💬｜general-chat

[left star]

rip

ok cool thanks for the info

[pale latch]

i've got 16gb and do alright

[left star]

8gb here, been strongly considering just going for a 24gb card

[pale latch]

totally just go for it. 4090 baybi do it!

||do it||

[left star]

if it fit in my case and my psu could handle it i would've already lol

[pale latch]

shit 3090 then

orrrr. riser cable and a second psu

just strap that 4090 like baboon heart hanging off the side of the case

probably safer for the power connector anyways

heohhhh

but seirously wtf is SD 1.6?

all my google dives bring up information about automatic1111 version 1.6

[winter iris]

I don't see a commissions channel either

Jobs for AI artists 🤷🏼‍♂️

[tender nebula]

WHAT?

[visual pier]

I can make money doing that? :D

[ionic oar]

Hey
I got error when I try to install dreambooth and try to run webui, it says "CUDA... bitsandbytes" something like that. I use Windows11
Does anybody know how to solveit?

[tender nebula]

Yes, I do.

[winter iris]

Yeah I mean....in other communities there's usually a jobs channel or similar where you can post job offers.

[ionic oar]

Can you help?

[tender nebula]

what do you jerk guard

[tender nebula]

Yes.

[ionic oar]

OK what should I do?

[tender nebula]

Ok?

[ionic oar]

Okay what should I do? You said you can help.

[tender nebula]

That’s correct.

[ionic oar]

How?

[tender nebula]

What?

[ionic oar]

Okay bro

[tender nebula]

???

Sorry I’m high

[ionic oar]

Nice, me too. I just need to fix this error

Do you really know how to fix this?

[ionic oar]

Hey
I got error when I try to install dreambooth and try to run webui, it says "CUDA... bitsandbytes" something like that. I use Windows11
Does anybody know how to solve it?

[visual pier]

Try adding --xformers into your commandline args ^-^

[vestal epoch]

me too

[ionic oar]

It makes me uninstall everything if it doesnt work, I’ll try

[vestal epoch]

Useless. I've tried

This error appears as soon as you add the dream plugin.

[ionic oar]

Yes

I dont know why?

[vestal epoch]

Bitsandbytes was not supported windows before, but my method can support windows.（yuhuang）
1 open folder J:\StableDiffusion\sdwebui，Click the address bar of the folder and enter CMD
or WIN+R, CMD 。enter，cd /d J:\StableDiffusion\sdwebui
2 J:\StableDiffusion\sdwebui\py310\python.exe -m pip uninstall bitsandbytes

3 J:\StableDiffusion\sdwebui\py310\python.exe -m pip uninstall bitsandbytes-windows

4 J:\StableDiffusion\sdwebui\py310\python.exe -m pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.1-py3-none-win_amd64.whl

Replace your SD venv directory file（python.exe Folder） here（J:\StableDiffusion\sdwebui\py310）

[vestal epoch]

I saw such a solution but I didn't understand it

https://github.com/TimDettmers/bitsandbytes/issues/858

[vestal epoch]

+source

[median escarp]

Anyone know how to make SD stop putting umbrellas in rain scenes? Tried umbrella in neg prompt and no umbrella in pos prompt...

[wanton karma]

why not giving me the right result????
#1100170312106127410 message

[teal pagoda]

Hello!

So I got a job offer for illustrating a kids book with AI and they asked me for a price per 10-20 illustrations. Since I do not really know the prices from that field, what do you think that are some good prices?

[thorny frost]

Apologies for jumping in here... but I've only just started with SD (using A1111), and I see quite a few posts about SDXL/SD 1.5. May I ask, why 1.5 is better for you?

[quartz hawk]

I can guess; It's a more mature product.

While SDXL is still very new so its addons, plugins, etc aren't as developed yet.

[visual pier]

For me personally my favourite chekpoint still uses that version

[next python]

Is this free to use or not?

[raven wagon]

hey guys
do you maybe know a way to remove metadata files from an already trained LORA?
it's a bit important and urgent

[bitter laurel]

I love SDXL. I use it for everything except controlnet and animatediff.

[thorny frost]

That makes sense. Thanks!

[thorny frost]

🙂

[thorny frost]

Are you using SDXL via A1111?

[bitter laurel]

I moved to comfy for SDXL.

Learning curve is low because you can just borrow json from others. Find a working flow and tweak it, instead of starting fresh.

[visual pier]

Lol I personally cant see how comfy with its geo node like interface is supposed to be easier to use than A1111 :D

As someone who has installed A1111 for some friends it has basically no 'learning curve' aside from the usual prompt crafting and such. Just having a prompting box and a big 'Generate' button is pretty intuitive :D

[thorny frost]

I must admit, I'm not a fan of nodes... I never know which one I'm supposed to use and where.

[visual pier]

I can recommend A1111 :D
Because prompt crafting and knowing stuff like CFG and Denoising is already 90% of the image, and I think the node noodles would only distract from that

Having a proper prompt is like already 80% of a good image

[visual pier]

Why do people like kitchen sink models so much? :O
What I love about Stable Diffusion over any other project is that you can change the checkpoint to be for the specific subject you would want ^-^

[bleak matrix]

Good morning, everyone! How are we all today?

[visual pier]

Moin ^-^

I will be checking out ComfyUI just because I dont like being biased without at least trying it out :D
But I have my doubts its the interface for me

[bleak matrix]

Waves

[bleak matrix]

Well, you can always use #🐝｜swarm-ui

It's a combination of both

[visual pier]

Currently I am using A1111 with great results. But Ill be trying it out :D
Theres no harm in at least seeing if they fit my needs ^-^

[scenic needle]

Can somebody teach me Animate diffvid2vid on stream ?

I've tried but there are few issues in output

[serene depot]

Hi guys, I have a little problem, maybe some of you had it too. I use ComfyUI and everytime when I start to render something, the pc freezes, right before the ksampler step starts, it freezes for about 5-10 seconds, then everything is ok. What is the problem? Is my RAM too low? I'm on 16gb.

[regal meteor]

Is there a function (that I can use on google colab, not a webui script) to merge checkpoints?

[weary quest]

This Riffusion?

Just making sure

Just with a D

For discord

[past summit]

so whats everyones opinion about sam

[visual pier]

How do I hires fix in that :D

[bleak matrix]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Basic Usage.md You can also find the tutorials on that we did on ComfyUI here if you want to get more familiar with it: https://www.youtube.com/@stablefoundationai/videos Scott has a lot of great videos on ComfyUI: https://youtu.be/CxB47DMEyYQ?si=d27J_HPDSLP8eWT0

[visual pier]

As far as I got with my own research its replicable with a 2 pass text2img, with requires a custom Comfy Workflow :D

[visual pier]

Thanks so far for the infos ^-^

In stableswarm I am always getting the warning that my embeddings do not exist :O
Even when I select them right from the UI
To elaborate on that: It seems in the prompt they can be loaded, but not in the negative prompt :O

[bleak matrix]

Np! If you have any questions, or problems, I'd refer to #🐝｜swarm-ui

[random cedar]

hello i saw today this server have text to video, is any one know is there any image to video with prompt?

[sharp zenith]

Hi! Has SDXL recently become more heavily censored? I'm a romance author and use it to create bare chested men for book covers. I didn't have any problems before but now it's not allowing any bare chested men to come through. If this is the new reality, SDXL had just rendered itself useless for my needs.

[sharp zenith]

Never mind. I've found what I need with SDXL using Night Cafe (and I have a ton of credits with them from before). They have an SDXL model. They've upgraded a ton of models, adding their own secret sauce.

[final egret]

Hi, I am trying to apply AI to replace a series of steps I need to do manually in blender. I start with a mesh representing a "sand trap / bunker" (golf game). What i want to do is alter the edges of specific mesh without changing the border itself just the vertices inside the border close to the edge. I am hoping to create variability in the edge of the bunker by doing this. Is this something that Stable Diffusion could be adapted to do? https://drive.google.com/file/d/11P3CQ1N8ro14jqWPI9l6CyiWNmmHTOAB/view?usp=drive_link The two matching faces (silver and yellow) would need to stay fixed. I would be interested in extra rows inside this "border" and altering z values etc. to make the edge appear to be a bity "wavy"

[pale latch]

if you're using it through a service like the bots here, yeah. it will always be censored. It's not smart business to create liabilities.

[limber gulch]

What is everyone using now? automatic1111 hasn’t been updated in months

[weak kettle]

TypeError: Accelerator.init() got an unexpected keyword argument 'logging_dir'

guys what is the problem

[sharp zenith]

I'm using the cloud website version. Thanks!

[tender nebula]

No problem! :-)

[visual pier]

A1111 works really well :O
And are you sure about no updates?

[pale latch]

weird bots are weird...

automatic1111 gets tons of updates on teh dev branch. The community started getting pissed because all these dumb 1 button install scripts would put "git pull" in their launch, so whenever the repo had any changes, they'd push to eveyrone's installs and break everything.

both sides to blame here. Automatic should've had it set up with proper releases, but also 1button install writers shouldn't have put "git pull" everywhere because most people have no idea how to manage python dependencies and discrepencies. Troubleshooting is a huge time sink on a community and it stalls momentum out. Those developers who wanted to create something easier actually caused a lot of damage.

So Automatic1111 started releasing a proper release cycle. 1.6 stable. Everything else is pushed to dev branch. Development still occurs daily, but most users don't catch it until major releases.

Extensions also always release, which are considerable updates to the UI

https://github.com/AUTOMATIC1111/stable-diffusion-webui/commits/dev actually, it does look like it's been a couple weeks for the dev version. but i'm also not sure what new versions or featurs need iterating at this point. there's tons of other branches people are working on.

