💬｜general-chat

[cerulean rivet]

The sample images give me 10min to upload them and send you a link

[old marten]

Sorry can't help u there 😦 I hate using lora for this reason

Their training is way different than what I'm used to, personally I avoid it if possible

[cerulean rivet]

Oh ok, good to know

I don't have to use lora

[slender kraken]

Good day,

Quite new to SD, so hopefully my wording is correct here!

I have 2 questions about SD (running SD on automatic1111, if it matters)

1. ---

I have played around with creating my own model with images of myself, trained in dreambooth. The results came out better than i expected for a first run!

I am just wondering, the model is 4gb, and typing myname in the street, it renders me ... in the street! so it obviously added my trained data into a specific model, as it understood street, etc.

I then went and created a merged model of my trained model and protogen, but AFAIK, it doesnt just dump my trained images in there, it blends them, and means some protogen is now missing, and some of my is now missing too, making the results look less like me in the merged one.

How would one go about, training directly into a specific model (like protogen, or f2222) from the word get go?

2. ---

When downloading a textual inversion pack from civitai, lets say i DL one of Sarah Michelle Gellar, and i have a prompt for "girl in the apocalypse", if i add the textual inversion phrase, will it copy the style, aka colors, etc, or will it make the model look like the textual inversion?

[cerulean rivet]

Just thought that it was good

[old marten]

Yah dont use it honestly

No, just do the classic 1300 steps .000001 training on fp16 and u should be gucci

[cerulean rivet]

Oh ok, and would you suggest that I do embedding or dreambooth? Eventually he wants me to generate images with him and his two kids so I would need to train three faces

[old marten]

u can train the protogen model directly

and for 2. it will copy the style

kind of like a trained model but more generalized from what I've seen, I dont use embedding cuz theyre not super good for 1.5

[cerulean rivet]

Ah protogen but not stolen dream. You think that I can't train with that?

[old marten]

dreambooth, embeddings are for styles

[slender kraken]

ok awesome! thats what i was hoping. when using dreambooth on google collab, is there a way to choose which model you want to train? or is that more advance and need to be done locally?

[old marten]

I think u can either train 3 models at one time or one at a time and merge them, this guy on YT called Aitrepreneur has a tutorial on training multiple subjects

[slender kraken]

thx for the replies!

[old marten]

Idk if it is still relevant tbh, worth checking out

[cerulean rivet]

Oh awesome. I'll look at it

[old marten]

depends on which repo ur using, I would just look and see where it loads the model from and just replace the model

[slender kraken]

thx man!

[cerulean rivet]

Ok, I was thinking with merging. I did not know one could train more than one model at once

[slender kraken]

one last question, i am running a 1070, its a bit slow, but works! i have seen some results people are getting are INSANE good though, mine looks, meh

is this all down to prompt magic, or does a better GPU come up with better images?

[steady arrow]

wellllcome. To the world of tomorrrrrowwww

[old marten]

its all prompts and models

GPU only makes it faster, and removes memory limitations

[slender kraken]

thanks, i better go study some prompts! haha

[trim yacht]

hey guys, anyone had a chance to try the new Ultimate SD scaler in img2img?

[mint hedge]

Guys i need any kind of info about embedding training that u might know, i've searched most of what i could search on reddit, if there's some more info or guide about proper embedding training, you would be great help to share it with me.

[thorn hare]

does automatic1111 update via launcher ?

[mint hedge]

u can add git pull to its command prompt in user-bat

[thorn hare]

just "git pull" in the command line ?

[mint hedge]

--git pull

[waxen igloo]

Hi what could be the reason for a extremely dotted/low res result using StabDif 2.0 PLMS 50 Steps 512x512 it looks like a very bad print of max 50x50dpi

[mint hedge]

it could be prompt, some prompts don't work good with some samplers and go reee from my experience.

[brisk bloom]

SD 2 works best with 768 I think

[waxen igloo]

i tried different prompts also at 768 only plms had this bad result

never got a working result with plms

[mint hedge]

PLMS generally doesn't like anything over 50 but at 50 it shouldn't be giving such bad results

[wide juniper]

Is there a way to get stable diffusion to generate lots of images for me over time into a folder? So for example I could generate 800 images overnight

[radiant linden]

depends on what software you use to generate pictures

[wide juniper]

Using automatic1111 and anything v3 model

[brisk bloom]

change the batch count number

[radiant linden]

right click generate, select generate forever in dropdown menu

[warm junco]

Hey @radiant linden how can we reset all settings for webui ? Which file to delete?

