💬｜general-chat

[reef epoch]

Now let’s hope everything is settled.

I still didn’t proof read it 👀

[zinc solstice]

I literally lost it at “Doth Churn out cookies with efficiency” 😂

Seriously hilarious and actually made some points. Only wisdom chatGPT could give I literally would not have been able to say it better myself.

Makes me wonder if comedians should fear for their jobs too?

[reef epoch]

Nope cuz the true comedy came from me telling it to do it in all that jazz

[zinc solstice]

Or maybe comedic writers would be the one with the problem now that I think about it 😆 also…TRUE lol @reef epoch

[reef epoch]

I’m gonna make it write a song now

For emad

For Christmas lol

[zinc solstice]

Haha! Too good. Needed the late night laugh

[reef epoch]

@wise stratus i hope you like this song created for u for Christmas.

NOOOO

excuse me... but how does a Christmas song "contains content blocked by this community."

DO YOU guys not want to see the song 😦

its actually a nice song.. i wont delete it until it gets posted

[mild jackal]

https://www.artstation.com/?sort_by=latest what do you guys think of this

[peak fern]

I only accept paintings made with body fluids like tears and blood. Everything else is a corruption of the essence of art by technology.

[uneven wasp]

pretty easy to filter out of a dataset.

I can appreciate the ones with a modicum of creativity that aren't just copypaste tho

[reef epoch]

is there mod here alive atm... we need to discuss my chatgpt Christmas song

[chrome owl]

Lol

Is there a widely acccessible free way to createunique AI art of a self portrait or sum yet

Or is it still the case of needing to rent a 4090

[reef epoch]

you can like, follow a youtube video to use colab if you'd like, you can like look at the mirror and put on makeup, you can serenade a person it doing it for you for free. look mans theres a lot of ways to get this done

[fervent thunder]

how would a layman use this chatgpt?

[reef epoch]

well first you need to explain to me what a layman is, and this explain to me why are you a layman, and go into the details of why do layman exist. and explain all of this with chatgpt for me here.

but to use chatgpt just google chatgpt

make an accounts

and get me my explanation

[fervent thunder]

nice. I thought i looked once before into it and it was a complex process to use it, doesnt seem to be the case now

[scarlet atlas]

Friends, I think I found what we really need for this. it's an anime waifu.
Look at hatsune miku. She's basically the same as SD. A software, you give it text, it generates art. The only difference is that she generates audio rather than images.
If you think about it, she probably has already replaced a lot of human vocal artists (singers).

However no one bitters about her, because she is an anime waifu. You cannot fight against anime waifu.

It's a mystery why we don't even have a Stable-chan yet, considering the fact we could have generated infinity of them. This war cannot be won without an anime waifu, just like how people couldn't have won the Hundred Years' War without the Jeanne d'Arc.

[mellow knot]

stable chan has 7 fingers each hand

[timber raft]

she generates audio was obtained by sampling from a goddamn real Japanese voice actress and singer and you forgot that

[upbeat tangle]

so i install wildcard extension,put the txt file in wildcard folder,and use double underscore when prompting, thats it right ? cuz im not sure if its working lol

[modest owl]

who we? Weebs?

However no one bitters about her, because she is an anime waifu. You cannot fight against anime waifu.
laughs in hyperrealistic academic art girlfriend

[timber raft]

at least Hatsune Miku credit the voice actor that made her voice, better than bunch of guys insult an artist on Twitter then still put his name into their prompts

[modest owl]

also, you can adopt my license for your future works:

I Don't Give a Shit Public License

Version 1, December 2022

Copyright (C) 2022 [Your Name]

Everyone is permitted to copy and distribute verbatim or modified
copies of this license document, and changing it is allowed as long
as the name is changed.

I DON'T GIVE A SHIT PUBLIC LICENSE
TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION

1. Do whatever you want with this licensed material.
2. I don't care what you do with it.
3. Seriously, I don't give a shit.

No, this is not a copy of WTFPL

[fervent thunder]

Anyone got a link for the v2-1_512-ema-pruned yaml file?

[modest owl]

https://github.com/Stability-AI/stablediffusion/blob/main/configs/stable-diffusion/v2-inference-v.yaml

check them all https://github.com/Stability-AI/stablediffusion/tree/main/configs/stable-diffusion

[agile spoke]

is there any way to increase CFG scale beyond 30 and sampling steps beyond 150 in the Automatic 1111 webui?

[modest owl]

here's one better

is there a way to unbind width/height from the specific steps?

i.e. make it truly editable

[warm canyon]

Wow you guys are very grea!..

[fervent thunder]

I had downloaded html but fixed it by viewing it raw and then saving the page. It wanted to do it as yaml.txt but I changed to all files and then voila! fixed.

[modest owl]

so many hoops

[icy pine]

anyone know the best guide for installing stablediffusion? Have tried a couple guides from the get started channel but they say they are outdated, then some youtube ones form the last couple weeks just dont work

[modest owl]

https://rentry.org/voldy

btw

is 80 too many steps?

[icy pine]

do i download 1.4 or 1.5?

[modest owl]

1.5

[icy pine]

ok and for python the guide just says 3.10 other guides said it had to be 3.10.6 is there a best choice?

[viral marten]

Hi, I have an issue when I open webui.user.bat, it can't install torch. I tried changing python versions and deleting the venv folder but it changed nothing.

[fervent thunder]

https://www.youtube.com/watch?v=8ksyuQG8j9g @tidal bough

[tidal bough]

Lots of discussion around this some time back; Check the pinned posts in this channel for more info if you'd like! 👍

[lusty mason]

Hey guys do you think it's possible to create a good looking anime using stable diffusion and other tools

[modest owl]

they will keep denying it

[lusty mason]

Maybe?

[silk robin]

did u add python to path when installing it

[viral marten]

yes

[silk robin]

did it give u an error or smth

[tidal bough]

Anime images, absolutely (check out #1010577750077210726 for some amazing stuff if you haven't already!)
As for a full blown animated anime, we're certainly getting there! Some big advancements in animation tools coming up. There's some really cool animation down in the #🎥｜animation channel right now; we're not far off!

[viral marten]

oh wait im supposed to ask in the tech support channel

oops

[lusty mason]

Ah pretty cool! Hope we get there, I always dreamed about creating an anime episode without having to put in years of work

[modest owl]

:weeb:

[modest owl]

https://hyptosis.newgrounds.com/news/post/1325058

opinions?

another butthurt artist mob?

(even tho it's just lensa's problems)

[main sparrow]

He should just make an embedding of his stuff that's better quality of what's available online, and he could make 9348573498x more images to sell. I'm putting thousands of my photography and art from over the last 30 years in one.

Maybe I should start attacking these people as being 'ableist' since I can't actually physically do the art I've been making with steel and stone anymore because my hands are old and busted. This new AI tech is a godsend for my old ass.

Or I just do what I've been doing and just ignore the whole drama and make what I want to anyway.

[storm shard]

You can change these settings in the ui-config.json

[modest owl]

turns out https://www.patreon.com/klokinator is a twat

that was an interesting find

[abstract pivot]

Has training toasted anyone's SSD? Like from a lot of training

[main sparrow]

I was worried about that so I use a regular hdd

[lofty parcel]

Artists have legitimate reasons to fear automation

Each advance in technology demands that existing artists adapt or leave the field to find employment elsewhere
People can only adapt so quickly and there may well come a point where the rate of automation outpaces the speed at which artists - and society at large - can adapt

The rapid advancement of AI will never not be controversial under the current system because it represents an existential threat to the livelihoods of millions of people, and that number only increases as AI achieves proficiency in more domains

As amazing as progress has been and as mindblowing as AI is, it's a double-edged sword

[keen belfry]

Hello, can someone tell me if it's possible to start from an external picture and how to do it? Especially if it's not on the Web.

[abstract pivot]

Yeah? Im thinking about getting an SSD for it, woke today with a SMART error on my NVME in my laptop, not even a year old and I had been doing quite a bit of training the past month or 2 since discovering SD, hypernetworks create so many files, I must have generated about 500gb of data(hypernetwork checkpoints) in that time, not sure if it's because im on a laptop and it might have gotten hot or there's just a lot of writing going on and it's very stressful on the drive?

not necessarily saying it is training but might not be a coincidence and some caution should be had perhaps

[main sparrow]

laptops are hot in general, but the new PC setup I slapped together a few weeks ago actually had a heat sink included for the 1st slot NVME drive. They get hooootttttt.

[abstract pivot]

reckon the stress will be reduced with distilled diffusion?

assuming you're doing the same work load

it will be a lot more efficient right?

[lofty parcel]

I can't wait for distilled diffusion
I hope the quality is still decent though

[marble lynx]

I'm looking to create illustrations for a children's book using AI generation - CGI character like imagery - what's the best platform for a beginner?

[modest owl]

oh no, the poor calculator people that got replaced by actual calculators

[main sparrow]

BAN THE UNDO FUNCTION IN PHOTOSHOP OR IT'S NOT REAL ART!!!!

[modest owl]

don't get me wrong, but i find the whole "i'm an artist who does nothing but exist and draw and scrounges for money on patreon or some other platform" wrong, get a fucking job until you can afford to sink in

yes it might get pretty miserable, but at least you have an income

[main sparrow]

I've been making a living as a self employed artist / craftsman for 35 years now. I get it. But adapt or die is how it's always been. It's just happening faster now.

[modest owl]

then whytf people are bitching and not learning?

[lofty parcel]

Probably the shock of it
A lot of people thought art was going to be the last thing touched by AI

[modest owl]

just assume everything will get automated

[lofty parcel]

And then what are people going to do for a living?

[modest owl]

if you a living in a mansion you can fuck off and draw whatever you like at your own pace

like japanese emperor cliques doing poetry

but when you are an average person with a job and responsibilities - you can't afford that without prior planning

[modest owl]

pray for a creative market

[main sparrow]

I took part in a multi artist art show in Austin Texas a long time ago. I brought my handmade stone inlay work and a friend of mine brought his masterful oil paintings.

Let me tell you. people selling inkjet prints and machine embroidered shirts from goodwill outsold us by a factor of 10.