[crystal edge]

hello where i can find good prompts examples

[visual pier]

I personally like browsing civitai on occasion. You can check the images that use your favourite checkpoint and see what kind of prompts people use ^-^

[crystal edge]

i did and didnt find anything good

i want weird type of style

basically a mixture of realistic characters and beautiful background(nature, buildings, sky, effects) and ancient medieval like style

like in movies related to medieval arabic wars

or vikings serie

[lavish lake]

https://discord.com/channels/1002292111942635562/1011743094309396631 can help, also experimenting, the good old trial and error, that's how you learn

[flat tendon]

Lol help

[visual pier]

What do you need ^-^

[weak kettle]

OSError: Can't load tokenizer for '/content/data/realistic'. If you were trying to load it from
'https://huggingface.co/models', make sure you don't have a local directory with the same name. anyone know why its hapenning on dreambooth

[pale latch]

you tried downloading it from hf through the api once, but didn't let it finish and it closed early, leaving a partial corrupted file. so now when it tries to load that file, it can't. delete it and try again . it might be saved to the appdata folder %localappdata%\huggingface but i dont know how db extension do it

i just load the models manually myself. i got my own folder

[pale latch]

civitai. seriously.. maybe not what you want but what you want isn't the only thing under the umbrella term of "good"

instead of just copy pasting prompts, look at how they're doing it so that you can figure out the language and get the prompt you actually want

[tidal magnet]

Hello guys! What's the best (and preferably fast) method to do regional prompting in Automatic 1111 at the moment? E.g. if I want to generate a picture with two distinct characters.

[pale latch]

HOLY SHIT, nevermind that. don't use civit to learn prompts. i just opened it up and it's all hardcore porn on the featured top 50

beyond hardcore.. like woah buddy.. we're talking about socks like how they turn inside out. top post this week. wtfffff

[tidal magnet]

There is a filter available at the top-right part of the screen, though some lewd art still slips through

[pale latch]

i opened the fitler and it said "videos or images" and that was it. so i turned on only images and it's all teenagers showin boobs

[trim magnet]

saw a girl showing her ankles there even with filter

[pale latch]

donnnnnnnnnnt send people ot civit owahhhwoahwoah i have not prompt explored there for some time

i'm not a fucking prude or nothing but WOAH

lots of butthole closeups

wonder if openart will last now that altman's ousted

[tidal magnet]

No, there is a Browsing Mode button next to your account icon in the top-right corner.

[trim magnet]

plus ppl who arent logged in cant see nsfw

[tidal magnet]

It has an icon of an eye crossed with a single line

[pale latch]

oo openart.ai moved away from openai long ago and is focused moreon other systems now

[tidal magnet]

@pale latch I have just found a picture of you on CivitAI: https://civitai.com/images/3587104

Lol

[pale latch]

problem with filters is they'll block softcore and hardcore and extremely hardcore indiscriminately. 13+ 17+ 18+ on the blurs, none of it is meaninful. trying it that way and it's a total fucking minefield. so you don't show the content at all and then you get zero babes at all because you're either into XXXXXXcore or nothing. no between. i've used civit with the filter off for a while. it was never like this

[trim magnet]

her legs are showin

[pale latch]

dude i got a pornstar model trained on civit. stop clowning yourself

[trim magnet]

no the panties part

[pale latch]

you know why i got this profile pic? cause i'm too sexy for this shirt. so sexy it fucking hurts

hammer time / / / / / / ...... \ \ \ \ \ \

[pale latch]

you know what isn't sexy? fucking goatse

[tidal magnet]

True. Unfortunately, these age ratings don't matter much. Often absolutely vanilla pictures (even without actual nudity) get 18+ for some reason.

[trim magnet]

i bet theres a lora for that thing 😬

[tidal magnet]

#1 thing I don't want to google for today

[pale latch]

https://i.imgur.com/xBfjy3K.png for the training data. risky click of the day

[visual pier]

The keyword (duo) works great, especially with landscape aspect ratios ^-^

[tidal magnet]

Kek. Is this really in Bioshock?

[tidal magnet]

I'll try that, thanks. But as far as I remember, it still tends to heavily bleed the traits between the characters. Gotta try Regional Prompter, as I remember it working fine for me. But I thought there are some newer methods.

[visual pier]

Might be because I use a furry checkpoint, but the posts at least can be inspiring for a few key words :O
But over all I would try to look into the basics of prompt crafting. Lots of the prompts on civitai are just badly constructed on top of the other issues

[visual pier]

Yeah its doing that alright :D
SD is not actually good for more than one character, so its great that you can do it at all ^-^

[pale latch]

it is but i think it's just a design coincidence. anotehr gem from the vaults. https://external-preview.redd.it/iheHchyBdzhQtGlu68za10uKhV878xc7Qm3iC6IGocQ.png?auto=webp&s=e93853cb441468fa0a9e0d210af570c22295e69a

[tidal magnet]

"Gem"

😳

As for this "regional promting" (or how would you call that), I've seen a pretty interesting workflow using the IP Adapter and ComfyUI: https://www.youtube.com/watch?v=vqG1VXKteQg

But I'm not sure if it can be 100% applied to Auto1111

Gotta look into that

[crystal edge]

what can i write to make the image shows more content

i dont like zoomed in portraits

[visual pier]

(far shot) is a great keyword. Other than that look into photography perspectives ^-^

[tidal magnet]

I usually use something like "full body view". Also, if you are generating a character, it often helps to specify the body part that is not getting into the picture. E.g. if you describe both footwear and hair of the character in your prompt, it's more likely that you get the wider shot as a result.

[crystal edge]

i tried far shot

they still give me portrait

[visual pier]

Haha I also use (full body view), but often times I specify (feet wide apart) to draw attention to that ^-^

[crystal edge]

i want more than full body

i want to show the whole scene

with environement

and buildings

[visual pier]

Ah then try using a landscape view aspect ratio

Like 16:9

[crystal edge]

should i write 16:9 in prompt

[visual pier]

Which turns out to 800 width and 450 height before upscaling ^-^

[visual pier]

Whatever UI you are using - there should be a direct way to control width and height

[crystal edge]

it still give me portrait haha

i dont want the face to be zoomed in

i want people to be in the background

[visual pier]

Which UI do you use?

[crystal edge]

automatic 1111

[visual pier]

Yeah that has easy Width and height controls :O

[crystal edge]

yes i did change the ratio

1280-720

[visual pier]

Nice

[crystal edge]

they still give me portraits

[visual pier]

Oh boi. I hope thats the ratio after hires fix?

[crystal edge]

check dms

[flat tendon]

If this is riftfusion

?

For music

[tidal magnet]

What model do you use?

[crystal edge]

colossus project xl

[gleaming spoke]

Anyone know where i can find a guide to make lora?

[tidal magnet]

Yeah, sorry, cannot really try it. My config doesn't work with XL models well. But in general, if you want to see something, you should describe it in the prompt. I know it sounds pretty obvious, but it works most of the time.

So if you want to see a city landscape, describe it as a "far shot of city landscape", or "wide shot of city buildings", or something like that.

[novel roost]

Satya: Will you return if I fire Ilya?
Sam: Yes.
Satya: Hold my Masala chai

[true vessel]

HI everyone
I am new in the field of stablediffusion,
I am seeing this thing when one can use the image nd change the outfit or style an all, lets say i have my image and i just want to make it like caption America clothing, same image or lets say
change the background or look,
please guide me through what should i check or learn, what models i should use , what prompt i should use ?

[granite fiber]

/dolphin

[visual pier]

Thats inpainting I think. You can do that in the Image2Image tab in A1111

[bleak matrix]

Good morning, everyone! How are we all today?

[bleak matrix]

If you want to learn more about prompting, I suggest checking out #📝｜prompting-help as well as #1080946152318443610 !

[visual pier]

Moin ^-^
I am bursting at the seams with enthusiasm

[bleak matrix]

Me too! I'm just having some fun with the bot and so forth. What's up?

[visual pier]

I am currently unraveling ComfyUI and StableSwarm and using some of their benefits ^-^

[bleak matrix]

Nice! How are you liking it?

[visual pier]

ComfyUI is almost brutalist in its approach to refuse having an intuitive UI. Its technical to a fault in my opinion

StableSwarm does a great job in providing ComfyUI with an actual UI. Its aaaalmost beginner friendly :D
All in all I am grateful that A1111 exists as an entry point into the technologies

[tidal magnet]

Yup, A1111 is still unmatched for me

I'm sure though that ComfyUI is great for complex workflows, e.g. if you are using different models at different steps during the generation

And stuff like that

Also, it is a great and fast backend for other projects

Right now I'm simply amazed by the Krita AI Diffusion plugin

https://github.com/Acly/krita-ai-diffusion

Regular generation is a bit too slow for me, but live painting with LCM sampler is absolutely great

[visual pier]

Exept for the very specific case of using two or more models during multiple pass text2img... I am struggling to imagine any work flows that actually benefit from all this technical stuff instead of focusing on better prompts and such

[tidal magnet]

Well, for me it was also interesting to get under the hood of Stable Diffusion and learn how it works (even just a bit)

But yeah, A1111 is more than enough for 90% of ideas

[silver inlet]

Ahoy
I'm working on my master degree diploma and I'm researching stable diffusion method
Currently I'm writing about sampling methods
so, are there any resources where I can read about them more?

and, yeah, I've seen this one https://stable-diffusion-art.com/samplers
but it seems too casual and too much of just an author opinion so I'd like to read something with more formal/professional attitude idk

ofc, bc that's a begginer guide 💀

[gloomy mirage]

whats the best way to use sdxl nowadays?

