💬｜general-chat

[fervent pebble]

If they could, they would..

no doubt there

[full lark]

In my opinion, there should be a big uncensored community model so that governments don't get in the way of stabilityAI.

[cerulean granite]

Who is "they"? Lol

[cerulean granite]

I just linked to it lol

[fervent pebble]

👀

[full lark]

Lmao im blind

Found it

[cerulean granite]

Really quite beautiful. Since it's built on 1.5, it also knows celebrities and art styles

I'm sure eventually stability and MJ will come around, I suspect they want to. They are just looking at the big picture and trying to keep progress from getting stunted by lawsuits and bad press

[full lark]

Does anyone know what is the minimum vram requirement to run 2.0?

at least at 512x512 or 480x480

[cerulean granite]

Once I got the right settings I'm easily and consistently able to generate 1024 by 1024 on 6gb 1660 ti on 2.0/2.1

[fervent pebble]

Stability, yes. MJ? idk about that one. I have listened in to there office hours and its largely a community of social warriors looking for the next thing to be mad at.

[cerulean granite]

Lol, there will always be those. But their time in the front will be limited

[fervent pebble]

I do hope youre right there and you probably are. The amount of times I have heard an AI art bot called racist is enough for 10 lifetimes lol

[bitter laurel]

i should make a chat bot hitler. would that get insta banned? yes it probably would. imagine reviving hitler just for funsies.

[cerulean granite]

Reality is racist, AI is training on reality. So...? The best those people can hope for is unrealistic AI.

Just yesterday, I was inputting song lyrics As prompts, and one line mentioned poverty. Not a lot of white people in those "poverty" images. It's not a hard case to make.

"...she lived in a rough old shack, poverty blew through, but she invited him inside..." if anyone wants to check my work lol

[peak orchid]

I wonder if Stability AI might contribute something to Unstable Diffusion, as they seem keen on helping community projects right now (according to what Emad said)

[cerulean granite]

I'd be shocked if some of the same people weren't involved. I wonder if they have asked for some "community grant" time on hugging face?..

[rustic rivet]

Thanks for explaining that so clearly. That fixed it for me too. One little space was the problem.

[upbeat tangle]

is using 200 imgs for a face dreambooth too much ?

[hushed edge]

sorry for the silly question, but what's the difference between SD 2.1 nonema pruned and SD 2.1 ema pruned?

[cerulean granite]

Does that actually do anything though if you haven't actually compiled it?

[quick narwhal]

Ema for regular use, non ema for fine tuning

[frosty marsh]

hi ppl, anyone knows what the difference between v2-1_768-nonema-pruned.ckpt and v2-1_768-ema-pruned.ckpt?

ow lol

XD that was a nice timing

[cerulean granite]

Probably. Better images > more images

I got great results training 1.4 with only 12 images

[frosty marsh]

what does that mean regular use vs fine tuning?

[cerulean granite]

If you don't know, you probably don't need to at the moment. Basically training your own model.

[frosty marsh]

like actually doing an actual separate model or just allowing the use of bias image to give the model some pointers on how to do stuff that it sucks as

[quaint pivot]

hey everyone, not sure if self-promotion is allowed here, but I run https://futurepedia.io - it's a directory of 400+ AI tools and also has a leaderboard along with it.
would love it if you guys had a look and let me know your thoughts 🙂

[midnight crane]

Does EMA-pruned mean that the EMA weights are taken out of the model? That would explain it, I thought Non-EMA meant that there was no EMA weights in the model, so I thought the exact opposite lol

@quaint pivot Yeah, I saw that on MattVidPro I think, nice site!

[knotty wasp]

what you thought first was correct. ema has ema weights and non-ema has raw weights

[frosty marsh]

BTW I'm following Olivia sarikas tutorial on how to install Auto 1111 and after downloading the model and copying the config link in the model folder, I executed webui-user.bat and it said "couldn't launch Python" and "Python can't be found"
What can I do to allow it to find and launch Python?

[strong tusk]

are you on Windows? make sure Python is installed and on your PATH

[frosty marsh]

yep windows,

let's see if I can find python

[strong tusk]

open a cmd prompt and enter python --version

[frosty marsh]

yeah about that..

[strong tusk]

if python is installed but that doesn't work, it's not on your Path variable

[frosty marsh]

what's a cmd prompt

[strong tusk]

windows key + R -> type 'cmd' -> enter

[frosty marsh]

python can't be found

[strong tusk]

did you install it? it does not come with auto1111

[frosty marsh]

ok silly me

3.11.1 is the right version?

[strong tusk]

i dont know if this is still the case, but when I installed auto111 a month or so ago, the latest version of python didn't work

I have 3.10.6 and still works

also keep an eye out during the install, I believe the installer will give you an option to add python to your PATH variable. make sure that's checked or you'll have to do it manually later

[frosty marsh]

it might actually solve another of my problems, I was using Noesis to export video game models but it didn't work

[tidal bough]

dropping a couple more prize giveaways! good luck everyone ^^

[surreal perch]

@wispy scarab

Congratulations Ivancho#2673!

You have won a raffle prize!

Prize

2500 DreamStudio credits!

Please contact Aether to claim your prize!

@tidal bough

@bright whale

Congratulations Horror#7777!

You have won a raffle prize!

Prize

2500 DreamStudio credits!

Please contact Aether to claim your prize!

@tidal bough

@floral hill

Congratulations jenben#3996!

You have won a raffle prize!

Prize

1 month Discord Nitro!

Please contact Aether to claim your prize!

@tidal bough

@west aspen

Congratulations BSRider#5937!

You have won a raffle prize!

Prize

1 month Discord Nitro!

Please contact Aether to claim your prize!

@tidal bough

[midnight crane]

Ok, I thought since there is an option in Dreambooth to extract the EMA weights from the model that the model with EMA weights would be better for training. That's what's causing my misunderstanding. I don't understand then why there is an option to extract the EMA weights if it doesn't contribute positively to the training.

[tidal bough]

Congratulations to our winners! Please contact me via DM soon to lock in those prizes!!

[bright whale]

somebody calculate the odds on that

[placid wing]

can i delete generated images from the txt2img-img folder?

[knotty wasp]

i've only seen that in the conversion script from checkpoint to diffusers, and then its for the 7+gb checkpoint that has both weights so you can pick if you want the ema or not

[steady arrow]

Why not?

If you don't like em, chuck em

[placid wing]

idk i read somewhere that it could corupt further images that u generate

[lilac reef]

I want to make, or attempt, a script but I have no idea what I have access to. Is there a list of what I have access to so I can manipulate them?

[placid wing]

the whole thing is opensource so you can acces everything

[lilac reef]

I need a frame work or something. I am used to C++/C and we need a frame work with what everything is called so we know how to call them and with what to change/modify them.

XYZ(a,b,c*2) sort of stuff

bit hard to call a function not knowing anything

[midnight crane]

@knotty wasp It's on the Create Model tab on Dreambooth, when you create the model you can chose to extract the EMA weights. Dreambooth revision is c2fc683ebcf4db23cfce22626aa6a7b565994c96.

Don't know why you would have the option to extract them if they aren't benefitial.

Posted the image of what I mean on general-with-images

[knotty wasp]

ah a1111 never used it in that

ema is best for inference so gonna guess it's for if you train with non-ema so you can get the ema weights after

not looked at what the extension is doing though

[midnight crane]

Ok, a lot of confusion regarding this, that's why I'm a bit suspicious of the whole thing.

And of course because I don't really understand it lol

[placid wing]

i'm sorry but i woudn't know where you would find those

[lilac reef]

Hard to make anything if there is no documentation on what shit is. Seriously, I am looking at this one script and how does he know
base_params = {
"sigma_min":sched_smin,
"sigma_max":sched_smax,
"rho":sched_rho,
"beta_d":sched_beta_d,
"beta_min":sched_beta_min,
"eps_s":sched_eps_s,
"device":"cuda" if torch.cuda.is_available() else "cpu"
}
??? Where did he get that info from?

There are tons more stuff we have access to but without a list, and how they are used/called, I can't do it.

[placid wing]

it might be in the vocab.json

[lilac reef]

Doubtful

This is programming stuff so the end user files wouldn't need to have that

[steady arrow]

I think not

[upbeat tangle]

lol,i should try with 20 then

[hushed tangle]

minimum VRAM for auto1111 dreambooth?

[crimson magnet]

Have there been any recent developments in terms of tutorials or setting/parameter explainers for training hypernetworks in Auto1111? More specifically, a stress test or semi quantifiable way to tell if an embedding was trained well? I've recently trained one using my own artwork as input data and, from what I can tell, it's working shockingly well. Thing is, relative to my other attempts, I have no idea why. I had to stop training due to low memory errors at around 3500 steps, (no idea, wasn't an issue before with my 8gb 3060ti), and assumed that + using more or less default settings and prompt templates it would be bad, very much not the case (I think?) lol. Idk any insight into this would be helpful.

[wise stratus]

No due to legality. It’s pretty much impossible to remove kids and nonconsensual stuff so we are focusing on SFW. Would advise them to take care.

Community fine tuned models really aren’t that compute intensive and the base with constantly improve from us at least.

Hope lots of folk release lots of open source models appropriately.

[fervent thunder]

how i can use the bot whit my pic?

[lilac reef]

@wise stratusWhat about my issue with the light in every single image 2.x makes now?

[primal marten]

Hi! What does "fine tuning" mean? In reference to the difference between non-ema and ema versions

[primal marten]

Oh I see it's already been answered lol

[primal marten]

Thanks

[midnight crane]

@wise stratus Hi Emad, I saw that you are seeking collaboration with Elon, is there something in the plans between you two? I know he's involved in OpenAI, so I thought it might be involving that 🙂

[wise stratus]

idk what issue that is

[wise stratus]

something maybe who knows

[peak orchid]

Google Cloud Compute hasn't had an available A100 40GB in days, so I can't test my new embedding datasets or even my MJ one 😦

[elfin stump]

Did a transcript of the recent Stable Stage Q&A ever get posted? Some of us prefer text to audio...

#💬｜general-chat message

[lilac reef]

The issue where you can no longer do low light photography because SD always sticks a bright light source in the scene. Was never there until 2.x so see this https://github.com/Stability-AI/stablediffusion/issues/93

[midnight crane]

@wise stratus Ok, well I'm eager to see all kind of progress into AI so it was intriguing to me. Collaboration is almost always a good thing. But I don't know, Elon has become a bit strange lately (and made things way too political), I just hope that it doesn't impair his sense of technological progress that he's been a forefront of.

[balmy compass]

Hey quick question(s): if I run stable diffusion locally do I have unlimited generations? How much storage does it require (approx)? How long does it take for a generation (I have an RTX 3090)?
TIA 😃

[midnight crane]

@balmy compass Yes you have unlimited generations, you need a couple of GB for the models (usually between 2 to 7 GB for each model that you want to have) the github itself is fairly small. You are definitely ok with the graphic card.

A generation will take a couple of seconds in your case.

[balmy compass]

Nice, thank you. Is there a guide among all the available ones that you recommend on how to set up everything?

[midnight crane]

I recommend either Automatic1111's WebUI or InvokeAI, they are both very easy to install (usually a one-click affair). You would need a model for Automatic1111's WebUI already downloaded I think, InvokeAI downloads it for you but then you need the token from Huggingface in order for it to do so. All is explained in their respective githubs.

https://github.com/AUTOMATIC1111/stable-diffusion-webui

https://github.com/invoke-ai/InvokeAI

Stability AI huggingface for models etc: https://huggingface.co/stabilityai

[balmy compass]

Cool! Thank you 💛 I'll try as soon as I get back home on the main computer.

