#💬｜general-chat

1 messages · Page 95 of 1

radiant tusk
#

in adetailer too

#

5 it/s without lora and down to 1 it/s with

#

Have you seen this error before?

nova zodiac
#

can't say I have..

#

in your webui-user file, do you have the line git pull in there?

wraith hinge
#

When will the bot be Fixed?

radiant tusk
#

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --opt-sdp-attention
set TORCH_COMMAND=pip install torch==2.0.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu118
set "TRANSFORMERS_CACHE=%cd%_cache\huggingface\transformers"
set "HF_HOME=%cd%_cache\huggingface"
set "XDG_CACHE_HOME=%cd%_cache"
set "HF_DATASETS_CACHE=%cd%_cache\huggingface\datasets"

call webui.bat

nova zodiac
nova zodiac
radiant tusk
#

Both lora and lyco has the same issue

radiant tusk
#

for some reason --medvram has actually made it go faster

#

up to 2 it/s from 1it/s

#

taking off --medvram and adding --xformers put it up to 2.8it/s

still glacier
#

why do you have all that stuff in your webui-user.bat ?

radiant tusk
#

Trying everything I can to fix my generation issue

still glacier
#

copy pasting random stuff from google wihtout understanding them is probably not a good idea.

radiant tusk
#

For sure

#

I'm at the end of my rope

#

I know I don't know enough so I can really only go by trial and error at this point

still glacier
#

start back with a clean .bat

 @echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers

call webui.bat
#

delete your venv folder

#

and launch

radiant tusk
#

9.53it/s without lora

still glacier
hollow oxide
#

I can't generate images, get an error "Something get wrong, try later“
So war yesterday, and now

errant pulsar
#

Още ли не са възкръснали ботовете?

true glade
#

I have encountered an error with DreamStudio while generating an image. The error message displayed is 'Something went wrong, please try later.' I have been experiencing this issue since last Sunday, up until 15:00 UTC. Could anyone provide any suggestions or status updates? #🤝｜tech-support

CC @sour lintel

nocturne pewter
#

Having a problem again with creating new face with InPainting and ControlNet Depth (midas and Zoe). For what ever reason SD creates an odd shadow over the face, this seems to happen every time, despite what settings I use. Ideas? Tried couple of different checkpoints, no change. I'm using SDXL pipeline all the way, including ControlNet models.

warm junco
sour lintel
bleak matrix
#

Good morning! How is everyone?

acoustic solstice
#

hi everyone, is there a tutorial somewhere of how i can run stable cascade locally on my GPU?

night spoke
gusty mist
#

Is anyone aware of a machine learning model that is capable of generating a Stable Diffusion-compatible prompt from an image description?

wintry stirrup
# bleak matrix Good morning! How is everyone?

Pretty decent! Any idea who may be able to help out those dreamstudio users among us that have been unable to use the tool since Sunday? Everyone who tries to use it is getting hit with a “error on our end” message

#

Actually, it appears to be working again lmao

bleak matrix
#

Yup! DS is working! 🙂

#

If you have an issue, I would reccomend that you e-mail support via the website!

night spoke
#

Anyone have a guide for altering python code from Stable-diffusionXL to stable-cascade?

unborn nexus
#

Does anyone have any good material on getting started with the newer version of dreambooth? All the turtorials i can find are a year old and seem alittle outdated

fervent thunder
#

Try chatpgt

mystic imp
#

Dreamstudio seems to be working. Just tried it and no errors

fervent thunder
#

Are there some decent web based stable diffusion text to image generators?

#

I am checking some, pricing etc

#

ok found some Xd

#

Why almost no one talks here? 🙂

#

Wonderss

leaden chasm
#

When will stable diffusion be back on discord?!

fervent thunder
#

Why

#

What kind of images are you generating?

#

I do political satire

pulsar lintel
fervent condor
#

someone said comfy UI has now suppoer for Stable Cascade but i don't see it working

#

it throws error trying to use official workflow

#

has anyone found a way to use it in Comfy ?

nova zodiac
zealous solstice
#

sd fills up my vram sometimes and hangs... interrupt button doesnt do anything i have to kill the cmd

fervent condor
pale latch
fervent condor
#

produces errors

pale latch
#

make sure to update

fervent thunder
#

hello guys i have a question

#

do stable diffusers like ai plugins on kritea etc online. so like i connnect to a big ass server and then the server does the image and then sends it back to me. so like all images are made off my pc and just sent to me?

fervent condor
#

oh i see those are different files

#

also why it doesn't use stage a ?

#

stage A is baked into b or c or what ?

nova zodiac
#

A is baked into B

fervent condor
#

ok

#

that makes sense then

pale latch
#

A is the VAE and i dont know why they gave it a ridiculously non descript name like "stage A" this time. they made the naming convention SOOO much worse

nova zodiac
fervent thunder
pale latch
#

and it isn't even an intuitive order. like, stage c loads first and a is last? what?

nova zodiac
fervent thunder
nova zodiac
#

Ah right - the krita one

fervent thunder
#

yeah my bad

pale latch
#

Immediate results from this research model reveals that the naming convention is just SOOOO bad and they shouldn't do that anymore.

nova zodiac
# fervent thunder yeah my bad

So theres two options with that. You can either use it locally which means the number crunching happens on your pc and nothing happens in the cloud; or you can use runpod which is a cloud pc and have it do the number crunching and it returns you the result

fervent thunder
#

local server managed by krita

nova zodiac
#

Yep - it creates a folder on your pc in the krita extensions area, creates an interface for doing the number crunching (its comfyUI under the hood) and uses the web server from that interface to do the work

fervent thunder
#

is it hard to do it locally?

pale latch
#

comfyui is local

nova zodiac
fervent thunder
pale latch
nova zodiac
#

It makes its own web server on your pc

pale latch
#

i can see how that is confusing

fervent thunder
#

damn im not knowledgable at this

#

why would it create a website

#

that is local

#

isnt a website meant to connect with the internet

nova zodiac
#

#oversimplified

fervent thunder
nova zodiac
#

Only need to connect to the internet to download models

fervent thunder
#

this is interesting man

#

i thought 10gb wasnt even enough to download a movie

nova zodiac
#

You can get a movie in 700mb these days fyi

#

The model isnt a collection of images, its a collection of numbers used to do calculations

fervent thunder
fervent thunder
nova zodiac
fervent thunder
nova zodiac
fervent thunder
cedar salmon
#

or a little brain

fervent thunder
#

that only the computer can access

#

all this time i thought these models worked with like massive gpus and you couldnt even use them on the best consumer pcs

nova zodiac
fervent thunder
still glacier
fervent thunder
still glacier
#

Turns out math and IT bachelors are not just solving 2nd degree equations and delivered something

pale latch
#

Will Smith is entering such a new comedy phase of his life. he immitated that spaghetti video! ITS SO GOOD i thought itw as AI!

fervent thunder
#

yo whys this ai making like 240p photos lol

#

nvm i think im mean to upscale idk. idk any photo terms

storm halo
#

im littel bit new to SD i have sd-v1-4. how do i get multiple versions

full lark
nova zodiac
rough comet
#

Genuine question, for those running higher end GPU's what's your average wait time with SD? Rather new to AI art generation so curious lol.

woeful maple
#

is there a prompting trick to have the character face a certain direction?