[silver inlet]

ofc, that's what i've said

[pale latch]

"of fucking course" why be hostile? forget it. gonna have to doubt that you're a phd candidate at this point. just weird flexing

[silver inlet]

bro, chill

[pale latch]

naw

[silver inlet]

I literally said that stable-diffusion-art is the first thing that popup, yet I can't really use it

[pale latch]

yeh and?

i should care now?

[silver inlet]

wow

💀 💀

[lavish lake]

let's keep it civil please

[pale latch]

they called a mod here because i revoked my advice after hostility? weird

[lavish lake]

nobody called any mods, I was just keeping a look on the channel and saw all this

[silver inlet]

Ofc is an abbreviation for of course.

I had no clue that it could mean other things tbh

[lavish lake]

@silver inlet there's tons of documentation and info about samplers online, a simple google search could help you get started

[silver inlet]

yeah, but I thought that asking for some help here could be really useful

[lavish lake]

that's fair

it's a somewhat complex subject and depending on what you are doing, it's best choosing ones over others

check graydient.ai guide

[pale latch]

ofc has alwyas been a hostile way of saying "no fucking way?!" in my vernacular and experience

[lavish lake]

let's get past this

[silver inlet]

yeah, but as I've seen it, there r a lot of literal developers here so yeah
thank u tho

[silver inlet]

got u, my bad
just chill, I didn't mean nothing like that

[pale latch]

we are past it. lets stop doing weird mod flexing. i was explaining where my understanding came from. now we understand . there wouldn't be that if i never posted it.

[silver inlet]

I didn't call for mods or something 💀

[lavish lake]

if you are looking about specific samplers or combinations used on the bots or apis, that's propietary information that's not disclosed publicly but as I said, guides like the one at graydient are good to understand the differences

[silver inlet]

ok, thank u, buddy

[lavish lake]

you can also check the guide made by one of Stability's QAs https://www.youtube.com/watch?v=N5ZAMa3BUxc

[livid ginkgo]

is there a tag to tell the bot NOT to do a video?

[lavish lake]

yes, if you press tab after prompting, it will let you choose image or video

[silver inlet]

thanks 💕
at least will be useful for better understanding

[harsh ether]

lolo

[wraith acorn]

Guys can we add a photo of our choice to generate?

Or just words?

[misty rune]

How can I get more roles in discord?

[clear otter]

How on earth a gay couple kissing on a bot room not considered as nsfw not been blurred out? 🫨

[pine fiber]

how is people kissing nsfw

[clear otter]

It's the idea of tolerating the general age here

[lavish lake]

https://discord.com/channels/1002292111942635562/1072236594930647050 😎

[sleek osprey]

literally never heard of "ofc" being anything other than ofcourse

[solid sand]

is there a particular model that's best for generating things like store interiors?

[bleak matrix]

Good morning, everyone! How are we all today?

[bleak matrix]

I suggest that you learn more about prompting--You can get help in #📝｜prompting-help We also have #1047197565365538826

[hearty moon]

does stable diffusion and stable audio has affiliate or referral program, i want to promote these amazing ai products.

[visual pier]

xD

[visual pier]

Wae. Since I can now do multiple model two pass text2img my images are waaay better quality than before. So thats a plus ^-^

My floofs are way more floofy

[bleak matrix]

Do tell! And fur=always good! I love drawing it, and creating more of it!

[visual pier]

Im gonna send one or two examples on #🏞｜general-with-images :D

[bleak matrix]

Awesome!

[errant seal]

I Hello. I just start using Stable Difusion 1.6 - specifically text to image. All my images are coming out blurred. How do I address this? Any help would greatly be appreciated by this newbi.

[visual pier]

Are you using A1111/ the general web UI?

[errant seal]

Not sure, I just used the prompt field that first appeared when I logged on. I am not a tech person; was using night cafe but was frustrated by their limitations. Do I have to know how to code to use this program. Thank you for the quick response.

[warm junco]

Hey, where do you use it?
On which website or program

[errant seal]

I did a search through Microsoft Edge, clicked on the Stable Diffusion link which took me to their home page.

I am using a Macbook Air.

[neon ridge]

can somone explain me how this works again?

[warm junco]

What's the site URL. The official stable diffusion website for generating images is
https://beta.dreamstudio.ai/generate

If your on
https://stablediffusionweb.com/
Thats a non official site that pretends to be from stable diffusion.
We dont recommend using it as its false advertising

You can use the bots in #1100170312106127410 for free here

[neon ridge]

its kinda a lot

[warm junco]

What exactly?

The Bot?

[neon ridge]

yes

[warm junco]

Check this guide for the Bot usage:
#1100170312106127410 message

[errant seal]

Thank you for the tip on dream studio, this is what I was expecting. Great help1

[warm junco]

No problem 🙂
Also here is the #1025467151206854736 channel

[bleak matrix]

My recommendation is to check out the FAQ here, where you can find the information on support directly here for StableAudio: https://stableaudio.com/faqs You can find more information about contacting Stability here: https://stability.ai/contact

[bleak garnet]

Hi folks 🤗
For using SDXL you only have to download any XL model , nothing to install right ?

[warm junco]

Hey, to use it localy you have to install a webui

[bleak garnet]

I already use A1111 with 1.5 (I forgot to mention it) I have to install another for XL ?

[warm junco]

Nope. Auto1111 works with sdxl but you may need some editing to your webui-user.bat Commandline_ARGS, depending on your GPU

For 8gb - 10gb vram use:
--xformers --medvram-sdxl --no-half-vae
For 6gb vram use:
--xformers --medvram --no-half-vae

[echo mist]

Heyas theere

Anyone having experience with AnimateDiff and able to help with it?

[lavish lake]

sure, I'd suggest hopping into the https://discord.com/channels/1002292111942635562/1025266140445933648 channel. lots of AnimateDiff users there

[echo mist]

thank youu, that makes total sense

[bleak garnet]

Ok thx I have 12gb

[warm junco]

Then you only need --xformers --no-half-vae and if it takes multiple minutes for an image then add --medvram-sdxl

[bleak garnet]

Last question bro (I swear) :

I heard XL is superior to 1.5 in prompt analysis, better output (these MJ Niji ripoff looks awesome) better consistency ,less need of controlnet but cons are more power need and less versatility due due much less LorAs is that correct ?

[heavy moon]

@warm junco hi can I ask if you know the solution for this
cv2.error: OpenCV(4.8.1) D:\a\opencv-python\opencv-python\opencv\modules\highgui\src\window.cpp:1266: error: (-2:Unspecified error) The function is not implemented. Rebuild the library with Windows, GTK+ 2.x or Cocoa support. If you are on Ubuntu or Debian, install libgtk2.0-dev and pkg-config, then re-run cmake or configure script in function 'cvDestroyAllWindows'

I did the solutions from google but I'm getting the same error

this happens when I try to use facefusion in stablediffusion web-ui

[bright ridge]

hello folks

#💬｜general-chat need some guidance.

[woeful breach]

good morning! I hope this is the right place. Is it possible to dl resulting images from the blackmagic websites that keeps the prompt text as the filename?

[grave sierra]

""Photon"" teaser https://www.instagram.com/reel/CxNu93cMlDq/?igshid=MzRlODBiNWFlZA==

Written and directed by:
@anya.koshka.neon.

@koshka.neon production, together with @osvaydercinematics, @brian.mitro (Spirit view) and @valentyn_grosu for the project @whitemirror_xyz, created a teaser for the upcoming film "Photon" using Unreal Engine 5 and artificial intelligence tools including the program @move_ai_ for animation and @stablediffusion for textures.

The film, which is currently being worked on, tells the story of Jeremy, a young guy from the USA. His strange visions give a special understanding of the nature of people's desires and motives and, through fear, lead him to discover the deepest mystery of life related to photons - elementary particles.

The teaser was presented at the 80th Venice Film Festival in 2023, and the film is scheduled to premiere at one of the major festivals in the United States in 2024.

[pale latch]

Ive always seen it as the snarky sense especially when it's in a one sentence post with no added context.

Witnessed a misunderstanding once before someone thought it was the calm version but their boss who said it wasn't calm at all. Ahhh text . That company was a shit show though.

[warm junco]

Nope sry, you can still ask in #🤝｜tech-support but include more info like what webui and extension your using

[warm junco]

Sdxl is very good in all that but as you say it lacks the lora amount and it needs more hardware performance

[pale latch]

So like, is stability going to hire some of these ML experts from the openai firesale?

seems like it's an everything must go situation over there

[prisma cairn]

How to add Stable Dreamer to my Server

[lavish lake]

you can't, it's only for use on our Discord Server.

[empty peak]

hey there, so I used to use the Stable Diffusion Photoshop plugin on mac and had a style nailed I didn't intend to modify, but I lost access to Stable Diffusion 1.4 and 1.5 due to updates. Is there anyway to access the older versions or are they no longer available anywhere?

[solar sand]

can we do img2img on here?

[lavish lake]

they are deprecated for anything that uses the API (like a plugin), you can still download and use them locally

[empty peak]

thank you

[bleak matrix]

You can also find models in #1047197565365538826

[weak kettle]

guys I have problem with my trained dreambooth model and converted to lora, its creating only face photos, when im giving in the prompt, standing, full body shot etc. What could be the issue:?

overtrained?

