 clip

took forever for the vae to actually do something

14 seconds for the actual rendering of the image

loading the model and the vae took so much time

[silver inlet]

How long is it taking per image? I'm also using a 4090.

[fervent thunder]

14-15 seconds on lowvram mode with fp8 clip and 24 gig dev model

if i load the fp16 clip the low ram i have is not holding and itll lag

[sage reef]

oh you using the full model? not in fp8?

[fervent thunder]

yeah full dev model

not quantized

https://comfyanonymous.github.io/ComfyUI_examples/flux/
the ones linked here

[silver inlet]

Hm, just compared our times, and yeah you're right compared to XL it is long, but after coming from a 3080 on my previous rig with SDXL taking 30 seconds to gen in the past I'm not complaining TOO MUCH. Haha

I'm sure it will get faster as we optimize.

[fervent thunder]

just the model is large, i hope it can be optimized and compressed more

[sage reef]

also, given how good the results are, it's really hard to complain 🙂

but yea we will get optimizations soon

[river sparrow]

yeah size is a factor. it currently is too large.

[silver inlet]

I'm sure the Pro model would make the 4090 look like an infant child.

[fervent thunder]

yeah, just the vram tho

[sage reef]

pro model secretly uses a 5090 card

[fervent thunder]

nvidia doesnt wanna give out more vram even tho they could

[pale latch]

they could, but they have larger markets that need supply too. supply and demand

the silicon forges are only cranking out so many chips at a maximum density

[sage reef]

i mean i dont expect that we will get 48gb vram, but i hope we get 32gb... i mean come on man... 😦

[fervent thunder]

i expect 48gb minimum for the 5090

if they dont its not worth getting

could get a second 4090 and get 48gb

[pale latch]

but why? 5090 cards are still technically consumer cards and what consumer commercial software uses even 24gb?

[fervent thunder]

bro have you ever played with pathtracing

[river sparrow]

flux

[fervent thunder]

pathtracing maxes out any gpu

[pale latch]

when the market is ready, cards with higher densities will come

[sage reef]

the next Cyberpunk sequel will use unreal engine 5 with realistic graphics they said, that means we need at least 32gb vram

[pale latch]

enterprise datacenters will probably drive purchasing in this arena more than home users

[pale latch]

delayed for 3 more years yet

[sage reef]

😦

just in time for unreal engine 6

[pale latch]

also for gpu market, and manufacturing supply, covid fucked up a lot of long term logistical plans. like building new factories.

we're still in the recovery period

[sage reef]

nobody talks about covid these days, because now they talk about the wars going on :3

[pale latch]

i'd recommennd saving for pro end hardware. you can get used a100s fairly easy on ebay

[fervent thunder]

yo wtf flux does nudity

[sage reef]

dont tell anyone

[pale latch]

lol. nobdoy had anytime to come up with any conspiracies about flux and it just showed up. its' such a shock bomb to the field

[sage reef]

woman lying on grass works too 😮 shocking the audience

[pale latch]

first woman in grass test i did failed. i like it a lot tho

[sage reef]

haha

that's because your woman was nude, admit it

[pale latch]

#🆕｜sd3 message

[fervent thunder]

15 seconds per full image is good, the images are high quality. but loading the model in and out of vram each generation to be able to load the vae

eeeh

anyways aint no way anyones gonna fine-tune flux if its this large of a model

[pale latch]

people will

[sage reef]

the top 0.1% will

[fervent thunder]

with what money

[river sparrow]

I tried woman laying on grass yesterday with the flux-schnell model, didnt really work

[pale latch]

all the good refines of sdxl are made by 0.1% anyways

you only need the best people to refine it

[river sparrow]

yeah, but some of the best people might not have the best hardware ....

so, size matters

[sage reef]

i personally like dev more than schnell, just gives me better looking pics, idk

[pale latch]

the money they have and you dont'? lots of people have enough money that they don't have to care about making sense of it

[fervent thunder]

for a good ecosystem to exist you gotta have users and with a 24gb model you only get users with the hardware

[pale latch]

this is a new field. i don't think there are any legacy approaches that MUST be done for a "good ecosystem"

[river sparrow]

good ecosystem is relative, either you are a company or person

[fervent thunder]

tools get developed by people and if the people cant use a thing they certainly wont make tools for it

[pale latch]

memory is tight i admit. i just dont see it as a deal breaker like you are portraying it

[pale latch]

you're projecting your own inability onto everyone.

[fervent thunder]

idk i just compare it with SD 1.5 that was easily available and because of that had people making stuff for it.