Meanwhile, can I try running it locally on my laptop? I'm running Ubuntu and have 8GB Ram, Nvidia 940mx (2GB Vram), i5 something - nevermind: If you have a GPU with at least 6GB of VRAM, you can run Stable Diffusion locally on your PC

[midnight crane]

You can try running it on CPU, but it will be very slow and I still think it requires a lot of RAM, not sure how much though. You can always try.

I don't know the instructions for running on the CPU though so you probably need to google it if it isn't explained in the repositories.

[raven edge]

I keep getting black images even after the xformers fix now that i did a git pull for Automatic1111

[balmy compass]

oh okay, I don't think it's worth it. I'll do everything tomorrow on my main machine. Thank you again!

[wise stratus]

How odd, sure devs will look into it for next version thanks for flagging

[chrome spade]

Anyone ever think about how many AI generations have been created so far with all the different tools? I can't even imagine the number

[wispy scarab]

Yeah, that's my problem as well, not enough VRAM on my laptop...

[uneven heath]

@wise stratus is there anyway I can reach out to stability for educational initiaitves?

an email or something

[raven edge]

I think something broke my img2img since I can't get any generation regardless of model

[modest owl]

PSA: https://www.kickstarter.com/projects/unstablediffusion/unstable-diffusion-unrestricted-ai-art-powered-by-the-crowd
is a scam. Don't give them any money.

[radiant bronze]

Why what happened

[modest owl]

they have produced 0 models still and have no demos

and at best they resort to mild erotica

no balls for other, more extreme forms of nsfw

[radiant bronze]

How is it a scam though

[modest owl]

all the money will most likely not go to the open source model production

[radiant bronze]

Why do you say that?

[modest owl]

because i know

i did my research

[halcyon granite]

anyone got some tips of making backgrounds more suitable for removal? For example I'm trying (blank background) and (empty background) but it doesn't seem so reliable

[lofty parcel]

Latest hack I've tried to is include "blue background" in the prompt, color the border of the image solid blue, and inpaint in the middle
Usually it's good enough, I find

The RunwayML inpainting model is probably your best bet

[halcyon granite]

do you notice any bleed of the blue onto the object?

[lofty parcel]

I think maybe a little, but I haven't found it to be too bad

I didn't spend too much time with it though, so I don't know about possible failure cases

There are also background removal AI models that can give really good results

You'll have to experiment but I think you can get something that looks good

[halcyon granite]

Cool, I'm gonna try stuff like blur, solid color, etc too

[meager nest]

Man, is anyone else extremely concerned about the ethical use of SD?

It seems like it has a lot of capability to cause a lot of harm

[tight quartz]

Many things in this world could cause harm, and there will always exist bad actors in every community. The best we can do is remain positive and stay responsible - make good choices and support others who do as well.

[meager nest]

Yeah, cuz like, even in my generations I do everything possible to avoid anything that's considered bad, but the fact that it is still possible makes me extremely uncomfortable

[halcyon granite]

anything capable with SD was already capable by state actors

[meager nest]

State actors?

[halcyon granite]

If you mean stuff like faking a president of something

[meager nest]

I'm more concerned about the nude images it can do

That in the wrong hands sounds extremely dangerous

[halcyon granite]

anyone can do that in photoshop, it's irrelevant

[tight quartz]

There's measures put in place to try to minimize the impact of that - and it's not isolated to SD - like Owen said; anyone with photoshop can do some poor editing.

[halcyon granite]

there's already fake celeb and stuff

[meager nest]

Thankfully, even on unstable diffusion they have rules to prevent communities from becoming a cesspool

[uneven wasp]

Hasn’t that already been legislated a bit? I think I remember reading an article about deepfakes for certain purposes being banned.

I’m not sure sd allows much new capability in that potential area of harm that existing deepfake tech was incapable of

[halcyon granite]

In a weird way maybe it's a positive because now celebs can just claim images are fake

[lofty parcel]

But also a negative because politicians can do the same

[tight quartz]

Friends, if you ever find yourself uncomfortable or uneasy about a community or a new technology - ask yourself where that fear came from. Perhaps its a narrative you read - or a famous idol mentioned their concern and you align with it - or even just the misinformation that exists. Do your research, ask the questions, speak to one another. That's the best we can do shine light on something (or anything) than can also be used in ill intent.

[peak orchid]

What research though? And Stable Diffusion is still so new that most large groups haven't created huge finetunes yet. They did help train Waifu-Diffusion though, which was a major finetune

[tight quartz]

Personal research is anything you do to further your own understanding of a topic. Even this conversation right here, is an example of :looking into something".

[halcyon granite]

To me the only different from traditional art is just the volume

[tight quartz]

@halcyon granite absolutely. Photoshop allowed faster creation of what we had to do in MS paint.

[uneven wasp]

Cherry-picking the best results from an incredible volume is a genuine quality advantage

[modest owl]

They did help train Waifu-Diffusion though, which was a major finetune
sure a buncha weebs will be happy, and the rest of the people will suffer

[halcyon granite]

people shouldnt be trusting just images alone anyway, look at ghosts and UFO communities lol. I actually warned those before the AI craze started lol

[modest owl]

mmm digital ghosts 👻

[peak orchid]

I don't really care for Waifu-Diffusion either, but my point was that they aren't completely inexperienced in terms of finetuning

[tight quartz]

new band name

[modest owl]

if there's no output - what's the point?

[peak orchid]

Its freely available: https://huggingface.co/hakurei/waifu-diffusion

[modest owl]

context - warped ☑️

i'm talking unstable diffusion output duh

[uneven wasp]

Not following the project that closely: are they promising to release weights?

[modest owl]

kinda, yes

[tight quartz]

Like I said - there will always exist operators that seek to push boundaries in different directions, sometimes unethical.

[peak orchid]

I've personally spoken with some of the UD staff, and haven't noticed anything sketchy about them. My main concern is whether or not they are budgeting for screw-ups. Stability AI's 2.0 mistake cost them more money than most Americans make in a year

[modest owl]

actually releasing a model that will push the boundaries and being a damn knowledge void is two different things

[peak orchid]

Not everyone can afford such mistakes

[halcyon granite]

mistake?

[modest owl]

most Americans make in a year
the rest of the world right now - 🙄

[peak orchid]

They set the punsafe (NSFW filter) values so high that it removed most of the humans from the dataset (they set it to 0.1, when it should have been set to at least 0.99). SD 2.1 was attempt to fix that by changing the values

Pretty costly mistake, but stuff happens sometimes

[halcyon granite]

I see, is it a pain because it takes so long to generate the dataset?

[modest owl]

and instead of rebelling you deal with it, what a foolproof strategy

[halcyon granite]

I mean 1.4, 1.5 still works great

[tight quartz]

i like this show

[modest owl]

don't like this tell tho

[peak orchid]

Yeah, training takes time and resources. It costs money to use their thousand GPU cluster of $15,000 GPUs, and training can take weeks or months

[lofty parcel]

Unironically better than reality TV because the subject matter is interesting

[halcyon granite]

ah they should've asked me, I would lend them my trusty gtx 1080

[tight quartz]

This is exactly what this channel is for tho. We discuss, we listen, we grow, we share - all perspectives, all concerns. We can't just shut our eyes to one idea, and not another.

[modest owl]

(still a stolen channel tho)

[tight quartz]

(a resolved issue)

[modest owl]

(nope, internet doesn't forget)

[tight quartz]

(neither do 📌 pinned messages)

[modest owl]

(and since discord is not indexable - yes, yes it does)

[tight quartz]

You play chess?

[modest owl]

https://www.youtube.com/watch?v=OL73o48db7s

[tight quartz]

[modest owl]

🧠

[tight quartz]

I wonder how that fancy chatGPT does with chess

[modest owl]

it does great with hanging mid-typing

[peak orchid]

How many Midjourney image-text pairs do you guys figure I need for a simple finetune of a 2.x 768 model? I've been collecting high quality ones as the OpenJourney team hasn't responded to my request for the dataset

[modest owl]

and lying

midjourney should be cancelled for discord gatekeeping alone

[lofty parcel]

I saw a YouTube video where someone tried and after 10 moves it started playing invalid moves, but still interesting

[tight quartz]

@peak orchid sweet spot on 200 (unofficial answer)

@lofty parcel opening move: queen takes both kings: field goal!

[peak orchid]

ooh, I've got over a thousand right now

[lofty parcel]

Victory Royale

[halcyon granite]

pawn nukes queen checkmate

[tight quartz]

@peak orchid there was some literature of over tuning

[peak orchid]

I should automate my scraper code though

[fervent thunder]

I wonder how much it cost them to create 2.0, not checked on it at all so it might be out there somewhere :P

[peak orchid]

Do you know where?

[tight quartz]

@peak orchid we have access to the same search engines. I can't recall myself, sorry.

[peak orchid]

lots of content being made on Midjourney tonight according to my scraper code

[tight quartz]

#weekend

[cyan temple]

seriously, dafuq happen with SD?

i use same prompt in SD 1 and SD 2 with MJ, SD look like c**p now compare to MJ, make me sad

[halcyon granite]

prompts dont transfer from SD to MJ

[cyan temple]

its like 10 step backwards by SD

[cyan temple]

yeah, apparently so, really saddening

[modest owl]

c**p
oh boy what a time we live in when people need to censor a damn excrement

[cyan temple]

and it just 2 months i never use SD, last time i used was 1.4 and its still okay

[peak orchid]

They actually do transfer with the right model: https://huggingface.co/prompthero/openjourney

Though its 1.5 based

[modest owl]

why do you need an imitation? You got gatekeeped and have no money?

[peak orchid]

that's no longer a thing

[modest owl]

not sure those PoS would make a free tier without restrictions

[halcyon granite]

I keep getting billed for MJ and forget to unsubribe, havent used in months

[modest owl]

check wallet for burn holes

[cyan temple]

lol

[tight quartz]

In 1.3, 1.4 and 1.5 - you could just type "moon" and it would give a wildly beautiful piece of art of the lunar surface in a starscape.
In 2.0+ if you type in "moon", you just get a moon. It's because it's literal , and it wants more description and tokenization. It's not less, it's more critical.

No other versions or datasets made a jump like this, so people seem to feel like there's something wrong - where as, you just have to re-learn to prompt better, and understand the model.

[cyan temple]

yeah, $30/month abit much whilst i havent generated any $$ from it. Just been experimenting for personal IP. used to using my 3060 laptop, but now just using the hugging face freebies. Abit sad when I see MJ result compare to SD. 2 months ago SD used to be so powerfull compared to MJ thou

[modest owl]

just don't touch it or it's imitations at all

it's that simple

it's too overhyped as dall-e anyway

[halcyon granite]

problem with MJ is I don't feel like i own it. At any time it can be taken away and then it would be impossible to recreate my series of images

with SD I know even if the government tries to ban I'll still have it lol

[modest owl]

gee i wonder why

[fervent thunder]

I think first impression is the most important thing about art. I believe that's why MJ is popular. It's not about how complex it is, but if you like what it just created. :3

[lofty parcel]

Unless they seize your devices

[halcyon granite]

same reason I will never buy a subscription to Adobe Photoshop, if my work cancel it I'll change to a different software

[modest owl]

besides mj's model is specific enough as an artist that shids into his palette and paints a final layer with that

i.e. not general like sd

[peak orchid]

I just realized that OpenJourney is likely a Dreambooth model. So my proper MJ finetune plans should be a lot better

[halcyon granite]

well they wouldn't have reason for that, it would be an issue with AI as a whole. I'm just using for game dev stuff so I want my style to stay consistent

[lofty parcel]

Oh I was thinking if they viewed it as a cp generator

[halcyon granite]

i posted a sample in general images

[cyan temple]

i tried using this negative prompt "disfigured, kitsch, ugly, oversaturated, greain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, , mangled, old, surreal >> and result is okay. Sorry the great SD developers for my previous negative remarks. You all have done Marvelous Job for humanity

[lofty parcel]

Side-tangent: Has anyone tried swapping their negative and positive prompts? I'm curious about what eldritch horrors this could result in

[halcyon granite]

I'll try it now with my current prompt

[modest owl]

nah, try doing datamoshing with sd instead

[tight quartz]

Negative prompts tokenize as a matrix against positive prompts. As in first, read prompt A+B+C+D, then remove "z" from each of those tokens.

