#💬｜general-chat

1 messages · Page 49 of 1

vast ingot
#

Not using artifact pictures would be best.
If you need to use some, i believe yes, adding just "artifact" would be best.

Each step of the training, an attention budget is given to your trained picture. Each token in the prompt gets trained, with the attention splitting along those tokens. It will try to put what it detects in the picture, in the most fitting token, and it will update the weights according to all that.
Adding more artefacts token means 2 things :

  • it reduces a little what attention was given to your main tokens, making the training slower
  • it prevents even more of the artifacts to be trained on the main concept token, good thing here.
unkempt hatch
#

ok, so maybe I'll just put 'jpeg artifact' on the pics with light artifacts, and add 'noise' as well to the ones with worst quality. Thanks

#

so the longer the caption, the slower the training...

vast ingot
#

Unless you put more learning rate, yeah

#

You have ways to fight that

#

Learning rate is the right to play with if you start to use longer prompt, but it's based on few tests so i can't be 100% positive either

unkempt hatch
#

I notice 'navel' and 'nipples' gets automatically captioned by danburu... I feel it's not needed, since everyone has these

vast ingot
#

It's different approach to training too. You can go with lots of tokens, this will be slower, or require to use as many tokens in prompt to have something similary good

#

Having lots of tokens has other good side effects too

unkempt hatch
#

like what?

#

what if some images have very long captions, and others short? Will it be a problem?

vast ingot
#

The more you have, the more weights you hit on the model, making it a little more fine tuning globaly. It helps as "inner regularisation" in that way, forcing the model not to just update a few weight and become imbalanced.

vast ingot
unkempt hatch
#

so I'm less likely to overtrain with long captions

vast ingot
#

But at the same time you need more steps

unkempt hatch
vast ingot
#

So it's a longer training with more monitoring of the loss, or comparing of the different checkpoints you made along the way

#

But a better global quality if all goes well, Since more token will fit the trained style

unkempt hatch
#

I've got some pics which are perfect, and they have shorter captions, and some with detailed backgrounds, lighting etc., artifacts, which require longer captions

#

around 1000 pics already captioned so far

vast ingot
#

Maybe putting the artefact pictures through SD ? When i have some blur on a good sample, i upscale it using latent or esrgan and then downscale it

#

It helps a lot

unkempt hatch
#

well, I didn't try upscale and downscale, instead I just did img2img....

#

I could try that... and maybe include both the 'clean' and original image in the dataset?

#

by the way, sometimes I have a large photo, and I'll include 2 versions of it in my dataset,

1 - downscaled, full body
2 - cropped head
3 - cropped body

Do I risk overtraining the image by doing that?

vast ingot
#

Yes you can very well have some overtraining there if dreambooth picks on the duplicated elements. Be careful about those

unkempt hatch
#

what if I caption them differently? e.g. with synonyms

vast ingot
#

you would spread even more the training

#

it should help with being its own regularization

#

just makes a lot of tokens to keep track of

#

but given the size of your dataset, you may be better going that way, like if you were full fine tuning

unkempt hatch
#

ya, I want to make something good

vast ingot
#

from having seen quite a lot of people around here with that same energy, one tip

#

don't wait too long for training. start baking. you won't ever really be able to understand beforehand what bias the AI will detect in such a dataset

#

it's already complicated on 20 pictures

#

I've seen people prepare around 12k pictures before trying, and finding that it didn't work for their specific case

#

and that going with 3 token prompts and 100 pictures landed better quality

unkempt hatch
#

problem is I send my pics to someone else to use his GPU... so it's another level of slowness

vast ingot
#

I don't want you to burn your wings if you know what I mean

unkempt hatch
#

but yes, I'm telling him to train with what we've got already 😛 thank you

unkempt hatch
vast ingot
#

because SD is already good at picking up things with 1 picture. when we human think there is variety, we are sometimes dead wrong, 2 pictures can be very close and trick us easily, not the AI.
So for example, taking 10 pictures of your face on different angles, that would be too many, it would overtrain/overconstraint the AI

unkempt hatch
#

even if you trained at a low LR and less epochs?

vast ingot
#

once you have given enough pictures for dreambooth to get what you need, every picture on top is multiple dangers :
1/ it slows the whole training, requires a little more step
2/ it can introduce a similar motif and create overtraining
3/ it can unbalance some specific concept that will have more example than others in the dataset

vast ingot
#

potentially going directly from "not enough trained" to "burned outlines on something that still isn't enough trained"

unkempt hatch
#

burned outlines.... so it's thinking vectorally I guess?

#

so what about training in batches? Like, instead of 1000 at once, 10 folders of 100

#

each folder containing a concept

#

until each concept is well-learned.

Then put them all together and do a few low LR epochs of all together for the final polish?

vast ingot
#

I used delay dataset batching once, to help balance the different tokens I was training. It does help but not exactly like you are presenting it

#

it first required me to train the whole dataset, and do lots of test to see on what epoch each concept was the best quality

#

and then start the training again, with only the slowest token at first, the one that took the longuest to train

#

and adding more and more other concepts when their time came

#

but those were separated concepts

#

your concepts are all mixed in the dataset

#

doing trainings one after the other (the first part of what you said), would make a model that moves and moves again

#

imagine it as a character on a world map. each training pulls the character towards where the sub dataset is

#

in the end, the little guy is on none of those and didn't really learn anything, having forgoten completly the first batch

#

I think this would just harm your model quality

unkempt hatch
#

oh no.... now I'm just really uncertain

vast ingot
#

well sorry, that wasn't the goal

#

I'm not sure what you are trying to train, so it doesn't help ^^

#

ok, you can remove it since I'm mod and the rules

#

but ok

#

well, I get the wanting lots of variety

#

and the problem of artifacts I guess

#

ok

unkempt hatch
#

and it's really foreign to the base model, which is vast majority female. It turns most renders into women, even if I write a lot of female words in the negative

vast ingot
#

training on only male will help here. how long is your list of important tokens ?

unkempt hatch
#

actually... I don't think I have such a list? I"m just individually captioning every image

vast ingot
#

this is one thing you are going to need/want

#

you have lots of concepts here

#

actions, ... no need to describe, you understand

#

you will want to balance the training on those

#

meaning you don't want 2 times more of 1 token compared to the others

#

so making a list, and having a count, subfolders, ... seems essential

unkempt hatch
#

ya, and even e.g. the face "open eyes, closed eyes, open mouth, hair color, eyelashes"

vast ingot
#

you can have specific pictures with this, on their own subfolder

#

I would treat this like it was a full finetune

#

take every token you need

#

create a folder

#

put around 20 picture presenting that specific token in it

#

and move on to the next folder

#

not overtag every pic

#

but I haven't done big finetunes like that myself

#

not on that scale

unkempt hatch
#

do you mean e.g. one folder with 'facial expressions'.

Or one folder with 'open mouths'?

vast ingot
#

so I'm also at my limits here

vast ingot
#

try to give each and every concept that same attention

unkempt hatch
#

I see, that makes sense

#

then train until it's got 'open mouth' perfectly trained. Then do the same for 'closed mouth'.

vast ingot
#

almost

#

train all in parallel, the whole dataset at once

#

no individual training

#

since every concept would be presented as much, with as many examples, it should do it

unkempt hatch
#

ah, and then give more attention afterward to the folder with the concept that was less-trained

vast ingot
#

don't use any duplication though

vast ingot
#

it's as if each concept was a cookie that took a different time to bake

#

so you need to add each one at the right time

unkempt hatch
#

that makes a lot of sense

signal dome
#

Where is sd LLM ?

unkempt hatch
#

if it's burned though, I could always mix it with another model to water things down, right?

vast ingot
#

but it would help yes

#

like, if you manage to erase the female bias

#

merging would sure bring it back

unkempt hatch
#

ok, thanks so much

cedar garden
#

This drawing I’m doing is turning out amazing (it’s of an ai art piece made by @ Yūū#2953)

left forge
#

I just trained my first lora and I must say it went pretty smoothly and the results were very satisfying.

cedar garden
left forge
#

It got me thinking about making something like a pipeline for training loras with specific characters. Like having set of images with specific shots and emotions. Kinda like separated character sheet images. I think that would make it easier to prepare references since you know exactly what kind of character shots have to be included.