[fervent thunder]

okay relax mr hobby psychologist this is not what its about

[pale latch]

i don't see telling someone they're projecting like pretending to be a day one psych. i see it more like "dude there's a boogie hanging from your nose"

[fervent thunder]

you're toxic and i dont even wanna engage with you

[pale latch]

we all got boogers but when they hanging out, i'd like someone to let me know

you just declared that tooling will never come for flux authoritively, and i'm the toxic one.

[fervent thunder]

fact is if the model remains large and slow for most user, people will stick to what they know. if there is less users, less tools get made its not such a big deal

i think you're taking shit out of context and making a hyperbole out of it just so you have an argument and an excuse for personal attacks

[pale latch]

sdxl did fine when it had far less users and people were saying it was useless cause they couldn't run it

this model will just encourage lots of users to buy new hardware. i'll probably get a 24gb card in the coming months now

finally seems like i've reached the limits of 16

[fervent thunder]

even with my 4090 i wouldnt use this model much because of the time. maybe making one high quality image is something others do but i like to just explore concepts fast and iterate quickly. with flux i can't do that rn because its considerably slower than XL and 3

[pale latch]

right. i'm the one doing hyperbole

got it

[fervent thunder]

you're projecting your own inability onto everyone.

[pale latch]

useful to know. now i can practice that inability and get gud

ty for pointing that out

[river sparrow]

lol, no skill issue there

i guess some patience and understanding is not on any skill lists

... in general, not aiming for anyone

[pale latch]

patience is a skill that's often tested

[river sparrow]

hm. undesired instant testing, yes.

my downloads are faster than I can clean my disk, arg. inpatient.

[pale latch]

i'm a big believer of the gauntlet style of test. no training. just go out there and do it. run the gauntlet

[river sparrow]

yes, im also for confrontational therapy and methods. The real thing is the only thing

[pale latch]

learn to fail. its good for you

[river sparrow]

up to a certain level, yes. people need to be confronted, but not burned to the ground with traumas for the rest of their lives

[pale latch]

but she turned me into a newt!

[river sparrow]

we give some training on the job, and my colleague has the style of throwing them into the deep (with stones) and see how they swim without help. Just to let them know they know nothing, and should listen to him. That is a great way to shut them down for the rest of the whole training.

I mean, yes! They need to be pushed into the water and get wet, but not drown

[pale latch]

ghahahahaha i love that. allow them to experience how bad they are

[river sparrow]

its a good motivation to start any training! 😄

[narrow kernel]

flux uses a 16 channel vae

[fervent thunder]

i'm sure its really fun to be around and work with you two and these highly transparent and professional work ethics

oh wow a second 16 channel model

[river sparrow]

we're not working at the beach or at a pool, it's just a visual story

[pale latch]

effective training is actually way better for an employee than it is bad. it empowers them

[river sparrow]

I guess everyone who had some training in business context, has a certain feeling about semi-commercial trainings

[pale latch]

training is a process. you obviously don't leave them in the water

oorah

[river sparrow]

right, that was just the teaser. then you help 'm out, explain some stuff, learn to swim together, etc

[pale latch]

semper fi!

[sage reef]

i knew it

looks too good

[fervent thunder]

Open Sora 1.2 is 67GB vram
so open source diffusion models with big VRAM use has already been a thing

[sage reef]

i really want to read the technical paper

[sage reef]

man the last time i heard that was in metal gear solid 2 LOL

[digital tundra]

Hi, can someone help with generating text in stable diffusion?

[pale latch]

wasn't deepfloyd huge too? not 70gb but it was significant

[fervent thunder]

ah I don't know about deepfloyd

[sage reef]

i only know pink floyd

[pale latch]

the us marines use it and they train in water hard.

[sage reef]

yea i think they mentioned that in the cutscene lol

[narrow kernel]

Yeah it does, wish I had more vram to run it at normal speeds though. 15 sec/it on this old 2080 :/

[sage reef]

yea i cant wait to upgrade my pc

just waiting for nvidia

[quartz siren]

it was good with short text and prompt following but image quality was ehhh. it was incredibly slow compared to sdxl as well. not worth it anymore now but was pretty impressive when it came out.

[desert dagger]

because it's got the same core issue that SD3 does, and obviously as it's not censored, it wasn't because it's lobotmized

[fervent thunder]

i'd only upgrade for flux if 5090 was released

and if i was rich

[river sparrow]

Uh, I think flux is censored? or trained not to show much nudity and other stuff

[fervent thunder]

another thing is that
the new model is ret flow again so no SDE samplers

[desert dagger]

sigh. no

[hardy nexus]

woman laying on grass is 50/50 on schnell.