#

seems like it only vaguely understands when i put something like 'facing left' or right or away from teh camera and everything by default faces forward

brittle heart
#

will bot ever return?

sinful sparrow
#

Or looking away, that usually turns the eyes not the face
But that's for models I use so idk if these keywords r universal across all models

woeful maple
#

i'm trying in cascade. i think it depends on what i'm asking it to do. if its standing/walking it does better in different directions

fervent thunder
#

@novel ocean

fickle siren
#

I doubt the discord AI system will return, I think they turn it off to force money payments.

sinful sparrow
#

Yea idk about cascade, that works for sdxl and SD 1.5 based models

pale depot
#

you feel me

woeful maple
sinful sparrow
#

Hmmm interesting PepeThumbsUp

pale latch
#

the sdxl bot was only for helping select which model remained

nova zodiac
#

the bot would have been costing them $$$ to run, and the gpu has probably been swiped to train newer models

forest trout
full lark
#

I wish it was possible to share some "gpu power", something like nvidia folding home to train open source models using the community gpus

manic latch
#

does anyone know how to setup a good face swap? willling to pay $

fervent thunder
dim surge
#

could anybody explain the difference between faceswap and lora training

nova zodiac
dim surge
#

Thx

radiant tusk
#

Does hi-res work with animatediff? it always just crashes for me

solid rock
#

hey guys can i ask if anyone here's on forge webui - and got cascade extension to work in there?

placid sage
#

Midjourney can produce that trendy "Dark Fantasy" style.. I can't post pictures here sadly.
Anyone have any ideas how to replicate this with stuff from Civitai on Auto1111 or ComfyUI?

fair cave
stuck gazelle
#

What's the best smaller 9:16 ratio before 2x Hires. fix upscale?

echo kite
#

dang what's with half the loras requiring a civitai account to download now

#

even ones that were free to download without one before

placid sage
#

It's crazy but I actually get faster speeds after I'm done gaming then switch to SDXL.

#

It's like it's warmed up or something

echo kite
#

did something change recently?

#

ugh, it's looking to be more then half the loras now

#

out of like 12ish I tried only 3 were able to be downloaded without needing a civitai account

placid sage
#

what's the problem with making an account?

echo kite
#

half the ones that now require an account didn't like 3 months ago

#

also I don't want to have to sign up just to download a few loras

broken smelt
#

whats wrong with signing up?

placid sage
#

i love how he didnt answer the question whatsoever

opal hedge
#

Does anyone know what's coming after cascade? According to Emad's comments on reddit stability is cookin' something good

wary epoch
#

Is access to generate images now gone? Logging in after 9 months!

echo kite
#

I already said why

placid sage
#

hahah

echo kite
#

this shit

#

is exactly why

#

I want to minimum interact with the community

trim magnet
echo kite
echo kite
#

how is it "unreasonable" to be annoyed that it requires signing up to download half the models that could be downloaded without doing it just a few months ago

placid sage
# echo kite but bro, I literally answered your question

since you seem to be obviously confused i will help you out
i asked you what the problem was with making an account for a few loras. the answer is that you are blowing this minor inconvenience wayyy out of proportion. you put more effort into whatever this tantrum is

#

also this is the end of our conversation consider yourself blocked

echo kite
#

are all the people in this community like this?

trim magnet
#

not all of them but a few are one of those redditors that like to poke ppl for fun

#

ty for the gold kind stranger 🪙

echo kite
#

did the community get spammed or something

#

is it an anti-scraping defense?

#

the reason I am confused is it wasn't like this a few months ago

trim magnet
#

yea some lora creators dont like it when ppl upload them to other place so they lock them

echo kite
#

ok

#

thanks for an actually straight answer

#

yeah I will probably just not use those loras then

#

this convo basically proved my point why I don't like to interact with the community that much

nocturne pewter
cedar salmon
#

the name is no lie, getting sub second generation on 2 step model on 2080ti

ionic crown
#

Why is the bot down and how long is it gonna take to start working again?

dim thorn
#

gmgm

undone latch
#

Heyyy is anyone awake?

jolly berry
#

any idea when will the bot start working agai?

#

and which other image generators can i use for the time being that are as good?

undone latch
#

IDK but i am pretty new... I dont even know how to download sd :///

#

Can u help me about it pls?

jolly berry
#

you don't need to download it, it works on web

#

i've been using it on web

undone latch
#

Wheree

#

Playground?

gusty mist
#

Hey guys i have a question about stable cascade.
The website says following about the licensing:
Today we are releasing Stable Cascade in research preview, a new text to image model building upon the Würstchen architecture. This model is being released under a non-commercial license that permits non-commercial use only.

But the GitHub repo is under MIT license.
So what is correct?

fair cave
#

Hey guys! You might want to checkout Picasso on practal.ai if you want to use another UI.

still glacier
gusty mist
still glacier
silver zealot
#

i got this error when i get first sample of train
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xe9 in position 252: invalid continuation byte

grim aspen
summer minnow
#

Is there any folder where my generated images are?

grim aspen
#

doesn't work with wine T_T

rain arch
#

Hi guys how can I use stable diffusion pls

thorn willow
#

yes

karmic cedar
#

this is insanely fast (from a human perspective lolol)

swift crown
iron anchor
#

where is stable diffusion v1 channel?

grim aspen
#

Does anyone know a tag manager for images that works in linux? Something like: Stable Diffusion Tag Manager or BooruDatasetTagManager.

stone quest
#

Hi

#

Can someone explain to me how do you do the art thing?

charred wigeon
#

explain

#

there are couple of them available

#

u can use fooocus, there are plenty of style models integrated

charred wigeon
charred wigeon
teal osprey
#

what are alternative websites of civitai?

charred wigeon
teal osprey
pulsar lintel
charred wigeon
# teal osprey oh you are one of those.. okay

wdym? there is literrally everything available - iam using the base model, without any refiner or everything else on top and i can create whatever i like. stop getting offended, explain what u are trying to do and we will help

pulsar lintel
#

U could use paying sites if that’s what you are looking for

pulsar lintel
charred wigeon
#

@teal osprey explain what u are trying to do and we will help mate

nova zodiac
teal osprey
charred wigeon
teal osprey
#

there is no such folder there. i put mine to stable-diffusion and can select them on checkpoint selection dropdown menu

charred wigeon
#

Fooocus? ComfyUI?

teal osprey
#

i have no idea

charred wigeon
#

what did u download to use sdxl

#

link me the tutorial or whatever site where u followed the instructions to setup sdxl

teal osprey
charred wigeon
#

automatic

#

ok

charred wigeon
#

check this out, everything is explained there, if u still have any issues, come back

#

and it seems like, u are kinda "new" - i can recommend you, using "Fooocus" instead of automatic, there are a lot of styles and models integrated and easier to use

charred wigeon
teal osprey
#

if thats easier to use why not

charred wigeon
#

iam using Fooocus for image prompting till today

#

and Comfy for everything else

charred wigeon
teal osprey
#

i want to learn to use this program for trying out different facade designs on works of architecture and also to create videos with deforum that are also architecture related

vestal dew
#

I can recommend not using online image generation tools...stick to something SD

charred wigeon
teal osprey
#

i tried learning controlnet for those facade renders but looks like i need to learn alot more

charred wigeon
vestal dew
charred wigeon
vestal dew
#

what do you mean backfire?