Should we have shook our fists at the clouds and tell those kids to get off our lawn? Or work harder and out-think those slackers. We did the latter and kept making a living.

AI is just a new disruption. Plenty of people are going to get run over by the AI train that can't be stopped, just like people got crushed by the last disruption.

[modest owl]

you should aim for more premium, handcrafted, artisonal direction

because that's what contrasts with mainstream machined goods

and adds to the price

[main sparrow]

One mistake that some people who fail (at any business, not just art) make, is trying to compete in price with the lowest price in the market. When I was at my peak labor-wise for stone inlay, it was impossible to compete with folks sending over ungrouted mosaics from Iran and India, glued to plastic meshes ready to be installed, grouted, and sealed. So I didn't compete. Value your time appropriately, and convey to everyone that that's what they're paying for. Plenty of folks want to buy local, and feel good doing it, even if it's more expensive.

[ruby lake]

I always tell my wife who has been working on starting her own business, dont charge what is competitive, charge what makes it worth it for you

It can also help you understand what's not worth your time

Where it would cost way too much, no one would buy it, so she stopped making it

[fervent thunder]

i personally think artists should recognise they've been rendered mostly obsolete with AI art and should switch to the new medium instead

but i came not for arguments but for questions, as I am too technologically illiterate for this stable diffusion thing

if i want artstyles and other things like that, should I use stable diffusion version 1.4 or 2.1?

i found out about stable diffusion through some guy on twitter making anime women with it and wanted to do the same

[main sparrow]

It was hard to leave 1.x behind, but I really do think 2.x is better in the long run for my purposes. If you're fine with smaller images, 1.x will do you fine. the 768 base of 2.x is too good for me to pass up, and there's still finetuning and embeddings.

[fervent thunder]

i heard something about textural inversion or something

idk anything about ai art other than it seems cool

i also heard there's like an upscaler for images that uses AI

also what's a 768 base?

[modest owl]

and why is that? Cause the tools are not dumbed down enough

[pallid dune]

Its good when you know how to use it, otherwise it's nightmare fuel and yes there are upscalers

[modest owl]

demand that from the devs

[fervent thunder]

lol

nah i'm fine

[modest owl]

you might be

other noobs aren't

[fervent thunder]

is it like dall-e or

[pallid dune]

Yeah its like dall e

[fervent thunder]

i meant in terms of quality

[modest owl]

but less anal

[pallid dune]

No useless you type a lot of words

[lofty parcel]

Slightly off-topic but isn't the Lensa app literally just Dreambooth?

[fervent thunder]

?

also what are the different "models" for?

[pallid dune]

Dreambooth but it collects ur data

[lofty parcel]

And they made half a million in one day...
I guess I'm trying too hard

[pallid dune]

im bad at explaining things but stable diffusion is like a Babel library of images and in this Babel librairy there is a lot of ugly images so to not see thoses images you must specify the quality of your image by adding words to the prompt like "8k,photorealistic,beautifull"

[fervent thunder]

ok

[pallid dune]

And uh adding the things you dont want to see in the "negative prompt" in some apps that uh have stable diffusion like "ugly, mutated, too much fingers"

YouTube tutorials explain these concepts better than me tho

[fervent thunder]

i understand negative prompts

[uneven heath]

latent space

mostly full of noise

latent universe

[fervent thunder]

so i have a guide for 1.4

and it recommends i install a model if i want something i like

will a model for 1.4 work with 2.1?

[jovial tinsel]

Yeah, you got it!

[jovial tinsel]

Is it really infinite?

2GB barely

or 7

I think its a subset of the library of babel

[elder sleet]

I cant get stable diffusion to work with a rx 580 4gb GPU 😦

any help?

[deft ingot]

is training/finetuning 2.1 a thing?

[peak sphinx]

work on google collab

[sudden marlin]

I'd definitely recommend the #🤝｜tech-support channel 🙂 they'll be a great help for you

[peak sphinx]

Btw can anyone provide a link for SD 1.5 Collab?

[elder sleet]

thank you 🙂

[elder sleet]

is it free?

[peak sphinx]

it's free but ther's some work to do which I don't know how

[fervent thunder]

I have trained two different models, with two specific people in dreambooth. Is there a way to merge them so I can generate both people in the same image?

[signal nest]

https://www.creativebloq.com/news/artstation-homepage?utm_source=twitter.com&utm_content=imaginefx&utm_medium=social&utm_campaign=socialflow backlash against generative art. I've already heard of artists trying to confuse public data with nonsense captions. I think everyone interested in generative art needs to consider directly constributing to clean public datasets. I've been helping someone out with such a project little and often for some years now

we used some of our own photos but also a lot of CC0 images. CC0 is fine! However , I am worried that some people posting those will rebel if they get this idea that all online data threatens jobs. I'm the otherway around, I see the efficiency of the opensource approach. But I'm a programmer. Most regular people dont think like us.

some persons open dataset project: https://imagemonkey.io/ it's a bit like "labelme" but I liked his UI. some sample data visualsed in this thread (human figure annotaitons and street scenes) https://github.com/ImageMonkey/imagemonkey-core/issues/300 a question I have is there somewhere this data could be moved and translated such that it can contribute to the bigger models being trained. and would anyone else like to help. are there any more popular services for doing this. I helped this person because he was being proactive, and i always had image generators in mind as the eventual usecase. I killed a fair amount of time during the lockdowns annotating there, lmao.

[silent torrent]

is it me or SD2.1 is even worse than SD2.0 ?

I cannot make it work

where I could at least get some decent images with SD2.0

I try many prompt for photo realism

it is blurry and unrealistic

even with strong negative prompt

[fervent thunder]

No, Sd 2.x is trash.

[silent torrent]

yeah it looks like

[fervent thunder]

They were too worried about weeding out the weebs to make a solid product.

Im on 1.5 for the foreseeable future.

[strange hemlock]

Yes. I just want to try out this ai but I can't and I don't understand why

[fervent thunder]

Why cant you?

[silent torrent]

I try SD2.1 512...

the images are better

but not great

for photo realism

it is blurry and grainy

[fervent thunder]

I recommend this: https://github.com/AUTOMATIC1111/stable-diffusion-webui

For anything thats photorealistic.

Basically, in 2.x the SD org decided that anyone that generates nudes or seductive things using SD are trash of the world. They spent all their collective brain power trying to do the lords work of preventing AI titties rather than making a better product.

As such, it seems that almost all "photo realistic" generation is in the shitter.

[ruby lake]

Oh hey, does the 2.0/2.1 model run on 8GB?

[fervent thunder]

Yes

it runs on just about anything

May not be fast...

[ruby lake]

With a 700x + resolution i wasn't sure

[fervent thunder]

or desirable.

I had SD running on a phone the other day.

[ruby lake]

Because I couldn't do over 650x650 ish on my 3070Ti

[fervent thunder]

It wasn't great, but it worked.

[ruby lake]

With 1.5

[fervent thunder]

Yeah, havent tried on 2.x

to be fair.

Again, I wont use 2.x

[fervent thunder]

This message is a LOT to unpack... But TLDR, artists that boycott AI are old fashioned and small minded.

[strange hemlock]

No matter what I'm asking the bot to make it says i want a forbidden thing. May be I have to pay money for it to work, but I can't find the place where my dreamtokens/coins/whatever balance is displayed and I can't find the payment thing to increase this balance. I am using discord with Chrome mobile browser.

[vital osprey]

yo

where is meme channel?

[signal nest]

doesn't matter what the counterarguments are - fact is we're up against an opposing faction with a conflict of interest . Precedent exists for locking tech & AI down - see China - we must be proactive to avoid that happening here. If you talk to them for 5mins you'll know there's no convincing them so long as Stable Diffusion is trained on scraped art, and even than it's most likely justification for the bigger issue of fearing job threat

[dawn yarrow]

what are you trying to generate

dreambot is free on here btw

[strange hemlock]

Sunrise over ocean of emeralds, view from space

[strange hemlock]

I guess I just don't know the command for the bot, what is it?

[cursive veldt]

What is the naughty thing in this prompt "blank writing paper, pastel colors, rococo style" ?

[fervent thunder]

Is there a generator for 3d models, like characters or environments?

[primal marten]

Rococo style could have references to naughty things. Rococo paintings by artists like François Boucher often displayed heavy nudity

[white mountain]

I can easily do higher than that resolution on worse hardware. If you're not using xformers already, then do

[fervent thunder]

Tech should be regulated. Just ask customers of FTX.

[primal marten]

What are xformers?

[fervent thunder]

Its a python module, that if used, can increase the speed of generation.

[primal marten]

Ohh

Nice

[white mountain]

It's an optimized library that both speeds up and reduces vram usage

[north stirrup]

I got about a 3x speedup from enabling xformers, plus I needed it for 2.1 to work

[white mountain]

I can get some pretty high resolutions even on 6gb, and you have I think 8gb

Everyone should be using xformers

[north stirrup]

What's "pretty high"?

[fervent thunder]

I need to get xformers to owrk.

Ive tried a few times, but havent had time to sit down and just debug the isntall failure

[white mountain]

I don't know the limit but at least double what they said they're getting

[fervent thunder]

Actually, nows a good time.

[north stirrup]

OK. I'm just remembering getting errors very early on that said my limit was probably 1728x1728 or something like that (3080 Ti 12GB) and wondering if I could go higher now that I've got xformers

I do a lot of generation at 1600x960

[white mountain]

Might be able to, I don't know

Increasing both dimensions does increase pixel count quadratically, so maybe not much more

Give it a try

[fervent thunder]

I need to learn how to activate my venv in windows

Im such a Windows noob.

Ah, looks like scripts rather than bin.

easy enough.

[signal nest]

finance should be regulated, not tech

[fervent thunder]

Both should be.

All things need balance.

Well, actually... I'd rephrase to say... the usage of tech should be regulated, not tech itself.

[north stirrup]

welp, 2400x2400 works, if I'm OK with ~3.4s/it

[primal marten]

Nearly all tech is regulated one way or another, either directly or indirectly, I don't see why AI art and AI in general would be excluded from that

[north stirrup]

or wait, it finished and it's destabilized anyway

Looks like 1920x1088 would work, but it might make it harder to get good images

[fervent thunder]

I used to be that "crypto should be fully unregulated" guy. My views have drastically changed on that seeing how stupid and toxic it has become. It's a steaming shit pile.