[modest owl]

question

[tight quartz]

answer

[halcyon granite]

I'm just getting blank colours with some noise

[modest owl]

haha, very funny

real grouping will come out when?

[fervent thunder]

I stopped using ugly as a negative prompt because it seemed to remove too much in-between type of faces on characters etc. Then I removed another, then another, now I don't use any negative words if I don't spot a theme I don't like in the generated images I get. It's actually just one negative prompt I have had to reuse a lot now a day and I love it! :D

[modest owl]

wanna group my prompt parts, no matter if it's blocks joined by && or AND

[halcyon granite]

I've found for cyberpunk using "blue purple hue" negative works amazing. It basically just uses normal colours

[tight quartz]

@modest owl there are syntactical difference in hyphenation, punctuation and compounding tho.

[uneven wasp]

Personally I find short prompt describing and negative with the exact opposite gives me pretty good results when I'm looking for something that can be described that succinctly

[modest owl]

try pulling this out in vanilla sd:

(red,cat)(blue,dog)

or

(red,cat) AND (blue,dog)

similarly try it for OR or ~

seriously, forgetting such basic syntax is unacceptable

[bold star]

Is this 1.5 or 1.6 dreamboat is on?

[modest owl]

boats don't dream, they swim, or rather, float

[bold star]

Yeah that too

[modest owl]

and yes, distilled SD when?

[lilac reef]

I wish we could run more than one script at a time in automatic1111

[vapid tangle]

2.1 is a huge step up. Awesome!

[modest owl]

i wish there would be a damn QUEUE in a1111

represented in ui as well

[lilac reef]

I really wish one of those devs would take a look at automatic1111 and fix the Hypernetwork training because it should not take that much ram (more than any other training) so I suspect it has no memory optimizations to it. Beyond my abilities to fix.

[calm herald]

yo. did the nsfw filter really dip out and then turn back on?? how do i fix this...jesus

[static lynx]

i think MJ is more understanding regarding what we want in the prompt and output compare to SD 2.1 , but it open source and free, so give and take

[lilac reef]

yes

I am starting to use less and less negative to now just one, or two tokens in negative. If I see something I hit it with my big ass wall of negative crushing power. a lot of times it destroys it and something completely different is made.

[tight quartz]

@river musk In some cases, it can skew results.
Example: garden = positive
But negative prompts like flowers, grass, trees, stones, sand, foliage, sky etc....
You'll just get walls and water fountains - since you disallowed everything else.

[untold ocean]

a lot of times negative prompts have unintended side effects

what you think is the negative prompt might not be how the model learned it

also prompt weighting in the text embedding space versus cross-attention is very different

[forest vale]

neg prompts being used like they are in sd2 is a crutch, nothing more tbh

[compact perch]

I'm running an rtx 2080, and I would like to increase the width/height of my generated images, but everytime I go passed 900x900, I run out of GPU memory. Is there anyway to fix this without shelling out for a second or new GPU? Or is it worth trying out basujindal fork? Don't really understand the trade off compared to regular diffusion.

[mellow knot]

seems intuitive to me

this is exactly what you get when you take the garden out of a garden

[daring niche]

guys do any of you have any suggestions of a good text-to-speech model?

[rich plover]

all my humans coming out like midgets 😔

[cold siren]

isn't that funny how we try to use negative prompts having no idea what they mean for SD, to fix positive prompts that do not mean for SD what we think they mean 🙄
i actually tried rendering some of these "standard negative prompts" as positive, and no way the results correspond to their descriptions

[bleak matrix]

Good morning, folks. How are we all this morning?

[fervent thunder]

yeah, I stopped bothering with negatives as much when I noticed how they didn't work. Felt more like a bias than a real help. Now I only add negatives after I notice something I dislike from the images being created. Like that stupid covid mask being placed on superheroes! >:I grumbles

[rich plover]

if i put bad art in the negative the computer gives me good art, its just that simple

[fervent thunder]

a negative prompt has to help more than not and it should be noticeable, even blurry seem to do nothing consistent for me, one I thought I'd never remove before :O

[cold siren]

i remember somebody putting some random gibberish in the negative, and getting better results than with a "sensible" negative prompt

[cold siren]

heh i wish watermark worked in the negative 😕

[fervent thunder]

Anyone using invoke in here that has been updated today?

The models arent being found and the location it says does not exist actually does.

[bleak matrix]

I typically rely on good prompting skills first. Build a base prompt that's solid and iterate on it. It's much easier because then you can understand flaws and interactions between words

[fervent thunder]

Nerdy Rodent on a recent youtube video had random things like wet spaghetti in his lol

[fervent thunder]

it does, it removes water puddles on roads! ;P
I've started to add detailed face to the negatives as well as so far it feels it helps with getting less google street view faces, I also have that as a negative prompt :P

[cold siren]

lmao 🤣

[bleak matrix]

I haven't personally updated but I'd check the server

If you're not in it, I can look and see if anyone has reported something for you

[cobalt cliff]

https://twitter.com/emostaque/status/1601919744944357376

will the god emperor grant us this gift?

[deep bridge]

yeah don't know about that elex

im not one to start drama usually. however. *pauses for dramatic effect*, the truck posting has gotten out of hand in #1019350310650646589 to the point where it's so saturated the mods should seriously consider adding a truck specific channel... @whole fjord

[worthy oriole]

that is probably tied to this thing, at least I hope it is
https://twitter.com/scale_AI/status/1582890586834489344?s=20&t=PFb6obdMj9l7yW2cN5zT4A

[cold siren]

only if it will be uncensored as SD 1.x 😐

[broken gull]

that's not basic syntax, this is still fearly new imo

[hasty badge]

Surely it'll get there eventually, but eventually might be a long time. Exciting times, I've never been so interested in new tech as the last year, first copilot, then dalle/sd/mj, now this, let AIs do all the busywork :). One problem with something like chatgpt is that it'll be a challenge to keep it up-to-date, so it needs a big org constantly feeding the internet to it.

[broken gull]

he already mentioned they're working on a language model. and I already spoke to the devs and they said they're working on it.

[cold siren]

besides... what if somebody actually wants to get a picture of a mutated, ugly, deformed, malformed, disgusting, surreal something with extra limbs, extra fingers and etc 🤔

like, can you describe a centaur with these

without using the word "centaur"

or any other weird mythical creature

[fervent thunder]

yes, those are not good negative prompts in my eyes, not ones you always want anyway

[cold siren]

or any hindu deity with ridiculous number of heads & limbs

[fervent thunder]

I get much better results having the negative prompt area as bare as possible from the start, and then you add stuff when you notice something in your images you don't want, something that often appear. Such as a color, type of clothes/hat, etc. If I'm gonna be brutally honest, then I believe stuff like extra fingers does absolutely nothing good :P

[cold siren]

though this might actually need a better language model 🤔 which might be a good reason for stability to do their own research there

[fervent thunder]

I believe everything will be changed in time, even the ai itself will probably need a "better" system in time, but we'll see, it's scary how fast this has advanced, well advanced in some ways, not so much in other areas

[uneven wasp]

They need better prompts for their clip model.

the caption/desc in laion arent that good from what I've seen

[fervent thunder]

I think they need tutorials, examples, lots of documentation for the normal idiots like me so I—I mean we can get good results quick :3

[cold siren]

from what i've seen, that's an understatement 🤭

[uneven wasp]

The problem with openclip probably might not even be a neutered dataset, it's probably just trash text pairing

Can't train a contrastive model on shitty pairs and expect it be good at prompting

[fervent thunder]

I'm gonna be brave enough to say that it feels like the image quality and text descriptors will move to a sort of crypto based thing everyone gets to be involved in, so even if there are bad actors destroying the training, it wouldn't hurt the overall quality

[uneven wasp]

I mean its not a hyper fault sensitive situation like a public ledger, you can have bad data in your dataset and still come out with a good model, it just needs to be a small enough portion

Only real application of blockchain/crypto I could possibly see here would be coordinating incentives for people to train each others dreambooth/TI/etc models

If you wanted that to be totally decentralized system

[cold siren]

at least i hope we get rid of bullshit tags like "trending on <a specific website>" at some point... 😐

like really what's next, "liked by <a specific person> on facebook"?

[fervent thunder]

I find it cute, it might the persons very first AI image :3

[lofty parcel]

Retweeted by Emad Mostaque

[uneven wasp]

viewed by <person #21515112> at 11:51 pm at 51.507351 -0.127758

[cold siren]

yep ☝️ and then everybody will copypaste this into their prompts, like "by greg rutkowski"

and the <person #21515112> then will sue them all to claim copyright or something

[uneven wasp]

lol copyright over ai is such a mess

It's most likely transformative fair use, except that the actual legal status of its copyright

It's most likely transformative fair use, except that the actual legal status of synth image copyright will end up being solidified by the law that will inevitably pass when somebody inevitably does something incredibly stupid

[rich plover]

any prompt you guys do to get a flat white or black background

?

[cold siren]

no, but i tried to prompt for a "blue grass, green sky" once

[lofty parcel]

Have you tried inpainting? You can try to force a white background by using a solid white border and inpainting the content you want

[sharp vortex]

Does stable diffusion scrap imagies from internet to study their style?

[lofty parcel]

Might not be perfect, but it's better than not being able to directly control the color

[lofty parcel]

The images were already collected and exist in the LAION dataset (which came from scraping), and Stable Diffusion is trained on them

[primal marten]

Hi! Question, does output resolution affect the created image? As in, does it change what SD will produce? Or is it always the exact same image (assuming the rest of the parameters are the same) but with a different resolution and that's it?

[vital osprey]

yes

thats why sometime you get text on your gen

[sharp vortex]

Gf mad at me cuz I love stable diffusion. she says it steals artist styles. says she who use uTorrent all the time🙃🙃

[cold siren]

by the way! have anyone seen the "seed rescaling" in auto1111 actually working?

[storm shard]

Is child a forbidden word? #1047760004280766474 message What is wrong with this prompt?

[magic citrus]

Hey guys,
I just published a new blog post on the LAION-Aesthetics image dataset and thought some of you might be interested in checking it out. In the post, I delve into the different parts of the dataset and discuss some practical use cases for it.
Give the post a read here: https://followfoxai.substack.com/p/exploring-the-laion-aesthetics-image?sd=pf
Let me know what you think!

[cyan scaffold]

Hey, I have some questions and needing help with stable diffusion and Anything v3 related for Img2img to get a certain outcome of art, which channel is the best place I can go to to get help? thanks

[broken gull]

In Belgium from what I heard from a lawyer is that you can't copy right the imgs but you can copy right the prompts.

[robust dove]

Question for you all: So lots of the time when I generate a full body shot of a person, the face is all messed up. What's the easiest way to get it to regenerate the face? I've been saving it, making a copy, cropping that copy, inpainting, then recombining in Photoshop but there has to be an easier way.