[west harbor]

hi guys, i am new fan of stable diffusion, i trained my model with an embeded name, how do i trigger it in my prompt?

it still shows someone else face

[pure canopy]

Hi I have a request. I want to upscale my old videos(noisy and dark) but I can't because I have an android and it requires softwares like topaz video enhancer or hiptpaw. please help me if you have oneone

[wind glacier]

what is the latest stable diffusion model ?

[twin monolith]

how to prompt to stop video generation?

[bright ridge]

#💬｜general-chat hi folks, new to stable diffusion. i have installed that in my local server. need to train that for one purpose. can anyone guide me ???

[bright ridge]

hi brother, i just want to train stable diffusion diffusion in my local for this kind of result, can you guide me?https://i.pinimg.com/564x/9f/b5/83/9fb583b48f7c171388738c8497627586.jpghttps://i.pinimg.com/564x/69/23/4c/69234c51b2b433405f774e5f4ad0ed65.jpg

[bright ridge]

https://i.pinimg.com/564x/9f/b5/83/9fb583b48f7c171388738c8497627586.jpg

https://i.pinimg.com/564x/69/23/4c/69234c51b2b433405f774e5f4ad0ed65.jpg

this to this

[weak tusk]

The Drama at OpenAi is interesting: https://workflowpedia.com/openai-mass-resignations

[pine wolf]

hello

can i send an image to stable diffusion and then tell him to inspire from it discord?

[trail pecan]

when is the sdxl update?

Are there any leaks?

[finite cloak]

SDXL

[rotund anvil]

Hey there, can anyone help me with a prompt? I can't make it look like how I want... :/

[finite cloak]

on the bot? tab after the prompt box and set 'format' to 'image'

[finite cloak]

there's a whole channel about that! #📝｜prompting-help

[finite cloak]

Not sure what you mean?

[finite cloak]

models are released when they're ready to be released

[finite cloak]

sounds like overtraining yeah. You might try generating without the lora and then inpainting the face with the lora?

or retrain with a better dataset

[finite cloak]

I believe you should be able to just switch it to using a more recent model and it'd be fine, right?

[trail pecan]

You seem to be the first to not immediately reveal leaks about the capabilities or insides of the model, but only after the model is completely ready for release. I thought you also show some leaks.

[trail pecan]

Where did controlnet go in the bot? I want to share my assessment that he did his job well.

[finite cloak]

it's disabled for now. May or may not come back in the future (hopefully will). The bot's used for testing things, so it's most likely return date is whenever we next have something involving controlnets to test

[simple comet]

Could you do variations of an image ?

[fervent thunder]

Anyone here got experience using CharTuner?

[crystal edge]

what is the best model rn?

[rotund forge]

Something getting released today or just hype here?

Emad
@EMostaque
What is being released tomorrow?

Guesses go here 👇
1:29 PM · Nov 20, 2023
·
61K
Views

[primal timber]

why does the bot generate videos?

[sudden ruin]

You have to set a format, otherwise it will randomly decide to make a video

[primal timber]

what idiot came up with this?

[sudden ruin]

No need to insult the people behind this. This is mostly happening to gather data as far as I know

[primal timber]

how to specify the format?

[sudden ruin]

#💬｜general-chat message

[lone rampart]

ahoy mates

[jaunty osprey]

anyone knows how to generate multiple images varying a single thing? something like
"a man's portrait at (KEYWORD) at night, epic, good composition blahblah" and the keyword points to a .txt file with a hundred different words like "the eiffel tower" or "barcelona" or "the mountains" or whatever? i KNOW there's a way but im not finding it on google because i forgot what this tool or option was called

im talking about A1111 btw, i don't use online generators or whatever

[cosmic crescent]

Dynamic Prompts extension maybe

[jaunty osprey]

WILDCARDS

i found it by searching what you said, so thanks

[sacred jay]

In one hour according to Emad. My guesses are:

• text 2 video
• semi-agentic text-2-3D scene builder
• stable audio V2 with V1 going open source

[umbral heron]

From https://fxtwitter.com/MysteryGuitarM/status/1726753676885377071 looks like another small LLM?

But I'd love to try Stable Audio on my own hardware

[crystal edge]

who is experienced with proxies and multiaccounting?

[real marten]

hello

I have been under a rock the past 4 months

I have noticed some new regulation about AI
And I didn't see the release of any new groundbreaking image AI program, other than DALL-E 3

could someone explain to me those two topics? thanks

[twilit gorge]

@signal coral could u check ur dms pls 😅✨🙏🏽

[livid basin]

Hello
I have been trying to upscale one photo, without success.
I have done many attempts with different upscaler and settings.
Does anyone with good experience want to give it a try ? And test his skills 😉 ?
And let me know what method he used
You can DM
Thanks

[flint harbor]

Have you tried the Clipdrop upscaler? https://clipdrop.co/image-upscaler

[plush knot]

Guys stable video diffusion announced

https://stability.ai/news/stable-video-diffusion-open-ai-video-model

[brisk mirage]

Oh wow, 40GB of vram

[cerulean bane]

wait list 404s

[silent gazelle]

🥱

[cerulean bane]

is this going to be open sourced?

the video model i mean

[silent gazelle]

https://tenor.com/view/whet-what-ew-can-you-not-stop-gif-9908266

[eager phoenix]

research only 😢

[charred mesa]

epic moment

[eager phoenix]

same question

[charred mesa]

thought it was gonna be 1.6 though 😔

[brisk mirage]

Need to get them workstation cards

[charred mesa]

guess that's staying closed source

or API only

[cerulean bane]

Im an idiot they already published the weights

[past summit]

is stable audio available to download and run locally? why not?

[rotund forge]

https://github.com/Vchitect/SEINE

[karmic brook]

#✍🏼｜rules-and-tos please keep it tame

[weary raven]

Sorry

Its very possible though

[karmic brook]

While things can be possible, we don't facilitate this server as a place to discuss the details of those things, as it can spiral very quickly

Thanks for understanding ^^

[forest trout]

The model being used with the stable audio website is trained on audio provided by audiosparx. Stability.ai has a revenue split with them so that model isn't getting released to the public.

[eternal wind]

i set format:picture or format:photo... voila: it's video

(adds format:video after my format choice)
**horned demon in mist, red eyes, cursed forest, poison grass, night time, cinema scene, volumetric lights aspect:21:9, format:photo** height:576 format:Video

[sudden ruin]

Did you do it like this? #💬｜general-chat message

[eternal wind]

i don't see that option

ops, i see it now

[livid basin]

I just tried the free option, doesn't change anything, but thanks
I would have prefer an option with stable diffusion

[lilac mortar]

does anyone in here know if something exists to show me the command being run when i click generate, so i can use it to build that in python?

[cerulean bane]

do you mean you want to build a plugin that shows the command?

[lilac mortar]

no im wondering if there is one, because im using so many extensions, i cant figure out what to send to the api

[cerulean bane]

is this in a1111?

[lilac mortar]

yeah

[cerulean bane]

if so i believe you can check there api docs

[lilac mortar]

i can see the api docs, but im trying to see how the whole command is built. like if i use inpaint and controlnet, and im using a specific checkpoint, how do i see how to format that in python?

or in api

[cerulean bane]

you could try a library like this to make it easier https://github.com/mix1009/sdwebuiapi

[lilac mortar]

so this repo makes sense, maybe i am asking the wrong question 😬
if i select all my options on the webui (inpaint, mask, controlnet w canny, etc) and i click generate, is there a way to see how that call was constructed so i can then build or use that same formatting when i'm submitting that through the api? or can i export that as a python script?

[cerulean bane]

the normal calls in the ui dont run through an api maybe there's an extension or something that you could use to convert the commands but i dont know of any conventional way to do that.

[lilac mortar]

ok cool tyvm 🙏 so if i get something on the webui setup how i want it, i then need to figure out how to build that in python or with the api calls on my own, it sound like?

[cerulean bane]

yes

[lilac mortar]

there's no verbose command for python maybe that will show the generate task in the command line?

[pale latch]

daily outloud wondering about what stable diffusion 1.6 is

[pale latch]

if you've got apple or an nvidia gpu laptop

[humble panther]

Is stable diffusion something that can be affected by what's happening at openai?

[pale latch]

yeah. they can acquire a bunch from the firesale

stable dalle anyone?

[primal shuttle]

6 months ago I would've shamed you for proposing such an abomination, but Dall-E is surprisingly okieday, recently.

[fervent thunder]

is there an api or anywhere i can use image2vid online

[pale latch]

dall-e 2 was pretty good. like it or hate it it would've been good to have it open still

[pale latch]

https://www.pika.art/ their discord server kinda

stability's app has a wait list

[primal shuttle]

Hate it.

[weak kettle]

are repeats important in kohya when using max steps?

beacuse i think its just calculating steps, but when you give max 3k steps for example it doesnt matter?

or nah?

[primal shuttle]

Repeats help, to a point. Kohya is on my shitlist just now and I want my damn A11 extension working again.

[weak kettle]

kohya is ass

for real

so repeats are important or nah? Beacuse when im giving 50 repeats, or 20 repeats or 10

it always 3k steps when im setting 3k max steps

[primal shuttle]

oopooh. Too many brah. How many images you training?

[weak kettle]

and repeats doesnt influence this i think

31 images, 500 class images

and 3k steps

[primal shuttle]

maybe 5 is what's needed?

[weak kettle]

5 repeats?

[primal shuttle]

Yeah, easily. For 30 images?

[weak kettle]

yeah

30-32 around

[primal shuttle]

I did 100 on 2. Fine results.

[weak kettle]

and how much steps

max

so idk how it works, if im changing the repeats it doesnt matter for the steps bcs its always 3k

[primal shuttle]

Are you using the dreambooth trainer? I've gotten better results with the LoRA trainer.