hazy pulsar
#

I can’t keep up with the daily changes. What’s the fastest and least resource intensive: Lora, Locon, Loha, lycos? Too much info.

grim aspen
novel robin
#

is it even possible to make an automatic1111 dc bot?

woeful pelican
#

has anyone been on DeviantArt lately? it's filling up with AI porn and AI anime porn.

fervent thunder
#

with AI, everything becomes more effective, even the unpleasant part of society catsprout

#

will certainly make collecting good data more difficult, a few years and we have more generated stuff on the net than anything else.
Reminds me of pollution, plastic is great in use, but probably problematic for future generations. Will be the same with the 'data waste'. maybe

vast ingot
novel robin
#

nice

#

ty

deep jetty
#

Hi everyone just came here 🙂

#

Is there any gallery here ? Or not ?

wild steppe
deep jetty
#

Thank you very much.

broken gull
#

what is a model.pt file? can i use them like any other checkpoint?

alpine bear
buoyant jacinth
#

if it's small (less than 1gb), it's likely a lora. If it's real small (less than 1mb) it's probably an embedding

#

and each of those is used differently. If it's over 1gb, good idea to convert it to .safetensors

twilit parrot
#

DOOD I'M SO HYPED ON NEW TECHNOLOGY #CHANGETHEWORLD WOOHOO

broken gull
snow ermine
broken gull
broken gull
snow ermine
broken gull
#

folders setup seem to be changed all of a sudden..

snow ermine
#

think the issue is the age of that model and repo. There has been tons of updates to the dependencies and colab itself

broken gull
snow ermine
broken gull
buoyant jacinth
#

it's almost 3 years old.

snow ermine
supple solstice
#

Hey Everyone I Really Very Urgently Need Help With My App.
Is There Anyone Just Anyone Who Could Help Me With Integrating Stable Diffusion Or Any API That Can Simply Take 3-5 Images Of Users And Create It's AI Avatar, Sort Of Like Lensa.
Please Anyone Just Any One Help Me With It It's Extremely Really Very Important.

ashen urchin
#

Hi, somebody know any trust site where download after effects or filmora free without ransomware or a Virus?

supple solstice
#

Even If Anyone One Of You Know A Person Who Could Help Me Please Connect Me To Them Just Please Help Me Guys.

supple solstice
buoyant jacinth
broken gull
#

Ah thx I am doing it now with the old one

#

What are the differences?

buoyant jacinth
#

dunno, try them out

inner merlin
#

Hello. I'm trying to get control net installed and i'm having problems with pip install open cv command. (It's saying that pip is not recognized as an internal or external command. Can anyone help?

supple solstice
inner merlin
supple solstice
#

Did You Download The pip While Downloading Python 311.

inner merlin
supple solstice
#

Delete The Python Folder And Re Download It From Python.org

inner merlin
supple solstice
#

Python 3.11.2

#

Make Sure pip Is Checked While Downloading

broken gull
inner merlin
snow ermine
broken gull
inner merlin
inner merlin
# inner merlin Hey. That worked. Thanks so much!

This seems to have changed the location from where my webui-user.bat location and it's pointing towards the old python. I changed the location in that file, but when i double click it, it seems to be still looking for the old location and file for my python.

supple solstice
#

Recheck Your Code If Somewhere You've Mentioned Previous Location.

unborn hatch
#

Has anyone ever get experienced get trojan while doing preprocessing images?

hallow plaza
broken gull
#

still trying the older version

#

I have a result .png file tho

wraith night
#

I'm curious if anyone else is looking at the best way to create a training set of image. I've been using bing downloader in python to gather images up but what I'm trying to do is have specific things about the image, like pose, or other factors and use another package to train or determine how similar the images are. Looking through thousands of images is challenging when the hit rate is fairly low.

#

I've been using sewar, tried a few of the others, but believe sewar is the best so far, curious what others find

winter pike
#

yall SDXL 2.2 alpha is being tested in pickapic (randomly tho)

#

its a russian roulette between SD 2.1, dreamlike photoreal 2.0, SDXL 2.2 alpha and SDXL 2.2 beta

past edge
#

I have yet to see a convincing 2.x model

#

that isnt censored to shit

winter pike
hallow plaza
#

Are there any websites that will take pictures and generate prompts?

past edge
#

although it isn't particularly spectacular

hallow plaza
#

I'll google clip interrogation

past edge
#

however every now and then it throws out an artist name that imitates the style of ur picture. thats how I found out about Dave LaChapelle and Annie Lebovitz and they have become staples of my prompts

#

check general with images, thats where u find the button

hallow plaza
#

Huh? where is that?

#

looks like there is an extension called clip interrogator

fervent thunder
#

Jesus

winter pike
polar elm
#

Hello, anyone can check if deepbooru is working for them when on most recent Auto1111 ?

opaque lintel
#

How do I use CIVITAI do I download a base modle first?

barren tundra
#

hi, i'm searching chat bot, could you show me please?

tame notch
#

anyone know where to download expmix_line? civitai shows me an error 404 message

dusty pewter
#

rawr

vestal pond
#

What are tokens? Is it another word for prompts?

wild steppe
vast ingot
thorny forum
#

Hey! If anyone is interested I made a dataset of twitch emotes and their names! Could be used for training (it's free so i hope this doesn't count as self promoting)
Just excited to see what people can make with it! :D

shy thorn
#

What is the difference between models and checkpoints?

warm junco
fervent thunder
#

mj5 is pretty cool. i hope sd can return fire with some new model eventually

fervent thunder
#

where should I post to get help with a image generation?

dreamy igloo
# winter pike pickapic.io

Thanks for the introduction. Its kind of fun to play with. Only positive prompts though, but for what I was generating (abstract art/objects), I don't need negative prompts. But I don't know how to tell if the images came from SD 2.1, dp2.0, SDXL 2.2 alpha or beta

obsidian geode
#

Unable to use color control net? did anyone get shape errors too?

dreamy igloo
dreamy igloo
#

Since some of the photoreal 1.5 and 2.1 models surfaced in the last month or 2, I have not seen a mj v4 or v5 image that I thought cannot be achieved in SD. The difference is with how easy it is to do so.

fervent thunder
#

I think people oversell SD. I wish it was as good as mj5

dreamy igloo
#

we can do a lot of things that can't be done on mj5 like control net, latent couple, plus the comfy ui multi-prompt

#

for example, i can draw a skyscraper with a stone age base, to classical mid, gothic style on high level and cyberpunk style on top...all blended in gradually at extreme res in a tall portrait image thru comfy ui multi-prompt. You can't do that on mj v5

#

I can't even imagine how many re-rolls are needed on mj v5 to even get that to show such change and progression naturally. On comfy ui, I can get it right in a couple of goes, plus change the location of each style, % of style, script it so these value changes in a loop, and keep re-rolling for dozens of alternatives. Try that on mj v5

#

If I want to use ai to help with professional design work, this level of control is crucial. The comment I hear from many MJ user since v4 is how much control being stripped away since v3, and how bias the models are (they gravitate towards certain style or imagery keep re-appearing).

past edge
#

Did midjourney nerf celebs?

dreamy igloo
# winter pike pickapic.io

Is there a way to clip interrogate the png from pickapic to find out what model they were generated from? From past experience, you need to know which model was used to begin with, in order to clip interrogate for a prompt.

dreamy igloo
past edge
#

its more so for people who just want quick and easy

winter pike
past edge
#

I'd love to try mj v5 to see if its as good as people say but u have to pay to use it rn

#

unless someone wants to share some users

dreamy igloo
winter pike
past edge
#

depends on what you consider good

winter pike
#

the style is recognisable

past edge
#

My own blend can generate insane prompts with 4-5 words

#

but yeah you need to mix a bunch of models for it lol

#

base models are garbitch

dreamy igloo
#

unless I do like a very simple dog sitting prompt, sd 1.5 2.1 can manage an acceptable stock photo type photo

past edge
#

nah base models are never good

#

theres no good reason to use them

dreamy igloo
#

using like realistic vision 1.3

#

not base

past edge
#

oic

dreamy igloo
#

i can show some examples

#

in a few minutes

#

oh i can't upload image here

#

I go to general-with-image

past edge
#

yeh plz

hallow plaza
#

How come when I do dreambooth the samples have completely messed up eyes?

hidden anchor
#

What's the group profile picture?

faint parrot
#

are there any realistic LoRA recommendations?

hallow plaza
#

What does it mean if my samples from dreambooth are blurry?

willow coyote
#

say, what ranks as the #1 most realistic ai photo you've ever seen?