[hollow raft]

Trying to combat misinformation with science --> I put together this explainer video to take some of the 'scary magic' out of the technology. Advice and or feedback / corrections / all appreciated. I'll put it on twitter tomorrow, but was hoping for a sanity check first https://youtu.be/THseWPhwceU

[midnight crane]

@hollow raft I think it was a good video, was hoping that you would paint something with that water lol

[frosty marsh]

Hi, I'm trying to install pytorch/torch/torchvision cuz stable diffusion seems to require it..
I installed anaconda, then Cuda but the command line to install pytorch is still not working..

[hollow raft]

Hadn't thought of painting with it hmmm, creating a new image from the noise... that would be a good demonstration!

[deft ingot]

is creating hypernetworks for 2.1 a thing?

[sinful talon]

hey! what is the difference between ema pruned and non ema pruned for SD 2.1 ?

[white mountain]

this seems quite plausible to me
CLIP trained with a high quality dataset + SD 1.X trained with a low quality dataset = good results
OpenCLIP trained with a low quality dataset + SD 2.X trained with a low quality dataset = bad results

low quality image-text pairs, that is

[main kestrel]

Hey!

[waxen sparrow]

hello, which channel would be most appropriate for questions about checkpoint merges?

[fervent thunder]

are there any ui that have added CLIP guidance yet?

[chrome spindle]

is it possible to combine several images with stable diffusion?

[white mountain]

not aware of it, but there is code here https://github.com/huggingface/diffusers/tree/main/examples/community

[fervent thunder]

Thanks Cheeze. But yeah for me that's useless. I'm lucky I know how to turn this thing on

[reef epoch]

is Stability considering taking the best trained styles / concepts from models or embeddings created by the community and finetuning them into upcoming models?

[river fox]

I rememabersome one linked me to a sire that generates texxt prompts from images

any one know a good site for this?

[white mountain]

clip interrogator is what you're thinking of

[iron pendant]

if you have a 4+ GB gpu you can run this locally en masse; https://github.com/victorchall/EveryDream/blob/main/doc/AUTO_CAPTION.md

[bitter apex]

Funny how yall steal peoples discord servers

[fervent thunder]

lol

[fervent thunder]

I'm so glad you came in here about 5 months late for this. And I know youre so concerned about this server because I can see that green leaf next to your name indicating youve never been here before.

so I'd say it's funny you have nothing better to do. Go get a hobby kid

[unreal kettle]

why do my dreamstudio images turn out so grainy?

[steep whale]

How do i use the Stable diffusion Inpaint model? do i just have to import it regulary?

[obsidian timber]

Is anyone else getting 'the application did not respond' errors?

[halcyon granite]

with img2img is it better to use low res images? I want to make variations of a previous generation. I have an original plus an upscaled version I could use

[halcyon granite]

is there a GUI version?

[iron pendant]

there's a notebook

actually two, one is just for autocaption and nawnie just submitted another one to run some of the other stuff as well, but autocaption is the only thing that takes GPU power

mostly just need to copy paste the commands if you use command line anyway, put images in the input and run it

there's a batch file that sets up the environment

[jaunty dagger]

hey, im new to stable diffusion and i have a question

how do u know what to write in the prompts? im seeing so much detailes and stuff idk where to even begin

when im trying to write something it looks shit XD

[compact perch]

Anyone knows if it's possible with the knollingcase module to make the knollingcase shattered/broken?

[storm shard]

Try lexica.art for ideas. Or copy the prompts in the dreambot chnnels and play with them. See what they do when you alter them.

[mellow knot]

utorrent
not the adware torrent client

atleast get her to use qbittorrent or something

[storm shard]

This is also very helpful to get more insight. https://openart.ai/promptbook

[rare fjord]

Hi, how can we import model to webui a1111 if there is no safetensors ot ckpt file?
I was trying to put bin file into embeddings and vae directories but with no success.
https://huggingface.co/nousr/robo-diffusion-2-base

[pale igloo]

I have noticed that the more the model is trained, the more narrowminded it gets. I was trying to train pictures of a girl in a boxing ring, but as the face gets more and more close to the input folder the more close-up pictures i get of just the person. Any tips on how to avoid this?

[digital ginkgo]

anyone have that picture search website that lets you search the datasets for stable diffusion?
i'm looking for the dataset used for SD v1.4 and cant find it

[white mountain]

they seem to to have deleted it (accidentally?) but it's still there in the commit history

[rare fjord]

I have checked it from previous commit, but It failed to load, so I guesses that it didnt work. Maybe I did something wrong.

[golden nebula]

is there a prompt list like of all available tags

[jaunty dagger]

ill try it thanks @storm shard

[reef epoch]

IM PROTESTING AGAIN

🔥 🔫 🔪 🍴 🪠 🪨

[midnight crane]

There is some confusion but people generally say that you should use the non-ema pruned version for training and ema pruned for generating images. EMA stands for Exponential Moving Average and pays more attention to recent changes which should be beneficial to training. But that's just my limited understanding of it.

Here's some tech about EMA: https://timm.fast.ai/training_modelEMA

Pruned means that they take away weights, so EMA pruned means that they have taken away the EMA-weights. Non-EMA pruned means that EMA weights are still there, at least from what I can understand of it.

I could be wrong though, so if someone else knows then please correct me.

[iron pendant]

yes generally you do inference on ema weights, and nonema weights are trained on

the base compvis/sai code should automatically pick the right version either way if you load it in for either training or inference

and fall back to the "other" one if the "right" one isn't found

can't speak to all the codebases/dreambooth tuners, etc out there, but the xavier DB forks and the pytorch lightning based code does this (robin/compvis code)

[midnight crane]

Ok, in Dreambooth you need to specify the version, it won't do any of that smart stuff. Seems I was wrong about the vocabulary, so non-EMA pruned means that it doesn't have EMA weights. Not that the EMA weights have not been pruned. It's confusing because Dreambooth specifically have a option to extract the EMA weights from the model as if that would be beneficial.

[fervent thunder]

has anyone named their png files for ben's fast colab? I just read you can do it, Im curious if anybody has compared it to just naming the pngs one uniform name 🤔

[midnight crane]

I would love to be able to train multiple concepts in one go by just naming the images with a description of the image. That would be excellent.

This one concept per model isn't enough I think.

[fervent thunder]

maybe something like that will happen when they release their own training stuffs

sounded like it was coming soon like most things lol

[midnight crane]

Yeah, I hope so. It's always just around the corner lol

Back in the day they had a lot of curved paths to fool the horses so they wouldn't see the far road ahead.

[fervent thunder]

they did release the CLIP guidance finally I guess (i was waiting for it at least) just nobody has implemented from what i can see with ui 😦

but maybe the training stuff will be next

[midnight crane]

Yeah, only online services have CLIP guidance as it is now. DreamStudio and NightCafe uses it. NightCafe gives out 5 free credits each night so I have collected quite a few 🙂

[fervent thunder]

Awesome. I hope auto adds it soon 🙏 Seems like a big thing to me, I dont know why nobody seems to care or in a rush to get it added 😭

I wonder if it would be an extension 🤔

seems like it only works with one diffuser on the official site

[wild hull]

Hello guys, there is someone in this world that have good results by training SD2.1 with his face ?!

[midnight crane]

I haven't been able to get it working on Dreambooth locally. I have a 12GB card and apparently people have trouble running it on that amount of memory.

[fervent thunder]

not my face but the results Ive gotten from ben's fast colab havent been nearly as good as what I got from 1.5

it's why I asked if anyone has taken to adding descriptions to the pngs as I didnt do that (and didnt need to in 1.5)

[wild hull]

with ben's one, no results ..

I'm really desperate, it's been a week of sleepless nights testing everything, everything is unstable. When you think it works, 4 minutes later it doesn't work anymore

Every time I use tools, there are unmanaged errors

Everything goes very fast, and what worked yesterday doesn't work today, due to an update

I'm really on the verge of ragequit and waiting for this to stabilize

[peak orchid]

Has anyone tried the diffusers embedding training? A111 is giving me grief when trying to train with it on SD 2.1

[midnight crane]

60C is like normal working temperature. The optimal range is between 70 and 85 degrees Celsius. It will go as high as 98 degrees before throttling. So don't worry.

Sorry 93 degrees before throttling.

[sinful talon]

thanks !

[midnight crane]

@sinful talon You're welcome, I'm trying to understand this EMA thing myself. In earlier Stable Diffusion models the smaller sized model had only EMA-weights, while the bigger had EMA-weights + normal weights. The naming itself is confusing as with previous versions only EMA-weights were enough for inference.

[frank moon]

What is the seed in SD? Not the number of the image but, the one in the set up example (seed: -1)

[midnight crane]

@frank moon I think the seed determines the psudorandom noise. The model is deterministic so you can recreate the same image by recreating the same noise, which is possible if you use the same seed. -1 just means that the seed is randomized.

[peak orchid]

@wise stratus Is the new Distilled version of the model sacrificing any quality for speed? Like was the main model trimmed down / converted to half precision to make it faster?

[brave matrix]

hello guys! What is the current state of AI generated animation? I've seen some demos but are any of it public use yet?

[young ivy]

https://metaphysic.ai/research-proposes-moral-sanitization-for-text-to-image-systems-such-as-stable-diffusion/ I strongly hope this never becomes a reality, that someones morality be imposed collectively on the entire world in an open source program. Though the one we have now will never really have it anyway....hopefully this is the equivilant of a "video games make people into murderers, we must ban immoral games"

[eternal jacinth]

Can you have multiple weights loaded in your models?

Oh, I think I figured it out. Just needed to change the checkpoint.

[cobalt cliff]

You have been visited by Prompta, the friendly muse of prompts. Many epic and amazing prompts will come to you today, but ONLY if you reply to this post with the prompt you are using right now. If you don't, Prompta will ruin your images with bad hands and weird eyes for the next 7 days.

[peak orchid]

Prominent anti-AI art activist Karla Ortiz is now going after Stable Diffusion projects on Kickstarter: https://twitter.com/kortizart/status/1601681381385699329 Currently she's trying to get Unstable Diffusion mass reported, but she wants a total ban on Stable Diffusion related Kickstarters.

[timid gate]

Prominent anti-AI art activist Karla Ortiz is now sucking my cock on Kickstarter: https://crymaldetc.com she's trying to get mass reported, but she wants a total ban on anything that makes her angry.

[balmy field]

Anybody got a backup of kurzgesagtish? https://huggingface.co/Imakemodels/Kurzgesagtish

[timid gate]

wdym

[weary lance]

Hey does anyone know if it's possible to generate minimal, vectorial logos with SD? I tried with some prompts but had no luck

[balmy field]

@timid gate https://www.reddit.com/r/StableDiffusion/comments/zdp84u/dreambooth_model_release_say_hello_to/ some model that got up 6 days ago and then deleted

[timid gate]

you want a model more than a prompt

[timid gate]

no idea

[balmy field]

@weary lance https://www.reddit.com/r/StableDiffusion/comments/xdz8o3/can_stable_diffusion_be_used_to_create/ doesn't look too good

[hushed edge]

Guys, did stable diffusion ever allow ascii art? what happened to it

[torn quartz]

do you know what is clip guidance? is like going back to clip-guided diffusion?

btw, anyone have saved the icons of the discord server of stable diffusion? I would love to see the evolution of it

[peak orchid]

Looks like OpenAI might break ChatGPT in an effort to "watermark" the text outputs: https://techcrunch.com/2022/12/10/openais-attempts-to-watermark-ai-text-hit-limits/

[peak orchid]

Can't wait for the Kurzgesagt video about how AI art is evil and should be banned everywhere lol (they are apparently really protective of their style)

[frosty marsh]

so I gave up on automatic 1111 cuz it was just impossible to install, I seem to have managed to install invoke Ai, at leats that's what the command thing said but it also said that i needed to run invokeai.bat and it said (in this folder) question being which folder?

now that I think about it, I could ask chat GPT

[grave cipher]

https://civitai.com/models/1093/kurzgesagt ?