[weak kettle]

kohya

but im training model dreambooth

then lora

without captions

https://rentry.co/lycoris-and-lora-from-dreambooth look ast this, and go to script (train_morgan.sh)

is it possible to replicate these settings in a1111 extension

just to copy this config or something?

[primal shuttle]

Check https://discord.com/channels/1002292111942635562/1004159122335354970

Kohya SS is incredible, right up until it can gobble a whole-@ss bag of d1cks and eat my @ss with a spoon. Just saying. Crashed compy harder than I've seen a compy crash just now. That was a whole world of BSOD bad.

[pale latch]

the world of pytorch can really bsod the shit outta compies

make sure all drivers are up to date. all overlays are closed. even a bios update helps

not just gpu drivers but anything touching the system

[spare spade]

Why did he generate a video for me???

[livid basin]

Does anyone knows why those models does not exist anymore ? Or if they've been replaced ?

• Deliberate
• Reliberate
• Lawra (LORA)
  Thanks

[lavish lake]

Where is this? They still up at civitai (i think), could be that they made newer versions and those got deprecated?

[livid basin]

that's the question because they are not there anymore

[halcyon crown]

is there any colab for stable video diffusion?

[lavish lake]

please follow https://discord.com/channels/1002292111942635562/1176560067961696276 they posted a bit up one

[halcyon crown]

thanks

[river plume]

hello. is there a way to use the bots in this channel for in-painting and out-painting? Thanks.

[normal pagoda]

I'm running A1111 and plan to transition to SDXL for some things... Should I switch completely over?

[bleak matrix]

Good morning, everyone! How are we all today?

[thorny frost]

Afternoon! 🙂

[bleak matrix]

How's it going?

[thorny frost]

All good here. Still trying to wrap my head around all of this... but I've had some great help from @visual pier

[visual pier]

I have been summoned
Moin ^-^

[bleak matrix]

Whatcha all up to?

[mild bone]

hello everyone

who knows why civtai.com can't open ??

[hasty hornet]

idk , it works for me

[normal pagoda]

Still puzzled about this...

[normal pagoda]

Heading home from work and trying to set up SDXL after a nice shower

[bleak matrix]

Nice! You can always use #🐝｜swarm-ui

[fervent thunder]

just use it additionally

SDXL can do some things sd1.5 cant

[thorny frost]

So, SDXL can be used with A1111? Or you need a different WebUI?

[humble ravine]

is it stupid that I joined and already have applied for mod

I been active with Stgable diffusion based AIs for ages so im not new to it

just nerver knew there was a discord for it!

[pale latch]

i've never understood masochists but i guess someone has to do the job

[humble ravine]

also how are you all doing

[humble ravine]

I have a LOT of moderation expirences

[pale latch]

thats my problem too. maybe i was moderating wrong communities though. (gamers)

I don't think it's stupid. KEeping communities safe is a great volunteer effort

[humble ravine]

I mostly did gaming and pride besides leonardo(an AI server im greeter for) but I did just join but then they did place the application on the first channel I see

[pale latch]

i'm not a fan of leonardo but things change. Their mistakes are past the tide line at this point.

they haven't kept licensing exclusive deals for community models since they were doing that when sd2 came out

[humble ravine]

there kind people tho

and I use invoke too

heck I started with invoke

[pale latch]

I'm sure many business school grads are but they also were signing exclusivity contracts with model trainers. Mistake in hindsight and they seemed to have stopped that behavior. it's been a while since that initial "wow i don't like these guys" happened

[humble ravine]

good to hear

can I sell models

[pale latch]

sure! my issue was they were approaching public community model authors and signing exclusivity rights to the already released models. if you release a model from the start and ask for $, i'd pay if it was interesting enough

i bought px8

[humble ravine]

mine is based on pictures from my father

who rised the world and knew steve jobs and bill gates

and also was a lead of the karamapa foundation

he passed away

due to the source images its very viarable

[pale latch]

careful about doxable info. that's just my inner spy persona being like "woah tell them to chll out on that!"

looks like pxl8 isn't being sold on the main shop site anymore. just the free version is available now. https://devilismyfriend.gumroad.com/l/drzhn

[humble ravine]

hence me using filtered images

and many people

even if most are dad and his GF

both deceased

so its not even doxable

[pale latch]

you'd know best. i'm just projecting i think

[humble ravine]

its fine

[pale latch]

welcome to the server experienced vet!

[humble ravine]

look in #🏞｜general-with-images @pale latch

I am currently training 5K+ of images

[finite cloak]

it works in auto, just a little wonky and lower performance vs comfy-based uis like swarm

[tidal magnet]

So is Stable Swarm good as a standalone UI? In terms of user interface, extension capability, etc. Of course I'm comparing it to Automatic 1111.)

I always thought it's main purpose was to use this GPU Network, or how is it called

Which I'm not really interested in

[humble ravine]

I am a member now

idk if that is new or just the default role

[finite cloak]

yes, yes it is

the multiGPU is just the feature that inspired the naming

(and is a really cool feature, but swarm is a lot more than just that one thing)

[finite cloak]

that's the default role

[tidal magnet]

Sounds good! And looks good too (watching some overview at YouTube right now).

What I want to try is to optimize some XL model for my 1060 6Gb, lol.

Don't wanna to dive too deep into Comfy again

[finite cloak]

swarm uses comfy as the engine, so you'll get the full speed of comfy's optimized core without having to dive into the noodles

[tidal magnet]

Well, actually it worked kinda fine when I tried it on Comfy, especially with LCM LoRA. However, LCM only speeds up the actual sampling steps. So it worked fast as long as I only changed the sampler parameters.

But on my setup even a simple change of prompts caused a very noticeable slow down in performance. And don't even talk to me about changing the model...

When it took about 9-10 minutes to load some random XL model, I decided to drop that idea for the time 😅 But now wanna try again.

[tidal magnet]

Sounds awesome

So as I understand, Stable Swarm takes all the power and customization capabilites of ComfyUI and streamlines it into a more comfortable and easy to use interface, with the ability to add additional custom workflows via Comfy.

Now the only logical step I see is to integrate all the common workflows into the base version of Swarm and get an Auto1111 with a better backend engine 😅

[humble ravine]

oh

how do I get the devs role or is it staff

unsure if devs is for all developers or just developers working here

I guess its a staff role

[finite cloak]

Devs in yellow is staff yes. You can get other roles in the #👥｜roles channel

[humble ravine]

ah I applied using the forum in #1072220168534642768

what role will I get if accepted or does it depend

sorry if I ask too much

[finite cloak]

if you applied to be a mod, you'd theoretically get a mod role

[humble ravine]

ah I hope I make the cut

if not I still be here showing my models/art

[finite cloak]

I'm not involved in the decision-making on that, but I don't think it's likely you'll get accepted if this is your first day here and you're unfamiliar with the channels & roles. You'll probably want to wait off on trying to become a mod until you're, yknow, familiar with the discord

[humble ravine]

they can always delay my app

im not here for moderator im here to work with stable diffusion

I am willing to help ofcourse

[finite cloak]

The main swarm Generate tab auto-generates workflows on the fly rather than using pre-set ones

trying to premake all common workflows would be pretty difficult just from the mix-n-match capability

(eg want to add a lora? that's a whole new workflow. Want to add 2 loras? that's also a whole new workflow. Thus, swarm auto-generates on the fly based on your params)

[humble ravine]

can I btw use multiple prompt templates

on training my model

[finite cloak]

Depends what software you're using to train, but generally your best bet is per-image captions

[humble ravine]

stable-diffusion-webui

[finite cloak]

As in, Auto WebUI? Are you... training a TI Embedding? that's the only trainer built into auto

(well, and hypernetworks, but, lol)

If not, you're using an extension not the webui itself

[pale latch]

wondering about a training thought i had. often i use generations in training data. what if i throw the prompt i used into the caption? thoughts?

[humble ravine]

pinged you showing the settings

[tidal magnet]

What if I want to add some more complex stuff like ADetailer or Roop/Reactor though? I guess I would still need to dive into the noodles, or find a relevant workflow online and run it once in Swarm?

[finite cloak]

oh, yeah, that's making a TI

[humble ravine]

I hope it will work my input is very variable

[finite cloak]

You... uh, should not be feeding 5000 complex images into a TI

[humble ravine]

its not one type of image

[finite cloak]

You'll want to train a checkpoint model or a LoRA

[humble ravine]

[finite cloak]

yeah TI won't work at all

[pale latch]

lora going to be better. that kohya like i was saying

[humble ravine]

welp I am training it so better finnish and see what it makes

im used to LoRAs but my LoRAs are always one artstyle

[finite cloak]

Swarm has ADetailer support built in (you just need to load the comfy extension to the comfy backend and download the models - in the near future that part will likely be automated away), you just drag an image to the prompt box and then select adetailer model on the left

[humble ravine]

these are many diffrent ones from irl photos to buddhist art

[finite cloak]

roop not yet but probably should be added. FOr now, yeah, noodles

[humble ravine]

well it seems to be working

[pale latch]

i wonder what i should make for my next published lora. maybe anohter celebrity face

[finite cloak]

LoRAs are commonly one thing at a time, but can be several. TIs have to be only one concept

[tidal magnet]

Cool, thank you for answering. I will probably try it today.

[humble ravine]

oh well most are IRL

[finite cloak]

wait sorry i had a brainfart lol - I just said "ADetailer" and described IPAdapter lol

[pale latch]

v2 of madball lora, i'm going to split it up into multiple concept folders

[tidal magnet]

Lol

[humble ravine]

most of the images are photographs as my dad toke them

[finite cloak]

ADetailer swarm actually has a better implementation of than anyone else and it's purely built in

[tidal magnet]

Huh?