[hardy nexus]

so half of the time is perfect

[desert dagger]

yup. Because the issue is in the core of the model, not something that happened after it trained

[quartz siren]

just do 6 steps instead of 4 and voila(no prompt enhancing), its pretty nice

[sage reef]

did any of you try if flux works with just clip and not t5?

[hardy nexus]

but otherwise with other shitty prompts like this anatomy is good. However this is the turbo model. Dev model is better

[hardy nexus]

T5 is his power

[quartz siren]

i was just gonna try that right now

[sage reef]

im not on my laptop now, will try tonight, cause i just want to see results and differences

[quartz siren]

quality will most likely drop a lot, the other text encoder is just 246mb i believe

[sage reef]

but of course t5 gives it the real power

[pale latch]

its based on al the research that sd3 was built on, so i suspect it could potentialy load with t5. i dont know if that's part of the distilled dev version though.

if it can't be done, but then someone makes the model run without t5 encoder anyways, i have to wonder if that violates the terms of use. restrictions include not circumventing limitations that are designed into the model

[desert dagger]

why would switching out the encoder violate anything?

[pale latch]

if the model doesn't run without the t5 encoder, but someone makes it do that, that could be viewed as circumventing a limitation. it's just an ambiguous case that the licence leaves open

[fervent thunder]

isn't it the case that loras often train the text encoder anyway?
which is why they have conditioning noodles in comfy

[pale latch]

you're not allowed ot circumvent limits designed inot the model it says

[fervent thunder]

could be seen as jailbreak i guess

[cedar salmon]

if it allows its not limited

[pale latch]

mmdit loras shouldn't need to train the text network anymore if i understand things right

t5 has self attention, and there is a text network running parallel to diffusion with cross attention

[fervent thunder]

is this the license for dev?
and schnell is a nicer license?

[pale latch]

schnell is apache license. truely libre. free beer

[fervent thunder]

I could still see some use for fine tuning T5 to add some vocabulary

a lot of these models are just missing terms

yoloworld thinks stormtrooper is hairdryer

[pale latch]

idealy, the model just has to learn the term and the token can associate to it. t5 might not know 1girl, but that'll get a token still

also, in an ecosystem where every model has a carbon copy tenc, situations like pony breaking compatibility and calling itself a new base model won't hapen. academically embarassing imo.

[desert dagger]

switching out the encoder isn't circumventing anything. they don't want you trying to jail break it

[pale latch]

ianal but the legalese is prety ambiguous there

there are restrictions that exist

[desert dagger]

the legalese is written the way the lawyers require. it's not ambiguous to a court. however i doubt they're going to be going into people's homes or checking their posts online to see if someone jail broke it. it just covers THEM if you do somethign that lands them in court

[cedar salmon]

but its unrestricted, there needs to be an attempt at locking it down to be a limit

[pale latch]

sd3 is designed to allow t5 to be removed. if flux doesn't allow it to be removed, they specifically changed that flexibility into a limitation. if soemoen else comes along later and publishes a way to remove the t5 and run it without, that is circumventing a limitation.

stability probably wasn't ever going to enforce 2000 image limits but they put it in there anways and it was a huge concern. restrictive licenses are concerning for mulitple reasons.

[silver inlet]

I gotta admit, it's strange the irony of Flux's release post SD3 is considering the people who made it. Almost poetic. lol

[desert dagger]

just pull it into comfy and don't give t5 any tokens. boom. you've removed t5. now what?

[cedar salmon]

go to jail

[desert dagger]

not even

[pale latch]

it is poetic for sure. life is beautiful.

[river sparrow]

I believe thats how I use sd3 anyway

[pale latch]

it's still in memory what do now?

[desert dagger]

let's back up and look at timeline here - SD3 is in process of training. Robin, dev lead for SD3, gets mad, and quits. emad quits. now robin releases flux. do you think this is an accident? wht tech do you think he took with him and used for flux?

[cedar salmon]

his brain

[desert dagger]

he took, and is using, the SD3 tech that he developed

[silver inlet]

All of those are really really good questions.

[desert dagger]

might be in memory, but it's not doing anything, so it has no effect

[cedar salmon]

well than surely we will see a lawsuit, till then it didnt happen

[desert dagger]

we won't. cause he owns it

[cedar salmon]

so he didnt take anything from SAI

[desert dagger]

the parting most likely went like this robin and sai agreed that sai could continue to develop with his tech, however they aren't entitled to anything new he creates and he doesn't have to give them anything else.