[bleak matrix]

Good evening, everyone

[frosty marsh]

any idea why there's no Invoke.bat command in my Invoke AI folder after I've run the Install.bat command?

[main sparrow]

I had to get it from the src zip download. the git pull didn't grab the invoke.bat.

[azure ice]

Posted a new PR for better color correction in A1111: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/5644

[white mountain]

I just realized

that unlike finetunes, textual inversion embeddings ARE compatible with the depth and inpainting models

[main sparrow]

I still want to make a mangled AI fingers embedding and see what happens when I use it as a negative prompt.

[warm panther]

what's the general opinion of 2.0/2.1 vs 1.5?

not as good?

[slim bolt]

Where Can I search for 1.5 embeddings? is there a site or something?

[torpid zephyr]

Still getting a feel for 2.0 + 2.1. It's 'different', but I think I'm more used to 1.5 at this point.

[desert whale]

Ok so if this is really happening https://twitter.com/EMostaque/status/1602001804677582850 that's a gamechanger, that's 30fps, that's video, and games in potentially real time coming far far closer to reality.

[torpid zephyr]

AI-produced animation would work really well in a horror game. A monster whose features change in real-time as you're interacting with him.

[main sparrow]

Just got to get everyone on a100's now.

lol

[raven edge]

I keept getting a black image with 2.1 I don't know what Im doing wrong I changed the userweb UI.bat

[lofty parcel]

As someone working on generating entire game worlds with SD, this is extremely good news since the compute requirements were looking iffy before

[wise stratus]

Be above that on H100s.

Just give it a year to mature and it will be quite surprising the speeds, where will we be August 23rd 2023

[desert whale]

I want in on this

I gamedev this seems next gen

lol

[wise stratus]

I’d look to new year for timelines, will be a quiet Christmas as giving team a week off given how crazy this year has been

So no big releases before then to make their lives crazy

[desert whale]

Hope you all have a wonderful holiday

Also not to mention that you could in theory make anime, esp with better attention

[wise stratus]

Something… separate on anime coherence

[green bear]

https://www.youtube.com/watch?v=RO7TYfMyPPw would be a good example probably

[desert whale]

Well I'm definitely looking forward to what you all have been cooking after the break.

[radiant bronze]

If there is an anime model with a depth finetune I'm pretty confident 3d animators will be able to make coherent anime using available tools, especially since you can pipe in your own depth maps on the input side of things. Please don't forget depth2img

[halcyon granite]

quick question but what kinda res is a 3090 capable of in landscape? I'm reading online articles and they say only 1024x1024 (which appears to be old since my 1080 handles it)

[halcyon granite]

can't you just use outpainting to scale up your team?

[bleak matrix]

They deserve it

[halcyon oasis]

Does xformers mess up the new accelerate program? Trying to figure out whether to install xformers

[brave matrix]

Anyone know how to make the "extra" upscalled images in AUTO1111 not save as BOTH jpg and png?

[fervent thunder]

do yall fuck around with chat ai here or is it just novel AI shit

[green bear]

wdym by chat ai

chatgpt?

[fervent thunder]

specifically gpt3

yeah

open source stuff i can feed data in

[halcyon oasis]

For sure

[fervent thunder]

is depth+img to img a thing for 2.0?

[green bear]

we do all types of fuckery

(mostly image generation AI though)

[fervent thunder]

is it just openAI or what

I want to make my own creature

I found something called EleutherAI but it's dead project and I want to find something trying to keep up

[halcyon oasis]

No it's called ChatGPT

[fervent thunder]

yeah can it be run locally

because last i checked these bastards wanted my number for some reason

found something nevermind

[lofty parcel]

I heard talk about this needing to be added in the AUTOMATIC1111 repo so maybe?

[green bear]

its a website

also afaik its not OSS

[frosty marsh]

src zip download?
what's that?

[main sparrow]

looks like they've updated since I've last installed. I used this a few days ago: https://invoke-ai.github.io/InvokeAI/installation/INSTALL_SOURCE/

but it looks like they just released 2.2.4 here: https://github.com/invoke-ai/InvokeAI/releases

[haughty mountain]

question, when using prompts if i were to use a trained model for something, for example 'samwall'

if i wanted sam wall as a pirate would i use the prompt

'samwall, pirate'

or

'samwall as a pirate'

[muted mesa]

hey what's up guys. am I using the correct site - https://stablediffusionweb.com/#demo. Also does anyone know a way to save the images in a one-by-one fashion similar to the option given in DALL-E

[uneven heath]

Is it videos?

[alpine sable]

why do i keep seeing giant tits on twitter with stable diffusion tagged.

[fervent thunder]

has anyone added the CLIP guidance to their ui? it seems like a game changer to me using it on the stability site but nobody seems to care its been released public now

[peak fern]

What does it do

[fervent thunder]

it makes the picture more accurate to your prompt is how I see it

like I said, its a game changer imho, but nobody even talks about it here. if the point of this whole enterprise is to make more accurate ai, id think more than a dummy like me would be excited its finally released but thats not the case lol

[peak fern]

i'm a total noob so have no clue about how i'd even add that but yeah

[alpine sable]

how to use stable difsuion

is midjourney better

[peak fern]

one is free and open source the other one isn't

[mellow knot]

whats this tweet

[finite hawk]

Stable Diffusion getting a shoutout on a Taco Bell commercial!

https://youtu.be/G84YqORpjHk

From the description:

fun fact: everything in this video had to be copyright free, I generated the snow texture with Stable Diffusion AI, and made this goddamn beautiful seamless skybox with AI too (hours of coaxing it and photoshop) https://i.imgur.com/oWHQbOO.png, if you're a mapper and want the texture files email/message me on snap, that'd be neat

[muted mesa]

DEFORUM OR DISCO DIFFUSION?

[rotund depot]

E

[reef epoch]

im very confused... so the distilled according to based on the interview is supposed to be next week? but here in the chat said no major announcements until after christler chrimsas Christman Charismas

[uneven wasp]

Depending on what was said later I'm going to default to after christmas

after next week is technically after christmas too

[safe sapphire]

Disco diffusion was before SD, it might or might not be improved by now. Deforum uses SD.

[glossy grove]

https://meet.google.com/exk-bdkv-fzt

[worthy oriole]

I feel that most People still prefer 1.5...But that might not change until there is some popular models trained on top of 2.x

[drifting fossil]

hey'

[cobalt cliff]

what is the difference between an embedding and an entirely different model?

if i make an embedding. or take 1.5 as a basis to make my own model

what does that change

[primal marten]

Hi! Whenever I try to use a model that isn't 2.1 or 1.4 I get these errors over and over in a few dozen lines.

RuntimeError: Error(s) in loading state_dict for LatentDiffusion: size mismatch for model.diffusion_model.input_blocks.1.1.proj_in.weight: copying a param with shape torch.Size([320, 320, 1, 1]) from checkpoint, the shape in current model is torch.Size([320, 320]). size mismatch for model.diffusion_model.input_blocks.1.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]).

Does anyone know what could be the issue? Thanks!

[next spade]

How do we check how many credits we have?

[gusty oasis]

Credits?? What is this midjourney

[next spade]

Do we have unlimited numbers of image generations?

[uneven wasp]

on what?

running on your own pc, absolutely

the bot in this discord? yes

some randos site that runs of sd? no way to tell without knowing which

dreamstudio? limited I think

[gusty oasis]

If youre running locally the only thing you need to pay is the electricity bill

[uneven wasp]

lmao if you had to pay to run it locally there'd be a million torrents for the weights

[next spade]

So I don't have unlimited generations on this discord server?

[gusty oasis]

You do

[uneven wasp]

One of the devs said you have unlimited on the server bot

[gusty oasis]

So just prompt away

Not like the little fancy generator over there that needs to pay and see your every steps

[lofty zephyr]

did anyone in SAI said anything about distilled models being trainable on consumer hardware?

[next spade]

Damn, just felt too good to be true

[gusty oasis]

Also my thoughts when i heard sd can do boobs

[next spade]

🙏

[lofty zephyr]

you don't have enough boob pictures online?

[gusty oasis]

It's not specifically tailored to my taste

[lofty zephyr]

well I'm intrigued now

[next spade]

So this is not for discord - "Today we're announcing an 80% reduction in price for standard image generation (from 1.0 credits to 0.2 credits per standard image). Each credit will effectively be worth 5x more!"?

[cinder quiver]

Can anyone tell me what is the use of this symbol ": :" in the prompt?

[sage forge]

am i the only one that doesn't know what any of this means?

[oak nest]

Hi

I have a really interesting question to the stability AI

Anyone can support me?

What is APIURL when connect with TypeScript

[young ivy]

I'd really love an actual software (not gradio webui) that can give you some really good functions. Or I guess it can be a modified Krita

imagine a healing spot brush that does inpaint

you mark an area and it generates just that

[bleak matrix]

Good morning, everyone

How are we all today?

[lofty parcel]

Probably going insane but other than that good

[finite hawk]

There's a Krita plugin and it kinda works already

It's not incredibly easy to use, but it works

[young ivy]

I know :) I've been talking to the guy in charge of it and he wants to make an entire Krita build based on SD

to write it into the software

[tribal oar]

@tight quartz I was thinking, do you think we could have a game like channel where one user posts an image, and others try to guess the prompt?

I think that could be a lot of fun, and also help teach people good prompting

[tight quartz]

@tribal oar like a game of telephone, yeah?

[tribal oar]

Oh that's good too!

[tight quartz]

I mean, people could use CLIP-int tho.

[tribal oar]

Yeah but that never matches the original prompt

I think the fun would be trying to guess exactly what words OP used to get the resulted image

Before the general with images channel turned into politics there's a little example with the bull topiary

But I like your idea better. One person posts a picture, and then everyone else tries to recreate the image, closest image wins, and then OP shares the prompt

I dunno I was just noodling on the idea, I think there's potential for some fun and learning there

[storm shard]

https://civitai.com/

[balmy field]

I somebody found https://civitai.com/models/1212/kurzgesagtish - cool site, not tracked by search engines apparently

[flint sleet]

can someone tell unstable diffusion that their head is up their ass and all they really need to do is follow what looking glass did

[lavish current]

I found a video about Dall-E Mini and it made me realize how old and primitive it is compared to SD and its finetunes

[white mountain]

There are some ways in which it's better even today

Definitely not in quality of the images, but sometimes it seems to understand a prompt that other models don't

[lavish current]

Yeah that's true

[vital osprey]

this thing really helpful

tyty

[frank moon]

How would you train a model to do multiple faces?

Can it be done on dreambooth or do you need something else?

[eternal jacinth]

Is there anyway when you save an image to export the exact settings used to generate an image?

[cobalt cliff]

is the hostility against AI growing?

how do you guys feel about it? feels like the takes are getting hotter and hotter

[leaden mural]

So... What do you guys think would be the closest competitor to ChatGPT? And I meant in the open source sector.

[leaden mural]

Maybe? Though some of the artists I know are for AI gen art at the moment.

[fervent thunder]

more people are becoming aware of AI so the takes will continue to get hotter

what I want to know is, is any ui integrate the CLIP they released? nobody seems to give a damn about it, but its a game changer for image gen if youve ever used it on the site

like you could fool someone it's a whole new model using CLIP the accuracy is so much better, Id think everyone would be excited its released like a new model, but peopel dont give a damn even around here

[keen harbor]

Does anyone know how to put in a seed and add onto it?