charred wigeon
#

like "roasted"

vestal dew
#

No idea

charred wigeon
#

k

vestal dew
#

I tried generating an image of George Washington - it won't generate the image : Gemnini, that is

charred wigeon
#

all censored anyways, yeah

vestal dew
#

I doubt there will be mainstream backlash.

charred wigeon
#

they thing they can "censor" something. lmfao, they forgot the people running anything on their owm machines

vestal dew
#

Im going to try fooocus. I have been trying to get controlnet to work in SDNext, with no results

#

SD or any run at home can decide to censor too

#

for now, it isn't

charred wigeon
#

imo - and iam prompting shit 24/7 (not even kidding) - fooocus is the best way to go and its faster then anything else

#

u can prompt previews, to know if u are going into the right directioon, 32 images in seconds

vestal dew
#

SDNext claims to work with control. However, when I try it, it just zips my GPU to 100% and hangs

charred wigeon
#

and then use the higher quality render

#

to do the the final renders

vestal dew
#

Im using A1111, but I can't get controlnet on there beyond pose, depth, and ganny, and other than pose, they make weird images with weird colors, etc

charred wigeon
#

dont want to offend any A1111 user, but its shit

vestal dew
#

It's been a nightmare since SDXL released. It was miles backwards, and it still hasn't caught back up

charred wigeon
#

go mate

#

u have set it up in minutes

vestal dew
#

I am DLing now

charred wigeon
#

lfg

vestal dew
#

can I select the models I want for it

charred wigeon
#

just check the video, there are shit ton of styles integrated

vestal dew
#

yes, but can I choose the model to use

charred wigeon
#

ofc u can

teal osprey
#

can i use controlnet and deforum with focus installed?

charred wigeon
nova zodiac
teal osprey
#

also whats up with seed option? do i need to write any number there or keep it at -1? or do i spam random numbers

nova zodiac
outer crescent
#

but not identically on another computer 🙂 seed random is super similar but doesnt translate generally.

#

my computers fuzz is similar to yours but not the same.

teal osprey
#

can i use a guide image for inpainting rather than prompts? i want a specific type of image on the painted area to be stitched. is that possible?

charred wigeon
elfin relic
#

what gpu has the fastest training time for dreambooth fine tuning?

vestal dew
#

foocus doesn't let me select a model

charred wigeon
charred wigeon
#

then on the uppoer right models

#

and be sure having the model in the checkpoints folder

vestal dew
#

I did, it only has one model despite me adding models tto the dir

teal osprey
#

should i download each model file for controlnet on hugging face? the ones that have 1.5gb for each model or are there any other way to get the models for it

nova zodiac
charred wigeon
#

mates, i really love to help, but most of your questions are sovable by searching on yt - "fooocus WHAT U YEARCHING FOR"

#

xoxo

vestal dew
#

good grief, foocus is sooooooooo simple

charred wigeon
#

xoxo

nova zodiac
vestal dew
#

add a prompt, hit generate, after a minute ERROR

nova zodiac
wraith prism
#

I have been away on holiday for a bit and just returning now - I noticed the old BOT1-10 channels have disappeared. Is that hidden under a new role or a different channel?

hallow sable
#

what holiday...

wraith prism
#

a personal vacation

vestal dew
#

Got it to generate images. Foocus doesn't have negative prompts. The output without certain negatives is bad, at least for the images I am generating. Negative is needed. For example, SD tends to make men big muscle men. If there's a prompt for athlete - I am trying images of tennis players - they look like body builders playing tennis unless it has a negative like "beefcake", which removes that big muscle magazine model look

nova zodiac
vestal dew
#

the same prompt, with negative, gives me tennis players who look like athletes in A1111. In foocus, I get topless women who look like bodybuildres

craggy scarab
#

Hi! Im new using Stabble diffusion and i have a couple of questions/problems with the GUI. It is ok if i ask them here?

broken smelt
#

whats the worst that could happen

glossy drum
#

Hello all I need some face swap help please can anyone help me please just some doubt

steady lynx
outer crescent
#

absolutely nobody

#

how do i report shopify accounts?

steady lynx
#

Hi where is the channel for use / dream please ?

karmic brook
outer crescent
#

anyway, gui questions

fervent thunder
#

@light foliou gay

haughty cloud
#

So what exactly is dropping on “22 02 - 24” (the date in the Xeet by @wise stratus )? Commercial-use weights for Stable Cascade? Stable Video XL?

modern crown
#

can someone tell me what the score_9, score_8_up, score_7_up, score_6_up tags do?

#

ponydiffusion model*

nova zodiac
outer crescent
#

also thats the second time someone asked that exact question, is it an actual question or astroturfing

#

actually, almost word for word. i AM going to assume astroturfing

nova pilot
#

hello, i'm wondering if anyone can point me in a good direction for figuring how to to impliment or generally use the generative-model?

outer crescent
#

what would you like to do?

craggy scarab
nova pilot
craggy scarab
#

1- Im having problems with automatic1111

#

specifically this:

#

venv "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 601f7e3704707d09ca88241e663a763a2493b11a
Traceback (most recent call last):
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
main()
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\launch.py", line 39, in main
prepare_environment()
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 560, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Press a key to continue. . .

#

No names found and torch is not able to use GPU

warm junco
craggy scarab
#

Ok thanks

outer crescent
#

ill meet you there thanatos 😄

noble pewter
#

Are there any good pixel spite models other than retro diffusion?

nova zodiac
forest trout
nova zodiac
forest trout
#

I'm just saying it's an extension with models you can download to pixellize the image in post.

nova zodiac
#

I'm guessing the former doesn't try and throw as much tiny detail in

#

whereas the other one would be trying to pixelise that detail and may not end up with a better result... dunno - never tried comparing them

forest trout
#

From my experience if you try to do pixel art, the blocks usually aren't square so you might want to put it through that extension anyway. The style is likely different though.

nova zodiac
forest trout
#

We had a pixel art contest at one point and some people used models and some used that extension but I can't seem to find the thread anymore.

nova zodiac
#

but yeah, I was here for that and made a lora from the resulting images 😛

#

I used the huggingface space to achieve the same effect

forest trout
#

Rip pow channel.

#

All we got is dailies now.