[desert dagger]

the format all along has been that the devs working at stability are really the owners of what they create, it might change now, who knows, but it's been more of an incubator than a corporation

[cedar salmon]

no one will work there

[quartz siren]

you have schnell which is apache 2.0 and truly open. Its fast as well. Faster then sd3/sdxl I believe.

dev doesnt have such a free license but its just too good so most people don't really care. hands/humans are great and prompt following is amazing with dev. Finetuners might care later on but most people who just want to use it dont.

if sd3 was like that, most people would not care too much about the license.

[silver inlet]

Yeah, the psychology of it at the end of the day is if SD3 had a better launch, the licensing issues wouldn't be overlooked but would be somewhat, "overshadowed", to say the least from the amount of interest it had overall, per se.

[fervent thunder]

most papers credit Peebles and Xie anyway
for DiT

a lot of this stuff was made in parallel by lots of different people though

[cedar salmon]

whatever model makes nicer overall stuff will always overshadow others, in 6 months again probably wont be using any of these

[fervent thunder]

I don't think the models matter that much yeah

on the scale of months

[silver inlet]

Considering how quickly AI progress is moving, whether it's LLMS or AI Art, stuff moves VERY fast. The whole scene can change within 6 months. lol

[fervent thunder]

SD 1.5 with ella, hi-diffusion, PAG, SAG, CADS, FreeU, Perpneg etc
is still really good

[cedar salmon]

im in a discord with alot of web generator people (midjourny, imagen ect), They pretty much didnt use SD1.5/XL/3. but now i see many using Flux

the draw of instant memes was too strong

[desert dagger]

If the comm7nity hadn't been toxic, an unfinished sd3 model wouldn't have been launched. They got what they demanded

[cedar salmon]

weak company

if thats the reaason

[fervent thunder]

to be fair stability did what open ai does and announce like 9 months in advance
which is naturally a bit annoying

its ok for open ai to do that because their actual goal is AGI and the products are just a means to an end though

but yeah waiting 9 months for Sora, for example, is also annoying

[silver inlet]

At least Zuckerberg is on the right track with Llama being open source.

Points for that one.

[fervent thunder]

ye

[sage reef]

will Meta start releasing image generators too?

[fervent thunder]

yes they already have

not open source though

[sage reef]

yea i was talking about open source stuff 😦

[fervent thunder]

maybe full chameleon

this one https://arxiv.org/abs/2405.09818

[upper panther]

The holy grail is image editing through LLM, of course. Chameleon would be a step towards that.

[silk condor]

Do I need something specific to run XL models or just accept that it will take longer time?

[fervent thunder]

there are quite a lot of ways to speed it up

[chrome herald]

Can anyone ELI5 how it is that SD3 and Flux use I guess the exact same stock T5 model? Is all the actual training in the transformer models then? I think I must not understand the relationship between them and the text encoders as I always thought the text encoders were the basis of knowledge of the original captions and whatnot

[fervent thunder]

in the feed forward layers

[sleek otter]

Just d/loading Flux.dev - is it worth all the "better than MJ" hype?!

[fervent thunder]

yes because bigger VAE

[desert dagger]

better is specific to the user. use it and see what you think

[river sparrow]

How many language models specialize in prompt generation?

[desert dagger]

none of them

[river sparrow]

they should hook up a language model in the feedback loop, automatic variations on prompt, guided by a feedback loop of some image quality - or rather prompt-adherence - grading model

[sleek otter]

FluxDev.sft and FluxSchnell.sft are both 22Gb. Are they both the same, or are both downloads required?

[forest apex]

no just one

[narrow kernel]

Schnell is a 4 step version

[river sparrow]

schnell means fast in german, it's the turbo version

[desert dagger]

"FLUX.1 [schnell] is a 12 billion parameter rectified flow transforme"

billion

[river sparrow]

from their webpage: "Trained using guidance distillation, making FLUX.1 [dev] more efficient"

[desert dagger]

i'm sure it did

[river sparrow]

yes, and I just ran a 12 billion model mostly using harddisk swap I guess ... 13 minutes for a 1024px img is not worth it. I need a bigger gpu

ill just go back and generate some ponies using sd3

[desert dagger]

you just need more vram, really

[river sparrow]

ill download some free vram

[desert dagger]

have fun with that

[oak latch]

is there a way i can get stable diffusion to play a sound every time it finishes a generation

[warped granite]

Is it me, or is Flux unable to do "abstract" art styles? 🤔
Even SD3 can do watercolor art, but Flux will always default to a detailed style. Especially for humans. (I'm using flux-dev with comfy on a 3090.)