#

wait, ill post in the images channel instead

jagged plover
#

And it seems to happen only from my iPhone

#

Would anyone k ow how I can reach for the support or something ?

winged sparrow
#

Does anyone know which model dalle2 uses for inpainting?

peak spoke
#

if i wanted to make a train lora model of sailor mars do i need to find the same artist, style, colour or can i use manga imagine?

#

thank you

crude cove
#

Hey all, is anyone at the GDC this year?

broken gull
#

I'd like to know how to properly use controlNet, I have this character I am making but I need her in a different pose so I can work on a figure better without interfering with her arms. I have the character created already but would like it to have in a T-pose. Then possibly change it back in a different pose after doing the transformations.

#

Are there any good short tutorials that explains how to do this?

novel robin
#

how do i get SD ultimate upscaler? in automatic1111

warm junco
novel robin
#

kk

#

hmm, is it a script?

#

i think I've already install it lmao

warm junco
#

Its in img2img under scripts

#

Like the normal sd upscale script

novel robin
#

What does it do? Regenerate the image or just upscale?

warm junco
#

Depending on the upscaler used

novel robin
#

kk

#

but in extra there isn't a script so just for img2img?

#

like inpainting and outpainting stuff

warm junco
#

Extras tab is upscaling only for the resolution

#

The upscale + fix is only with Highres fix or the sd upscale script

novel robin
#

kk

unique glacier
#

how to repair the face when use full bodywaow

broken gull
quick holly
#

I have a question: does Stable know Chris Cunningham?

broken gull
sullen walrus
#

Anyone here got gpt plus/4?

austere coral
#

hey

#

I have an old gpu (1070 gtx)

#

if I will use a pod cloud service and rent a better gpu, will it make the rendering faster for animations?

midnight hornet
#

Hi sorry for my newbie question, whats the better way to do a SD install (For dummies)

pastel fern
#

why does emad hate openai so much

broken gull
pastel fern
#

but thats another company

#

he seems to really hate them lol

#

is that really the extent of his hate?

sudden lynx
#

venv "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Commit hash: a9fed7c364061ae6efb37f797b6b522cb3cf7aa2
Traceback (most recent call last):
File "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\launch.py", line 380, in <module>
prepare_environment()
File "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\launch.py", line 287, in prepare_environment
run_python("import torch; assert torch.cuda.is_available(), 'Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check'")
File "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\launch.py", line 137, in run_python
return run(f'"{python}" -c "{code}"', desc, errdesc)
File "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\launch.py", line 113, in run
raise RuntimeError(message)
RuntimeError: Error running command.
Command: "C:\StableDiffusion\SD512_TUTOROGRAD_EDITION\venv\Scripts\python.exe" -c "import torch; assert torch.cuda.is_available(), 'Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check'"
Error code: 1
stdout: <empty>
stderr: Traceback (most recent call last):
File "<string>", line 1, in <module>
AssertionError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

#

help

pastel fern
#

well

#

what gpu are u running

sudden lynx
#

rx 570

pastel fern
#

you cant use cuda for amd gpus

#

since they dont contain any

sudden lynx
#

and how to run without cuda

zealous yacht
#

My computer is blocking Interrogate DeepBooru in stable diffusion as it seem to be consider as severe threat. Never have this problem before and Interrogate DeepBooru still work fine for me few days ago. Did they change something in Interrogate DeepBooru files after update recently?

broken gull
# sudden lynx and how to run without cuda

You can't Xd Stable Diffusion needs those, you can however install Stable Diffusion on linux and then run it on your AMD device but it will be slow and idk how to do that.

twilit willow
twilit willow
#

It'll be very slow

sudden lynx
#

@twilit willowI can change where to do sho, I didn't have this guide

austere coral
#

maybe better to run on a pod?

quasi rune
#

hey, I just started making images with SD and I want to create more images with object/conceptual continuity. For example, a specific object / person / animal appears in an image and I want to make more images with it. Not looking for visual similarity necessarily - doesn't have to be same composition - but want more of that specific object.

Is there a technique to this? Or even a term I should look up? Haven't found anything yet. Closest thing I have found is fine tuning in dream booth and textual inversion, but those seem like you need several images to use for training.

Is there a way to start exploring an object or concept starting with a single image?

austere coral
#

any tips of how easliy I can turn footage into animation? I want to be able to take an mp4 video and turn it into animated version
is it possible?

twilit willow
dense anchor
#

hiiiii 😄

crude forge
#

hi

dense anchor
#

how are youuuu

#

you well?

fallow cape
#

Hi

rose bridge
#

Hello everyone, i'm trying to use the t2i style adapter, i'm using auto's ui, and a 1.5v model, i have the yaml into the model folder, but whenever i try to use it, i get this error:

" typeError: init() got an unexpected keyword argument 'linear_start' "

Can anyone help?

fervent thunder
# twilit willow It'll be *very* slow

yeah it is indeed pretty slow and frustrating on at the same time but the images that are generated are satisfying. It took me more than 3 minutes to render a 'cute cat' prompt for me

upper pawn
#

I keep getting ' Could not allocate tensor with 268435456 bytes. There is not enough GPU video memory available!' but I have 8 gigs of VRAM and a good GPU

upbeat tangle
#

what model should i use for animal and buildings?

broken gull
#

those bytes

upper pawn
#

258 mb

#

but I have 8 gigs vram

#

like even when I put it on low -vram

#

it still crashes

broken gull
#

are you sure you don't have any other programs running in the background that use vram

upper pawn
#

yes but the stable diffusion takes up most of my VRAM

#

even on a low setting

fervent thunder
#

any ways to use AMD gpus to render image in S.D ?

upper pawn
#

yes @fervent thunder

fervent thunder
upper pawn
#

Clone this repository

#

instead of the normal one

#

it works but sometimes there is issues

delicate dust
#

Can someone help me install all the software i need to start using stable diffusion

upper pawn
#

Darkshadow what is your gpu

delicate dust
#

no cap i was watching same vid

upbeat tangle
fervent thunder
#

😦

broken gull
oblique star
#

Can someone help me with GPU? I just need an advice

#

Like if that GPU would be ok to run SD on it

novel geode
#

how do I do to creat images? please, help me

astral goblet
astral goblet
# broken gull https://youtu.be/xf8JNX3sKKk is there a way to combine this with ControlNet?

that has nothing to do with stable diffusion. i'm sure their "AI auto tools" don't even use deep learned networks and are just coded an old fashioned way.

That video is just an advertorial. the guy produced it to look like a tutorial or a review of software that is real content on his channel, but it's actually just an affiliate marketing advertisement. He gets paid for everyone that uses the program.

#

dont believe the hype. he says it's "free" but that's only for 300 frames tops. 10 seconds

#

$300 and it expires after a year

broken gull
astral goblet
#

it's just an informercial/advertorial

#

stable diffusion isn't even a 3d system, and that software is for animating 3d models

sleek venture
#

Hi guys, is this the best chat to ask any questions?

vast ingot
#

depends on the question, there can be more specialised, but this is the main general one, yeah

gentle rose
#

Hi

fervent thunder
#

hey!

#

Yea , that's like one of shittiest things on youtube....

Dude just made an ad video and tried to hype it up with "AI" talk, while all AI does is...trying to make pose look more natural?
But it doesn't stop there, he made an ad inside ad lol and liked shitton of comments telling what he wants people to see.

That might sound a bit rude, but how do I react to it otherwise?
We can't see dislikes on the video, we can't read comments cause actual comments somewhere far down after all nonsense he liked...

I'd bet he removes comments he don't like manually too, if he has so much time to like so many.