AI Image generation is kinda the same. My initial knee jerk reaction was "AI generation should be fully unregulated, let it evolve", but I see it's pretty quickly deteriorating. I was kind of annoyed that they nerfed nudity in SD 2.0. I learned to train models using my wife and sister in law, with their consent. But I also don't share my work. It was just to keep shit interesting and motivate myself to learn. Having said that, I see a lot of toxicity of people doing terrible things without consent. So I'm willing to "take one for the team".

[signal nest]

crpyto is finance. people considered it an investment, or money (..lol) that should certainly be regulated.

[fervent thunder]

Any tech that can cause irreparable damage to someone else, when used recklessly, should be regulated. if image/video generation isn't that, I dont know what is.

I would argue that it should be regulated over crypto. Those who get fucked up in crypto are willing participants, for better or worse. There are victims of AI image/photo generation that didnt get the choice to be participants.

[north stirrup]

Crypto ought to be regulated where it's used for things like money laundering and such at minimum—there is always spillover when it's money-related

[fervent thunder]

Yep, 100%

[north stirrup]

I might prefer that crypto die in a fire, but the cat's out of the bag until/unless the crypto markets crash

[fervent thunder]

Yeah, they won't go much of anywhere.

But they are kinda hamstrung by other factors like China's regulation.

[north stirrup]

We've got an opportunity to police ourselves before regulation comes into play

[fervent thunder]

Yep, it sounds good in theory.

[north stirrup]

I'm not sanguine that others will actually follow that idea, because most people are dumb as rocks

[fervent thunder]

Regrettably that would require 100% participation.

[fervent thunder]

This.

One can still be a deviant and be responsible about it. lol

[north stirrup]

I can at least have my hands clean, I suppose

I keep trying to think about what the "simple rules" might be for avoiding harm to those that diffusers can harm

To start, I avoid prompting with artists unless either a) they're long dead or b) I'm wildly genre-shifting away from their oeuvre

I also avoid using real people or copyrighted/trademarked characters, but … is there a place for AI "fan art" in the grey area?

[fervent thunder]

But I think a lot of those are already covered by law.

Law already allows for use of original work so long as there are significant, notable changes that provides artistic value and qualities.

[north stirrup]

True! Copyright law covers a lot already (just many people ignore that part, to our detriment)

[fervent thunder]

Yeah, im with you.

[north stirrup]

Right, the use has to be "transformative"

[pallid dune]

I generate images with the use of artists prompts when its for myself

[fervent thunder]

And I don't think thats an issue.

[north stirrup]

Right, I've done that too; I think the line needs to be drawn at sharing those

the instant it's not purely "personal use" it's a lot ethically hairier

and legally

[royal pumice]

Any way to generate a greenscreen background for images?

[bleak matrix]

Well, what kind of image are you trying to create?

[modest owl]

I once had some luck with invisible

But it generated an alpha/black bg

[fervent thunder]

I see using an artist name to make ai art in the same way as if you were to try and explain to an artist what style you would want them to paint the picture to you :3

[royal pumice]

hm I could work with that aolko

[modest owl]

Also try transparent

[royal pumice]

i'm no prompt wizard. I feel like inpainting should theoretically be better than a greenscreen option but that's sort of black magic as far as i'm concerned at this point

[fervent thunder]

But again, thats already well covered. If you happen to be a world class artist, and you illustrate a nude of a celeb, its really the same. If you share it, there may be consequences.

[north stirrup]

A big part of what needs to happen is spreading norms around the various communities that have sprung up around diffusion
It's literally impossible to outright prevent people doing the bad stuff, but policing ourselves could mitigate the problem

[fervent thunder]

I agree it will help for sure.

I dont think it will stop the inevitable; however, it may slow it, and it will provide a large group of reasonable people to help advocate for moderate and reasonable regulation rather than outright bans. Everything that can be used for mass destruction in America will be regulated eventually… Well, except guns.

Solution! Lets start calling "diffusion" the term "guns" and then no one will regulate it!

[royal pumice]

Oh interesting, the depthmap plugin for webui has a "contrast cut level" that looks like it would give me an adjustable background mask I could work with. Not technically a green screen but close enough I think.

[fervent thunder]

How are you creating your model?

[royal pumice]

All my attempts at training do the same thing, Logie. I think it's because my training set is mostly head shots

I can get a hell of a good version of my head/face but it goes to hell if I try to include it with anything else

[frank moon]

Dreambooth locally on automatic1111.

[fervent thunder]

Got it.

[tribal oar]

Not sure if anyone here uses wildcards, but I've got a collection going https://huggingface.co/spaces/Redw04/wildcards/tree/main

I just uploaded about 20 different ones for architecture, and cameras, and lighting, etc

Feel free to take whatever you want, and contribute your own if you have some to share!

[unreal cargo]

Awesome, thank you so much!

[robust coral]

I use wildcards a lot. thanks for sharing them!

[unreal cargo]

I have a bunch, but I hesitate simply because I want them to be attributed to the person that made them and I totally forgot/don't have the link

[tribal oar]

Eh, we're a community, I think these things are meant to share

[robust coral]

yeah created some of my own as well. but very specific. like european animals you would find in a forest etc

[tribal oar]

What's food for the goose is good for the gander, etc

[unreal cargo]

is there a way to wildcard the wildcards?

[manic canyon]

Can I render 2048x2048 on 2070? It says CUDA out of memory, some fragmentation bs

[unreal cargo]

yeah you should be able to

[tribal oar]

Yes, there is haha

You can nest wildcards

[unreal cargo]

I'll send a google link drive, I found the link

[manic canyon]

Hmm I think my vram is not being freed after renders, can I force it somehow?

[robust coral]

dynamic prompts is a lot of fun 🙂

[tribal oar]

It is honestly the only way I prompt now hahaha

Right now I'm running "analog style, {an amusement park|a Ferris wheel|A rollercoaster} {at night | }, __camera__, __coloring__, __lighting__, {2-3$$__camera_mods__}"

[unreal cargo]

ahh okay, I'll do that

[tribal oar]

I love setting a script and then walking away

[unreal cargo]

for nesting is it just like __foldername/__?

[tribal oar]

Anything in the /wildcards folder will be read, even things in sub folders, so that's not necessary

You can just call __wildcard__ in whatever .txt you're currently in

[unreal cargo]

Here's the wildcard folder:
https://drive.google.com/drive/folders/1Vz5wwrXuhuUzfwWzzQ9B4MZmbeRLN-CM?usp=sharing

ooo cool!

There's 210 ish

also no NSFW pictures but there is text that is less than SFW in some of the txt files

so

heads up

[robust coral]

yeah - same here. also putting wordlists into {alternating} brackets etc... I have build some templates for portraits and other scenes. what I use a lot is portrait at __time_of_day__ from __countries__ during __seasons__ also like adding __eras__ and if something interesting comes up, I start exploring that location in the latent space

[tribal oar]

ERAS IS SUCH A GOOD IDEA

WHY DIDN'T I THINK OF THAT

[tribal oar]

Right now my photography superscript for subjects is ((subject name)), {2-3$$__subject_mods__}, ((__camera__)), {2-3$$__camera_mods__}, ((__new_locations__)), __season__, __lighting__, __coloring__

[robust coral]

1700 futuristic london during winter 😄

[tribal oar]

My biggest hurdle was figuring out what should be a camera mod, and what should be a style mod, and what should be a subject mod

[robust coral]

nice!

[tribal oar]

Subject mod is things like "highly detailed face, textured skin, etc"

Camera mods are like "rule of thirds, good composition, leading lines" etc

[robust coral]

yeah. I'm still sorting lenses, perspectives, focal length etc.

[tribal oar]

Wait a minute

[tribal oar]

These are all from my huggingface 😛

Did you just put them in a G drive link to make sharing easier?

I was like "Oh wait a minute I recognize these!" 😛

[unreal cargo]

Yeah, I grabbed yours and popped them in there, I didn't overwrite any of the ones I had already

simply because I've added more info to some of them

we probably pulled from some of the same sources too

[tribal oar]

Oh probably

I stole all the "dan" ones from a guy on the unstable diffusion discord

He might be in this discord too I'm not sure

BTW do not check out that discord...not worth it

[frank moon]

What are wildcards? and what is dynamic prompting?

[tribal oar]

Ethically that discord is a dumpster fire

[tribal oar]

Are you running locally?

[frank moon]

Yes

[tribal oar]

Dynamic prompting lets you use wildcards, the text files we're sharing. What it does is you call __wilcard__ in your prompt and then it will randomly choose one of the lines in the document

You can have them for whatever you want

I have a list of cameras, a list of telescopes, clothing, professions, countries, etc

[robust coral]

and you may find combinations that work quite good that you may have never tried. very helpful and creative

[unreal cargo]

Also nice to just batch like 400 pictures and get a coffee and come back to a large collection to sift through

[tribal oar]

That's precisely why I do it hahaha

[unreal cargo]

good for inspriation for new project ideas

[tribal oar]

Well, originally, the idea was to model my own face, and then take "me" on a photoshoot around the world

I've had..I'd say a medium amount of success with it

Oh that reminds me I need to try LORA today

[unreal cargo]

was that updated on the dreambooth extension? or seperate repo?

[frank moon]

How do you use the text files? Drop them into the scripts "prompts from file"?

[tribal oar]

Are you running 1111?

[tribal oar]

https://huggingface.co/spaces/ysharma/Low-rank-Adaptation

This is what I was going to use

Haven't really read too much into it

But results have been promising

[frank moon]

Yeah

[tribal oar]

*results from reddit

[tribal oar]

Check the images channel for a second

Once you have the extension installed, check out this link https://github.com/adieyal/sd-dynamic-prompts

It will explain the syntax for your prompts

And then check the google drive that @unreal cargo posted above and grab whatever looks appealing to you 🙂

Mix and match, and have fun!