#

I think I'm currently winning this daily 🙂

warm junco
#

I've done pixelart with models, can work good. Upscaling is tricky.
I also tried the Extensions, but I prefer the direct image generation

nova zodiac
forest trout
#

I mean it's good at pixel art for sure, the grid does have a "homemade" look to it since it's not pixel exact. It's fixable with that extension though.

warm junco
forest trout
#

Last time I used the extension you could adjust the pixels resolution too, but I haven't used it in months.

dim minnow
#

stable diffusion

nova zodiac
dim minnow
#

thank you

hallow sable
#

xbox please make me a elon munk making a sandwitch on the twitter, best quality, 4k, nsfw

#

xbox... ?

nova zodiac
barren walrus
#

I have been trying to understand the various policies regarding image generation and have run into a problem. I want to depict events that happened but were not recorded. Some of these depict war or catastrophic events and they don't seem to make it through filters. Here is an example prompt "scene of a roman crucifixion where everyone is going about their business as if this happens every day". That is blocked by Bing. Or "Adolph Hitler and Winston Churchill seated at a table playing a shell game while Josef Stalin and Franklin Delano Roosevelt watch". That never happened obviously, but it's purely metaphorical (you could imagine it as a sketch in a newspaper in 1943 accompanying a story). Suppose I want to tackle serious topics as an artist? How is this usually done?

forest trout
#

Any text to image application will apply all the text to all of the image so usually more complex processes would require you to assemble the image more precisely through tiled diffusion: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/

Basically it works like this: you select a part of the image and apply local text to that area (frog, pitcher, flowers, man reading book, etc) and this will blend the sections together into a single image. This is an extension for automatic1111, but there's ComfyUI versions of tiled diffusion out there.

barren walrus
#

ok. I will look into this. Thank you!

copper spoke
#

is civitai being slow for anyone ?

nova zodiac
ionic locust
#

Hello all, new here, hoping someone can point me in the correct direction. I am loooking to do some AI generated images for my nieces birthday, but am a complete noob when it comes to AI image manipulation and generation. Where is the best place to get some advice on how to go about this? My PC has already proven inadequate to do the images myself using stable diffusion, so i am at a loss as to where to go to even figure this out. Any help would be greatly appreciated.

prisma lava
minor iron
#

Hello everyone, please does anyone know of a good tutorial or workflow they can share/recommend of using stable diffusion to generate videos. I have seen some tutorials where videos are converted into png files and the files then transformed using imagetoimage, but I have not come accross a video that explains the process in details. I will be very grateful of someone can help me or recommend resources for this. Thanks. I am using stable diffusion automatic 1111 on google colab

hidden wagon
#

look up deforum

#

there's heaps of tutorials on it depending on what you want

minor iron
hidden wagon
#

No worries

#

Does anyone know how to make the last frame of an animation in deforum the same as the first frame

nova zodiac
nova zodiac
hidden wagon
#

ahh bugger was hoping to loop it in deforum as I've already set everything up with prompts and keyframes

nova zodiac
#

There are also cloud pc options to run your own ai generator in the cloud - runpod, vast.ai, rundiffusion, thinkdiffusion, paperspace, aws sagemaker, google colab

fair cave
minor iron
#

Thanks

nova zodiac
nova zodiac
# minor iron Thanks

Dont know enough about deforum myself to help sorry. There is a really good tutorial series on civitai doing a intro to deforun by Harrowed

hidden wagon
#

yeah I've seen that one, it's a good one!

fair cave
nova zodiac
#

Will do

fair cave
nova zodiac
fair cave
#

Any suggestions? SD's REST API masking seems to be a little quirky and returns a completely different image every time, ignoring the mask.

fair cave
nova zodiac
#

Oh - you asking about using the mask with the stability API?

#

Id have to look at the docs later to see if its even possible

dim thorn
#

gmgm

fair cave
static lance
#

There are also cloud pc options to run your own ai generator in the cloud - runpod, vast.ai, rundiffusion, thinkdiffusion, paperspace, aws sagemaker, google colab

nova zodiac
fervent thunder
#

Hello

wicked charm
#

somebody can help me to draw a scheme diagram about macromolecule material?

nova zodiac
hollow summit
#

I've trained a LoRa model for logo on Runpod, but I'm not sure it learned properly. Sometimes the model generates the exact logo, but it seems more like luck. I trained it using 7 high-resolution images. Here are my parameters:

Steps: 20

Epochs: 10

Batch size: 1

Rank: 128

By the way, I also did captioning and I use the SDXL base model. What could be wrong? Is it the dataset or the parameters?

grim coyote
nova zodiac
fervent thunder
#

How do I get rid of Shiny plastic skin in the Anime styled pics I make?

warm junco
spark garnet
#

please fix the bot already im bored

nova zodiac
glossy temple
#

Is there anyone who can build me a Workflow in ComfyUI? I will pay for it to make it happen, please DM me <'3

bleak matrix
#

Good morning, everyone! How are we all today?

spark garnet
#

🙏

fair cave
#

If I use stability.ai's REST API for text-to-image generation, will adding LoRA placeholders to my prompt text help, or are they just ignored, such as "lora:AndyLau001:1"?

forest osprey
#

If you like I could create you a custom one

glossy temple
nova zodiac
south flame
#

Heya, I am curious if there is a specific thread for self-promotion or offers for commissions and such here guys

stone hazel
#

SD V3, whos excited?

umbral heron
#

Need more examples of what it can do 👀

stone hazel
umbral heron
#

If it's just some check in the code like it was in 1.x then it's fine. If the model believes nothing exists under the clothes it's a different matter...

#

Though that also could be a sort of conspiracy that if the base model can't do NSFW it's impossible to fine tune it into it

#

It probably is possible, just needs more work

#

I think that's the worst aspect of "AI safety", the utter vagueness of it. Because if you say explicitly what you consider unsafe and will do your best to prevent that unsafety from happening, it might cause even more damage.

tight glen
pliant jasper
open flicker
#

SD 3 will eventually be open sourced right

pliant jasper
#

what i'm alarmed about is the closed initial release + the trned toward centralization + censorship

pliant jasper
umbral heron
umbral heron
trail lion
#

none of the models could do nsfw, that was trained in later by users

pliant jasper
umbral heron
#

Well, same happened to SDXL though, closed beta release first, then it leaks 😅 and then open for everyone

umbral heron
pliant jasper
umbral heron
#

I never did full scale training, only loras, so I don't know how actually hard it is to add fundamental novel concepts such as breasts if the base model doesn't know about them at all

#

I'd love to run a huge model that takes up all my 24 Gb and the output quality is just as increased

thorny widget
#

Stable Diffusion 3 is out guys!

#

how exciting is it!

#

where could we try it?

near wave
#

where that has been posted?

umbral heron
#

Nowhere, it's not open yet

brave matrix
#

Not out yet, there’s a preview waitlist

storm shard
knotty wasp
#

the waitlist seems a bit bugged btw, there is no feedback if it actually got submitted or not, so hard to tell if you actually signed up

tawny crown
#

yes boys

#

new drop

primal star
tawny crown
#

i got feedback

knotty wasp
#

what did it say?

tawny crown
#

i cant put an image here?

#

Thank you for joining the Stable Diffusion 3 early preview waitlist. You'll notified by email with an invite to our Discord server when you've been granted access to the preview. To learn more about Stable Diffusion 3, visit our blog here.

#

surely i get access tmrw for being so quick

knotty wasp
#

multiple people didn't get any, weird

tawny crown
#

lmao

#

can you resubmit?

knotty wasp
#

it blocks you from submitting the form again, but can probably refresh and try again

tawny crown
#

yeah try deleting cookies

trail lion
#

fine letting others be the early adopters, took forever for xl to get controlnet, community contributions, etc, it was painful for a while

tawny crown
#

you mean you wont bother with waitlist?

trail lion
#

yah, why bother?

tawny crown
#

fair enough

umbral heron
#

It's been 84 years

broken bloom
balmy rune
#

catwhaaa There's a waitlist?

buoyant moss
#

Diffusion transformer eh

#

This just got released as an open model with weights in the hugginface + diffusers support

wispy jasper
#

SD3 be like ""Our safest safety minded safe model yet, here are stable diffusion we really care about safety, and so we're so happy to that we can finally introduce a model that has safety in mind. Please be safe""

steady bolt
#

a

buoyant moss
#

I think that's a great thing to say

#

And if statements about safety are what it takes, then they should say it 100%

wispy jasper
frail wind
#

hello newbie here, what is the best settings for rtx 3060 on stable difusion

tawny crown
tawny crown
#

i.e. do u want performance or quality

wild leaf
#

Hello everybody, i am looking for free uncensored Image Generative AI, is someone aware about them ? what happened to unstability.party ? Thanks you all

cerulean carbon
#

The new model looks fricking awesome

#

And why are yall complaining about the safety? It took like 5-6 months at most for SDXL to be able to generate simillar stuff with 1.5.