#

Nothing against tool , just youtube guy and youtube itself 😦

dapper willow
#

is it me or 3/4 of anime girl models suddenly disappeared from civitai ?

vast ingot
#

I felt a great disturbance in the force, as if millions of voices suddenly cried out in terror and were suddenly silences

fervent thunder
#

idk, I still see tons of them on top

delicate dust
vast ingot
delicate dust
#

nah it was in clone wars\

#

i coulda sworn

fervent thunder
#

damn, people using so many words in prompts now based on civi prompts

vast ingot
#

Well tbh I copied text on a meme

delicate dust
#

oh

vast ingot
#

It was a meme of Obi-Wan in a new hope

#

But i didn't do my research for real lol

delicate dust
#

oof

vast ingot
#

I'm a bad mod. I'll do sepuban

#

bans himself

#

I did some furies as dj the other day though, it may be a diminishing factor in my ban

fervent thunder
#

waow don't commit sudoku , we need you

vast ingot
#

I could ban myself, i do have the rights. Not to revert it though so I'll keep calm

#

And keep making furry

fervent thunder
#

I need non-fury idea, wanna gen something too waow

vast ingot
#

Analog diffusion, make great vintage pics of old cars

fervent thunder
#

all my ideas either done or I can't be bothered to gen it more...got >1k attempts on sword-woman lol

vast ingot
#

Go out of your comfort zone, learn new tokens on a new subject. Ask chatGPT for random stuff maybe even

#

It helped me a little the other day

fervent thunder
#

waow I want to do creatures, but I don't have model capable of doing so weird things

vast ingot
#

That one maybe

#

I made it using the communities around here

fervent thunder
#

That's kinda creatures , but that's too cute, I want like...dragons, griffons, idk...big, realistic or fantasy-ish looking

#

those looks like toys

vast ingot
#

Yeah they are... Elder ring diffusion has some good monsters with tokens like demon or mimic, giant,...
We did some nice things around it on Halloween here

#

I'm not finding any monster specific model in my mind though

fervent thunder
#

I have lora I downloaded some time ago, it can do existing creatures, (well, I guess model can, idk what's lora purpose actually), can't even look it up anymore, can't find it

#

Tried to mix it with boss fight embeding on dark model...

#

That was weird

#

Boss fight embeding makes humanoids too often...like 99% of the time, even with "human, humanoid, man, woman, girl, boy" in negative 😦

broken gull
#

heya how can I get controlnet to work properly? I have a img from posemaniacs but my character doesn't wanna take the pose, also I noticed I have no controlnet models installed does that matter?

dull forum
#

yes, you need to have controlnet models installed

astral goblet
astral goblet
#

i say internet culture thing since richard dawkins hadn't coined "meme" yet

mystic moon
#

How do I make the character not change in img2img? Is it possible to just change some part of the image while keeping others? I don't want to change the face, but I want to change other parts of the body...

foggy wadi
#

does anybody know if the windows version of python 3.10 works for automatic1111 ui

fervent thunder
foggy wadi
#

i keep getting an error that python isnt there but its not even reading the correct path

frosty tide
#

why there is no one in vc?

foggy wadi
#

i was in vc

broken gull
#

Also could you share settings that has worked for you to keep your init img pretty much the same and just be able to change the pose?

lavish current
#

Been glued to Deliberate Diffusion all day making bad photos

#

I like generating but upscaling is so boring for me 😛

#

It's comparable to the inking phase of drawing I guess

foggy wadi
#

i solved my issue all i had to do was delete venv file

dreamy igloo
#

Another update to instagram censorship on a.i. art. A week passed by, and the censorship of my posts have gotten worse. I've removed all words related to the series for 2 weeks, and instead I made a title for my series. However, anything that I post bearing my own title, gets shadow banned 5-8 minutes from posting. The posts disappear from all hashtag spaces. I also removed references to the show from hashtag spaces to avoid their radar. Regardless of the content inside the post, whether there are characters that resemble a little bit of the characters on the series, or something entirely of my own creation that do not exist on the show, they get shadow banned. As long as the post contain my own show title, have a similar looking string of hashtags, either the show's company bots report them, or instagram automatically ban them after 5-8 minutes.

This is unacceptable kind of censorship. I've deliberately put up content and stories unrelated to their show to test the limits of their reach, they would still ban it if their bot mistakenly thinks I am using their i.p. The only posts that I can put up and survive would be non a.i. generated robot, monster, superhero related. Otherwise the posts get banned.

So be careful what a.i. art you put on instagram. Once their bots lock on to you, they become more and more aggressive. Complaining to instagram is no use, they will not respond. If you become targeted, new a.i. art that you post, completely unrelated to an i.p. franchise may still get targeted once they got their radar on you.

At first, they were selective in banning a post, only banning 1 in a few. Then they got more aggressive, 1 in 2. But begining today, every new post I post up, with just a slight reference to my own show title and a few similar hashtags got banned.

#

The only consolation is that they did not remove the old posts that for some reason they allowed them to stay, but it appears I can no longer post any new content related to my own show. Even some totally unrelated a.i. art pieces got affected too.

I am considering posting the entire series on Twitter instead, with the show name and characters referenced. It is a parody, and it is a not-for-profit fan project, I do not agree with the reason for my posts to be taken down, and the way how they handled the taking down of those texts and images.

lavish current
#

I think I'm done bunnying for today. Oof, that was a lot

#

I think I was generating bunny girls for hours 👀

tight totem
#

Anyone know what EXIF tag Stable diffusion uses?

hallow cove
#

hello

vivid knoll
#

Hello, is anyone working on using SD on KYC/AML?

regal snow
haughty knoll
#

can anyone help me. I got an error saying:
RuntimeError: Detected that PyTorch and torchvision were compiled with different CUDA versions. PyTorch has CUDA Version=11.7 and torchvision has CUDA Version=11.6. Please reinstall the torchvision that matches your PyTorch install.

tight totem
dire fable
#

hi, I'm using diffusionbee on mac (m1), can I update the list of suggested text prompt listed from cmdr2? wanted to add specific model prompts such as nvinkpunk or mdjrny-v4 style

quaint solar
#

why do no ones photos have the prompt along with them anymore

vestal dew
#

Some time ago someone posted drivers or something, some fix to make 4070TI run well with SD. Anyone know what it is?

#

LOL out loud

vestal dew
vestal dew
#

anyone have a link to the Auto1111 install? I thought I could just transfer files to a new system, but I have to install

fleet hamlet
#

Hello! How can i post images here?

vestal dew
#

general chat with images

nocturne citrus
#

Guys, which server would you recommend for stable diffusion models training?

nova silo
#

Hi. Are any SD based models okay with fingers yet? Don't offer midjourney

wide bone
#

Hey, I'm using my own trained models, and in the last two days the results got significantly worse, using the same model, prompt and seed. Did anything change in the repo that I'm not aware of? Did anyone else see some oddities with the results?

alpine bear
alpine bear
# wide bone Hey, I'm using my own trained models, and in the last two days the results got s...

Model, prompt and seed aren't the only determining factors in output. Clip Skip, Hypernetworks, LoRAs, any extensions you may have had active, your graphics driver, your python version, your xformers version.. if I were in your shoes I'd be looking for local changes like that.

If your images have png info or embedded metadata you should be able to check that for a complete list of prompt / setting parameters and ensure your surface-level stuff is identical. Going deeper, could be any number of things affecting your output. I haven't seen any issues - I'm on the latest gitpull of automatic1111.

nocturne citrus
#

Guys, I want to train Stable Diffusion with my artworks and try to mix it with another artist to get some inspiration. Is it possible with Astria or is it mainly for more photographic-like stuff?

solemn sky
#

man, what was the AI art generator for AMD users that was in development ?
if i'm not mistaken Stable Diffusion only works with Cuda cores...

warm junco
#

The automatic1111 webui works now with AMD too with the directml Fork of it. But its not good performance wise.

solemn sky
#

very nice, thank you man.
automatic1111 is for Stable Diffusion ?