[unreal cargo]

I've read that a bunch but I must have missed the sytanx for nested folders

[tribal oar]

Let me give you an example

"a {small|large|dwarf|supergiant|binary} {__color__|colonized } moon"

[unreal cargo]

oh nvm that's a different link

thank you

[tribal oar]

😄

So this will choose one item in the brackets randomly

[unreal cargo]

erm

[tribal oar]

And then if it rolls __color__ it will then choose a line from my color wildcard

[unreal cargo]

If I want something random from a folder, not just a specific keyword. just simply __foldername__ would work?

[tribal oar]

That's a good question

I don't know

But you can do {__wildcard1__|__wildcard2__|etc}

[unreal cargo]

is quite the large hammer

but yosh, that would work

[tribal oar]

Oh shit

I didn't notice just how many of your own you added to this

Thank you @unreal cargo

[iron basin]

What is this server about?

[tribal oar]

We're avid enthusiasts of the TV show "lost"

About 95% of the posts here are fanfiction episodes of the TV show "lost"

JJ Abrams keeps trying to shut us down

[inland pewter]

Whats the best UI for running local SD2 now? Been away a few months used to use AUTOMATICs

[tribal oar]

I still use auto

Works good for me

[deft ingot]

does AUTOMATICs support training v2.1?

[inland pewter]

Can you feed it the new 2.1 model? @tribal oar

[deft ingot]

ya

[tribal oar]

Yes

But I can't get custom models to work

Like I trained a model on 2.1, on my own face

And it works online

But locally my prompt isn't working, it refuses to make my face

[robust coral]

have you added the 2.1 yaml to your custom model?

[tribal oar]

Yep, and it's x768 so I even made sure I'm running V yaml

It works, like it can make pictures

But I say "a picture of redw04" and it doesn't do it

Super weird

[robust coral]

hmmm

[deft ingot]

wait was that yes @ me @tribal oar ?

[tribal oar]

And the same model hosted on huggingface, using their web model, works 100% fine

Makes my face perfectly

Oh, sorry missed yours, yes

trains on 2.1

And LORA

[deft ingot]

whats LORA

[tribal oar]

Low Rank Adaptation

[robust coral]

only saw the repo but since I'm still not that much into training (yet), I haven't really researched it

[tribal oar]

https://huggingface.co/spaces/ysharma/Low-rank-Adaptation

I don't really know how it works , but I have seen results that look very promising

[robust coral]

so where does lora fit in compared to the other methods for you?

ah ok I see

[tribal oar]

I'm actually testing it right now

Wait it finished in like 5 seconds

That cannot be right

It gave me a .pt file

What the hell do I do with that?!

[unreal cargo]

I think you have to use one of the scripts to merge it?

https://github.com/cloneofsimo/lora#loading-merging-and-interpolating-trained-loras-with-clis

[tribal oar]

Thank you 🙂 Fuck that is so much work haha

So, if I understand it properly

You train a LORA model

And then you can fuse that model with anything else

[unreal cargo]

would be interesting to see if it can be used as a hypernetwork though I know that hypernetworks will activate off of embeddings pts, but they do very little so

yeah, according to that repo

[tribal oar]

As per reddit:

Lora DB is based on diffusers, so you first set up a diffusers environment, and then you add Lora into it. With conca, this looks like that for windows/python 3.10

call conda create -y --prefix venv python=3.10

call activate venv/

call pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116

call pip install git+https://github.com/huggingface/diffusers.git transformers accelerate scipy ftfy

call pip install git+https://github.com/cloneofsimo/lora.git

once this is done, you can activate that venv and Lora will work into it, using calls very similar to Shivam or other DB tools. You'll need to add 8bitsand for that option to work, this is just like classic DB.

And to use the results, since they work on diffusers you need to convert to something auto can use (unless he added lora support and I didn't hear). Lora has a script to merge Lora pt file into a unet, and then you can make a ckpt out of those diffusers. But I'm sure there is a simpler way of doing it

[unreal cargo]

$ lora_add --path_1 stabilityai/stable-diffusion-2-base --path_2 lora_illust.pt --mode upl-ckpt-v2 --alpha 1.2 --output_path merged_model.ckpt

seems easy enough

[tribal oar]

I'm new to coding

I like pretty GUIs

Give me a CLI and I start to get overwhelmed haha

[frank moon]

When I create a model using about 15 pictures and then render something with a prompt I just get the face. Even if in the prompt it reads something like outside or at a lake. If I remove my model I get renderings with the landscape.

What do I have to do to get my model to render what I ask? Do I need to add photos of outside?

This is with dreambooth (locally)

[tribal oar]

That sounds like overcooking

In the oven too long, basically you pulled the base model too far in one direction and now all it wants to do is re-create the pics you used

Solution is fewer training steps

And a lower learning rate

[tribal oar]

I cloned the space to my own HF repo and rented a card to train on, trying again

The examples say 5-10 pictures

I used 140

Maybe too much? We'll see

So @unreal cargo as long as I have LORA cloned locally, I can use that command to merge two models?

I want to merge this with analog style, and 2.1, and see how those come out

[unreal cargo]

appears to be so!

haven't tried it yet though

[tribal oar]

Well I guess we're going to find out haha

[unreal cargo]

[tribal oar]

I wonder how long training will take

[unreal cargo]

probably the same as it takes to convert diffusers into a ckpt

[serene swan]

that emote looks like nina kosaka lol

[tribal oar]

I already feel like this isn't going to work out well

At least it's only like 2$ to train

[hallow knot]

Hey, any Artstation war thread?

[tribal oar]

?

[fervent thunder]

artstation don't understand how the art ai work, so they're doing something someone who don't know how the art ai works on their site :P

[tribal oar]

So they're banning AI art?

Whatever, there's a million sites to post your work to

[fervent thunder]

yeah, and they are only hurting themselves as the ai has already trained on their stuff :P

[tribal oar]

Well, some sites need to be for real artists

I totally understand some sites wanting to draw a line, and that's cool

[fervent thunder]

pun intended?

[tribal oar]

Artists work really hard to learn their craft, they'll want a safe space to post their work

Omg no lol

But that's awesome

[fervent thunder]

how do you know they worked hard?

[tribal oar]

Because I'm friends with a lot of artists

And they've worked for like 10 years now to hone their craft

I'm on the AI side of "AI art isn't art", I think the debate is silly

But I understand why real artists want spaces where they don't have to "compete"

[fervent thunder]

sounds like any job out there really, and most artists I've seen don't have much variety on their style ;P

[tribal oar]

Yes that's because they've worked hard to /create/ that style

[fervent thunder]

like the rest of us, with everything we like to get better at :3

[tribal oar]

Personally

I think it's pretty uncouth to be derogatory towards real artists, just because we can make stunning art for free now doesn't mean they don't have value in society, and we don't just get to write off all their training and experience

Artists are massively important, as are their contributions to their culture

[fervent thunder]

yes, we are :D

[tribal oar]

and this is why I think we need a new term for what we do

I love AI art

But we are not artists

If anything, the computer is

[fervent thunder]

what makes someone an artist to you?

[north stirrup]

We can be artists with diffusion, but the technique belongs to the computer, and not all “AI art” is art

[tribal oar]

Someone who applies their knowledge and skill in their specific field to contribute a specific piece with an end goal in mind, to contribute to their society

[north stirrup]

There’s a lot of shitty diffusion out there

[tribal oar]

I agree with nihl, txt2img CAN be art

But the ability to put words into a text box and have your computer produce something for you in that style, does not make us artists

I saw a funny comic yesterday where some dude gets his spotify wrapped email, and it says "we've curated a playlist for you based on all of your tastes for the year" and the dude goes "wow I'm a musician!"

[fervent thunder]

I think someone's an artist, if they say they are, or if someone else say they are :3

[tribal oar]

This debate has literally been going on for centuries

And the obfuscation that AI art introduces is /not/ helping

But no, I do not consider myself an artist just because I can make pretty pictures

I think that's a misrepresentation of what I actually do, which is much closer to programming than art

[fervent thunder]

Too, bad you made some nice stuff. I like being an artist myself either way :D

[tribal oar]

I just think it feels like a shortcut

[north stirrup]

I think art requires creative intent; that’s weak at best for lazy prompting

But some people make amazing things with diffusion as a tool

[tribal oar]

And saying "wow I'm an artist!" when you haven't really done anything except enter text into an image, allows you to fool yourself into thinking you don't need to grow and learn a skill

[tribal oar]

Yeah, exactly, it's a tool

[fervent thunder]

it's good to know I'm not one of those people you're talking to then :D

[tribal oar]

I'm an architect, and I am trying to learn how to use SD to render buildings I create in different styles, so I can then pick and choose what I like and add those to my model

In that situation, I consider SD to be a tool, like a paint brush.

[south thorn]

I think focusing on the term "art" at all is a disservice. It's basically anything.

[tribal oar]

I am confident that society will adapt and create new terms to distinguish

"dreaming" is catching on

I could see us being called "dreamers" or something

[south thorn]

Maybe not lol. In the 80s an RPG was a role playing game. Nowadays it means level ups and fight monsters... Sometimes language just gets dumber

[tribal oar]

Well, it evolves, dumber I don't know about

[south thorn]

I agree we need more words though...

[tribal oar]

You know this is actually kind of a good example

An RPG still describes a role playing game

What a "roleplaying game" is has shifted

[south thorn]

It does... But only to old farts like me 😭 haha

[tribal oar]

And now the 80s definition of "RPG" has become "tabletop games"

[south thorn]

Role playing itself is just pretending to be in character as someone other than yourself

[tribal oar]

Role playing is when I pretend I'm a 17th century nobleman and my girlfriend is a peasant who I have taken pity on

[south thorn]

I think I just get in a twist because I keep reading "ai isn't art!" all over. I could go poop on a sidewalk, stick a dandelion in it, and even that would qualify as art (tastelessness being another matter entirely haha)

I remember reading a story about a guy who hung a bunch of goldfish in IV bags from the ceiling of a museum. Animal rights people managed to save a few before they all died. The artist felt his work was ruined

[deft ingot]

@tribal oar do you have a .yaml for training v2?