[desert dagger]

no. that would be a system thing anyway

[oak latch]

well i figured it out

u can 🤦‍♂️

[river sparrow]

you really dont want to

[oak latch]

why is that

[river sparrow]

well maybe you like it, just try it. I would get mad with all the sounds.

I rather have some background music or youtube video

[oak latch]

ah yea ik ima get mad at the ding sound but i need it cuz every time i gen a image i forgot to check back

[river sparrow]

sounds that interrupt are never good, either you are focused or not

if you are focused on other thing, then that is ok too. they might need your attention

multitasking is a fairy tale, both for people and cpu(*cores). that's why we work in groups, or have multiple cores

[oak latch]

im focused on a few things but i just need to go back to a1111 change 5 words from a bot i made and back to what i was doing

[river sparrow]

once you generate a few images at a certain resolution and amount of steps, you'll know the typical duration. you can adjust to that.

it doesnt hurt if you're a few seconds late

[oak latch]

im usually half a hour late

and my runpod is just eating up funds

[river sparrow]

your what is doing what??

[oak latch]

what

[river sparrow]

Just kick the dog back into its cage or something. no just kidding. What is runpod?

[oak latch]

https://www.runpod.io/

online gpu hosting

for a 4090

70 cents a hour aint bad

[river sparrow]

so, you pay for being logged in, instead of paying per gpu/cpu cycle or what?!

[oak latch]

yea per hr im using the gpu

[trail lion]

I usually get somewhere between .22 to .24 on vast for a 3090

[oak latch]

im using a 4090

but also i tried vast shit wouldnt work well so i just used runpod and it works good enough

[trail lion]

yah, for me the tradeoff for a 4090 isnt worth it, it's double the cost or more but not double the performance

[oak latch]

not wrong but now i cant even choose a worse gpu

the worst one i can get is a 4090

keeping the same storage solution

[trail lion]

runpod is a bit more polished experience, but I tend to use vast, just because it's cheaper by at least half

[oak latch]

yea ik i started valuing my time and effort over that small money increase so i just took runpod

couldnt be bothered wit that shit

[dry raven]

can anyone help me out and list the top models for photorealism

[trail lion]

why do you want to creak it? what does 'creak' even mean? what are you trying to achieve? please provide more context before posting links

[forest trout]

Guys, don't click random links...

[trail lion]

or post them

[dry raven]

fr

plus bro spelled it all wrong

can anyone help me out and list the top models for photorealism

[river sparrow]

thank you for this puzzle!

it was great to see chatgpt try to crack it a few times

lol

chat relax, it's fine. it was a fine puzzle

ill give the encoded text here straight away, if not too much text

but wait, first some text above it. it might give some hints on it.

Listen closely, and you may hear the faint echoes of those who came before you—seekers of truth, adventurers of the mind, perhaps those who ventured too far and found more than they bargained for. Their stories are woven into the fabric of Vault Delta, a testament to the peril and promise that lies within.

"Dawg uq fql jnjq vr aaa Aaaaaa? Pufe xgnzmebk fhe sbjtbdck cokkhdt atn tcrp, bzjulsroqb mwk ouoaskqm bj pqayqlf. Geq addl jnjq vr aaef qlpsvhmep Fgtq Shnq dctmrul bkq vr aaa rzgcqazx’f splmcllp jkqaqapxo, m zqlyxp hzmdz vghl fm m zxhrzf mqf. Dvapcu klaeaa roq abpyb “Kalahn,” roub lkyffyyk akwibxcy’b gwzb gz wloae envwnu, napgus ah geq wdxmhqaa kfeclku qtya bbknbrzbz coxi."

[trail lion]

that seems like a subjective question. what is best to me may not be best to you. what qualifications are needed to make it the best? edit, best/top

for sdxl you can try realvisionxl, crystalclearxl, nightvisionxl, polyhedronxl, and realvisionxl...that's by no means an exhaustive list

[dry raven]

for me it means simular to Jugg V9 and Realistic Vision v5.1 and v6.0. something that looks more like an iphone shot and less cinimatic, more amature, capturing skin texture

@trail lion I appreciate the clarification on my question and I agree that it does mean a great difference to each individual

[trail lion]

cinimaticredmond is also pretty good, if you like skin details, etc (and btw there's also a lora for skin details'

[dry raven]

whats the lora called

[trail lion]

search "realistic skin texture" on civit

[dry raven]

thanks

[trail lion]

also consider using the style selector extension in auto1111 as the cinematic and hyperrealism styles can often help achieve some of that photorealism you're looking for

[tulip bison]

roop extension no longer can be use?