#

aight, i'll check it too

austere coral
#

which one faster to render a video? png2png or video init?

lone dune
#

I made a game using artwork generated with stable diffusion. is there a rule/faq on whether I can post a link for it here? or is there another place I should go?

vast ingot
#

it gets the most visibility while not scrolling too fast, compared to general chats

lone linden
#

Hello there 😄

lofty parcel
#

It's really interesting to think that AI image generators seem to indicate the beginning of the end of media scarcity

The development of generative AI has the potential to usher in an era of endless personalized media in a way that appeals directly to the individual in a way that nothing else can

It's a double-edged sword, however, because the end of scarcity necessarily requires the decoupling of output from human labor and the elimination of its economic value

potent spire
#

Id like to see how lawsuits end up first

#

Sp we have more clarity on all of this

#

so*

fervent thunder
# lofty parcel It's really interesting to think that AI image generators seem to indicate the b...

Right now I think good artists only becoming more popular, cause people making models based on their art or making images using their style.

Good artists still will have a value for professional work, at least for now.
I'm sure some people will still buy and appreciate human work even if AI will get really good.

But yea...that's still somewhere in the future after law suit situation settle down and AI becoming better and gets more possibilities.

sleek osprey
#

Did you ever try it? I attempted it, wasted about $50 to get a model which just spits out a bad version of the previous frame

#

I think my dataset was bad

lone linden
#

guys I have a question, about something I saw on TikTok, I think the person is using Stable Diffusion and I wanted to know how to make it, can I share it here ?

vast ingot
lone linden
earnest light
#

Where can i get a step by step of how to install sd locally?

vast ingot
lament steeple
#

Currently feeding an iClone render through realistic vision with denoising of 0.25 and instructions to turn it into a photo

lone dune
lament steeple
#

and 0.25 was too high. needed 0.1

clever stone
#

where to try the stable diffusion bot

vast ingot
#

sorry, there is none currently

surreal perchBOT
#

@clever stone

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

clever stone
clever stone
vast ingot
clever stone
vast ingot
#

well, I can confirm this isn't us, Stability AI. we host Stable diffusion on dreamstudio like I linked through the bot, or you can even install it
I know lots of other websites will also let you implementations of Stable diffusion for free

clever stone
vast ingot
#

like hundreds

#

the thing is

clever stone
vast ingot
#

Stable diffusion is an open source tech

clever stone
#

yea

#

ik

vast ingot
#

so lots of sites propose it too

clever stone
#

so thats why i think its cool

vast ingot
#

yes you can install through github

clever stone
#

mkk ty

vast ingot
clever stone
#

mk ty

#

is there mac app for it

vast ingot
clever stone
#

wow this is the first app which i meant to run on linux

#

lmao

#

and not windows or mac

vast ingot
clever stone
#

ty

vast ingot
#

I think that one should be the easiest one, but yeah, it will require linux to run.
There is another SLACK client that came out too, and that should run ? I'm not sure, it's not my strong suit

clever stone
#

oh ok tysm

clever stone
vast ingot
#

when you make a picture, SD takes a very noisy picture (think of old TVs) and tries to remove all the noise, to see your picture under it, based on your prompt.
It does not very good, but we add a little less noise back, and ask the same thing again, and repeat that process

#

this is a step

clever stone
#

how come the quality of images on dream studio is bad

vast ingot
#

bad would be quite subjective, I imagine you compare to something else

clever stone
vast ingot
#

sorry

#

it's mostly the quantity of work that SD will do

clever stone
#

They're pretty good

#

but the image quality

vast ingot
#

512x512 ?

clever stone
#

yes

#

Is my prompt bad xd Olympic rings far away floating in a black sky near an ocean with nothing else

vast ingot
#

yeah, that is a problem of the base models, 1.4 or 1.5 for example.
If you use 2.1, you can use 768x768 instead for a start.
Then people tend to use upscalers, because those dimensions are what the AI is trained on and going a lot over those, even if possible (check the sliders), lands duplications in your pictures

clever stone
#

whats the best version

#

and best image size for that version

clever stone
#

It's doing a good job but not able to generate the olympic rings properly

vast ingot
# clever stone Is my prompt bad xd `Olympic rings far away floating in a black sky near an ocea...

in part, yes. you only want to describe what will be in the picture.
So remove that "with nothing else", it will confuse SD.
add more details to what could be in the sky or the ocean, like waves and clouds instead, describe how the sunlight should be, what the camera angle could be, what style of image it is (painting, photo,...)
Also, "far away" could be better as "long shot" or "ultra long shot", photography terms

clever stone
#

I had to add with nothing else

vast ingot
# clever stone Oh is 2.1 best?

2.1 is currently the model trained on the biggest images, and the last one that came out.
there are also lost of fan models (check civitai.com) but those won't be usable on dreamstudio

clever stone
#

without it was adding some random things

clever stone
vast ingot
clever stone
#

Is there a recommended word count for prompt

clever stone
vast ingot
clever stone
#

oh

#

whats credits

vast ingot
clever stone
#

man v2.1 much slower but i guess its better

vast ingot
#

dreamstudio gives you free credits to try SD

clever stone
vast ingot
#

but it costs after a little while, you'll be asked to buy credits to continue in it

clever stone
#

so bascially dreamstudio is just SD hosted online

vast ingot
#

there are other services and ways to use SD though if needed

#

yep

#

with some tips and tricks

clever stone
fervent thunder
vast ingot
#

there are other fun features in other SD webui too

clever stone
clever stone
#

cant i post images here

vast ingot
vast ingot
#

it depends on the client

fervent thunder
#

oh

vast ingot
#

automatic1111 has a logic that reduces the impact of each token but makes it possible for infinite tokens

fervent thunder
#

hm , interesting

delicate dust
#

If i want a fullbody portriat instead of head portriat what should i type

alpine bear
delicate dust
#

alr

alpine bear
#

Also try a vertical aspect ratio, something like 512x768

wide juniper
#

I've been using stable.art but it has some annoying glitches, and they've also not been updating it recently. Anyone using a better one?

cedar garden
#

Tf

vast ingot
#

this has no place on an AI discord, we make and share pics here

#

you were already warned and timed out for that

#

this is a ban my man

cedar garden
#

Haha drop the ban hammer on him

vast ingot
#

yeah I just had to do my checks, I never want to hit by mistake

cedar garden
#

Also hello guizmus👋🏻 how are you doing

vast ingot
#

well good but strange introduction, sorry for that

cedar garden
#

😂😂

vast ingot
#

we do get quite a lot of "tf" moment

#

I do swear,I rarely have a low day around here

cedar garden
#

Lmao

vast ingot
#

how are you doing ? first time I catch your username I think

#

anime user 🙂 I'm less there, I'll admit

cedar garden
#

Lmao I changed my name I used to be called Space_Bunny|Cosmopus

vast ingot
#

ho yeah I remember 🙂

cedar garden
#

Haha

#

I changed my pfp and hence the name change

vast ingot
#

yeah. you're liking the stuff ? it's kind of a drug at first, SD

#

I couldn't stop generating pictures

cedar garden
#

It will either be space_Fox or space_Bunny

#

Haha thanks to ai art I have been able to get loads and loads of reference materials

#

I posted in the #🌶｜off-topic Of what I have done thanks to a fellow member that gave me permission to do it

vast ingot
#

you've got a project you'r working on ?

cedar garden
#

Anything anime related 😂

vast ingot
fervent thunder
#

Anyone listening nvidia talk about AI?

cedar garden
#

Lmao

vast ingot
#

gtg for now, time to go do the groceries

#

see you later mates

cedar garden
#

Cya later bud be safe

zenith thunder
#

that died

fervent thunder
#

yea we killed it instantly

rapid atlas
shell mortar
#

XD

indigo ermine
#

The 'pickapic' thing is clearly overwhelmed...

pliant summit
#

stability can not keep taking these Ls

fervent thunder
#

I mean...that's just a way for them to collect data, I don't see how it's an L

#

but yea, that doesn't look prepared for many people, what is it, at 32 requests rn?

mild sundial
#

hi guys, i installed stable diffusion but when pasting the URL, it does not connect, what seems to be the problem?

orchid lance
#

you installed what?

mild sundial
orchid lance
#

...