#

The main question I have is that it seems we are going to have like a whole bunch of models at different parametre sizes.

broken bloom
broken bloom
cerulean carbon
#

That will probably split the community up too much. I wonder if the LORA's or CN's trained for one would be usable on the other ones.

lusty beacon
cerulean carbon
#

Half of the community still cannot adopt XL due to hardware limitations

buoyant moss
#

Yeah, stable diffusion is releasing their models for free

wild leaf
buoyant moss
#

What more do you want? The super duper censored DALL-E and Google Bard generators?

#

At least with open weights you can fine tune

thin stump
#

any model can create a cyberpunk cat?

buoyant moss
broken bloom
buoyant moss
#

Type cyberpunk cat and generate

fallen gale
#

omg stable diffusion 3

buoyant moss
#

ikr

cerulean carbon
#

This scares me

buoyant moss
#

the biggest SD 3 will be bigger than SDXL

lusty beacon
#

How can I access it?

buoyant moss
#

Sign up for beta

fallen gale
#

i want it so bad

lusty beacon
cerulean carbon
#

I wonder what the spec reqs will be

fallen gale
#

well its between 800m and 8b params

buoyant moss
fallen gale
#

so like anywhere between 6gb to 24

cerulean carbon
#

I was thinking about purchasing a 4070tisuper

buoyant moss
sacred jungle
#

gn

fathom island
#

Huge news, congrats team!

buoyant moss
#

For also a transformer diffuser

cerulean carbon
#

Now I am either gonna wait until the 5000 series or gonna purchase a 4090, depending when the model drops.

buoyant moss
#

I have a 4090 and it can run SDXL 50 steps in like 3 seconds

#

at 1024 res

cerulean carbon
fallen gale
buoyant moss
#

8B model is like 20% bigger than SDXL

fallen gale
#

unfortunately

buoyant moss
#

SD3 8B model*

cerulean carbon
#

Yeah I just wonder if Nvdia will have 24gb vram with any of the lower models

#

Ooh, then even the 8B model should be runnable with 16gigs given enough optimizations

buoyant moss
#

Maybe

#

It is unclear if the transformer part of diffusion transformers will mean higher memory use

#

As transformers have quadratic memory cost w.r.t. resolution

cerulean carbon
#

I mean in 5000 series.*

#

We shall see indeed

light patio
#

have anyone granted access to stable diffusion3?

cerulean carbon
buoyant moss
#

I hope they watermark the models this time, so we know who leaks them

umbral heron
#

Need 48 or even 60 Gb consumer hardware now

#

Hopefully 5xxx will offer that

cerulean carbon
umbral heron
buoyant moss
#

5xxx might offer 32GB if the rumors are true

cerulean carbon
#

I'm just hoping for a 24gb 5800

buoyant moss
#

50% in bus width for memory

#

So I'd bet 5090 is 32GB

umbral heron
#

Would be awesome

buoyant moss
#

But CPU offloading works

#

If you use the diffusers library

cerulean carbon
#

Damn cant wait

#

Offloading is too sloow

buoyant moss
#

Yeah it is slow

umbral heron
#

It's mostly for the LLM stuff, RAM offloading becomes a bottleneck

buoyant moss
cerulean carbon
#

Damn this week has been exciting

buoyant moss
#

Oof, I had made 30 thousand SDXL images

#

Guess that's down the drain

cerulean carbon
#

Ahahahahah for what?

fathom island
#

down the drain? No, time enjoyed wasted is not wasted

versed yoke
buoyant moss
#

For a dataset

cerulean carbon
#

Ooh that is bad tho

#

Rip

buoyant moss
#

I've collected 300 thousand Stable diffusion prompts from this discord

versed yoke
buoyant moss
#

Every single prompt that made it to showdown

#

And every single prompt that got voted #1/2/3 in showdown

#

300 thousand from last july until Jan this year

cerulean carbon
#

I mean you still can train that model

spark vale
#

Is there the private discord for subscirbers by now?

buoyant moss
#

u know what

#

We should scrape all of deviantart

lusty beacon
cobalt frigate
buoyant moss
#

That's Ada generation prosumer card

#

Slower than 4090

#

With 32GB ram

spark vale
manic narwhal
#

Was there a message / announcement here for SD3?

#

I found out from HN

buoyant moss
#

Waitlist here

lusty beacon
lusty beacon
buoyant moss
#

Biggest change is

  1. from conv diffusion to transformer diffusion model.
  2. Biggest SD3 is roughly 20% bigger than SDXL
  3. will be open weights
#

And a slightly different flow + score matching training objective

robust coral
#

wow SD3's prompt adherence looks fantastic in the demo images

manic narwhal
robust coral
#

hey sandbit! 🙂

buoyant moss
#

Hello masslevel!

hard kestrel
#

holy sh! its sandbit

buoyant moss
#

Transformer diffusion models should be cheaper to train. Pixart-alpha https://pixart-alpha.github.io/ claims just $26000 training cost to get to SD1.5 ish quality

robust coral
dawn tartan
#

Has Stability abandoned this Discord? SD3 announced but nothing in here??

buoyant moss
#

😮

#

I'm going to write a paper on the prompts

manic narwhal
#

That's really good wow

buoyant moss
#

Ooof I pray it isn't that good.

#

We have a paper on compositional diffusion

#

Where you have controlnet not restricted to the segmentation labels from ADE20k

broken bloom
buoyant moss
#

There's a reason, due to safety they used a dataset with all human faces blurred out

broken bloom
#

Oh

buoyant moss
#

... it is what it is, I'm not an author of PixArt, but know the authors

balmy rune
buoyant moss
#

Due to "safety" they couldn't train on too many pictures of human faces

balmy rune
#

☔️ one day they will realize safety is a myth

buoyant moss
#

To clarify, some of the dataset has human faces, just a large part of it is blurred out

broken bloom
#

Safety is indeed a myth

buoyant moss
#

I only thought it was interesting due to SD3 also using a diffusion transformer, and PixArt is the first open source diffusion transformer model with weights

paper mesa
#

Will SD3 be released as an Open Sorce model? Or it will be closed?

balmy rune
#

SD3 should be open-weights hopefully. Otherwise I have little care for it

#

Any non-open AI model is useless to me

buoyant moss
#

We just had a giant meeting about SORA yesterday. Discussing the likely architechture based on what we know in public and private details

#

50+ people, our conclusion was that SORA likely spent less than a year training, as one of their top 3 authors only joined OpenAI within the last year

#

And we agreed the likely architechture was diffusion transformer, likely with a hierarchical and first image conditioning like NUWA-XL

balmy rune
buoyant moss
#

A bunch of people working in machine learning professionally that know many people from openai personally

#

You've likely read our papers at some point in time if working on image synthesis

balmy rune
#

Very interesting.