[reef portal]

Guys what do you think about this I need real feedback philosophical, ethical, in light of artstation ban, in light of reddit hating AI more than the unabomber hated people etc.

https://www.galleryofai.com

[hollow lichen]

Hello friends. I want to try to run stable diffusion locally. I have a 12gb gpu. I'm pretty new to this can anyone point me in the direction of a good guide to get me started?

[north stirrup]

https://github.com/AUTOMATIC1111/stable-diffusion-webui is what many are using. and the linked page includes installation instructions

[near knot]

Is there a website/timeline of all the AI stuff that is coming out? Seeing all this AI stuff pop up towards the end of 2022 makes me think there was a new tech breakthrough and wanted to learn more about it. Like what made stable diffusion/GPT/etc.. possible that wasn’t possible in 2018?

[hollow lichen]

Wonderful thank you 🙏

[noble lily]

Anyone having prtoblems with the bot adding "number"? It keeps adding the word "Number" to the beginning and it doesnt register as actually the image number function

[tribal oar]

https://github.com/Stability-AI/stablediffusion/blob/main/configs/stable-diffusion/v2-inference.yaml

[deft ingot]

thats for running inference though? don't i need the non-ema pruned model for training?

[tribal oar]

Oh

I'm really not too sure

[frank moon]

Really? I was checking some of the posts online regarding learning rate and steps and it seems mine are much lower.

I did 15 pictures (1500 steps) at the default learning rate. the ones online where like 9000 steps.

Would them all being photos from the shoulders up be an issue?

[chrome dagger]

does anyone know if there's a way to input 2 images in img2img or some script, and have it create some sort of amalgamation of the 2 ?

[tribal oar]

I'm honestly not too sure, I'm still figuring all this out as I go

#🔧｜finetune would have more answers

[broken gull]

RPGs still means roleplaying, just nowadays game companies don't put a lot of their budget in the story ig. But HL is RPG openworld. If you love RPG and Wizarding world then I am sure you gonna love it.

And yes level ups and fighting monsters is a part of it

Also character customization

Etc etc

[fallow remnant]

I remember seeing a tool on Twitter to generate repeatable patterns (like these: https://www.vecteezy.com/free-vector/print-pattern). Does anyone know if that still exists and what it's called? Can't find it anymore.

[inland pewter]

help, running local instance of SD using AUTOMATIC and v2.1cpkt, all images coming out as brown gradient square

is v2-1_512-nonema-pruned.ckpt [d1672608] busted?

should I have done the pruned one?

[winter pike]

youre using the wrong yaml file

[fathom iris]

I'm trying to make these diagramatic botanical illustrations in SD and it just seems impossible

https://prompts.aituts.com/19th-century-botanical-illustration/

SD doesn't want to generate things within the border of the image, the plants always run off the page

[peak tinsel]

Yo can I download stable diffussion instead of using the credit based dream version in the browser?

[cobalt cliff]

so many people whining about copyright laws, wtf let's cancel sampling and hip hop now then

[white mountain]

Transformers and scaling, both of which turned out shockingly powerful

[fathom iris]

I don't know what GUI's other people recommend, but I've been using NMKD's GUI and have recently started using InvokeAi

[weary yacht]

I have a really dumb question.

Why is midjourney better?

[timber raft]

what idiot said midjoutney is better?

[tight quartz]

be nice, friends

[timber raft]

sorry, I haven't sleep, my brain is pretty stress right now

[winter pike]

midjourney its a product and was made to sell as much as possible, so makes sense that they care about the quality of their model

[fathom iris]

I mean they have a point

You don't need as detailed a prompt, and it usually produces some very aesthetic outputs

[tight quartz]

They didn't specify better than "what?"

better than myspace?

Midjourney > msyspace

It's also subjective. Like "why is bass guitar better?"

um....

[timber raft]

the first negative when I use Midjourney is they ask me use Stripe to pay

[weary yacht]

Hence the "Dumb question" part.

[timber raft]

and Stripe not support in my country

I drop

[weary yacht]

Midjourney seems like it produces better quality output for the prompts.

[fathom iris]

I guess that's the question then. Why can Midjourney get away with charging people? It must be a product people value over free alternatives

[weary yacht]

Yeah, but isn't it using the exact same software? That's sort of the reason for my question

[winter pike]

the model is different

[tight quartz]

AH! So the question is "Why do I like the outputs from midjourney better?"

[weary yacht]

Sure? That works

I'm new to this AI art generation. But people are willing to pay for midjourney.

[timber raft]

Idk, depend on taste?

[weary yacht]

So I'm just curious what they're doing different. On some basic level

[peak tinsel]

can u link me InvokeAi

please :=

🙂

[tight quartz]

The TL;DR - picking only beautiful artsy images as your collection of data will yield a biased result.

[weary yacht]

Ah, thanks. That helps.

[timber raft]

when someone tell me they are making professional concept art with only AI on Artstation right now, I ask their portfolio, they silent, I laugh

[fathom iris]

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

[weary yacht]

It seems like all this stuff is super cool and interesting, but far from generating useful stuff

[fathom iris]

Jump into #🏞｜general-with-images

And you can see whether it's generating "useful stuff" - weird phrase, is art meant to be "useful"?

[fervent thunder]

is SD 1.5 good or should i skip straight to SD 2.0? i've been using 1.4

[timber raft]

or can be use in the industry, it's just for hobbyist right now

[fathom iris]

I mean sure, but I personally know people that are using it on a professional level

[winter pike]

im personally enjoying sd 2.1, sd 1.5 is also very good and an improvement of 1.4

[warm canyon]

Hello I'm tranna join up the talk pardon

[timber raft]

for commercial Art, honestly no, it is the future of course but still very far from it

[fathom iris]

What do you mean by commercial art? Art that people pay for? Again, I know multiple professionals that are using SD and selling it in different forms. One is using MJ and selling biblical artworks, one is using SD and making a board game entirely with SD art

We all know AI art has won art competitions where the judges didn't know it was AI

[weary yacht]

If you, for example wanted to make a star wars poster in a specific style for a specific movie, I don't see the AI doing a great job. It'd probably come up with some very star wars looking characters, but would you be able to make an 80s style rogue one poster?

[fathom iris]

Lmao

[timber raft]

Most people talking about SD or Mid will in the Industry soon have no clue what production art actually looks like and what the purpose of it is. They think you get hired as a concept artist and you are going to paint portraits of anime girls or spectacular dynamic poses rendered with background just like the ones they churn out with AI, all of this is far from true.

[fathom iris]

You obviously haven't seen many artworks

[weary yacht]

AI seems like it struggles really hard with specific characters

[timber raft]

Zoom into any of your AI character paintings and tell me what the materials are, how they are connected together, you will find most of the time its a bunch of noise that needs an artist to be able to explain.

Lots of shapes the AI makes look really good, but if you actually had to 3d model or create that character someone is going to come ask you how the hell that 5 layered piece of cloth around her neck folds and why is it sparkling and then you will actually need a real concept artist that has a design education allowing him to understand what is working in the AI piece and make it look intentional.

Even for illustration though, AI (at the moment) is only able to match the first glance of a real illustrator piece (example Riot key art) because it has learned contrast and shape from great artist and replicates it in its pieces.

[fathom iris]

I've tagged you guys in some examples that I just generated today (not the best I've made) in #🏞｜general-with-images

Happy to chat in voice too

[shut saddle]

hey! does anyone know if there is a photoshop plugin that allows me to run SD locally? I’d like to have it running on my laptop, and I’m hoping to use it when I’m not connected to the internet. please @ me with responses if you’d be so kind, thanks all and happy holidays!

[fathom iris]

https://www.reddit.com/gallery/z6g49d

[timber raft]

what's good in the AI is a nice enabler for non-artists to express some kind of visual idea they have.

[golden citrus]

https://twitter.com/i/spaces/1YpJkgjjPzyJj

Debate on ai art happening now !

[timber raft]

For a real commercial workflow right now it is basically useless, because it need a lot of intention to detail

[fathom iris]

I'm confused about your critique. AI art currently does not replace the human curator if that's what you're wondering

AI art still needs hardware, software, good prompts, and the right eye to select and curate

[uneven wasp]

It’s not useless for commercial but it does take a lot of finesse to get something working

[fathom iris]

Quite simply I don't need to work for hours on canvas or in photoshop in order to make very good artworks that communicate exactly what I'm trying to

[timber raft]

yea, that's what mean, like making visionary projects, if you spend 30 minutes fixing up an AI generated image then I know dozens of artists who would have handed me a sketch in a shorter amount of time

[lilac reef]

What does this mean: Seed resize from: -1x-1

[timber raft]

basically with current commercial arts workflow, you guys still go through hell like normal artists

[fathom iris]

It just doesn't take anywhere near as long and you're fooling yourself if you think it's equivalent

[uneven wasp]

It lowers the difficulty to make something look pretty by a couple orders of magnitude. If you’re dealing with tweaks on something abstract like composition or getting the right feel, automating actual drawing that intersperses decisions doesn’t automate the whole process.

[timber raft]

or dealing with hard clients in my company

what I want right now, is AI in UV map and topology, I don't care about those drawing things anymore, they always say "why it's not clean and symetry enough"

[uneven wasp]

For clients they want to look at an intermediate picture like it’s a finished product it’s gotta be a massive timesaver

[fervent thunder]

Did anyone else get kicked?

[near knot]

I see. I found the github for transformers, any article keywords for searching Scaling?

[white mountain]

"scale is all you need" is the catchphrase

literally just "what if we used more GPU"

bigger model

etc

not sure if GPT-2 was the first of that paradigm

[timber raft]

damn, Artfol just announced banning all AI generated content

same policy with Newground

[wanton rivet]

Can someone tell me if I can give some command to the SD to use more VRAM?