#

I don't think so.

fervent thunder
#

Assumming we're talking about automatic1111

mild sundial
#

yes, its installed properly

orchid lance
#

yeah, that's what I was trying to get him to say what he installed

pliant summit
mild sundial
fervent thunder
gloomy anvil
#

lol, we just hugged Pickapic to death

orchid lance
pliant summit
quasi leaf
#

pick a pic is actually quite decent fun

pliant summit
#

Hopefully things change

fervent thunder
#

That's not exactly true, but aight, you do you , I guess

mild sundial
quasi leaf
pliant summit
#

Releasing 1.4 was the last meaningful/impactful thing Stability did

exotic zealot
fervent thunder
#

I don't get your point...
1.4 wasn't "quickly superseded" ?

Basically anything will be "superseded" eventually, it's just a matter of time.
The only thing that didn't change yet was a1111?

#

Do you even use base 1.4 model? or maybe any base model at all , at this point? I don't

zenith thunder
#

ai is something that needs a lot of testing

pliant summit
#

No I wouldn't say so, everything that has come following 1.4 has been essentially a continuation of the same technology. It hasn't been "superseded", imo. But now we're getting into semantics.

#

It's just that all the meaningful progress has come outside of Stability

#

i mean stability didn't even make SD, if I recall correctly

zenith thunder
#

.> also wait one of the devs is an otter

pliant summit
#

RunwayML made latent diffusion, which became stable diffusion

clever stone
#

what is the point of user karma in pickapic

pliant summit
#

Runway released 1.5, which if you guys remember, StabilityAI tried to TAKE DOWN

zenith thunder
#

why does SD have soo many problems making otters?

fervent thunder
#

So...what's your point?
"1.4 good, everything else bad" or what? what are we talking?

Progress is a progress, whatever it is, meaningless you think it is or not

pliant summit
#

I am simply expressing my discontent with the state of their efforts

fervent thunder
#

waow I never heard any promises, so whatever 😄

pliant summit
#

Oh?

orchid lance
#

all I'm hearing from you is your entitlement "I want it now, where is it?!"

pliant summit
#

Do you remember the upgrade of SD that was going to allow real time generation? 30 pictures per second?

mild sundial
#

AI will eat the earth - Aloy: horizon -forbidden west

pliant summit
#

That was supposed to be released at the end of last year

#

Do you remember Deepfloyd? That was teased several months ago, was supposed to be out by now

fervent thunder
#

no idea about those

orchid lance
#

yes and I'm a developer myself and I know how much work that does involve.... that always take time.

daring wagon
#

do you guys remember when we thought deep_dream and big_sleep were the best ai txt2img models ever

pliant summit
#

Do you remember the slew of models Emad himself said would be coming week after week?

orchid lance
#

blah blah blah

pliant summit
orchid lance
#

don't care

pliant summit
orchid lance
#

so you don't get all upset? cool

mild sundial
#

hey, how to solve it guys, i got the part on "running on local URL"

fervent thunder
pliant summit
#

Disagree, you are not up to date

fervent thunder
#

lol

pliant summit
#

and that's okay

orchid lance
#

even I'm guilty of this too... sometime I'd talk about some ideas and plans on what I'm gonna create and such... but we often forget the reality of actual development, which can take longer as the more bugs and issues come along the way.

#

no matter what, we'd always wish to finish something quick and done with.

#

but this isn't new.

pliant summit
#

As early as August of last year, Emad teased a text to video model from Stability, for example

orchid lance
#

so? he was having fun

pliant summit
#

My point isn't to be like "why is it taking so long!!" it's that, I believe he's quite literally making things up

#

I think no such thing is in the works, or ever was

orchid lance
#

or maybe there is but it's not simple enough for someone like you to use it

broken gull
#

I've got 0 bytes left on my drive PepeHands

fervent thunder
#

I mean...you can just ignore all promises and get things when they are actually released and won't be upset
I don't care about promises , cause I don't even see them lol

broken gull
pliant summit
#

I mean, we have text to video open sourced from someone else now

#

its not hard to use

pliant summit
#

I'm talking about Stability AI

orchid lance
#

um...

#

you do realize a lot of people are using SD

broken gull
#

How is it not relevant?

#

I think they used SD as base as well

orchid lance
#

yeah

fervent thunder
#

I feel like he just want to be upset about something at this point 😄

pliant summit
broken gull
#

What does it matter if they created it or not?

fervent thunder
orchid lance
pliant summit
worthy tangle
#

I'm having an issue with the pickapic thingy, where do i report it?

orchid lance
#

oh yes... very late, much late.

#

whoop dee do

pliant summit
#

My claim: Stability has not released anything of note since 1.4. They consistently overpromise and underdeliver. All advancements and developments in the field have come from Runway, Midjourney, and opensource.

fervent thunder
orchid lance
fervent thunder
orchid lance
#

yeah

pliant summit
#

I simply hope that things change moving forward

fervent thunder
#

I'm fine with developments being made by other people or community ¯_(ツ)_/¯

worthy tangle
pliant summit
gloomy night
#

I do wonder why dreamstudio doesn't produce pretty pics as easily as Midjourney does

pliant summit
#

unfortunately

gloomy night
#

🤔

mild sundial
#

yow it now works, thanks for not helping lol

gloomy night
fervent thunder
#

Isn't mj older then SD?

gloomy night
#

But Midjourney v 5 does better pics than dreamstudio

pliant summit
gloomy night
#

I just wish stable diffusion was as easy as mj

fervent thunder
#

nah, that's whole point I use SD for - cause it has more sliders to play with , you can control more shit

gloomy night
#

I still remember how much much better mj got Once they integrated stable diffusion in

fervent thunder
#

while MJ better at gen and understanding prompts

pliant summit
#

MJ V5 is a far superior model and it's a shame that closed source AI currently winning

#

And that is the source of my frustration

gloomy night
winter pike
#

midjourney is a commercial model, aimed to less tech people who wants to make pretty images

#

but every image they make have the same looks tho

pliant summit
#

MJV5 with controlnet and other tools would be insanity

gloomy night
fervent thunder
#

Deep in my mind I hope that one day mj leak...lol

gloomy night
#

That's the dream

winter pike
#

and SD is a general model meant to be as broad as possible

winter pike
gloomy night
dry venture
#

SDXL? whaats

gloomy night
#

I have only ever used dreamstudio as I am not familiar with github

#

I am no coder / developer. It's a bit daunting to me

wild steppe
fervent thunder
# winter pike me too, i want to know how it really works

I'm thinking about other thing, it's just MJ's commertial part, what gives them alot more resources to do shit.
It's harder for opensource and even community to keep up.

Yea we have more tools and people, but we can't put huge server size machines to train or do other things

#

So...leak will help in this case, less reasons to use mj...even tho, there still will be "hosting" part of mj...yea...

winter pike
broken gull
#

no matter how small the prompt in MJ you get pretty picture even just cute dogwill give you someting nice, while with SD you will have more work

gloomy night
bleak matrix
broken gull
clever stone
#

sorry to say but midjourney is so good\

#

I didnt even have to set anything up

broken gull
#

I mean it is good, but also not it's not free and that's why SD will always be superior to me

clever stone
#

I gave the same prompt and omg SD and MJ arent even comparable

broken gull
clever stone
#

MJ too good

clever stone
winter pike
clever stone
#

like omg

broken gull
#

You have SD models that are trained just like MJ and give similar output

#

and are free to use

clever stone
#

The result I got is amazing

clever stone
#

Omg its so good man

#

Wow oh wow

#

guys try MJ

young ivy
#

midjourney has their own frontend, backend and network structure. Which they don't make available to the public for the most part. They're set to make pretty images, but for the most part. Can they even produce the quality and consistency of images that are exclusively owned by me and I don't need to pay an extra 30 dollars or more to not h ave made auto-public and posted? What makes it worth like 60 dollars a month?

#

Ive used it before back in september last year

#

and it just got more and more restricted and censored

winter pike
gloomy night
clever stone
gloomy night
#

And sd costs monnee per iteration

fervent thunder
# clever stone guys try MJ

I'm here for more options and possibilities MJ can't do and also for being free, so SD that is.
I'd do MJ for quick easy prompts , but that's about it.