[trail lion]

I never used it but my understanding is it was no longer supported by the dev, and was succeeded by reActor, which I've also never used. I prefer ipadapter personally

I shouldnt say prefer, I'll say instead that I've been happy enough with ipadapter that I've never tried to seek out an alternative

[tulip bison]

I see, if there is an alternative then is ok, thanks

[trail lion]

ipadapter with the insightface model is really pretty amazing if you've never tried it

[desert dagger]

no. use ReActor instead

[tulip bison]

ok, thanks

[forest ice]

How can I upscale flux generations? Does supir work?

[thick epoch]

Is there a test/dataset like the MMLU (yes, I know it is flawed) of prompts for text-to-image models?

That or just a list of prompts ranging from simple to complex to test an image model?

[tribal fractal]

I'm trying to download requirements for stability AI's generative models rn. Currently trying to download invisible watermark. Is that a super long process to run? I found a github or something for using gpu vs cpu, but would it matter too much? Idk how long that specific process is.

Are there other requirements for the generative models that would be better to download for gpu instead of cpu? Just trynna see if I can speed up the process.

I posted a longer text asking for someone specific to help in the stable-video-diffusion channel if ya'll determine gpu use is better here and are willing to help with that.

Kinda have the same question for fairscale setup. Found installation instructions here: https://github.com/facebookresearch/fairscale/blob/main/docs/source/installation_instructions.rst

It didn't make much sense to me though cause it was kinda written for people who have more experience than a first-timer like me.

[astral snow]

Heya friendos

[fading crypt]

Guys, when integrating SDXL into an iOS:Android app, if there any policy in what a user can create? Limitations?

[echo drift]

You would need a certain model that has those limitations I think?

If that's your idea, not letting them create something fucked.

[desert dagger]

you're going to try to run SDXL on a mobile device?

[cold scarab]

guys I am using Xformer flash attention, generating 1024x1200 images on pony model, 12 steps dpm2m++ sde, it takes 3s to generate 1image, is this the maximum speed I can get? 4090

[desert dagger]

faster that it runs for me

[cold cape]

im trying to use SD3 on StableDiffusionWebUI and the quality is awful, im not sure why, ive given the exact same model in comfyui the same prompt the quality of the image is exponentially better

its the same SD3 model for both of the interfaces

[cold scarab]

it is awful, i tried yesterday, can't finetune to generate any good images

[cold cape]

webui?

[cold scarab]

yeah

[cold cape]

like bro

wtf is this

wait i cant post pictures

ill put it on imgur

https://imgur.com/a/yuhw1d4 look at this

exact same model, exact same prompt

but webui just cant generate a good looking image

[cold scarab]

maybe just because SD3 sucks, and webui is now not that compatible....

[cold cape]

tbf

ill try it with 1.5

also, what do i do if i want to use custom models with comfyUI?

[fading crypt]

Yes

[cold cape]

i tried downloading a custom model from civitai but it doesent come with text encoders so i cant use it with comfy

[desert dagger]

not sure if you ahve the same settings as i do on comfy, but if you do, set cfg at 4, and use the huen sampler and do NOT use karras - it doesn't work with sd3

[desert dagger]

i wish you lots of success trying

[fading crypt]

Is it hard?

[cold cape]

no comfy looks great

its webui that looks shit

[fading crypt]

Why what’s wrong

[desert dagger]

the settings are wrong. what sampler and scheduler are you using?

[desert dagger]

jsut let us know when you get it to run if you would

[cold cape]

not sure, ive just started it lmao

[fading crypt]

Is the API bad?

[desert dagger]

no, the APIs fine. but if you're going to use the API, then get sd3 not sdxl

[fading crypt]

It’s the cheapest tho for users, at $10 for 1000 credits I believe

[desert dagger]

okay, SD3 doesn't use the same neural network as the other versions of stable. you can't use most of the samplers with it, and you can NOT use karras as the scheduler

so check that first

[cold cape]

alright i set the scheduler to the same as comfy and the quality instantly improved

how do i use custom models with comfyui?

ive tried downloading one from civitai but it doesent come with text encoders so doesent work

infact, when i tried it it crashed my GPU 💀

[desert dagger]

fine tuned models are for various base models. so you can't jsut switch them around. you look on the page on civit and see what model they are for.

same with loras. you ahve to use them with the base model they were trained for

[cold cape]

the model i downloaded was for SD1.5

[desert dagger]

then you load an SD1.5 workflow, and find the model, and use it instead of base 1.5

[cold cape]

right

alright, thanks!