[pale latch]

i was about to say "didn't know adetailer was in comfy now!" but then i assumed i just didn't know as much as you and shouldn't question things

[finite cloak]

just add to your prompt: <segment:face> a perfect face or whatever to use "ADetailer"

but you can also change what object it's matching

[pale latch]

oh i like that

[finite cloak]

(which adetailer requires training a classifier model for. Swarm uses clipseg to allow text specification)

https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7236821

[primal shuttle]

ONLY 24 HOURS OF TRAINING TO GO! WOOOOOO!

[tidal magnet]

I see, that's actually clever. I just was thinking about something like that today.

[pale latch]

one feature i'm dying for is lora autocomplete in prompts. i've got an extension to do autocomplete in auto1111 but its so horribly buggy and has bad ux that i keep it off

[humble ravine]

well seems my model works I will keep traiining it

[pale latch]

type <lora: and get a little drop down of them all that'll filter as you type. just like visual studio or any ohter autocomplete would.

[finite cloak]

Swarm has an actual interface for it

but yeah tab-completer is on the planned features list

[pale latch]

NOICE

i like swarm's interface. nice tiling

[finite cloak]

already have that on the grid generator interface

[pale latch]

squeeeee

[humble ravine]

Well seems the model is making sense

even if its crude art wise

[tidal magnet]

Haha, I just realized you are the dev of Swarm (if I understand that correctly)

[pale latch]

i'm very liquid with UI's lately. i been using stability matrix to try different ones, but i'm going to abandon using that. i have more flexibility without it and it the launcher itself will always crash and close the process AND the browser

[humble ravine]

did not expect a model with 80% photos to make actual art

[pale latch]

TI's aren't models. They're textual embeddings

[tidal magnet]

Regarding that "ADetailer++", can I use something like segment:face without any additional prompts to use the main prompt? As it's done in ADetailer itself.

[pale latch]

you're turning all those images into one catch all token that represents them

[humble ravine]

ah

works good for this ammount of variables

[tidal magnet]

I never really used additional prompts for ADetailer's face fix, just it's capability to quickly inpaint faces at a higher resolution

[finite cloak]

yes

[finite cloak]

oo, not currently (other than, yknow, just copypaste the prompt text) but I could probably add that

[tidal magnet]

That would be cool, as I suppose I'm not the only one who uses ADetailer that way, lol

But using the CLIP Seg is absolute fire. Wonder why haven't anyone implemented this into some extension for A1111 yet. Or maybe they did, lol

[whole cliff]

Hey, can we no longer use images as base for generating?

[finite cloak]

In what context? The bot here? #🗣｜artisan-support-feedback ... the Stability API? #1042896447311454361 ... a local UI?

[whole cliff]

Yup, here.

[finite cloak]

if the bot here, then #🗣｜artisan-support-feedback

[thorny frost]

Ah, thank you!

[humble ravine]

watching your own AI model grow each training epoch is truely amazing

slowly but surely its getting better

it already is making some amazing enviromental pics

it also is good at text and UIX

[humble ravine]

I forgot to filter out the ID cards and the AI is trying to make fraudulent IDs now

fortunately it failed

and fortunately only 1 of them was ressembling a ID card

on the bright side its getting better at faces and finally figured out how to make a server

[fervent thunder]

Jesus loves you all

[pale latch]

he hasn't returned for over 2000 years. exempliary father figure huh

[tidal magnet]

https://www.smbc-comics.com/comic/2008-11-24

[primal shuttle]

He bin hurr bruh.

[amber hazel]

whats a good resource to get more knowledgeable about SD stuff? samplers, refiners and stuff?

[tidal magnet]

Do you mean practical guides or more indepth technical stuff?

For the former I've used the tutorials from https://stable-diffusion-art.com/ for quite a few times

For the latter... Well, I would also like to know that, lol. The articles and videos I find on the Internet are most of the time either too complicated, or, on the contrary, too shallow (again, delving more into the practical side of things).

[amber hazel]

well really anything, my knowledge about SD at all is quite poor

will check out sd art com tho

[tidal magnet]

There are a couple of pretty cool YT channels I check from time to time, though they also mostly cover the practical use of Stable Diffusion. Look for @OlivioSarikas and @sebastiankamph.

@Not4Talent at YT is my most recent discovery (though I've stumbled upon his earlier videos before). One of his latest videos deals with the topic which a lot of SD enthusiasts tend to literally sweep under the rug. Or, well, at least hide as much as possible. https://www.youtube.com/watch?v=oPcQzhhwsGU

Yeah, it's about generating hands

[swift heron]

Hey 👋 just wanted to say hi and that I'm happy to be here. I started looking into stable diffusion recently, and playing with it and it's so fun. I love that stability.ai made it open source too. 💯

[tidal magnet]

Hi! Yeah, working with SD could get a bit frustrating at times, but it is incredibly satisfying to finally produce the picture you had in your mind. Even better is to produce a very good picture that you least expected, lol.

[swift heron]

out of curiosity, I'm planning to try run it locally and do some fine-tuning. Are most people here doing fine-tuning as well?

[swift heron]

Yes! I'm blown away by the output. I also looked around in this community - the art work is pretty amazing.

I just started and haven't experienced the frustrating part yet, that will probably still come lol

[pale latch]

i finetuned some madballs this week

[swift heron]

Just so I know what is awaiting me - what can get frustrating?

[pale latch]

https://civitai.com/models/207437?modelVersionId=233559

[pale latch]

out of memory errors

[tidal magnet]

Honestly, with the sheer amount of LoRAs out there, and with stuff like ControlNET, IP Adapter, Reactor, etc. I don't really need any finetuning most of the time.

[swift heron]

What is madball?

[pale latch]

collectibles from 80s

[swift heron]

oh wow

[tidal magnet]

Or a hardcore punk band from 80s

Lol

[swift heron]

they look cool

almost looks like they're real and as if I could order them lol

[pale latch]

lol unrelated to the band

there were series made recently i know about them cause a friend collects a lot of stuff like this

[tidal magnet]

Well, I'd rarely get OoM errors @chilly furnacecold has mentioned, as I mostly use SD1.5. And it works perfectly even at my 1060 6Gb card (albeit a bit too slow if you start to increase the number of ControlNET models or other processors).

[swift heron]

ah.. is there an ideal hardware setup for finetuning -- one that's still within reasonable budget? Not sure if there's been discussions about this before. I'm looking to renew my equipment.e

[pale latch]

turn that batch size up and you'll find them. training especially

[swift heron]

I saw the minimum reqs on the website, but they seemed a bit low to me

don't want to run into too many frustrating moments lol

[tidal magnet]

I was talking mostly about the times when SD doesn't want to generate the picture you want it to, heh

[pale latch]

fine tuning is usually what we call the deepest levels of model refinement. you need 24gb to unfold all the layers and refine it all.

Dreambooth and LoRA are other methods that we use to refine a model, and they require less memory

[tidal magnet]

Like completely ignoring a pretty simple pose you are trying to apply.

[pale latch]

attention problems are a big deal too. While these models take natural language inputs, they're REALLY stupid and don't pay attention

or if you're using sd15, a 512x512 pixel trained model, and generate a 1024x1024 image, it'll only give attention to 512x pixels at a time.

[swift heron]

oh okay.. I wasn't planning to go that deep. I was more thinking of making the outputs a bit more personalized to start with. Good to know that 24gb is required to unfold all layers though.

[pale latch]

centaurs and siamese twins

[swift heron]

Yea, I haven't tried that myself yet, but I saw some examples from other users.

[pale latch]

sdxl is trained at higher resolutions, so it's base attention is much higher

[swift heron]

do many people go that deep btw?

[tidal magnet]

You can look into the names I've mentioned in my reply (IP adapter, etc.)

Chances are it would be more than enough for your purposes

[pale latch]

most people do lora's i'd think. thats all i stick to. if i were to do full models i'd want a 24gb card. i've gota 16gb one

[tidal magnet]

There are quite a few guides and overviews on YouTube

[swift heron]

I'm sorry, I might have missed it -- which reply are you referring to?

[tidal magnet]

Sorry, this one

[swift heron]

Gotcha. The requirements for lora seem quite reasonable too 😄

Difficulty: easy-hard (less likely to affect the base model)
Style training: good quality
Subject training: good quality
GPU cost: medium (8GB or more)

[tidal magnet]

ControlNet and IP adapter are basically a bunch of additional models that allow you to better control the generation. E.g. you can use depth maps to control how far the object in your scene should be located from camera, or use OpenPose to control the pose of the character you are generating.

[swift heron]

Ah thanks, I still need to look up what these all are. I'm only jsut getting more familiar with A1111

[pale latch]

aaaahahaha sam altman's trollololol video making rounds omg that's a properly timed remix for this meme

[swift heron]

link? 😄

[pale latch]

https://www.reddit.com/r/StableDiffusion/comments/181fgjt/sam_happily_returning_to_openai_after_this_weekend/

[tidal magnet]

IP Adapter allows to replicate the style and/or composition of a certain reference picture. It also has some models for face replication, but if you need a fast and reliable face swapping tool, then definitely look into Reactor. (I'm not even talking about deepfakes, it just really helps with character consistency, lol.)

[swift heron]

LOL 😂

[pale latch]

personally, if i want to do a person, all those guidance models don't compare to a lora

[tidal magnet]

All of the things I've described exist as extensions for A1111, so yeah, get aknowledged with the basics, and then you can dive deeper... and deeper... and deeper.