[eternal jacinth]

You should be able to just paste the seed

[dawn yarrow]

are there any good online resources for image 2 image? i cant run locally with my crud gpu

[muted skiff]

is it possible to prompt "nightclub style" dim lighting? The portraits are always so bright

[muted skiff]

if you can run txt2img you can run img2img

anyway, the TLDR is that it's basically txt2img, except in addition to the prompt it uses the image you give it as inspiration for the image it generates. And the Denoising value determines how much (or not at all) like the original image it should be.

[grim aspen]

Is malicious code in stable diffusion models really a thing?

[fervent thunder]

anyone know of a good model that does img to img cartoonization good?

[manic yoke]

Does someone got clip aesthetic embeddings to work, it always gives me AttributeError: 'FrozenOpenCLIPEmbedderWithCustomWords' object has no attribute 'tokenizer' ?

[dapper nest]

Hi guys. I want tl train an AI with my drawings so the AI can draw like me. What path should I follow?

[midnight crane]

CLIP guidence should help the inference to guide the image to the prompt, but I'm not exactly sure how it does it.

[still sequoia]

ckpt files contain pickled python objects which can run arbitrary code, so they can be malicious, but I haven’t heard any reports of this in the wild. Auto1111 by default scans for malicious code in ckpt files.

[still sequoia]

I’ve heard that textual inversion is good at copying styles https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion

[muted tartan]

Are there guides for solvers?

[woeful pelican]

what is the best free web image to image?

[frosty marsh]

how do you accept the license agreements of stable diffusion 1.5 and stable diffusion inpainting?

[timid gate]

playground ai

[signal raptor]

Quick Question, I built a free iPhone app that helps people come up with prompts to create AI Art, what channel should I post that in?

[delicate oxide]

whats up

[fervent thunder]

hi

i heard ur discord got stolen

by discord

[golden nebula]

using ai as an hybrid to set up the art piece and drawing over it with human effort, problem solved

[fervent thunder]

is that true

[fervent thunder]

ur a bot?

[golden nebula]

no

[fervent thunder]

i was gonna say

why u got nitro if ur a bot

[peak orchid]

The guy on Stackoverflow who spends all day replying to questions

[wise stratus]

https://open.ai 👀

[austere pulsar]

Once stability releases their super fast distilled model, how long do we think it'll take to be integrated into the diffusers library (as opposed to merely being downloadable from huggingface)?

[vapid tangle]

Good question. Is it hard to implement I wonder? I definitely noticed too that clip guidance is an absolute game changer

[worthy oriole]

Wait, what? Open.ai?
And they say:

"Imagine if the best AI models were open and free.
What would you do with them?
Open AI is working to make the world's best AI models open."

I'm confused.... what is going on..?

[wet vine]

https://github.com/VoltaML/voltaML-fast-stable-diffusion
Has anyone tried the VoltaML accelerator released today?

[haughty sentinel]

@wet vine yes I did and it works great, but I'm excited for the eventual add to the automatic web-ui, seems like dynamic shapes support as well.
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/463#issuecomment-1345642637

[gusty oasis]

https://open.ai/ regarding this. You're telling me that a Company Named Open AI is different with the company that's also named OpenAI what the heck is going on

[radiant bronze]

Is it different? It says trademarked 2014 which sounds sort of like the open ai we all know and love..

[gusty oasis]

OpenAI doesn't exist back in 2014

Also OpenAI being open ??? Lol not even

[uneven wasp]

OpenAI is the actual company, not really open but they have been doing grounded breaking research into ai for a while. open.ai is some random ass shit, check their archive.org history. They were doing some crypto stuff, then reccomended stable diffusion, etc

the site used to direct to openai.com but it looks like they lost the domain earlier this year

[gusty oasis]

crypto
Well, now i know i can avoid that company

[worthy oriole]

But now is not from OpenAI?

And why do they use the name? And Emad links it? 😄

[gusty oasis]

Yeah and why did Emad post it on discord

[midnight crane]

It does generate images pretty quickly, don't know what model it uses though, but maybe something good will come out of it eventually

[worthy oriole]

I would say that it looks like Dalle 2... If not, is pretty similar

[compact violet]

Hi! Had a question. Is there a website like Lexica for SD2.0? Learning what works and what doesn't has proven a much bigger challenge than i first imagined with SD2.

[worthy oriole]

Oh
https://web.archive.org/web/20221203135203/https://open.ai/

Ok...Hm....Isn't this people gonna get sued?

[hollow mortar]

Anyone have a quick and easy way of gating a model behind a web frontend?

Spending a bunch of time making the model, and don't want to have to double my efforts here, and make a whole website from scratch XD

Wondering if there are any open source ones

[uneven wasp]

you mean like gradio?

[queen river]

Is this a Crypto

[ruby lake]

How is Automatic1111's gui like right now?

Does it support all the new 2.0 features yet?

[fervent thunder]

yes and also 2.1

[vapid tangle]

do they have an optimized model for 2.1?

like they had for 1.5 so weaker gpus could run it

[full lark]

I wonder the same 🥺

[weary jackal]

anybody here know how to use depth mask well ?

i ve got everything but cant create usefull results

[left mango]

trying to figure out how to do insert a fine-tuned image into another masked image... something like dreambooth + img2img... anyone got a clue for me?

[winter pike]

we gotta wait till the distilled stable diffusion gets released

[bright salmon]

Hi I have a question. Are there any scripts that allow prompts to change at a specific frame for IMG2IMG? Im looking for specific frame control sort of like deforum but for img2img. is this even possible. I am using Automatic1111 web ui and on the latest version

[timid gate]

gotta wait for new year

[timid gate]

some of the guys at #1010577750077210726 may have your back, ask about parseq

[bright salmon]

dope! thank you

[winter pike]

will be my best birthday gift

[oblique garnet]

Is there a way I can tell SD not to generate multiple people when making taller images? (512x896 or something like that)

[muted mesa]

does anyone have a guide on resources I can use to help me make better quality images and art

[crystal chasm]

How do I use my newly-generated dreambooth model?
I'm using autmatic 1111 and its not showing in the list of models

[full lark]

Maybe if you put something like "crowd" or "many people" in the negative prompt 🤔

[wet spear]

Any Stable Diffusion developers in Bangalore, India (DM for weekend catchup)

[woven spire]

algum br pra ajudar?

[thin nimbus]

😹

[urban tusk]

When we getting stable diffusion on our phones

?

[quick basalt]

Another day, another Streisand effect https://civitai.com/models/1305/sda768

lmao

[primal marten]

Wait what's the difference between a model and an embedding?

[crimson abyss]

I'm running it the same way as the 1.5 model and others, works fine with 6GB here using AUTOMATIC1111

[vapid tangle]

oh thats good to hear. did they make a new tutorial for using 2.1?

[crimson abyss]

Nothing has changed really, it just doesn't interpret prompts like 1.4 and 1.5 but the way to use it is the same

If you have ever used other models like the elden ring one it's kinda the same thing

[fervent thunder]

has anyone tried ben's fast colab for training 2.1? curious how the results are

[muted mesa]

im trying to generate an image but it has something that resembles a watermark on it
and regardless of what I type in the negatives it remains, is this prompt specific, can someone help me?

[upbeat tangle]

hey could someone link me the vae/embedding that improves hands ? i think it was released by SD

[quaint yarrow]

are they going to continue working on SD 1.5 or is SD 2.0/2.1 the primary focus?

[fervent thunder]

Hello,
Can you redirect me to a person responsible for partnerships/marketing for your project?
Thanks

[ember copper]

What us “mucha”? I see it in a lot of prompts but no idea what it is

[peak fern]

https://en.wikipedia.org/wiki/Alphonse_Mucha

[vague mirage]

when training the ai on creating consistent characters, such as myself, is it better to provide 512x512 or 512x768 samples?

[ember copper]

pretty sure it's 512x512. i mean haven't tested it, but everything i've seen says that the models it references against are 1x1

[ember copper]

thank you! odd they don't separate it by commas, just putting artists mixed with mucha.... didn't know it could do that. thank you!

[onyx sapphire]

hello

[muted mesa]

are there any tips and tricks when it comes to img2img

[grim aspen]

Does it make sense to use something like this then: https://github.com/zxix/stable-diffusion-pickle-scanner

there is also this one: https://github.com/mmaitre314/picklescan

[worthy oriole]

It probably makes more sense for the community to move away from ckpts and use safetensors....

Those things can fail.

[leaden mural]

Weird question guys, does Unstable Diffusion get their datasets in an ethical manner (like taking from people with prior permission, datasets that are free to use, etc)?

Yes, it's weird, but I am making a post.

[worthy oriole]

With prior permision? You mean like... Asking the autors?
Surely not. That's impossible anyway,, But honestly speaking, a lot would say "no" if asked

But you could probably ask them

[thin gyro]

what is your favorite img2img model?

[grim aspen]

So should I use this?: https://www.reddit.com/r/StableDiffusion/comments/zhozkz/safe_stable_ckpt2safetensors_conversion_toolgui/

[cerulean rivet]

Could anyone tell me where to fine the SD 2.1.ckpt file for auto1111. I only found this ema file and it did not seem to work: https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/v2-1_768-ema-pruned.ckpt

[worthy oriole]

If you create a model and want to share it....

If you take a pickled ckpt and you run the conversion you Will run the pickled code, so it wont save you haha

[grim aspen]

ic

@worthy oriole but the resulting safetensor file should be safe

I can convert them on a 2nd pc that I can just format afterwards

[fervent thunder]

What happened to the meme channel

[primal marten]

Follow this tutorial https://www.youtube.com/watch?v=e3vcYVwEkW0

You gotta add the yaml file as well, its all explained in the video

[cerulean rivet]

Ah awesome, tks

[ionic copper]

yoo whats new

[feral raft]

What version is dreambot running on rn?

[bleak matrix]

Good morning, everyone.

How are we all today?

[lofty parcel]

Good and stressed at the same time

[bleak matrix]

You know, I feel that. Had a power and net outage after this morning already lol

What's up with you?

[lofty parcel]

I kinda don't know how to market my AI game when it's done

Mostly because I've never done marketing in my life

At least not properly

[bleak matrix]

Well, that depends

Marketing while you are creating can be a double edged sword, especially as a small developer

There are a lot of fantastic small communities, though

Both on Discord, on art sites, on Reddit, etc, where you can get feedback, interest, etc, as you develop

That said games like Hello Neighbor and We Happy Few are perfect examples of releasing while developing, which, while it gathers interest, can cloud your division as a developer (or small team)

Because you can get far too caught up in maintaining interest over making a good game

[soft basalt]

is it fine if I invite dreambot to my server?

[muted lake]

Hi guys, how can I use one IA, i new at these stuffs

[bleak matrix]

A good game, in a lot of ways, is a market itself. But I would say focus on creating your game first, and making sure it's enjoyable to play. How do you plan on selling your game? @lofty parcel

[bleak matrix]

Are you trying to use the website (Dreamstudio) or locally install?

[lofty parcel]

Probably a subscription like most of the other AI services

[compact perch]

Just wondering what the law says about making fan art by using celebrities faces?

[bleak matrix]

With your own site as a mmo, what?

[north stirrup]

In many jurisdictions, they could very well sue on the basis of personality rights, but 99.99% of the time won't bother because it's a civil matter

[bleak matrix]

Laws vary from country to country. I'm not a legal authority by any means, so I reccomend you check with your local authority/country.