[desert dagger]

welcome

[cold cape]

https://imgur.com/a/TIHYuwy

take a look at this

exact same model and same prompt but comfy looks a lot more realistic

idk why

ive set the settings to the same as comfy

[pale latch]

hyper realistic is an art style. youve also got high cfg and steps on auto. your settings aren't at parity

[cold cape]

i just noticed

im setting them to the same now

[pale latch]

i kind of think both aren't very photorealistiic, but the auto one feels more hyper realistic style

[desert dagger]

your cfg is way too high

[cold cape]

ye i just brought it down

they are starting to look similar but comfy still looks more realistic

[desert dagger]

yes, well - comfy is a much better interface - why are you using the other one?

[cold cape]

because i was told to use it with custom models

guess not

ill load a SD1.5 workflow and use comfy

[desert dagger]

just put your models in your models folder and load them into comfy like you would any other model

who told you to do that?

[cold cape]

not someone on here

[desert dagger]

good

[cold cape]

yeah comfy seems a lot better

i guess its not as user friendly but i like it a lot more

it looks more polished

reminds me of that tab in blender that you do all the shaders with lmao

[tribal fractal]

For what kinds of prompts do you decide to use different schedulers or samplers? Do you use the same one for everything or vary it depending on what you're trying to go for?

[cold cape]

where would i get a SD1.5 workflow from

[weak smelt]

Has anyone actually tried the video and audio feature

[desert dagger]

i don't use the same samplers or schedulers for everything.

[cold cape]

to add onto this, i cant find one on the SD1.5 repo like there was for SD3

[desert dagger]

i have workflows i can DM you if you want

[cold cape]

if you could that would be great, thanks

[tribal fractal]

Ok. I'm still trynna download the models I'll be using but could I dm you with details about what I'm trying to do to see which samplers or schedulers you use? Workflows could be useful too, just downloaded the tripleprompt json you sent in another chat.

[desert dagger]

sure

[tribal fractal]

Ok

[lunar barn]

High-end British restaurant logo

[tribal fractal]

I'm trying to download ultrapixel, have some questions and wondering if anyone can help. I don’t have some of the libraries listed under requirements for stable cascade, I’m assuming I should download them?

[low moon]

I tried it

it froze on me

[tribal fractal]

Ok. How did you install it? Did you download the models listed in the readme file for cascade diffusion? And did you download everything under requirements? I'm trying to figure out if I need to or not.

[low moon]

i did everything as i was supposed to

but when it came time to generate it just hang...

fomcfyui just gave up

[tribal fractal]

Dang

[dusky current]

hellow

[floral oar]

am available

[vagrant turtle]

how do i use the ai????

im new…

[low moon]

Well

What do you want to do

[fervent thunder]

do you have a graphics card

[main snow]

When the diffusion is stable

[main snow]

Assuming you wanna make pron like 90% including myself: get web ui, pony xl checkpoint then a bunch of loras

[lusty beacon]

https://www.reddit.com/r/StableDiffusion/s/Sby5nw5tei

[fervent thunder]

we're on discord you can just link to the discord comment instead of linking to a reddit post with a screenshot of a discord comment

[sweet token]

I think that screenshot is from a different discord server

[fervent thunder]

do links to comments on other servers not work?

didn't know that

[sweet token]

yeah they don't

[fervent thunder]

ah ok

it worked in DM but maybe that's different

anyway I don't want to start any drama but the pineapple is wrong it is not impossible to fine tune

its a large ret flow model distilled from a second, larger, ret flow model
that should not be impossible to train

distilbert is distilled and it is one of the most fine tuned models out there, in LLM land

[primal crater]

so is flux worth downloading ?

[abstract quarry]

it's great

[low moon]

yes

[abstract quarry]

it's not as flexible in styles as SD 1.5 and XL but besides that it's awesome

[wet grotto]

can someone pls show me a controlnet inpaint for sdxl? I don't found it

[fervent thunder]

powerpaint v2 with brushnet

[sick flint]

It's kind of amazing that such models can be runned locally, even if it is quite slow. Makes optimistic that consumers GPU's can produce a crazy good stuff eventually, don't need Dall-E for that level of prompt adherence (it often follows prompt better than Dall-E).

[wet grotto]

thanks!

@fervent thunder man there an a lot of controlnets model that I don't know, I know only the basic like controlnet, T2A and CoAdapters

[fervent thunder]

yeah same I am not a big controlnet user

my favourite nodes are FreeU, PAG, SAG, CADS, vectorscope, rescaleCFG, dynamicthresholding, vectorscope
and I just play around with those

[desert dagger]

They will show up as unknown for anyone that hasn't joined the server

[fervent thunder]

ah ok thanks

[harsh fable]

Can someone help me?