For something else I won't pay and can't be bothered to struggle with discord abuse nonsense each 25 gens

clever stone
#

bruh

winter pike
clever stone
broken gull
clever stone
gloomy night
young ivy
broken gull
#

how much dollar per credit was it again? a few cents

bleak matrix
clever stone
broken gull
fervent thunder
clever stone
#

but if ur gunna just outright compare SD AND MJ only based on prompts, MJ is better

clever stone
vestal dew
#

SG is better

young ivy
#

but I'm going to gues that yes it is

clever stone
#

MJ absollutely deserve to be paid

fervent thunder
vestal dew
#

You have no control over MJ. You have to pay per pic. No control of the output, the images, and it can change at a moment's notice.

clever stone
#

I got the perfect image in one prompt

winter pike
#

gonna cite myself: "the problem is the lazy prompting and as i said the majority of MJ users arent the most tech leaning people, they want a pretty image with 0 research and it works with MJ"

clever stone
gloomy night
#

The thing is, recently mj because of boobs gutted their model

clever stone
broken gull
#

while MJ is like 30/month, while with SD it is when you use it. so if you use it lot then yes DS will end up being more expensive but if you don't use it that often like you are subscribe to them and have only generated like 5imgs that's a waste of money imo

bleak matrix
vestal dew
#

they can censor your images - make a good image of a person they don't like - they can censor.

young ivy
#

they censor words too, or did

#

in the end I could barely write a prompt like "young woman in sexy dark fantasy armor and long dual bladed katana, licking blood from her lips, blood stained white curtains and bed, bloodied corpse beneath her feet"

#

well I couldn't at all, not even close

broken gull
# gloomy night Mmm i doubt that

At default settings, it’s 0.2 credit per image. Depending on the image resolution and step count you choose, your pricing will vary. Here is a detailed chart for explanation:

Steps 512x512 512x768 512x1024 768x768 768x1024 1024x1024 15 0.1 0.3 0.4 0.5 0.7 1 .0 30 0.2 0.5 0.8 1.0 1.4 1.9 50 0.4 0.9 1.3 1.6 2.3 3.2 100 0.7 1.7 2.6 3.1 4.5 6.4 150 1.0 2.5 3.9 4.6 6.7 9.5

gloomy night
young ivy
#

which means the images aren't owned by me

clever stone
gloomy night
young ivy
#

I'm never going back to MJ. Well not unless they go open source and charge nothing for nothing

gloomy night
#

Let me say it again: I am not bashingSD. But becauseof thereasons I talked about, at the moment it's not more attractive to me than MJ

young ivy
#

MJ uses SD though

clever stone
gloomy night
#

I remember clearly the jump of quality between v 3 and v 4 and in between there was the stable diffusion powered modes

clever stone
#

well they can use code from SD and they can sell it

fervent thunder
broken gull
clever stone
#

ig not wrong

young ivy
clever stone
#

but

gloomy night
broken gull
young ivy
#

as long as there is even a hint of censorship in MJ, not going there

gloomy night
fervent thunder
# broken gull not with 2.x

as far as I know not in 2.0 cause 2.0 model doesn't have nude stuff in it, but still can be additionally trained
(they couldn't use dataset cause of lawsuit or something like that)

clever stone
young ivy
#

ooh boy

gloomy night
young ivy
#

let's not enter philosophy

#

or we gotta move channel

broken gull
#

if they only censor NSFW I am fine with that

clever stone
clever stone
young ivy
#

I guess people can't help themselves

broken gull
clever stone
clever stone
fervent thunder
#

eh, idk, I'm not even against nsfw tbh, art is an art, even if it's nude art.

bleak matrix
#

A lot of our community members have made great guides to help understand SD. Styles are a new way to bridge the gap, and make it easier for newcomers to prompt. One of the easiest ways I have found to prompt is just to go "subject, background, style." Like, "a cute kitten, laying in the sun, oil painting."

young ivy
#

if you want to talk to me please do so using actual words

broken gull
bleak matrix
#

But if you need any help, you can just ask.

fervent thunder
clever stone
young ivy
bleak matrix
broken gull
# clever stone then whats the issu

I am not making any???

You asked what's wrong with censor, I said depends on what and I give as example if they only censored NSFW then I would be fine with that....

clever stone
gloomy night
broken gull
#

Like I said two times before now not that ik of

young ivy
#

while some prefer things like "insert style here, insert brush here of subject"

bleak matrix
bleak matrix
fervent thunder
young ivy
#

censorship is only acceptable with heavily illegal things that I should not need repeating

broken gull
young ivy
#

ah my lora at 5000/6700

#

wonder if it'll work

bleak matrix
#

The "Friendly Stable Diffusion Guide" is geared more for 2.1+, since I haven't been able to update all of my prompts to 2.1+ yet. But the informational still be useful, as there is a lot of it

young ivy
#

1.5 is superior for art still

#

mostly for photorealistic too

clever stone
broken gull
young ivy
#

I dont know what they did with 2.0 or 2.1

broken gull
broken gull
bleak matrix
young ivy
clever stone
broken gull
broken gull
young ivy
#

the people having their own models can

clever stone
fervent thunder
broken gull
clever stone
#

whats the best sd model

broken gull
young ivy
#

they can release a new model and if people like it, they can work on creating new models, loras etc

fervent thunder
bleak matrix
young ivy
#

the official SD model from the stable team doesn't need to be NSFW

#

it just needs to be open for modification

broken gull
young ivy
#

without enforcingg SFW, in which case people will prpobably backwards engineer 1.5 and release a awhole new thing

#

if they force the model and all trained on it to be SFW

broken gull
#

What stability.ai does for us is open the way, it's up to us to create a city. kekw (tried my best for the allegory)

fervent thunder
#

yea...2 biggest comminuties are nsfw and anime...

young ivy
#

yeah thats why I say that as long as they don't try to tell us we can't do things and even program etc so we can't

bleak matrix
# clever stone whats the best sd model

As you go through #1072220168534642768 you can read more about SD, and how it works. SD allows you to customize and train your own models. Like, do you have a favorite item that you would like to use with AI? Great! You can do that with SD. You can learn more about this by just reading gradually, and when you're finished, you can also check out #1047197565365538826

young ivy
#

I understand the need for the open source version of an AI image generation software to be SFW, as it opens up a lot of possibilities for kids to use it to generate...whatever they want. If their original model SD x.x has no datasets that are nsfw, and keeps it legal. Then they will have no agencies getting involved

clever stone
bleak matrix
clever stone
young ivy
#

hmmmm 5850/6700

#

aaaah

#

gogogo

broken gull
#

?

clever stone
#

but for now, i got my desktop wallpaper frmo MJ

bleak matrix
#

Seems good to me--it's your desktop, after all! 🙂

clever stone
#

lol

#

although the quality is low

fervent thunder
#

You can upscale it in mj

#

or you can upscale it with SD

frank delta
#

hey guys, i'm new to stable diffussion and have just installed it locally, sometimes it says runtime error, not enough memory, lower the resolution.

what memory is it talking about? my ram? or my gpu's ram?

also, i wanna use a model from civitai, where do i put it and use it?

drowsy crown
#

I hate discord. Its not enough to mute #📣｜announcements and set mentions to none. Why do i have to block a person.

fervent thunder
#

yea that's weird

fervent thunder
drowsy crown
#

there is no at everyone immunity, thats one of the small complaints. i literally hate discord and wouldnt have a sense of community without it

fervent thunder
frank delta
winged flare
astral goblet
# fervent thunder as far as I know not in 2.0 cause 2.0 model doesn't have nude stuff in it, but s...

filtering nsfw was a voluntary choice. a strategic direction. they intend base models to be in libraries, schools, public centers, etc. The filter was accidentally set too strong and that wasn't picked up tiill after it was all trained. so lessons were learned that 0.9 threshold is still far too low to get only pornographic material.

base models shouldn't be smut enabled and i think that's kind of obvious when you consider the goal of getting these systems into every single class room. NSFW can be refined on top of that after. the drama all came up out of the unstable diffusion community, who conveniently were using it to fuel fundraisers

#

all of the nsfw dramarama is one gigantic social engineering con and this community is just way to young and fast growing to have any sort of immunity against those classic scams

hallow plaza
#

Are there any good walkthroughs showing what are the best settings in Dreambooth to get realistic results?

wispy berry
#

Ah, not sure about it, but when I typed in pickapic.io manually in the browser, it went to http://pickapic.io - and then, logging in with Discord didn't work. I don't know if it is possible to redirect to https always when the user tries to login with Discord, but some sites redirect to https anyway, so that might also be an option?

fervent thunder
#

yea it is possible to always redirect to https

astral goblet
#

i dont think the site is intended to be used without https since they don't link it anywhere without it. the web admins probably just did an oversight and don't have a redirect policy in there

#

theres never any reason to run without https these days. don't give any middlemen any ground to spy

wispy berry
#

Also, for pickapic, should you rate the picture quality or whether it follows the prompt? Because some pictures absolutely do not

astral goblet
#

the one you think is better

fervent thunder
astral goblet
#

its intended to train better looking outputs. so the idea is to vote aesthetically. but really thats entirely subjective so thats why its a crowd sourced system

#

idealy it should just average out and you'll get results that the crowd consistently ranks higher

frank delta
#

hey, how can i use "checkpoint merge" models from civitai in auto1111 gui? i tried to put one in the models/stable-diffusion folder, but it gave me some error

fervent thunder
#

What error?

frank delta
wispy berry
#

oh damn, I think I should rescind all my votes ... is that possible?