[north stirrup]

so say, putting an actor's head onto a kangaroo for lols probably won't get you sued, but selling a print of it is much more likely to

in other news, I am not a lawyer, this is not legal advice, and you should consult a lawyer about anything you're concerned about

[lofty parcel]

MMO? As in a multiplayer game? My game is singleplayer so idk what you mean

Basically it's like NovelAI's text generation in that every adventure is entirely separate from the one before it; a brand new world is generated from scratch every time instead of one world being shared amongst all players

The ongoing cost then, is mostly the compute required to generate new worlds for every adventure

[strange hemlock]

I'm sorry for asking such an obvious thing but I can't understand how to use those bots. Every time I get a response that my dream can't be posted cause it contains forbidden stuff even if i am asking for a sunrise. I am using discord in my phone browser and can't find those "dream credits" menu. Please help.

English is not my native language so I struggle to understand the instructions

[woeful pelican]

will you maintain the stats of the character throughout the game?

[lofty parcel]

Basically, each adventure acts as its own self-contained RPG, with different characters, story, enemies, bosses, items, and environments with its own separate progression

It might be confusing because the "game" I'm developing is more accurately described as being a game generator

[woeful pelican]

oh ok interesting

[lofty parcel]

We may not yet be able to generate games with AI from scratch, but generating infinite content with AI is more than doable at this point

[woeful pelican]

dose sound interesting

[lofty parcel]

What I'm doing is programming the technical part of the game and creating a system that prompts SD and GPT-3 to fill in the content in a hopefully interesting way

[broken gull]

Stability.AI is gonna release a model for games

[bleak matrix]

Yeah, multiplayer!

[lofty parcel]

We'll see
If they pull it off in a compelling way I and millions of others will probably have an existential crisis

[lofty parcel]

Yeah I'm just doing singleplayer, multiplayer would be an afterthought, not a core part of the service

[bleak matrix]

You mean here on the server, right?

[desert lily]

Stupid question, is it possible to use SD as a just a denoiser?

[bleak matrix]

You mean like, put in an image and have it....adjust a photo?

[lofty parcel]

I think so, you could try using img2img with very little AI influence

(i.e. high image strength)

[woeful pelican]

try this https://huggingface.co/spaces/lnyan/stablediffusion-infinity

[desert lily]

yes

[woeful pelican]

it's what I've been using

[white mountain]

Forgot to turn on hires fix https://www.reddit.com/r/Damnthatsinteresting/comments/zkuvar/identical_twin_sisters_briana_and_brittany_marry

[bleak matrix]

I mean, you could take part of an area and try blurring it or adjusting it. You may or may not have good results, depending on the photo, strength, what you prompt.

The worst you could do is try.

I haven't personally tried it myself, so I can't really vouch for anything.

[desert lily]

Well using a prompt would defeat the purpose

[bleak matrix]

A prompt such as smooth skin or blur

[tawdry forum]

I just released a new model: https://www.reddit.com/r/StableDiffusion/comments/zl03dl/seekart_mega_a_new_general_model_for_stable/

[queen palm]

Hello. Could some please explain me how to use the ai with my face or a picture or something, I'm new here and I'd be really grateful if someone offered to help

[still plaza]

I won free Nitro, so that's cool.

[quaint coyote]

hi yes kindly someone explain since im new here

[zinc solstice]

Also does anyone know any good communities on IG or otherwise where I could share/tag my artwork? So many scammers on Instagram just looking for a quick buck with promos…it’s frustrating! I get 20 messages a day with pricing (scam) I love discord but want to reach as many people as possible.

[frozen trench]

I'd love a quick rundown on what code to download/run to start my own stable diffusion

Not sure where to start

[fervent thunder]

I believe most people use the AUTOMATIC1111 webui when using the sd locally

[broken gull]

He talks about starting their own stable diffusion so ig he means training models?

[fervent thunder]

ah, then just ignore me :3

[broken gull]

Since if he wants to recreate stable diffusion from scratch that's not gonna happen unless you got plenty of money.

[broken gull]

Well you can use automatic1111 for training so your answer isn't unvalid.

[compact perch]

So, if I want to train the AI on my face, I should take atleast 20 photos on different angles and then crop them to 512x512? Does it matter which model I have active in the webgui? Like if it's stock SD or a custom one?

[spare egret]

I have been trying to get SD2.1 model to work in Automatic1111. I got the model file, renamed the .yaml file, bu tevery time I try to load the model it is trying to download a large file and fails after few percent progress.
File "C:\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub\file_download.py", line 495, in http_get
for chunk in r.iter_content(chunk_size=1024):
File "C:\stable-diffusion-webui\venv\lib\site-packages\requests\models.py", line 760, in generate
raise ConnectionError(e)
requests.exceptions.ConnectionError: HTTPSConnectionPool(host='cdn-lfs.huggingface.co', port=443): Read timed out.
Downloading: 3%|██ | 122M/3.94G [00:30<06:56, 9.18MB/s]

What is it trying to download? Can I download the file manually?

[rose sedge]

Is there a website to look 1.5 embeddings?

[fervent thunder]

Sup kiddos?

[modest owl]

Power user question:

is there some model orchestration available?

In case if you have a dozen of models and want to switch between them via keywords mid-generation

[bitter olive]

Does someone know a model that's good with fantasy worlds/ creatures?

[modest owl]

try searching here https://civitai.com/

[serene swan]

hey all Ive been gone for two days what's new lol

[modest owl]

or maybe try https://huggingface.co/nick-x-hacker/bad-artist

(not just for weeb bs)

[bitter olive]

uh,so that's also a place where I can get packages for free, like on Huggingface?

[modest owl]

kinda

influence here https://github.com/civitai/civitai/discussions/categories/ideas

[bitter olive]

What are ".PT" files? Is that something that downloads the real SD package`?

[modest owl]

those are the embeddings

as far as i know those live in embeddings folder and are triggered with a keyword

like (bad_prompt:0.8) for fixing hands

https://huggingface.co/datasets/Nerfgun3/bad_prompt

[serene swan]

do parenthesis help for weighting a prompt in 1.5?

[modest owl]

same story as usual

afaik

[bitter olive]

Ah ok, so those are basicly like guidlines when you load a model.

[modest owl]

but hey, it can't generate invisible people for some reason

at least hassan's model can't 😠

[modest owl]

basically, here's an example https://civitai.com/models/481/hydrasuit

[bitter olive]

ahh, so they help you Tweak the outcomes before they are generated. interesting

so instead of searching for SD Models that have been specifically trained in one area. I can use a model that's good enough and embed some ".pt" files to Tweak it to what I need?

[modest owl]

like introducing some specific concepts

[grim aspen]

does someone know a good discord server for gpt like ai?

Not deploying a chatbot but talking to others about the topic like here

[modest owl]

uh, i run my small server 😛

called AI Shack

[fervent thunder]

My office is legit 12 degrees hotter than the rest of my house because of model training. lol

[serene swan]

omg

[modest owl]

gotta have a warm winter

[fervent thunder]

Yeah

[modest owl]

drop some lumber on that gpu

[fervent thunder]

just waitt il I get the bill... lol

[modest owl]

🔥 🔥 🔥 🔥 🔥

[fervent thunder]

my average it/s right now are 20-22

Would be the same here...if my computer didn't escape and fly away after I installed all those extra fans to cool it down 😭

I have two rigs. They were side by side, but I had to move them because they were heating each other.

One has a 3090Ti the other a 4090

[serene swan]

do I have to upload a checkpoint to gdrive if I want to train it with thelastben colab?

[modest owl]

train in huggingface instead?

https://huggingface.co/spaces/MirageML/dreambooth

[serene swan]

hm I'll give it a look

If you would, what are some advantages of this?

[modest owl]

https://www.reddit.com/r/StableDiffusion/comments/zh00jo/simple_dreambooth_training_web_ui_on_hugging_face/

[serene swan]

hm I would like to use a specific model based on 1.5 not 1.5 itself, how woould I go about doing that?

[fervent thunder]

Progress:|█████████████            | 52% 206/400 [0:51<01:4,  21.01s/it, loss=0.0239, lr=1.02e-6] 234d2332 

[fervent thunder]

are there any 'must have' 2.1 models yet? i mean something like that f222 which was a good photo real 1.5 model

[signal nest]

https://twitter.com/arstechnica/status/1602766327067705345. Scary level of control over AI in china . If we dont get people on board (and perhaps police ourselves regarding blatent plagarism) - this is whats possible. Seems in china the level of state control over internet apps (dont they have one do everything phone app or something) lets them police what the state does and doesn't approve of :/

I'm quite worried that from artists and other walks of life people who feel threatened by AI will call for similar in the west, and justifications might be : [1] eco/energy use (restirct AI users like people wnated bitcoin regulated during the mining crazes) [2] misinformation (eg especially with world conflict) [3]privacy [4]protectionism of human workers

we must be proactive in egtting non-contraversial models out there

AI is an amazing tool and it would be awful to lose the freedom with it

[serene swan]

they already are and will try to get ai art banned, and there's no amount of self-policing that will stop them. The crux of it over here is a misunderstanding of copyright here methinks

can't stop progress though

[little plaza]

I do not think generating images on your GPU with SD uses that much more electricity than normal. you'd have to be running it constantly ... even then it's not going to be anywhere near what a single NFT uses.

[sage sky]

@supple jasper

[signal nest]

well to be fair , until the energy crisis hit , I did want to run dual RTX3080's 24 hours a day every day (for graphics related machine learning generally) 🙂 (the crypto GPU shortage held me back from that) . I've scaled back to planing for 1 rtx3080 class GPU x 8hours per day max. I would personally argue that using AI can usually save energy in other ways (eg needing to move people to specific workplaces, and save people time to do other things more efficently). but I can easily see how an ecological argument can be made against it.

[little plaza]

I dunno how accurate this is but "It takes an estimated 1,449 kilowatt hours (kWh) of energy to mine a single bitcoin. That's the same amount of energy an average U.S. household consumes in approximately 13 years"

I guess it's a possibility if like there are huge services offering AI generation on a mass scale for like consumer apps they might be regulated. Guess it depends on just how much we can scale them down (like if smartphone generation becomes a thing)

[ocean mango]

"In an interview for Fortune, Emad said that next week Stable Diffusion will generate 30 images per second instead of one image in 5.6 seconds. The launch of distilled Stable Diffusion should be as early as next week."

[fervent thunder]

"distilled"? :O

[signal nest]

if there was such a model - 150x faster - i'd just generate 150x as many images, and do more ambitius projects. (2d -> higher res/more pohtoreal -> 2d->3d, 2d->animated , 3d+animated, texturing large scenes..)

I can beleive a distilled model thats much faster for applying a specific texture theme to a world would be possible , I guess all these use cases I just described could use them.

[late dawn]

something happened in the Midjourney server? saw an exclamation point and now it doesn't appear even in reinvites

[winter pike]

it will be close to new year last time i check, so stability employees can enjoy christmas

[winter pike]

a "temporary outage", gotta wait till its up again

[fervent thunder]

their community is huge so maybe it's just an overload :P

hi

someone

can help me

how i can use the bot

for make a imagen ?

[sudden marlin]

Hi there, if you head to any of the #1047760914008522782 to #1047760279267721236 channels then you can use the /dream command

[serene swan]

wow, just realized I have to download a ckpt and THEN link it to colab for it to train a custom model. lol oh me

[midnight crane]

Does anyone know what model Lexica Aperture uses? https://z.lexica.art/aperture

I find it very strange that there isn't any information about it. But it produces excellent photorealistic pictures of famous people.

[frank moon]

How do you outpaint on SD Automatic1111?