[uneven wasp]

they can ban it all they want, but there's legitimate value in a technology like this and history has shown that it does nothing to prevent the adoption of valuable technology

being a luddite doesn't work

[timber raft]

no, they still allow using AI but only in some parts, same with newground, you have to tell which part is

I don't like only type and click either, doing that for days make me bored

[uneven wasp]

its like vr: extraordinarily fun at first but it wares off quickly

Though the new possibilities of what this allows has been extraordinarily fun

[timber raft]

imagine in the future kids in art class instead of touching a pen they type prompt on pc

[uneven wasp]

Imagine a kid in future English class instead of typing an essay they juts have chatgpt create one

[timber raft]

so boring

I still don't know how to combine using the AI generate to created reference then draw hand base on it either, because my brain want everything must be perfect including my prompt and turn out I was trying a perfect a reference for hours

[uneven wasp]

You photoshop/kitbash stuff together until you have the composition you want then img2img it to look consistent/pretty

[lilac reef]

Question for anyone who knows but automatic1111 on colab doesn't use yaml files so what is it training with? All it says is model.ckpt

[tribal oar]

I am having the exact same issue

Is your token not working for you?

[lilac reef]

it is because the model file on colab is different as it has a different hash

[tribal oar]

Training works great for me, it even generates previews in the colab that look just like my face

But in Auto1111 my token is not working at all

[lilac reef]

I got it to load now by creating a yaml

Because you used a token already known to it.

[tribal oar]

No I used a unique identifier

[lilac reef]

Ran into this with someone else so I had him change the token to something else and it worked fine.

[tribal oar]

I used (Julcto)

[lilac reef]

You think you used a unique

[tribal oar]

Which works with all of my 1.5 models absolutely no problem

[lilac reef]

different clip

[tribal oar]

2.1 trained model of me? Refuses to work.

Hmm

[lilac reef]

CLIP is completely different now. Nothing carried over.

[tribal oar]

Okay, kind of making sense, but then why does the preview in the colab make my face with the same token no problem?

That's where I'm getting tripped up

A photo of ((julcto)) works great in the colab preview

A photo of ((julcto)) returns random stuff like buildings or cars in auto1111

[lilac reef]

same happened to him and it was HS for the token for Harry Stiles. Worked in the preview but locally no dice HS = HighSchool so it was always a bunch of hs kids in the pic.

[tribal oar]

So should I try like a hash instead?

What's the fix here?

[lilac reef]

a new token will solve your issue unless it is known as well

[tribal oar]

That makes sense in his case, but Julcto isn't a word that could be confused with anything else

[lilac reef]

this is why I despise DB cause 3h to find out my token was known so do over until you get one that works

stop fighting me and try it

[keen pewter]

Just got banned from MJ discord, for saying its disappointing that they did a 180 on custom aspect ratios outside of 3:2 or 2:3, after they said a month ago they would come in "a couple weeks". I mentioned how other systems do it just fine, then BOOM, banned. Crazy

[tribal oar]

Oh I'm sorry no I appreciate the help big time

I'm just trying to understand

It makes no sense to me

I'm paying to train my models (renting a card on huggingface) so the more I understand now, the less I have to pay haha

[tribal oar]

Actually I can't access it either right now

Discord tells me it's down

[lilac reef]

Look, please, I understand the frustration as I don't like spending money, nor time to find out the token CLIP knows so fuck me start over. This part of SD is BS but apparently there is no way to know before hand? I find that amazingly short sighted, and pretty stupid if true.

[keen pewter]

Damn lol. The timing made it seem like a ban. I dont even see it in my server list. Thank god 😂

[lilac reef]

I typed Julcto and it knows it

[tribal oar]

Man this is really pushing the limits of my knowledge haha, my brain hurts thinking about this

[lilac reef]

There, I just showed you it already knows.

I mean it is even spelling it out as best as it can.

[honest spear]

guys, is there a way to input a precise color as prompt? Like the RGB code or something

[peak orchid]

So what are people's thoughts on Emad letting artists opt out of being included in the training dataset for SD 3.0? Sounds like another experiment that could easily lead to another broken model like 2.0, or it might work okay

[peak orchid]

MJ goes down literally every day at the moment

[keen pewter]

I think it's dumb. It's just a style (barley)....Nothing actually includes their work.

[lilac reef]

Not precise but in 1.5 blue #00000f red green, etc.... did work but no longer in 2.x

[keen pewter]

If artists complain, to bad so sad. No need to cater to them for such a silly thing

[tribal oar]

I keep saying it, EMAD is a hype man. People are looking to him like he's solely responsible for where we are right now. SD is a company that exists for profit, of course he's going to do something to improve public opinion, even if it means making the model worse

Personally I don't really care, 1.5 is good enough

[honest spear]

oh, well I guess I'll just have to find a creative way or pick the color from another image. Thanks:)

[keen pewter]

NSFW filters, catering to certain crowds, will Gimp this new technology into nothing eventually. I say just make it amazing and forget anyone who complains, especially people that dont even use AI art generation

[little plaza]

I think most artists apparently don't actually want to opt-out given how weak the response to it has been on twitter. They're not gonna be happy even if they're all removed from the data set lol. They obviously just want to be mad.

I am not sure what I expected

[tribal oar]

Wtf is an SKS token?

[peak orchid]

They back down and change things though when they make the model worse, like with the 2.0 fiasco. Part of me wonders if this could be a repeat of that

I welcome experiments, but some experiments also can slow our path to having a better model

[uneven wasp]

These aren't really good faith criticisms that once addressed the people making them will no longer have an issue with ai art, its that many artists see the tech as an existential threat and this is the best current argument against it. Once addressed they'll start looking for a different one. However, the primary target of the appeasement isn't those artists, but everybody else who hasn't made up their mind about it. It doesn't matter if the artists are happy with the solution, just that the opt-out effectively neuters the criticism in the mind of everybody else.

[tribal oar]

I just wonder where we start drawing lines

Like, okay I can /understand/ why an artist wouldn't want a bot to copy his style, from at the very least a financial standpoint I understand and empathize

But where is the grey area?

It's the infinite monkeys and infinite typewriters problem

Eventually I'm going to generate something that looks like another artists' work

Regardless of training data

[little plaza]

It's just frustrating because i've been seeing non-stop people saying that there need to be a way to opt-out and they don't have a problem with the technology otherwise, then they completely ignore me when I point out that option exists now. They don't even know why they're upset.

[uneven wasp]

I can definitely understand the mindset, it's just that I don't agree with it. If any image generated is close enough to conflict with an existing copyright that can be trivially verified. However, I don't think anyone should be able to own a style, so arguments that its stealing their really fall flat on me.

[fervent thunder]

i always find it exciting to point out from time to time that this technology has its origins in medicine, e.g. in the early visual detection of cancer.
every technological development that is triggered by the 'hype' of this technology naturally flows back into the medical sector.

[tribal oar]

This reflects my own opinion

And I personally feel like it's irrelevant anyways

Maybe. /maybe/ SD is able to remove individual artists styles, hypothetically

Tomorrow artist diffusion comes out, trained on the same data set

[winter pike]

im pretty sure thats the default token for dreambooth traning

[tribal oar]

All those styles are right back in

[tribal oar]

Thank you

[keen pewter]

So, i keep seeing people talking about 1.5

Is 2.0 good?

[uneven wasp]

And improvements here will probably improve medical imaging ai by proxy

[tribal oar]

AI-based protein folding is going to be huge in the medicinal world

In fact

[fervent thunder]

this is simply a random combination of letters that they used as an example, resulting in many often receiving weapons in their prompts because SKS is a weapon. 🤣

[tribal oar]

I was just reading about "protein diffusion" earlier today

[uneven wasp]

We've already used it to read minds, which is about as terrifying as it is cool, but it means there's an some level of applicability to the human body thus far

[uneven wasp]

Anything cheaper than nmr will change the world immeasurably

[tribal oar]

My friend is very big into "folding at home"

Imagine if I could use my graphics card in the same way?

[uneven wasp]

20 years down the line, being able to design the genetic code for a new protein based on its purpose has consequences the imagination can only dream of

[tribal oar]

txt2protein

But to conclude my thoughts on the art conversation. I think the genie is out of the bottle and there's no going back now, the only option artists have to not have their works included is to not post it online any more

[uneven wasp]

"An enzyme which inhibits the excess production of X"

[keen pewter]

Im confused. Does 2.0, have artists that are excluded?

[tribal oar]

And even that won't work, archiving exists

Yes, to improve their public image SD is slowly removing things from their training sets

Started with porn (which, turns out you really need for just regular SFW human anatomy) and now I'm hearing 3.0 is removing artists

Select artists?

[fervent thunder]

many artists who are now outraged are students and earn an additional income - if you read studies you have the impression that the job loss caused by AI will take place in the professions with higher education.
That they are threatened in their art is still the smallest evil for them.

[tribal oar]

I was just talking about this earlier actually

I don't quite think governments are ready for the amount of people about to lose their income

There's hundreds of millions of coders on the planet right now

Chat GPT is not perfect, far from it

[keen pewter]

If i have a local install, is there a way around having artists removed?

[tribal oar]

But it can already help me write code, when I never have before

A model trained specifically on writing code is going to wipe out the "code monkey" job

That's a HUGE section of the economy

And when the model releases, companies are going to scramble to adopt them

Millions of people are going to be laid off in very short periods of time

I see this as an inevitability we are not prepared for

[fervent thunder]

just think of the management of businesses, finance, etc.

[uneven wasp]

Job loss only happens on a sector by sector scale with technological advances.

[tribal oar]

I'm an architectural technologist, my job is to take buildings that architects make, and translate them into workable drawings that can be submitted to the city

[winter pike]

yeah i saw that, then they recommended to use something like >[]¨#$[¨ as a prompt but no way i can remember that 💀

[tribal oar]

The second Autodesk can create a model that does that automatically based on revit input? I'm gone

[tribal oar]

Yeah, but that's a huuuuuuuuuge sector

[fervent thunder]

i live in a country where, thank god, this is already openly part of the social discourse - in others, especially with few and no social systems, this could lead to collapse

[tribal oar]

The person I was talking to today feared it will

UBI is the only solution I can think of if you want to continue to prop up a capitalist economy

Or, we need a fundamental change in governance and economic policy

[fervent thunder]

if you fire all the people thanks to AI, you make more profit, but you also have no more customers who can buy your products.
unfortunately, i believe that greed triumphs over reason, as always.