[pale latch]

i wasn't able to get madballs from guidance models either

[swift heron]

ah that's interesting, didn't think of using face swapping in that way before

[tidal magnet]

True

[pale latch]

the face swapping models are nice too, but they're based on 256 resolution transformers and then fixed with codeformer

there's a lot of limitations there

[swift heron]

sounds like it's going to be a deep rabbit hole 😄

[pale latch]

use reactor on anime generations for instance

a lora will make an anime versionm of the subject trained. reactor will face swap their photographic face onto an anime face

[tidal magnet]

Also true 😔

[pale latch]

and with reactor, because it's 256, you can see the pixels. so codeformer fixes it , but then sometimes what reactor swaps and codeformer fixes, don't line up and you get a big bunch of pixely hair

[swift heron]

I think I might go with 16gb as well, feels like something safe to start with. It will probably be a while before I will think about upgrading, but why would some people do full models though - for what kind of use cases?

[pale latch]

yeah 16 is nice. i want 24 but i've lived a lot with 16

[swift heron]

Ooh I would like to try that out 😁

[tidal magnet]

I don't think you do, haha

It looks... weird

[swift heron]

Do you have any experience with not doing it locally but renting servers?

[tidal magnet]

Definitely not me, but I know a lot of people are using rented GPUs for training

Maybe other people here could elaborate more

[swift heron]

Am thinking that IF I really get that far along, and my hardware isn't sufficient anymore, perhaps I can rent GPUs too, but not sure how the UX is.

Or I don't buy new equipment now, just get familiar and play with things while renting GPUs -- and decide later what equipment to buy

[swift heron]

Do you have an idea if more people here rent GPUs or run it locally?

[tidal magnet]

Nope, I have just joined this Discord a couple days ago myself, lol. Though I have been playing with SD for quite some time already.

[swift heron]

Cool, we're both new here! Although I'm much more of a novice yea

Have you played with other models before joining this server?

I want to try out so many things, but after hearing you both talking about SD, I think there will be enough for me to learn and try out here before I go exploring on hugging face again.

[tidal magnet]

I think I tried Midjourney for some test generations, but decided to stick to SD in the end. Just because of how configurable and modular it is.

@finite cloak Hey man, sorry for a pretty basic question, but where does the Stable Swarm UI keep the server logs? Cannot get IP Adapter to work, as it says the following: No images were generated (all refused, or failed - check server logs for details).

[arctic sedge]

I heard something about Stability updating SD1.5 to 1.6?
Can anyone tell me if they know what the differences are yet? 👀
Curious.

[pale latch]

noone knows /guitar

its on the api. some comparisons were shown. that's it. thats all we know. some have said it won't be compatible with any of the loras and controlnets is why its not released

[arctic sedge]

Ah. I see. Thanks for letting me know. haha! xD

[limber gulch]

Q* is here AGI is here

Its over

[pale latch]

it's probably not agi

a generalized model? sure. general intelligence? nawh

[humble ravine]

okay the model is okay enough

time to work on my woman/anime/dental LoRA

why those things AI fails a lot at putting teeth in mouths

and I wanted it to be better at faces and have anime style options

I also included NSFW but I have my doubts those images make anything good there like 3 of the 1K

but well while making a LoRA for humans it would be interesting to include some questionable ones

as I did try and AI has failures in nakid pictures as well as when it holds items

so might make a LoRA for hands

oh also my LoRA has a bunch of PD paintings

so im truely wondering what it will make

also making one with all of the pics from dad

after this one

so one for mostly woman faces and cartoon

and one with anything my dad stored on his drive

[pale latch]

i looked Q.ai up. they sound like a theranos scam

[humble ravine]

ai scams

[pale latch]

https://www.reuters.com/technology/sam-altmans-ouster-openai-was-precipitated-by-letter-board-about-ai-breakthrough-2023-11-22/ oh they mean this clickbait nonsense exclusive being passed around recently

first claimed that altman was fired over q* then backtracked it because they rushed to publish something saucy

[vale shell]

Hello, I am new to all of this, is there a specific subsection to discuss and receive tips on LoRA training?

[lilac crypt]

is there any youtube video to get training to get started on Stablefusion along with some sample prompts. I am a newbee wanting to learn

[sterile maple]

can I use stablediffusion to create a new image from a base image?

is there a subscription needed and what prompt do I use to start with a base image the then alter into a new image?

[verbal halo]

I am finding lots of very high quality Gifs made with Prompt Travel in Animatediff -- but no one mentions how to get that crisp quality. Any one has a good guide?

[rancid hedge]

Hi, anyone have experience about animateDiff in ComfyUI connection lost issue?
ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None)
handle: <Handle _ProactorBasePipeTransport._call_connection_lost(None)>
Traceback (most recent call last):
File "asyncio\events.py", line 80, in _run
File "asyncio\proactor_events.py", line 165, in _call_connection_lost
ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host
Prompt executed in 3839.24 seconds
ERROR:asyncio:Exception in callback _ProactorBasePipeTransport._call_connection_lost(None)
handle: <Handle _ProactorBasePipeTransport._call_connection_lost(None)>
Traceback (most recent call last):
File "asyncio\events.py", line 80, in _run
File "asyncio\proactor_events.py", line 165, in _call_connection_lost
ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host

[untold lantern]

is this an official sever?

[humble ravine]

I think it is

[untold lantern]

Can we generate videos where the characters move? I generated a video but it's only the camera that's moving.

[humble ravine]

stablediffusion is a IMAGE AI it is possible to make a gif from the images but you got to do that manually

[untold lantern]

wdym?

[humble ravine]

can however use adobe animator on AI characters but there still images

[untold lantern]

but i've heard of stable video diffusion

[humble ravine]

no videos as videos are not easly fully made using AI

[humble ravine]

it may make videos but only without sound

and only if you make a tool that makes it frame to frame from images

[untold lantern]

ik, but how can i access stable video diffusion here?

[humble ravine]

and due to AI being unpredictable this is might fail

[humble ravine]

#▶｜stable-video-diffusion

my bad

I forgot there was stable diffusion videos

[untold lantern]

but is there a stable video diffusion bot here to genrate them directly?

[humble ravine]

I only use the image stuff

[humble ravine]

I recommend doing it locally anyway

if your GPU can handle it

[untold lantern]

my colab isnt powerful enough and im broke

[humble ravine]

I feel your pain

my 3080 barely holds up

[brazen plover]

Do I only have the free version since the dream bot Channels are locked for me?

[humble ravine]

and that thing is pretty beefy💀

the only thing I know better then a 3080 is a 3090

[untold lantern]

Will there be a svd bot here? plssssssss

[humble ravine]

idk if there is even one for images here

[untold lantern]

there are for imgs in bot channels

[humble ravine]

oh well it wont take my custom LoRAs anyway

[untold lantern]

what are custom LoRAs?

[humble ravine]

LoRAs are like plugins for the AI

they make it better at specific things

there like plugins to models but unlike plugins for applications there trained using AI like a model

so like mini models that get injected into for example SD2 or any other model

so you can see LoRAs as expansions for AI models

[thorny frost]

Even a 4090 won't cut it for Stable Video Diffusion locally... you need 40GB of VRAM.

[humble ravine]

im talking about image gen anyway unlike LoRAs finetuning is making a new model based on existing ones

[thorny frost]

This is all very new to me. I'm still trying to understand what does what...

[humble ravine]

it's fine if you compare AI models to games

LoRAs are DLCs and finetuning is a new game in the same engine

im currently testing my LoRA

but first im making a normal gen using SD2

its sad to see that my old LoRAs are incompatible

[thorny frost]

What is a normal gen?

[humble ravine]

with that I mean just the base model

no LoRAs or control nets(extensions)

I forgot I put way to much anime in my LoRA

need to change the prompt to even make it ressemble the same thing

[thorny frost]

Ah... OK. This may be a dumb question... but how to you make a normal gen? I don't mean the specifics (I'm sure that's fairly complex?), just the outline of what you'd need to do.

[humble ravine]

just select a model and do not touch the LoRAs or control nets

idk if it is the actual wording for it

[thorny frost]

I think that's half the battle for me. I mean, remembering the names, and what thing does what to an image. Haha. The bit I really don't get when you say you're making a normal gen is that if you're using a base model (in this case SD2, is that right?), and you're not using any LoRAs or other extensions etc. then how do you influence the base model?

And apologies for all the questions!

[humble ravine]

I know the basics

like controlnets and LoRAs

controlnets are code based extensions LoRA are expansions of the model

[thorny frost]

Right... I think I understand. So what are you doing to make a normal gen then?

[humble ravine]

I made a anime LoRA by mistake

[thorny frost]

OK... I think I'm getting this. So you want to to 'de-anime' your LoRA? Or you are starting from scratch? If it's the latter... what process do you use?

[humble ravine]

im making a new one

[thorny frost]

So you're making a new LoRA for SD2?

[humble ravine]

I made one but Im making a new one for 1.5

I do wonder about AI copyright

especially seeing that this new SaRA model is made from images I do have rights on due to the one taking being my deceased father(I inherit his copyrighted work as a child of him)

not to mention most humanoid ones in it are of him and sari both them are deceased and I can garentee if dad was alive he would 100% aprove of me doing this

I know dad he loved tech and he would love AI

and most are enviroments

this SaRA LoRA will be amazing in variations

dad toke so many pictures of diffrent enviroments,people and even some stuff normal people do not have access to like sacret buddhist statews and server rooms

[thorny frost]

That sounds amazing!