[ocean mango]

gotta install script

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts

[wise stratus]

Sd 1.5 finetuned on high res images I’d imagine

[muted mesa]

would an m1 macbook air be sufficient for a local Sd download

[fervent thunder]

Is this a general Stable Diffusion discord server or its specific to 1 model/organization?

well Emad is the ceo of Stable and he has a message two above yours

so its the place to be

[astral parrot]

Hey everyone - if you've got a project around Stable Diffusion, and are thinking about turning it into an actual startup, you might be interested in coming to an event that Y Combinator is hosting on Saturday.

We're hosting an online meetup for people interested in starting a startup doing something in generative AI to meet potential founders. It's free and open to everyone.

You can sign up for it here: https://twitter.com/ycombinator/status/1601246100685406208

[wise stratus]

This Discord is run by Stability AI and focuses on our part of the ecosystem mainly. So place to be for 2.x in particular plus all our stuff around that

[edgy basin]

does someone know if there is a local GUI for SD with infinite canvas function ?
AUTOMATIC1111 has it?

[fervent thunder]

anyone know where i get a vae for a 1.5 model?

i did a new install and cant find the file

[uneven wasp]

its in the object once you load the checkpoint under first_stage_model

[dapper ferry]

omg i found how to make 2.1 hands down beat 1.5

[waxen storm]

really hope the next model is far more progressive...2.1 is the worst yet

[dapper ferry]

dude 2.1 is amazing right now

are you using dream studio?

[waxen storm]

ummmm...its not just me man 🙂 https://www.youtube.com/watch?v=H13_q6mYclI

[fervent thunder]

thats because it has the CLIP guidance

[dapper ferry]

i know i totally know that 2.1 is shit, but i just found a setting that makes it easily beat 1.5

[fervent thunder]

they released it publically but i guess CLIP requires lots of vram to run local

[dapper ferry]

come to the chat with images

[north stirrup]

There's this, which I haven't tried yet: https://github.com/lkwq007/stablediffusion-infinity

[rain basin]

https://www.reddit.com/r/Cyberpunk/comments/zk2n8g/dysons_moderately_horrifying_airpurifying/ looks like Dyson’s designers been messing with stable diffusion

[wise stratus]

Oh, you think dorkness is your ally. But you merely adopted the dork; I was born in it, moulded by it

[unreal cargo]

highly recommend InvokeAI for infinite canvas + solid GUI

uploading custom models is a bit of a pain but... I digress

[untold ocean]

its a custom model they trained. They made a couple tweets about it

[untold ocean]

invokeai released a cool one recently iirc

[fervent thunder]

would you recommend creating flipped copies?

[unreal cargo]

it can help if your training images are low

[fervent thunder]

if I have 15 would you?

[unreal cargo]

honesttlyyy I try to avoid it if I can

it might help! but, yeah~

They're pretty fast to train, you could always try both who knows

[fervent thunder]

good idea

do you think 8 vectors for 15 photos is good?>

[unreal cargo]

I'm still figuring out vectors personally LOL

I have zero idea

[jade meadow]

eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard so you not going to eat a turkey sandwich with no honey mustard so you're not gonna eat a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard eating a turkey sandwich with no honey mustard

[uneven wasp]

you good?

[summer haven]

My muppet hypernetwork if anyone is interested. have fun! https://huggingface.co/doctorderp/muppet/tree/main

[fervent thunder]

Im confused, embeddings are .pt files? i thought they made ckpts to load

where do i load an embedding in that case/the pt file?

[jade meadow]

eating a burger with no honey mustard

lol

[quick basalt]

Guys what is the biggest difference between the anime models

I use EimisAnimeDiffusion, but there is also waifudiffusion and anything v3

[uneven heath]

Twitter, if you actually interact with AI folks there

It can be either gold, like absolute gold

or hellfire

[peak orchid]

Artstation users are currently trying to sabotage webscraping with fake tags, and demanding that the site ban AI content while going after anyone who trains on Artstation content: https://www.artstation.com/?sort_by=trending

[jade garnet]

lol
This only makes sense in the alternate reality where AI models are actually just querying artstation to photobash

[scarlet atlas]

trending on artstation
im sorry

[unreal cargo]

I mean, I don't blame them. Change is scary and humans fear the unknown. Especially when part or all of your income relies on it, as well. There's a lot of misinformation / disinformation on those posts about AI Art which makes me sad

It'll eventually go away with education but, welp, until then

[jade garnet]

It is not going away lmao

More and more platforms are getting seized by this

[unreal cargo]

AI/ML isn't going away

some people will always hate on anything and everything, that will never go away

[jade garnet]

to the other people reading this: please do not expect a politically motivated force to simply vanish with time. The organised opposition will not disappear by simply waiting

[unreal cargo]

There are people that still hate on photoshop

or the camera

or accounting software

automation is a big one

Coal

Technology literally in general

I'm not sure I understand your point

[bold cave]

Question, what’s The software I can use to make a model of my own artwork datasets ?

[unreal cargo]

Textual Inversion, Hypernetworks, Dreambooth, EveryDream

[bold cave]

Ok thanks 🙏

[unreal cargo]

Hypernetworks are a bit of a gray area atm, but... Dreambooth or EveryDream might be your best bet, are pretty high quality too

[bold cave]

Nice I’ll give it a try

[fervent thunder]

can you make ckpts in auto ?

[steady arrow]

What does everyone like most for sampler in Auto

[unreal cargo]

the dreambooth extension lets you

[fervent thunder]

oh that explains it

i havent added that

[unreal cargo]

mmmm

[fervent thunder]

are there any extensions worthy for 2.1?

[unreal cargo]

I thought you were trying Textual Inversion first?

which doesn't make a ckpt

[fervent thunder]

I was I wasnt too imrpessed

but maybe just didnt do it quite right

[silent torrent]

Stupid question but what is the difference btw SD2.1 768 and SD2.1 512

768 is for width ?

I cannot generate images under this dimension ?

[fervent thunder]

have you ever used Dreambooth in auto? Trying to figure out what Lora model, source checkpoint and scheduler to use/are

[unreal cargo]

Yeah I use it quite a bit

not lora though, though I will as soon as it's officially added to the extension

[zinc solstice]

Is a photographer not an artist because they didn’t build the buildings, make the food or create the humans, animals they are capturing?

Are digital hand drawn artists not artists because they didn’t create the tools/brush types they inevitably use to make their art?

Every artist is inspired by other artists. Sure…they didn’t write it down in a prompt in some generator, however subconsciously or consciously artists extract from other artwork and styles. Hell we were taught to damn near paint by numbers in art class.

You remember a style, are influenced by a style and in the end you combine all of that inspiration into your own style.

Are musicians not artists because there’s no denying every single piece of music and every musician has been influenced, shaped by one another:

“Beethoven did it. Even Robert Johnson and Ray Charles did it. Creative masters all, they each appropriated music from others in their works and were borrowed from in turn.”

Should the artists with health issues and disabilities not partake in this new way of creation? What about the creative mind with carpal tunnel syndrome, arthritis or the amputee? Should they hold out on their passion to please some of these professional artists in the community because to them it’s “not art”

It’s sad and disheartening but at the end of the day AI Art is fast growing, addicting and helpful tool that is only going to grow larger. It seems some of the protesters I’m seeing are very angry, almost threatened.

I am a painter, I use decoupage in some of my hand made artwork, I am somewhat of a calligrapher, graphic designer and just an overall creative person. My young son introduced me to this way of creating just a couple months ago. We spent all night thinking up creations together, laughing, ooo’ing and ahhh’ing…just enjoying life. Being happy. Emotion is a big piece of the art pie. I love this way of creating and I will continue to support artists regardless of their medium.

[north stirrup]

I completely agree with the powerful pathos of your argument. We still need to treat the source artists as people, and that means limiting ourselves a little when using stuff that's close enough to their style to not be truly original. In copyright terms, I'd say that some uses risk not being transformative. We should actively seek to respect the talented individuals that inspire us.

[zinc solstice]

I agree with you, but to ban all AI Art and dismiss it all together is a tall order of kicking a can down the road with a hefty side of bitter beans. There will always be a way to incorporate that artist without using their names within the prompt. Is our Art going to then be labeled as plagiarism? I’ve yet to see one person who is anti-AI have a mature mindset and embrace the positives. Is it the name? The way the art looks? Because again, I’ve created so many great pieces without adding trending on art station and yet it looks like I could have.

Now on the flip side there is abuse and corruption within so many communities of this world. Am I against the jerk who uses an artist in their prompt and passes it off as an original? Yes. Or the sad person who pretends they didn’t use the generator at all, pretends to be all hand drawn and reap rewards in vain of true artistry? Of course!

But there doesn’t seem to be a happy medium on “that side.” Hopefully that will change when they see the very thing I’m talking about.

[north stirrup]

A lot of the arguments I’ve seen for banning it are due to people’s poor behaviour with recognizable people and characters that belong to others. We could neutralize arguments for banning the technology by behaving unimpeachably in the first place

[fervent thunder]

exactly as I see it. Another part is that I've yet to see someone say that the ai straight out copies the same style of an artist, it's always "close to, but not quite." So to me that says even more about the argument as can't AI even come close to someone else's style? Is it the name? No, that can't be, because using an artist name is extremely normal I'd say. What is the easiest way to explain to someone what something looks like? "Oh, that reminds me of Van Gogh painting!"

I believe it's just the sudden shock of thousands, if not millions of AI images suddenly appearing everywhere. Where there was few, now there are many. And I believe people are scared of being overtaken by technology. But in the end, this will help everyone. Artists will turn to AI for assistance to make better art quicker. Teachers will be able to teach more students. People may have more unique art in their home. Before there were small pockets of people making art, now the rest of humanity can join them as well.

[zinc solstice]

Love that response @fervent thunder great points. Also great points by @north stirrup This is still fresh, only time is going to tell. If I had a magic wand I’d want all artists to be happy, not just what I believe makes art fantastic for me. Until then I will keep creating and experimenting.

[reef epoch]

Okay fine I’ll settle THIS FOR EVERYONE

gimme a sec lemme go ask ChatGPT

[brave matrix]

hello guys! how would i ask stable diffusion to use the same "style" idea for each frame of a img2img animation?

would it have to do with the varience setting?

[reef epoch]

ChatGPT prompt: Write me a very detailed Explanation about how a text to image diffusion model, works and learns. Explain the difference between it and art created by the traditional hand created art. Talk about the ethics and importance and if copyright is an issue between them. And I want the entire explanation to be half written in Shakespeare style, half in an argumentative Aaron sorkin screenplay style and include metaphors of cookies

IN SORRY EVERYONE BUT PREPARE FOR THIS LONG ASS EXPLANATION

I didn’t proof read it

To speak of a text to image diffusion model, doth require one to delve into the realm of artificial intelligence. This model, when trained on a vast quantity of text and images, doth learn to discern patterns and correlations betwixt the two. And with this knowledge, the model can create new images based on a given text description.

But in contrast to art crafted by human hand, a text to image diffusion model doth lack the same level of control and intention. Whereas a human artist hath complete dominion over their creation, a model is beholden to the algorithms that guide its hand.

And so, with this in mind, one must consider the ethics of such a model. Can it truly be called "art," or doth it fall short of this lofty title? And what of copyright, doth the model's creations infringe upon the rights of others? These are weighty questions, worthy of consideration.

Perhaps a metaphor can aid us in understanding. If a hand-made cookie is the product of a baker's skill and care, then a text to image diffusion model can be thought of as a cookie factory. The factory doth churn out cookies with efficiency, but lacking the personal touch of the baker.

In conclusion, while a text to image diffusion model may be a wondrous thing, it must be wielded with caution and due regard for its ethical implications.