[sudden isle]

You know I'm seeing writers starting to get nervous about AI too, from ChatGPT

[tribal oar]

That's a very good point that I somehow brushed right over

[uneven wasp]

The increase in efficiency increases the aggregate labor demand by as much or more than the aggregate supply. On an individual scale, the portion of the economy you need to dedicate to creating a good/service goes down with efficiency faster than demand for that service goes up, but on an economy wide scale, this does and has always evened itself out.

[tribal oar]

GPT4 will pass the turing test

I'm sure of it

[keen pewter]

So, if im using 1.5 on a local install, i wont have these restrictions right?

[uneven wasp]

Same reason we didn't all lose our jobs in the first industrial revolution

[tribal oar]

I could just be very tired, this is not an insult or a put down, but your syntax is very hard to follow @uneven wasp, I don't understand what you are trying to communicate to me here

[winter pike]

yeah, only you and god will know what youre generating haha

[keen pewter]

But having 2.0, on a local install, will have these restrictions?

[fervent thunder]

chatgpt is stupid, I told it to make a roleplay and it is a yandere from an anime now - god what a disappointment, it will take many more years

[keen pewter]

Im asking, because im attempting to install them as we speak

[winter pike]

the thing about 2.x is that it wasnt trained on nsfw images so its incapable of producing explicit content

[keen pewter]

I see

[winter pike]

you can still generate sexy images tho, i tested it

[tribal oar]

"When efficiency increases, it means that it takes less effort or resources to produce a good or service. This means that the cost of producing goods and services goes down, which can lead to an increase in demand for those goods and services. As a result, more people are needed to produce the increased amount of goods and services, which leads to an increase in the overall demand for labor.

Overall, the increase in efficiency can lead to an increase in the demand for labor, as well as a decrease in the cost of producing goods and services. This can benefit both individual workers and the economy as a whole."

I asked chat GPT to ELI5 lol

[uneven wasp]

The job loss from technology does not exist on a macro scale because the reason why it exists on a micro scale does not apply.

lol I guess gpt can parse my rambling

Sorry, I tend to write what I think as I think so the overall structure of what I'm saying can suffer

[tribal oar]

No you're good it's on me, I'm really tired and parsing basic sentence structure is tenuous for me today haha

But economically, especially in the last...four decades? We've seen time and time again that businesses will value profit over humanity. And if you approach the problem purely from an economic point of view, it makes complete sense to me that a company looks at the new "coding bot 1.0" that can replace their most basic low-level workers, or maybe 90% of them with a few managers for the output, and decides to pull the trigger

We've also seen how businesses will cut their nose to spite their face, I don't think they're thinking about long term success, they chase short term profits

And, short term, laying off 30,000 people in your work place and replacing them with a few servers makes pure economic sense

Multiply that by all the main code employers

[uneven wasp]

You get structural unemployment

[tribal oar]

Boys it's 10pm and I just put on a fresh pot let's go hahaha gonna talk about AI all night

[uneven wasp]

Yes and no. Did amazon think only about short term profits when they were reinvesting all their money into themselves while being criticized for not being profitable? While the the reason pharma companies exist the way they do is an entire discussion of itself, are they thinking short term when their product development cycle takes a decade or more?

[wise stratus]

Folk are gonna be so confused when they realise the release of stable diffusion is just part of a giant modern art piece

[tribal oar]

Well well well, if it isn't the Emad himself

[tribal oar]

Both of these examples relied heavily on government funding to assist them though

I know you directly mention that with the first part of your pharma example

[keen pewter]

Which one should i use for 1.5?

v1-5-pruned-emaonly.ckpt 4.7 GB
OR
v1-5-pruned.ckpt 7.7 GB

[tribal oar]

But that gets to exist outside of a short term market specifically due to the fact that they are heavily subsidized

That can't be the case for Microsoft, and Adobe, and Apple, and Spotify, and Facebook, etc etc.

Not at the same time

[uneven wasp]

It certainly stirs up some strong emotions lol

[tribal oar]

Can you expand on what you mean by this?

[uneven wasp]

"If it makes you feel something, it's art" vibes

[wise stratus]

I think this is a bit rude tbh, I do do a bit more than hype and have been reasonably instrumental in making foundation models open source with a great team.

One day folk will stop saying I am an ex hedge fund manager just like they did with Bezos perhaps

[tribal oar]

I knew this would blow up in my face

I never expected to explain it directly to you hahaha

What I mean, specifically, is I see a LOT of people describing you online in the same way they do Elon Musk

I don't belittle your contributions to the scene, I apologize for that

[wise stratus]

All ok just joshing

[tribal oar]

When I say you are a hype man

[keen pewter]

Nothing wrong with critisism. I don't see the point of catering to anyone who isn't going to use AI art generation

[wise stratus]

Elon is interesting

Like he’s the richest man the world

[tribal oar]

What I mean, specifically, is you are the face of a new technology people don't understand, and they are rallying around you for answers and a roadmap

[unreal cargo]

the 7 GB one appears to be better for training and the ema-only is for inferencing (prompting) you can use both for most things honestly

[tribal oar]

You don't always have those answers

[wise stratus]

Created a bunch of great companies

[tribal oar]

As things are always growing and evolving

[wise stratus]

Folk hate on him but he can be a dick

[wise stratus]

Yeah it’s my job to take good and bad for now until we get proper infra in place

Most efficient mechanism and saves devs

[wise stratus]

Use smaller one, isn’t there an fp16 one somewhere?

[unreal cargo]

yosh

[tribal oar]

Exactly, and it's like Musk with SpaceX or TSLA, people hang on his word and get upset when promises don't get delivered, despite not knowing the behind the scenes workings

Like the improvements to the rendering times that was teased a while back, people were upset that it wasn't in 2.0, there's most likely a reason it wasn't included then and needs further improvements

[winter pike]

he was the richest man last time i checked or i got clickbaited

[uneven wasp]

according to forbes live list Bernard Arnault is, tho I don't know how accurate a live can be for something like net worth. Debt and private companies muddy the waters more than daily change.
https://www.forbes.com/real-time-billionaires/#195a783d788b

[wise stratus]

Hehe timelines accelerated

[uneven wasp]

Full self prompting when?

[fringe zealot]

I'm extremely disappointed about the recent news of SD3 training, the team hasn't hit the mark multiple times in a row since 1.4 imo

[wise stratus]

When we get self driving cars https://youtu.be/zhr6fHmCJ6k

[wise stratus]

I mean why not let artists opt out not like model depends on them

[tribal oar]

@wise stratus we were discussing this earlier, I am curious for your opinion

[wise stratus]

When distilled releases soon will be like 100x faster and cheaper than release

[uneven wasp]

Ok next year, got it

[wise stratus]

Easier to tune too

[tribal oar]

Could another model not be released that just re-incorporates those artists?

[wise stratus]

Reality is nobody will use any of these models in a year

[fringe zealot]

we've seen with 2.0 and 2.1 what happens when you remove good data from the model, I am expecting to be disappointed again

[tribal oar]

As in, from a separate company

[peak fern]

Isn't it concerning that corporations would decide what should or what shouldn't be a part of AI Art ? Considering the impact and the growth it will surely have ?

[keen pewter]

Yup, cant get the best of both worlds. SD is getting gimped with artist removal and MJ V4 cant have custom aspect ratios

[wise stratus]

Other companies can do what they like

We just build foundational architecture

[tribal oar]

Sure, I just mean the whole conversation seems silly. Artists will rally around SD to stop their art from being included, until tomorrow when another model comes out with their art

[uneven wigeon]

wasn't sd 1.x and 2.x supposed to be a sort of test anyway? And the larger, "imaginator" models with more advanced architecture will be the main event

[tribal oar]

That cycle can only be repeated so many times until they give up

[wise stratus]

Yep

Wonder when that drops 🙂

[fringe zealot]

and the loss just keeps increasing 😅

[wise stratus]

Also licensing datasets and more

[tribal oar]

Listen guys I'm sorry to derail this but I have been having issues literally for weeks now and nobody has the answers, and I'm surrounded by the smartest people in the field right now.

[wise stratus]

For now folk should be happy with the 100x speed up if they think about it

[tribal oar]

Why does my trained 2.1 model, that works just fine on the huggingface preview, not work in auto1111 with the same token?

Makes no sense and it's driving me crazy

I've trained like 10 models on my face and none of them look like me consistently except for this /one/ that I can't even get working right

[wise stratus]

Tbh release cycle was messed a bit with the kid nsfw stuff devs understandably decided not to release as it hit photorealistic

[quaint tapir]

Man, Wayne Barlowe really hates ai art lol

[wise stratus]

Even then this has released insanely fast

We have a bunch of models we won’t release due to that but that’s ok

[peak fern]

Wouldn't it be possible to prevent a combination of those instead of straight banning NSFW ? Not like I'm a Waifu enthusiast but the human body is quite central to the history of artistic expressions

[wise stratus]

Like the 1024 and original v3 1.4bn

[wise stratus]

It’s really not folk just postulate that

[peak fern]

What?

[wise stratus]

See analog diffusion or Lexica aperture

[tribal oar]

This is what I keep saying, you need anatomy to recreate the human body

[winter pike]

did you checked the resolution? it should be 768x768 at least

[tribal oar]

Oh no I haven't thank you will try

[wise stratus]

Nsfw tuned usually have higher res images

[uneven wigeon]

will the larger models be 1024+? and will they be trainable on consumer gpus, meaning 24gb vram or under?

[peak fern]

Are you saying the human body isn't a core part of artistic expression since forever?

[wise stratus]

That’s most important part

[tribal oar]

I have a lot of friends who are professionally schooled artists, and most of their training is studying anatomy

[peak fern]

I'm confused 😄

[wise stratus]

No I’m saying these models can do humans well without nsfw

[winter pike]

auto1111 is by default 512x512, i did the mistake of not changing the resolutions a lot of times

[quaint tapir]

With great power comes great responsibility.