#

That changes the whole perspective

fierce musk
balmy rune
buoyant moss
#

I don't believe the crap about SORA training on unreal engine, we think most likely is movies + short clip dataset, likely they labeled themselves

balmy rune
teal osprey
#

are there any downsides of merging similar contextual models with each other? i dont want to switch between them frequently

buoyant moss
#

To be a good world model, you need object and spatial consistency, SORA doesn't have that (yet)

balmy rune
#

Something about the SORA captions looks very GPT-4 Captioned. It has GPT-isms (utterly and hopelessly verbose)

buoyant moss
#

Well, I think there are probably many teams globally with the talent and compute to do what OpenAI did with SORA, probably 10+ teams globally.

But.... No one has the OpenAI video dataset lol

broken bloom
#

I believe sora has gpt4 prompt enhancer

balmy rune
buoyant moss
#

I doubt they'd make money with a world model (for robotics)

#

If they do text-to-vid, they can at least charge subscription fees

balmy rune
buoyant moss
#

My 2 cents, if it can fool 50% of humans then it is what they claim lol

balmy rune
#

Very good point. I keep forgetting that most people don't spend enough time at their computer to notice patterns in how GPT writes, how t2i has that AI sheen or anything like that

#

The way of the programmer with too much time 🙏

robust coral
#

I beg my clients not to send me emails expanded by chatgpt.

balmy rune
#

Use their weapon against them catlurk

robust coral
buoyant moss
cobalt frigate
#

been using SD1.5 and then SDXL for some time, few days ago I have finally subscribed (more to support them, I never expect to earn anything, since I don't nor plan to sell anything). didn't like that their later models weren't open in open-source sense (even if I mean only weights and ignore content limits), but they are only, well not open, but probably something like weights-accessible capable image models. after reading about SD3 and how important is the censorship for them, and discovering how Stable Cascade doesn't understand chest size (even for sfw images), I am disappointed. I read from several guys who train (both loras and finetunes), that excluding this kind of images is detrimental to how model handles poses, clothes, legs and arms in majority of images, not just nsfw. and sufficiently censored model might be very hard or impossible for community to uncensor. I am starting to consider unsubbing 😞 I don't like using discord, so just leaving it here

balmy rune
buoyant moss
#

I just subscribed too, just to support them (well, via clipdrop, which is owned by stability.ai)

hasty badge
crimson beacon
#

What's the VRAM req for SD3?

broken otter
fierce raptor
#

I am assuming that fine tuning will get over the censorship bump. Stability are just protecting themselves here considering the growing wave of political and social pressure building up around deep fakes

hasty badge
#

They actually sold clipdrop, that announcement is right under the SD3 one 😉

#

Good riddance i think, clip drop never made sense and their outputs did SAI more harm than good

runic temple
#

sd3 is bigger than xl right?

civic basin
radiant burrow
#

thats a big boy

runic temple
#

no😭

fierce raptor
#

Cannot wait to see how the new architecture supports video.

nimble matrix
#

gonna have to wait a few weeks/months until the community manage to make it run on 10-12gb

#

without having to wait 1+ minute per image

radiant burrow
#

it would be awesome if we can add data to it. For example, give it images of my dog and make it do all sorts of stuff

fierce raptor
nimble matrix
#

well, that's what LoRAs are for, and controlnet

glad oasis
nimble matrix
#

we'll just have to stick to sd 1.5 forever it seems

fierce raptor
nimble matrix
#

or wait until training an AI becomes cheap enough a small group of people can do it

final summit
#

release the weights 🗣️

nimble matrix
#

SDXL is censored too, isn't it? You need to use another model on top to generate NSFW images and even then the anatomy is usually worse than in SD1.5

dire perch
#

Howdy, ya'll. I was deep into SD for a while and tapered off around SDXL release. I was mostly exploring ControlNet, Parseq, Deforum, and building a workflow of animated texture sequences into Ae and Blender.

Anyone know a good place that curates updates and new releases so I can catch up on what's happened since?

#

Not just in SD, but the tools, addons, and utilities as well

fierce raptor
#

This is why I think SD 3.0 will have as its base model a censored product, but it will be fine tuned the hell out of to make it actually useful

nimble matrix
#

Yeah, but finetunes don't work as well as it simply not being censored

#

SD 1.5 still generates better anatomy most of the time

#

even if SDXL's overall image quality is better

#

Controlnet doesn't work as well with SDXL either

nimble matrix
dire perch
#

Thankfully there are lots of places to search

fierce raptor
#

You might want to have another look again, I can generate what I like from SDXL and while there are fewer specialised LoRAs than 1.5, they are being developed every day.

nimble matrix
#

I suggest trying to look for every "News" or "Resource Update" post that has atleast 200 upvotes, since those are usually the important ones

#

You can definitely do good images with SDXL, but that doesn't change the fact it struggles with anatomy more often and using Controlnet is harder than in SD1.5

#

also it struggling with anatomy more often doesn't mean it can't do anatomy, just that it's more likely to fail

fierce raptor
#

Anyway, we have to suck it up that 3.0 will be censored - the political environment will not allow otherwise unfortunately.

So, we will have to rely on others to create the foundational models we need.

nimble matrix
#

Yeah

#

the real problem is if SD3.0 isn't really open source

broken bloom
#

It should not be censored 😑

fierce raptor
#

Would that be the case?

#

I did not know it may not be open source

#

That would kill it for me

nimble matrix
#

In theory it is, but there's open source and "open source"

fierce raptor
#

Hmmmm

pine fiber
nimble matrix
#

if the code is open source but the weights aren't then it's not really open source

#

Removing NSFW images from the training data

fierce raptor
pine fiber
#

yeah but if they release the model weights we can just remove all of that

#

finetune it for nsfw and stuff they didnt include

nimble matrix
#

filtering inputs and outputs can easily be disabled when running the AI locally, but there's nothing we can do about poisoned training data

fierce raptor
nimble matrix
#

finetuning isn't as good as just training it with the correct data from scratch

pine fiber
#

even if we need to retrain the model entirely there are people willing to do that and have the datasets already

nimble matrix
#

but yeah we can do that

#

retraining the entire model would cost millions

pine fiber
#

for example there are companies like novelai that dont really censor anything

fierce raptor
#

So let them censor it (as long as released open source), then let’s encourage them to create a video service off the news architecture (similar to Sora). Then … we fine tune it

pine fiber
#

I think sora is a lot bigger than 8b though

fierce raptor
#

It will be via fine tuning and LoRAs of course

pine fiber
#

so video diffusion based on sd3 wiill be bad

fervent thunder
pine fiber
#

thank my gpu

#

I feel inclined to buy the nvidia 5000 series after investing in them so much lol

fierce raptor
#

Sora is going to be a massive gpu sink - just have a look at Open AI purchasing from Nvidia

pine fiber
#

yeah even dalle 3 is probably losing them money in chatgpt

#

they had to reduce the generations per prompt from 4 to 2 (or 1? havent used it in a while)

nimble matrix
#

i don't doubt nvidia will be a bunch of assh*les again and have RTX 5000 series have little to no vram

pine fiber
#

yeah probably

fervent condor
nimble matrix
#

just so companies and enthusiasts are forced to buy either the 5090 or AI-focused cards

granite abyss
#

Hi there!