I have a running instance but I want to use it in my python program

[fervent thunder]

huggingface diffusers library

[desert dagger]

explain, please

[pale latch]

a webui server? you'd just socket into it. it probably has an api mode that your program can interface with.

[harsh fable]

actually, it's not up yet

I'm in an ldm enviornment and I don't know what to do from here

I just followed the requirements on the github

is there a llama AI that can generate images

[fervent thunder]

are you in the original compvis repo maybe?

cos they tended to call it ldm

[harsh fable]

Yes?

https://github.com/CompVis/stable-diffusion

[fervent thunder]

ok I would not use this repo

[harsh fable]

Should I use the automatic111 instead?

[fervent thunder]

I would recommend huggingface diffusers library

but it depends on what you want to do

[tribal fractal]

Anyone in here have ultrapixel up and running and familiar with this error?

"When loading the graph, the following node types were not found:

UltraPixelProcess
UltraPixelLoad

..."

[fervent thunder]

it might be the case that using a pre-made UI in API mode is what you are looking for
rather than a library like diffusers

[harsh fable]

I just want to host an AI art generator for my python script

[next ether]

hey everyone, I finally have SD installed and have generated some stuff, mostly the "bad" stuff so far lol... I know this group is not for that, but how do I find groups where I can share and discuss that type of content?

[warm junco]

Then use Auto1111

It also has an api

[raven laurel]

hello is everybody using the inswapper_128.onnx?

[cold scarab]

may need some negative prompts and embeddings

[dim light]

Hi, please help.
#🤝｜tech-support message

[feral fox]

Has anyone tried running Flux on 16GB VRAM (4070 TI Super)? I'm in my holidays so I can't test it, but I want to know how long a image takes.

[fervent thunder]

i run it with 2060 6gb+16gb ram cuda set to system fallback + virtuall memory to 40gb

took 50 sec+ to run an image, rip ssd.

[floral oar]

hello any one into roblux

am into those

[feral fox]

Thank you so much!

[feral fox]

Do you use the Schell or Dev Model

[fervent thunder]

Schell

they can't make a channel for their main competitor lol

"why no midjourney channel"

[feral fox]

Ok Thank You!

[sudden ruin]

I dont think thats true lol

[sweet token]

No it was run by a community member before SAI took it over, and it caused a lot of drama

The server was being ran by Kaarssteun

[fervent thunder]

intersting news.. for how much got sold ?

[sweet token]

They didn't sell it, SAI contacted discord and discord handed them the server

[fervent thunder]

by force no consent ? really ?

XD

[sudden ruin]

When was that

[sweet token]

Just before the server launched to the public

During the beta stage of 1.4 if remember correctly

[fervent thunder]

this is funny and wierd , what would be intersting if they offred free life time ai generation with a super fast workstation.

to the org discord owner.

[sudden ruin]

[sweet token]

Kaarssteun did see it coming and didn't mind it, what caused the drama is that SAI didn't contact him and straight up just went to discord

Becuase SAI wanted the invite link

[fervent thunder]

this is crazy 😦

[sudden ruin]

This seems to be solved and there is no reason to bring that back really

[fervent thunder]

do not tell me they even blocked/kicked him from server without even making him a vip member ?

[sweet token]

Yeah it has been resolved for a long time now, but its just nice piece of history of the server

[sudden ruin]

No hes still around and has a special role

[sweet token]

Reminds me of the time where we would go to a special channel to request our prompt be ran by someone who had access to a early version of 1.4

The images weren't good or coherent for that matter, but we were super exicted for it since the only other T2I ai at the time was Dalle-2

[nova lily]

Besides hires fixing, anything you could do to like boost up 1.5 to sdxl+ standards. Sometimes im dissuaded from using 1.5s because of the low detail levels but they still have great creative ability or theyre just awesome checkpoints.

[fervent thunder]

Ella, Hi-diffusion, Deep shrink, FreeU, PAG, SAG, CADS, vectorscope, rescaleCFG, dynamicthresholding

you should quite easily surpass the details of the average SDXL image using SD 1.5 if you use these

[chrome monolith]

hi chat

i have stable diffiusion set up on my pc

if i want to delete it do i just delete the stable-diffusion-webui folder?

[nova lily]

those things are above my brain-grade. extensions?

[fervent thunder]

which UI are you using?

If its A1111 then I don't really know it

[dim light]

What is fp16 and fp32 model difference? Which one should I prefer?

[nova lily]

yep,
thats alright

[desert dagger]

how much vram do you have?