[wise stratus]

It’s basically impossible as we’ve found it to remove kids from these vs nsfw at the moment

[peak fern]

Well the definition of "nsfw" seems quite large. Greek statues for instance I don't think are porn

[wise stratus]

Well let’s put it as concepts that could get you in jail under proposed laws if an idiot puts it and kids in prompts

[keen pewter]

What will end up happening, is with all these restrictions. Lots of people will eventually have good gpu's and people will more often have local installs and there will be community models to get passed the restrictions. When that happens, there will be no need to rely on these more official releases which means less money for them. Local installs are getting easier to use everyday.

[tribal oar]

This is a very, very fine edge to walk, and I hope people understand what I am trying to say, as I am NOT an advocate for child pornography. But do you think there's therapeutic potential in having an AI generate those types of images? If someone is going to look at those anyways, from an ethics standpoint at least no real children are being affected

[wise stratus]

Tbh we don’t really try to monetise

[peak fern]

Yeah I understood that part. Don't mean to derail the conversation or engage you in some debate, I highly appreciate the tool that is at our disposal. I just feel like applying current, modern, morals to the ability to generate art is quite counter intuitive.

[wise stratus]

No

[tribal oar]

Again, to be really clear, I think CP is horrible, I am not advocating it or saying it's a good thing, at all

[uneven wigeon]

It was a good call imo, better to cripple the model a bit and work to fix it, vs dealing with legal/political fallout and get the whole thing shut down

[wise stratus]

Just follow the law and give input where we can is how we are doing things

[tribal oar]

I just wonder, it is technically better that no children get hurt when AI generated images exist

[wise stratus]

It’s not that it’s modern and upcoming legals

[tribal oar]

Am I making sense or do I sound like a monster right now

[uneven wasp]

From a legal standpoint, it's still a problem. Harm reduction policies can really only be carried out by governments. It the same thing as a rando handing out free needles vs a social program.

[little plaza]

Not necessarily, if its photoreal it could impede investigations, spend time looking for fake victims.

[keen pewter]

Nobody likes the idea of murder, but im not gonna take away some ones gun cus of the possibility they might. I think this is a good analogy

[winter pike]

i think its like fuelling the philia more than a therapeutic use

[wise stratus]

It’s a complex topic so better to err on side of caution and work with authorities here

[quaint tapir]

Serious question: say I put “by HR Giger” in the prompt and I get something that resembles his art. It’s not actually sampling images of his art right? Isn’t it just directing the pixels to imitate the style? I’m not sure that makes sense. I’m tired

[tribal oar]

All good points

[wise stratus]

Really don’t get involved here unless you have to

[peak fern]

I understood your point about children pornography. I was asking if there wasn't another way to fix it rather than remove any sexual depictions at all

[wise stratus]

It’s really serious

[unreal cargo]

I feel like that was mostly down to the pUnsafe 0.1 thing, I don't think eliminating NSFW would "cripple" the models if it was done differently

[wise stratus]

Not that we could find maybe we will in future

[winter pike]

i got your point and i remeber being debated at some point here when sd was in beta

[wise stratus]

That was erring on the side of caution, now adjusting. 2.0 can still do accurate photorealistic humans

[unreal cargo]

Oh so that was intentional?

[peak fern]

In my opinion AI art should have the same restrictions as any other type of media but not any more or any less, but it's understandable that being so new it's preferable to be prudent

[quaint tapir]

Generating Child pr0n and making fake news images are pretty dangerous and wrong

[wise stratus]

There was a big internal debate and devs went extreme just in case

Tbh I still don’t know why it can’t do Emma Watson

[peak fern]

What's with Emma Watson lol?

[wise stratus]

Did it like merge all the Watson’s due to deduping

[unreal cargo]

she might show shoulders in a lot of her pictures

[tribal oar]

I feel like I need to say one more time I am very much against it, and those people need to seek help. It's just something I've been thinking about. It extends past CP too, like could murderers use this to generate graphic images and "get it out of their system" ? I see the potential at least for alternative means of offending that doesn't actually hurt any real victims

[uneven wigeon]

I guess better datasets are needed too, really curated ones..something like insane resolution scans of classic artwork in museums like smithsonian etc

[quaint tapir]

And Christina Hendricks

[wise stratus]

Idk she was like 5% of all the beta prompts or something

Think it may be deduping

[quaint tapir]

Or whatever her name is

[peak fern]

People are horny for Emma Watson, is that what you're saying

[wise stratus]

Fine tune in and it fixed

[uneven wasp]

An entire gamut of styles and abstractions are encoded within it, and it associates 1024-vectors for each token/word. So when it gets an artist name, the most efficent thing to do is associate it with various styles. So it's more of learning a style and associating artists to it. Someone pointed out that most old italian sounding names produced the pretty much the same thing regardless of the actual art style of any of the painters.

[tribal oar]

I am so glad I put that pot of coffee on

This is the best conversation I've ever had here

[quaint tapir]

I have seen samples of two artists and the image was almost exact

[tribal oar]

Monkeys, typewriters, etc

[uneven wasp]

there are less styles than artists imo

[quaint tapir]

It was a painting of something but came out almost identical

[distant kiln]

hi!
first time here.
is this pure Stable Diffusion 2.1?
#1010580410884964383 message

[peak fern]

The person answered that he trained his own model

Check messages below

[winter pike]

i mean, someone out there is probably already doing that, but they needed to train the model themselves, stability is not going to facilitate it

[quaint tapir]

How many images are recommended to train a model?

[winter pike]

i see SD as an starting point

[uneven wasp]

I still think dataset prompt quality is a bigger factor in model quality than punsafe filtering

[tribal oar]

Another great point

[unreal cargo]

No

You can filter by pUnsafe to see the various images that were filtered out

[little plaza]

It can do this when the dataset has too many of the same image in it. But it's not supposed to do that.

[unreal cargo]

That's super interesting! Is it replicatable? like did they provide the settings?

[uneven wasp]

Yes it reduces the dataset, but a lot of quality can be gained back by the proper embeddings. So the capability is encoded into the network, but the descriptions they're contrastively trained against are bad at picking them out.

[unreal cargo]

after all, someone could just img2img it

[unreal cargo]

I wouldn't consider embeddings prompts though

[little plaza]

I'm not sure of the technicalities of how it works tbh, but it will re-create like, Starry Night/The Mona Lisa very very closely to the original because there's thousands of duplicates in the training data. It's called overfitting.

The smart technical people in here can probably explain 😅

[unreal cargo]

oooh that? I was thinking you were talking about an artist and someone generating something that looked almost identical to their piece

[little plaza]

Ah well maybe I misunderstood porkchop's comment.

[uneven wasp]

No, but I would consider a large component of openclip a full embedding for every default token

[unreal cargo]

for the most famous artists or art pieces, it makes sense since that art is everywhere on the internet

even after they anti-dupe it

or, maybe the technology is pretty good at that

[little plaza]

Yeah i'd be curious to see examples of not-super-famous artists/art pieces where this has happened.

[unreal cargo]

I'm actually not sure

[unreal cargo]

mm mm!

[uneven wasp]

windows xp background a little bit

It's not as accurate as the mona lisa but it's got the same exact vibe

[little plaza]

haha well if i search windows xp in LAION its got ... many

Idk how much of that is in the training data for SD though

[uneven wasp]

also if you put in "gta 5 screenshot" you'll get a photo on mt chiliad pertty quickly

[final holly]

There seems to be some kind of luddite storm on Artstation and twitter etc.

[keen pewter]

is Automatic1111 using Novel AI?

Im genuinely so confused about some of the stuff im reading on reddit

[unreal cargo]

Automatic1111 doesn't come with any models

[uneven wasp]

Automatic1111 will use whatever ckpt you feed it

[winter pike]

they think theyre feeding the model with those anti-ai images 💀 whos gonna tell them the dataset is static

[unreal cargo]

I think the idea is to prevent future models being created on it

[uneven wasp]

iirc there was some drama between automatic1111 (the person) and leaked novelai stuff but I'm not to clear on it

[final holly]

or how simple it is to remove a huge number of images that look exactly the same in the future 🙂

[unreal cargo]

Who's going to tell them there's an anti-dupe script

[keen pewter]

"NovelAI had a bit of code that effected how their model generated images. This code was then found on automatic1111s fork of Stable Diffusion"

This is why i asked

[winter pike]

literally... this is proof that they dont know how this actually works

[unreal cargo]

Clip Skip is a feature of Automatic's repo, yosh

[uneven wasp]

Was it the exact code or a feature resembling the overall process

[final holly]

So many uninformed people. Probably the biggest thing is its over. It is to late. AI art is never going away will get 1000 times better tand take over everything. Sorry GG>

[uneven wasp]

if its the latter I'm never going to gaf, and if its the former oh that sucks I'm using it anyways

[unreal cargo]

For clip skip? no idea

[winter pike]

the pandora's box is already open and it cant be closed again

sd has been downloaded 1m+ times already

we all have a copy on our hard drive

[final holly]

Exactly. Now I am just using ChatGTP trained with some examples in what I think is best for MJ V4 and then given instructions about what parts to be random and in what way. Random 1960s polish illustrators and a random portrait related subject etc.

[final holly]

I really really want a open source same level ChatGTP to come out

[uneven wasp]

The only problem is that its a large language model, so although an open source version would be great I'm not sure it would be able to run on consumer hardware

[winter pike]

stability is working on it

[torpid zephyr]

What is the DreamStudio "Just A Moment" screen for? Is that something new? Is it to check for bots or something?

I get it every once in a while, but not all the time.

[final holly]

The main thing is it would be out there. We can figure out rest after that.

[torpid zephyr]

I can't wait for them to do a text-to-music AI.

[winter pike]

maybe its that thing to protect against cyber attacks

[final holly]

Test to music would be great and a smart one like ChatGTP

[winter pike]

in the harmonai server theres some demos of it

[torpid zephyr]

Got it. I could see how they may see my input is botty, but there's generally a method to my madness. I'm often trying the same prompt multiple times, to farm the results, and I like it doing it one image at a time rather than generating sets.

[torpid zephyr]

That is awesome. I eagerly wait for the finished product.

[winter pike]

this is a very early model of text to music that uses stable diffusion: https://www.riffusion.com/?&prompt=rnb+music+soul+the+weeknd+gospel&denoising=0.75&seedImageId=og_beat