fervent condor
#

fintuting will fix it though

nimble matrix
pine fiber
nimble matrix
#

another country won't

fierce raptor
nimble matrix
#

the EU probably is the one that will always be insufferable related to that

#

oh no you can generate a naked cat that's criminal

pine fiber
#

always gonna be a problem with open source sadly

fervent thunder
fierce raptor
#

Yeah, but I see it as a neutral tech. Since it is used for evil, we either accept the base models will be censored, or, we embedd tech to trace who created what.
Prefer the former

fervent thunder
#

OpenAI makes MSFT billions, just from the stock valuation

#

they can lose all the money in the world, and still make the MSFT stock go up

#

it's all about the image/perception

pine fiber
#

true

#

soon trillions perhaps

#

already happened for nvidia lol

fervent thunder
#

also Google has way more compute than MSFT/OpenAI

fierce raptor
#

Just disappointed I did not buy stock in Nvidia a year or two ago

fervent thunder
#

Google will soon have more than 5x compute

#

TPU

pine fiber
#

im up like 50% so pretty good

nimble matrix
#

That's not good for StabilityAI then

#

for the government and authorities they have a bad image because they are against open source AI

#

for the community they have a bad image now because of the censorship

pine fiber
fervent thunder
nimble matrix
#

Gemini 1.5 looks much better than GPT-4-Turbo

#

from what was shown till now

fervent thunder
#

internally, I bet Google's models are very close to OpenAI's internal models

pine fiber
nimble matrix
#

From what some ex-google employees their researchers were frustrated

pine fiber
#

I wonder what openai has internally

nimble matrix
#

that Google would never release anything

fervent thunder
nimble matrix
#

but apparently now that's changing since they don't want to lose any more researchers to Meta and OpenAI

#

Gemini 1.5 was apparently done all this year

pine fiber
#

yeah google sucks at actually releasing stuff

nimble matrix
#

since some of the papers were released in 2024

fervent thunder
pine fiber
#

makes sense

fervent thunder
#

I doubt it

nimble matrix
#

AI Explained has a video on Gemini 1.5 Pro

fervent thunder
#

Gemini looks like outdated research

nimble matrix
#

he shows that some of the cited research papers used on Gemini

#

are very recent

#

and they are fundamental to how the AI's context window works

fervent thunder
#

I would believe if you said Gemini just finished up being "aligned"

nimble matrix
#

so they couldn't just be "attached" to the AI later

fervent thunder
nimble matrix
#

yeah

fervent thunder
#

interesting

#

I still believe the leakers

#

At least for OpenAI, the leakers are saying that Sora is very old

pine fiber
#

nah they were being intentionally obscure

#

probably the initial model was first trained about a year ago

#

when gpt 5 tho

opal hedge
#

Any predictions on SD3 release dates?

versed yoke
gray wyvern
#

from Emad: Some notes:

  • This uses a new type of diffusion transformer (similar to Sora) combined with flow matching and other improvements.
  • This takes advantage of transformer improvements & can not only scale further but accept multimodal inputs..
  • More technical details soon
#

sounds super exciting

opal hedge
#

Hopefully it releases soon than 3 months from now so the community can make SD3 useful

versed yoke
gray wyvern
#

multimodal sounds very interestign

versed yoke
#

Yes

opal hedge
#

I'm conflicted with SD3. Either it's going to be totally awesome or another SD2.1

#

Hopefully there's no forced censorship

gray wyvern
marble plaza
#

how big is sd3?

frosty swallow
#

im trying to use this repose workflow i found but it says I need anything everywhere

#

I cannot find that anywhere

#

in the manager or on hugging face

dry stratus
#

Hopefully for a quick release

frosty swallow
#

there are a few bits I am struggling to find tbh

opal hedge
#

Yeah, the prompt following abilities of SD3 are insane

dry stratus
#

Does anyone here had his hands down on SD3 ?

opal hedge
#

I wonder how we're supposed to train it and if poor training ruins the prompt following abilities

gray wyvern
# marble plaza how big is sd3?

" The Stable Diffusion 3 suite of models currently range from 800M to 8B parameters. This approach aims to align with our core values and democratize access, providing users with a variety of options for scalability and quality to best meet their creative needs. "

marble plaza
#

800M ≈ sd 1.x
8B ≈ 3 × sdxl

reef epoch
#

@wise stratus out of curiosity with the model prompt ability, could you please generate for example " a clown balancing on a hug red ball with one hand, and two monkeys standing on each of his feet juggling green balls, and on the left side of the painting a tiger dressed in a velvet red tuxido playing on a jet black piano, and on the right side of the painting a baby elephant balancing on a small red ball wearing a playing the trumpet " .. this is for research commercial purposes obviously

nimble matrix
#

Or just generate two female hands but each individual nail has a different color or art, "1° nail is painted blue, 2° nail is painted yellow, 3° nail has a rainbow drawn on it, 4° nail has a drawing of vines" and so on

#

but i doubt any current ai can do that without it looking like a monstrosity

lusty beacon
#

anyway to get fast access after sign up

reef epoch
keen relic
#

That would be testing it to the extreme... even regular pose would be great. Like a two fighter holding correctly their 2-handed weapons while fighting.

opal hedge
keen relic
#

Or a boy playing correctly with a rubik's cube.

reef epoch
#

i think my prompt is perfectly reasonable

hasty sierra
keen relic
#

they didn't post will smith eating spaghetti nor a horse riding an astronaut, so I guess it's on par with D3 but not that good at prompt adherence

reef epoch
hasty sierra
keen relic
#

Or better, the triangle, sphere, cat and dog example was already better than D3

reef epoch
#

dalle3 allows me to be dyslexic comfortably

#

but dalle3 aint stable diffusion

keen relic
#

Isn't tht because they rewrite prompt? Such a layer would "guess" hug is huge

reef epoch
#

fuck it lets keep it as hug then

keen relic
#

(or they would have banned you for daring to ask for someone hugging a non-consenting object.)

opal hedge
reef epoch
#

when you ask it for van gogh style and it says noh 😦

fierce musk
#

@wise stratus yo bruh. Got a question if you don't mind. Is SD3 gonna be as responsible as Goody-2 is? If so then count me in!

keen relic
#

seriously about van gogh? Omg...

opal hedge
#

Does each copy of SD3 come with a helmet and a pair of shin pads? I can never be sure if it's safe enough

#

More seriously, I hope the censorship is like SDXL where it can be trained out

reef epoch
#

can it make a sandwich with specific layers?

fervent thunder
#

why are people pinging the server owner

balmy rune
#

GOODY-2 is the end-stage of Anthropic and OpenAI

reef epoch
#

because emad sometimes is cheeky and he likes to drop in sometimes

fervent thunder
#

pinging the server owner, in a server with 318,000+ members

#

dude

reef epoch
#

plus you gotta always shoot ur shot you know

#

or you miss 100% of the opportunity

reef epoch
pine fiber
marble plaza
#

I meant parameter size

pine fiber
#

800m-8b parameters

reef epoch
pine fiber
#

sdxl + refiner is like 10b

tranquil surge
#

T.T where download? :c

marble plaza
#

I heard SDXL alone is 2.3B

#

sd3(biggest version) would need 24gb vram

rough sail
#

I think the main thing for the community is if SD3 is censored or not. If it isn't, goodbye SD 1.5. If it is, we'll stay at SD 1.5.

pine fiber
#

if its censored we will uncensor it

marble plaza
fervent thunder
nova fable
#

It's not just about porn. It is about being able to generate gore, "disturbing" imagery, "offensive" topics, and a variety of other content that were also censored heavily in post 1.5 releases. They filtered out artistic styles, images of specific celebrities, etc which is why image quality sucks for so many topics. Censorship = missing information = bad image output

gusty oasis
#

Look lets stop beating round the bush, Do the wider audience really truly bother using SD if its not for pornography ? Lets not kid ourselves here and see how the internet uses it. What are you trying to achieve if SD dont have it ? Dalle is literally better in every aspect so what gives and why bother trying if you keep denying your purpose lmao

nova fable
#

I came here wondering the same thing re: censorship. I really hope they release an uncensored model and let us decide how to implement guardrails ourselves in our own applications

keen relic
# gusty oasis Look lets stop beating round the bush, Do the wider audience really truly bother...