[humble ravine]

it is

and legally I own the images most are of humans that passed away,wont mind me training AI with there face(monks would not mind) or of stuff that would make the taker of the picture the holder

in the case of dad the only copyright holder is me due to his passing

let me check just in case

Like any other property you own, what normally happens is that ownership of your copyrights is transferred to the heirs of your estate.

I guess this includes photos otherwise copyrighted by dad

[thorny frost]

That makes sense. And I'm sorry to hear about the passing of your dad.

[humble ravine]

its tragic especially as he passed away suddenly

but AI can make him not die in vein

also im training the data and already see an avr_loss of 0.12

likely as its extremly viarable

AI LoRAs/Models love data of diffrent stuff

[past summit]

can you apply q* to stable diffusion

[humble ravine]

what is Q*

also SD is just a model

so you can in theory use other AI to train it

if you make a image(and own the rights to it) you can train AI to remake it

for example

if you draw anime and cram that into a model

the model is yours to control

but if I cram all of highschool DXD into a model that is infringing

as it can remake all of the characters and uses one style

this is ESPECIALLY a danger if the person is real

I say its okay to use people and anime art but it has to not be one style/person

it makes for better AI models

im also making a LoRA for hands

[azure mica]

Hello!

Are we allowed to promote our services on this server? Because I couldn't find a rule against that. It's about generating images with stable diffusion on cloud GPUs.

[past summit]

so if you hide the theft it's better. got it

[humble ravine]

no its still recommended to use copyright free or your own images

as as soon its regonizable as something it is automaticly infringing its rights

for example if you ask a AI to make homer simpson its the same as overdrawing homer from a screenshot

but if you draw a bunch of stickmans

and cram those in a model

then ask it to make a stickman

then its still yours same if you use picaso paintings or other PD art

same with Voice cloning

If I use my voice to cover adele set fire to the rain

and I have optained lyric and beat rights(and the concent of adele to cover her song)

to cover the song

then I can

even if I use a AI version of myself

as my voice = my voice

but If I take adele's voice and make her sing bumble bee then I am able to be DCMAed by adele

the waters are murcy with dead people

same with mixed work

hence why im using royality free content and dads images

dads images are mine copyright wise due to his passing

[exotic rover]

How do I unpack a model, its a safetensors?

I have a model that I created a while back and it worked quite well but I want to make an updated version but would like to know what images I used to create it in the first place to avoid any kind of duplications.

[humble ravine]

you will need it's origenal dataset

with description and images

else you can only retrain/finetune it

or LoRA it

[bleak matrix]

Good morning, everyone!

How's it going? Happy Thanksgiving!

[exotic rover]

Ok. No problem, I was wanting to check which images I used as I deleted soon after. Nevermind

[solar mesa]

Hy guys
I'm kinda trying to get sum overall knowledge of the SD models and the generational process
Could you correct me if I got it wrong pls.
The trained SD model is a complex model which consists of multiple components:
CLIP - nn to ensure the corresponding between images in training dataset and their text descriptions
U-Net - nn trained to define the details in the latent space representations of training dataset images, and to evaluate the noise in that latent space representation
VAE - consist of encoder and decoder. nn encoder converts images into their latent space representation and nn decoder that converts the latent space back into pixel space image

I feel like I have some grasp, but there are still a lot of dumb gaps
I have read a lot of things, but they cover different aspects and don't give the whole image

[lapis mortar]

star wars is bad

star wars the force awakens is the best movie

anyone of u wannna doooo my homework......

[plain crane]

Hi, im trying to git clone comf ui, but it keeps giving password error. I made a ssh key already, and i tried to paste it into github, but it says invalid key

[lavish lake]

I don't think you need passwords or ssh keys, are you on the real comfyUI repo?

[plain crane]

yes, i tried to use my personal token, but it still says fatal error

[lavish lake]

why do you need a token?

you just need to git clone this: https://github.com/comfyanonymous/ComfyUI.git

onto whatever folder you want

[plain crane]

it is asking for my password, but i typed it is asking me to switch to currently recommended modes of authentication.

[plain crane]

I tried to git clone, but i need to authenticate first before i can clone it

[lavish lake]

let's start from the beginning, where are you installing? OS?

[plain crane]

Comfy UI, mac

[lavish lake]

Ok I understand, the MacOS might be asking for your admin password in order to let you install thru the terminal

that has nothing to do with comfyUI

you need to have permissions

have you installed other things like that before?

[plain crane]

no it is not my admin password it is asking what is my username on github

[plain crane]

usually i use github in colab. Occasionally I use homebrew

[lavish lake]

it shouldn't ask for a password to let you do that

go here on a browser> https://github.com/comfyanonymous/ComfyUI

click on the green "code" button

and download the repo as .zip

unzip it wherever you like on your HD

and then follow the instructions on that page to install the rest of the stuff you need on a mac

pytorch etc

[plain crane]

I thought just downloading the zip but in the directions, they told me to manually install it

[lavish lake]

dont worry about not cloning it, you can install the manager and update everything to the latest version from the manager

[plain crane]

download zip as https or ssh?

[lavish lake]

https I guess

from the browser

[plain crane]

yeah i need a ssh key for ssh, i tried to add one, but it didn't work

Lol it is 1.3mb

[lavish lake]

yeah it's a bunch of text files

you have to download the models and controlnets etc later

[plain crane]

How do i run it?

[lavish lake]

there's a mac section on the github

you need a couple more things for it to work

pytorch nightly

without that it wont run

[plain crane]

i have pytorch nightly

[lavish lake]

did you isntall the dependencies?

[plain crane]

the models not yet

[lavish lake]

no, the dependencies

[plain crane]

yes

[lavish lake]

pip install -r requirements.txt

[plain crane]

i did that already

I will run it now

[lavish lake]

then on the terminal run it with the python command

which mac is it? M1, M2, M3?

[plain crane]

m2

where do i put the comfy ui folder?

[lavish lake]

it doesn't matter, you can have it wherever you like

[plain crane]

I tried to put the comfy ui folder into the terminal, so i can run the main py

but it says access denied when i tried to put it into the terminal

[lavish lake]

do you have experience with the terminal? it's unix based

you have to /cd and then the path to the comfy folder

[plain crane]

sort of, i forgot a lot of the stuff already. I only use it sparingly

[lavish lake]

once you are in the comfyUI folder at the terminal you just run this: python main.py --force-fp16

[plain crane]

It is not working, i tried " /cd/User/.... " and "cd/User/...

[lavish lake]

/ls will list the contents of a folder, /cd will change the directory

depending on where you are you need to navigate, or do the path from root

/cd / is the root

/cd /user/xx/documents etc

[plain crane]

I use /ls then

no such file or directory it says

I used to know this stuff lol, but i forgot. Years ago I learned it.

[lavish lake]

do this "/cd" after the cd do a space and drag your comfyUI folder into that space

it should give you the path

then you can /ls

and should show you the content of that folder

[plain crane]

still says no such file or directory:

[lavish lake]

let's please take this to https://discord.com/channels/1002292111942635562/1002602742667280404

ping me there and we'll try and figure it out

[vapid berry]

Are there discord servers that are more general purpose about AI art tooling and techniques?

[fickle siren]

Seems the AI is struggling again with simple stuff like faces and detail

My images are getting harder to get the way I want as the algorythm seems to have either forgotten stuff or been reprogrammed.

[pale latch]

doesn't work like that

the model is basically frozen outside of the training process

[fickle siren]

But, surely it changes as it learns?

[pale latch]

digital neural netoworks aren't like the brain at all. People lie all the time about that. THey're a mathematical construct. Not a biological brain

[fickle siren]

Oh ok

[pale latch]

training is a very intensive process and you won't be doing it accidentally

[fickle siren]

But it draws upon experience noes?

[pale latch]

no. it's frozen. they only learn when someone runs training code on it and saves a new weights file

[fickle siren]

Oh ok

And then does it use the data from users feedback to learn?

[pale latch]

with chat gpt, you'll see that it'll draw on a particular session . the weights acting on the previous text. sort of like feeding the output back into the same algorithm

doing that on stable diffusion is img2img.

[fickle siren]

So if you bias a convsation with suggestion into what you want to hear, it will tell you what you want to hear?

[pale latch]

user feedback is used in training, but what happens is it's a spreadsheet that an operator coallates and then provides to the training algorithm during runtime.

[fickle siren]

Yea ok

so me spamming the same thing over and over and pressing image a or b as good isn't affecting it in live time?

[pale latch]

Oh yeah. LLM's (chatgpt) love predicting what you want to hear. That's their whole jam. Sort of like Stable defusion is a denoiser trained to predict what would be where noise is. LLMs predict what the next character in a sequence will be

[pale latch]

yeah thats right. It won't start picking up that you love apples because you prompt for them a lot. you might think it is because you start seeing better apples, but you probably were just tweaking knobs towards better results progressively over time

[fickle siren]

Yea ok.

[pale latch]

automatic1111 and comfyui and most UI's aren't set up for training. Well, sort of. A1 has a training tab and some extensions. Either way, you need to run a specific script for it to change, and then you get a new file

[fickle siren]

I see! I made some amazing stuff before but it is like doing the same thing now produces poorer results, but I am using the samed image filname, which has been truncated so has lost some of the details I inputted.

Ok

[pale latch]

you might've had different sampler parameters before

[fickle siren]

So it is far from self awareness yet 😄

[fickle siren]

Yea

[pale latch]

more steps. different cfg . something like that

[pale latch]

that's right. Super intelligence might come before self awareness though. We just don't know. Engineering awareness is a tough concept