#

I used probably too complex prompts which weren't fulfilled most of the time

fervent thunder
#

It didn't do even 1 of my prompts decently

wispy berry
#

Ah, so I should probably try text prompts

fervent thunder
#

probably just simple 3-5 words would do, but eh...

charred mesa
#

Where can I post SD-XL results? I can't find the old channel for it

wheat hazel
#

what's a good model to work on a comic book ?

gleaming summit
#

Has anyone tried making an actual logo with AI yet? As in a logo with text rather than just objects? If so what was your favorite AI that created it ?

chilly pasture
#

Uhm do i have to do something in picapick for sdxl? "Model: Currently, we use SD 2.1, dreamlike-photoreal-2.0, SDXL 2.2 alpha, and SDXL 2.2 beta."

charred mesa
#

No it's on by default to the newest

#

Idk if it's the alpha or beta but it's SDXL for sure

#

then generate something and pick the image that looks better than the other (or click the button below if they are both good/neither is)

chilly pasture
#

Ok thx! I already generated couple 100 karma 🙂 not sure if it can really show off with those small pictures but i try some more if it helps

fervent thunder
astral goblet
#

I think you could still do logo concepts with ai. it can be part of the process

fervent thunder
#

yea, as a concept why not

astral goblet
#

any good logo design starts with many thumbnails

#

from there you work with the client or target marketing to get the right symbol and start refining it down. AI could even assist with other parts of the process. like once you got the sillohuette done, you can fire off 10000 colorized versions of it to see where it communicates best

#

once in a while you get lucky and pay some lady $100 for one of the worlds most recognized logos. just do it that way maybe

sick fractal
#

I don't know why so many people reacted with the Russia emoticon to the latest ping on the announcements.... it surprises me that so many people support this terrorist country.... stable diffusion should stop providing its services to russia.... Fu ck Russia! Every Russian is guilty of this war! Slava ukraine

astral goblet
#

sorry i was wrong. they paid her $35

astral goblet
fervent thunder
vestal dew
#

Anyone know how to optimize a 4070TI for SD? I remember seeing something about optimizing the 4000 cards, but I can't find it now

astral goblet
#

here guys, have some FOSS, but you! YOU OVER THERE! donnnnnt you use it

#

hey, i said don't!

fervent thunder
#

yea...

vestal dew
#

er, North Korea

astral goblet
vestal dew
astral goblet
#

i can confirm the cudnn files are still needed. encountered that 2 nights ago

vestal dew
#

So I just replace the DLL's?

#

that link is from Dec 2022. Seems like old stuff

astral goblet
#

its on the pytorch side. until we catch up to pytorch 2, ada cards need the dlls

sick fractal
astral goblet
#

and because of licensing, automatic1111 webui can't redistribute the dlls themself

#

you have to sign up to nvidia developer account and get it yourself

vestal dew
#

Well, I'm not clear on what to do.

fervent thunder
astral goblet
#

resistence is futile

vestal dew
#

go door to door to examine everyone's computer

#

Those who want to use SD can get a free train ride to a camp where they can use SD

blazing lotus
frigid orchid
#

Is there any information or estimates of VRAM requirements for SDXL yet?

vast ingot
#

not that I saw, nope

buoyant jacinth
#

it's gunna be 1mb more than my card can handle.

astral goblet
#

it's a giant conspiracy to prevent poor people from using it for a couple of years before tech is cheap enough for them! i tell you! :zoidberg:

young ivy
#

sorry

astral goblet
#

someone is eventually going to invent a way to turn it into fp 1bit and any device in known history will run it

summer moth
#

How do you use weights?

buoyant jacinth
zenith igloo
summer moth
#

Like how do you make the generated image resemble the original image more/less? Is that what weights are Villager_Think

frigid orchid
past edge
#

If it cant run on 24gb vram its trash. Thats my motto

frigid orchid
#

I only have 16GB.

#

So qq.

#

I hope it can run on 16GB. But you never know. I know originally SD said they wanted a 8GB model and then a 16GB model before moving onto like 30-40GB model etc. However I feel like plans may have vastly changed.

past edge
#

What exactly is sdlx? I'm too lazy to Google.

frigid orchid
#

I forget the exact meaning but it's like large image.

#

Basically what it stands for I could not find to much info at all.

#

I don't even know what large means? SD 2.1 has like a 768x768 model or something like that right?

#

I know you can go above that or below that but I think the images it was trained on were mostly 768x 768 or something like that.

#

Something tells me SDLX will be larger than that? However I will be honest I'm not that well informed. Taking guesses here.

#

I can't wait until we have a model that's trained on only 4K images. And can also do perfect text... now sure it will likely need 128GB of VRAM. lol

past edge
#

Nobody needs a 4k model

#

1024 is as much detail as the human eye can detect

#

you can see that in current homebrewed 1024 models

#

we're not generating jpegs, we're generating high fidelity raw images, therefore they scale much better than any compressed image pulled out of the web

#

if you want 4k, all we need is a 1024 model and a semidecent upscaler and there ya go. I've been begging the powers that be for a 1024 model but due to all the controversy it doesn't seem to be happening anytime soon

astral goblet
#

its parameter size that makes it XL, not the resolution it was trained at

#

it'll probably gen at 1k though

past edge
#

fuck i cant wait

#

my gpu is throbbing as we speak

frigid orchid
astral goblet
#

um, i'm not entirely sure where it will benefit. I think it has to do with the pretraining phase and therefore it should have stronger zero shot capabilities and better multi subject focus. deeper and longer neural pathways through the latent spaces

frigid orchid
# past edge nope.

Yes I was kind of joking. I too believe 4K is overkill and ultimately very difficult to train even for the foreseeable future.

astral goblet
#

adobe will push it up to 8k models in no time at all

frigid orchid
#

The thing I love about SD over most or all other models is that it's open source. No matter what silly controversy or laws that follow. Once a model is released it's released. Basically always going to be a way to use it. And even improve it. Now new base models with more overall quality just add to the user models that we will see released. Basically it will never get worse than it is now.

astral goblet
#

its some of the most powerful opensource i've seen in a long while. i hope the trend continues

#

have a great day fellow generators

frigid orchid
#

Yes every day we see new AI news. Things like Control net etc. Pix2pix... every day stuff is being developed on top of just new models and embeddings etc.

#

The more advanced GPU's get the more ease to develop. I can't wait until the average GPU or at least the average high end GPU is 48GB or more

#

I want video game developers to push games into higher VRAM. lol So there is more of a need to release higher VRAM GPUS. lol But it's entirely possible that the AI market place drives consumer use anyways over the next 1-5 years.

buoyant jacinth
vestal dew
#

I read that the 4 series Nvidia card tweaks someone posted earlier are already installed in the l;atest update for A1111 11 111 . Is that true? Do I have to install something like those DLL's to optimize?

frigid orchid
vestal dew
#

You may notice you can read my messages faster than before. That's because I'm now using the new faster computer.

frigid orchid
vestal dew
#

there was some update and new DLL's