Honestly, customization of the image through img2img and inpaining in ways D3 can't. D3 produces great images if you're lenient about the content. When illustrating my RPG campaigns logs, NPCs are already detailed and I need the image to match every aspect of the description that was given during the game. I can't do this with D3, I need the framework of tools around SD. All my NPCs are clothed.

#

D3 can make impressive images of a towering orc leader on the battlefield wileding a 2 handed axe, but if I can't have him wear the crescent-badge of his clan, it's useless. I prefer having to work harder on SD and be able to inpaint the badge.

#

So, there are use case (maybe niche) outside of porn.

crude willow
keen relic
#

Also, were I live, it's not illegal to make caricatures of political leaders. D3's censoring is appalling in regard to this.

crude willow
keen relic
crude willow
#

its confirmed

keen relic
#

SD3 certainly, but D3 I don't think so

crude willow
#

oh I though d3 was a typo for sd3

keen relic
#

We're speaking about dall-E 3 (as to why we need alternatives to it outside of porn)

crude willow
#

my bad

tall ivy
#

Is SD3 posted in Discord annoucements yet?

vague pond
#

what I want to know is if the smaller models are distinct models trained from scratch, or if they're distilled models

gray wyvern
#

Imagine BCI (Brain Computer Interfaces) with SD3 and SORA like models ... that would be wild 🧠 🧙‍♂️ ✨ ✨

"...We've got some really exciting experiments running in the coming months using fMRI and ECoG, and then plans for multimodality + hopefully applying to new BCI tech..."
https:// x.com/humanscotti/status/1755021725077504027?s=20

tall ivy
fervent thunder
crude willow
fervent thunder
fervent thunder
#

Seems bad for your mental health

velvet fossil
#

hmm

#

you guys got me wondering

fervent thunder
nova fable
# fervent thunder Generating all that gore

I might want to generate imagery of a combat scene. I might want to generate a picture of a surgery.

And again, it's not just gore and porn and racism and whatnot. I literally can't generate pictures of celebrities, or emulate artistic styles because these fools removed everything that could potentially offend/upset anyone.

fervent thunder
#

I'm not pro censorship

crude willow
fervent thunder
#

I'm pro extreme decensorship

#

I'm pro absolute free speech of every kind

buoyant moss
buoyant moss
crude willow
nova fable
crude willow
fervent thunder
crude willow
keen relic
#

Free speech doesn't prevent accountability. You can yell fire in a theater. It's not the act of yelling fire that should be prosecuted. It's the fact that people were stampeded to death because of your false alarm.

fervent thunder
#

i.e., content involving hurting humans (which is not to do with reporting/news)... especially hurting minors

nova fable
#

Anyways, I digress. If censorship is going to happen it should be by application developers, not baked into the model

fervent thunder
#

literally

crude willow
fervent thunder
#

the law is a funny thing

nova fable
crude willow
nova fable
#

Not too long ago, segregation and slavery were legal.

charred mesa
#

point is, will we have nipples in SD3 or not

keen relic
#

If you're in a dictatorship, you're in a bad place. If you're in a democracy, I'd suggest not voting for them

nova fable
#

Eugenics was legal

crude willow
#

it's like you've never seen the modern world and is are just spouting the most aggresive stance you can because you think you're edgy and cool

fervent thunder
keen relic
#

It was legal because most people were OK with that. Then we progressed in our collective enlightenment.

frosty swallow
#

can anyone help me?

fervent thunder
#

What Elon did, was constantly perma ban people on Twitter/X, who never even came close to breaking US law.

nova fable
keen relic
#

And we successfully voted out eugenics.

frosty swallow
#

I am struggling to understand why when I clone a repository only the frame is copied and not the files

crude willow
frosty swallow
#

why is this not working like it works for others?

keen relic
#

Slavery, too, though I have heard that some countries had a little fight over that

crude willow
#

just name calling

nova fable
fervent thunder
#

Never trust someone who says he is a free speech absolutist.
Never trust anyone's word. Only trust people's actions, not words.

crude willow
nova fable
keen relic
#

if free speech absolutism can get me an invite to SD3 faster, I hail my new Elon Musk overlord.

vague pond
#

why are yall comparing a base model that doesn't have nudity to slavery, eugenics, dictatorships and segregation? 🤨

nova fable
#

moving on now ... are we going to have an uncensored version of SD3? Or will censorship be baked in again

fervent thunder
nova fable
vague pond
fervent thunder
#

Lets not devolve into using ad hominem, people? buh

nova fable
#

@vague pond Someone made the claim that ethical choices should be made based on the law. These things were brought up as examples of things that were unethical but legal.

fierce musk
opal hedge
ebon axle
#

damn this server and its gif thingy

fervent thunder
keen relic
#

Model censorship (I don't know about SD3 yet) is eminently political. I guess a model made by a Chinese company would be much more open than OpenAI's on some aspects, and ruthlessly forbidding on others. So it's an outcome of the political climate.

wintry stream
#

shh embed

#

don't spoil it

fervent thunder
opal hedge
#

I do want to try SD3, but it's not going to be usable for a few months because of the censorship

#

I can already do better with midjourney or Dalle

fervent thunder
#

Ever heard of Wuhan's animal meat market?

#

Freedom isn't always a good thing.

#

lol

#

I mean China does have certain extra freedoms.

#

Like selling random ass animal meat. (Yes I know they've started to police this a bit more.)

gray wyvern
fervent thunder
#

The Chinese government is very scared of AI though. For many political reasons you can think of.

keen relic
opal hedge
keen relic
#

I'd be scared of what the chinese government could do with AI, if I were there.

fervent thunder
opal hedge
#

Hopefully AI will teach people not to trust anything on the internet, as they should have been doing for years

#

But I know people will still be taking twitter posts' truth at face value even as the world around them is generated with a prompt

keen relic
#

It took centuries for people to realize that anything written wasn't necessarily true (OK, most people were illitterate)

craggy granite
fervent thunder
opal hedge
obsidian inlet
#

Trying to get my discord id to register in stable diffusion 3. Any ideas how from the phone?

past jay
keen relic
#

My take is that critical thinking toward news lowered in the last decades

past jay
#

From what I've seen they seem pretty confident they can censor things that they don't like with filters on their inference platforms.

fervent thunder
craggy granite
fervent thunder
#

You can be scared of nukes while building your own nukes