[radiant linden]

config.json

[severe field]

I swallowed like 6 kids

[warm junco]

Ah thx!

[radiant linden]

is that a euphemism for something

[rancid osprey]

doesn anyone know some places to find images with the exact prompt and settings used to create it? something like lexica.art

[warm junco]

Publicprompts.art

[rancid osprey]

awesome ill have a look

[radiant linden]

that one is a pretty small amount of prompts from a single person

[hardy zinc]

In automatic1111, what the difference between 10 batches of 1 image and 1 batch of 10 images?

[radiant linden]

latter is overall somewhat faster but requires a lot more VRAM

[hardy zinc]

Oh, so it does it in parallel. You know, I think I knew that. 😆

[radiant linden]

if you have live previews on you'll see it made in parallel too

[warm junco]

Also look at openart.ai

[radiant linden]

https://www.krea.ai/

too

lexica.art, though, probably still the best

[hardy zinc]

Yeah, I actually tried that weeks ago and noticed that. But I didn't really know if it was parallel or a little of this one, then a little of that one. Recently I've just been doing a number of batches of 1 image, and then I forgot about trying it the other way weeks ago

[radiant linden]

you just need to switch rthe filter away from their model and to stable diffusion

[rancid osprey]

alright, its indeed the best one till now. In this discord there is not a section where people also share the prompt right? i could only find channels where they share the end result

[radiant linden]

sometimes you can find the prompt in the image

[hardy zinc]

It's kind of fun to leave it unattended while using wildcards and then come back to see what you got.

[rancid osprey]

ah as in the name of the image it is saved as?

[hardy zinc]

Load the image in the PNG info tab

[radiant linden]

no, the image can have entire generation parameters in PNG fields, if the creatore did not disable that

you can drag the image into PNG info section to check

[rancid osprey]

what is the PNG info section?

[warm junco]

Its in the exif data of an image. Most Gallery programs can show it

[hardy zinc]

That's one of my favorite features. Drag in an old image and load everything with one click to txt2img it's a great feature

[radiant linden]

you can also drag the image into the txt2img prompt

[hardy zinc]

Yes but that just stuffs all the info in the prompt field

[radiant linden]

press rthe blue button

[hardy zinc]

Hmmm.

[rancid osprey]

ah alright, so when i drag the image into SD i get the prompt

[hardy zinc]

I've never noticed this button

[radiant linden]

very useful

[hardy zinc]

I don't have my PC on now, but I'll take a look tomorrow

[radiant linden]

you usually copy someone's prompt shared as a text

paste it into the prompt field

[rancid osprey]

but then i have to use the google colab version right? I use playgroundai to create my images

[radiant linden]

and press the button to put everything where it belongs

[hardy zinc]

Ah, that's super cool. I was wondering if that was possible since some sites post the PNG info in that same format.

[radiant linden]

can use sites like https://jimpl.com/

they'll show you parameters as text too

[warm junco]

So when cfg scale is in the prompt it changes the cfg slider and delete the tag from the prompt ?

[radiant linden]

in a nutshell, yes

it has to be in exact certain format, not just somehwere in prompt

[warm junco]

Very cool feature, gonna try that later

[rancid osprey]

great, ill try it out thanks!

[warm junco]

I think still PNG Info ist the easiest way, also to watch the efix data on Windows i use imageglass+efixtool and on Android Aves Gallery

[radiant linden]

in number of clicks, easiest tis to drag into prompt - 1 drag + 1 click

png info is 1 click, one drag, 1 click

[warm junco]

Yea okay then it depends if i want to generate fast or just look at the prompt

[main sparrow]

I still save almost all of my stuff as jpgs until I find one I like and want to start working on it for print quality. pngs take up a ton of space after a while.

i use the save txt file for paramaters option excessively

[radiant linden]

jpeg can also have params embedded

[main sparrow]

ah, cool. Thanks for the heads up!

[hardy zinc]

I want to make an extension some day. I want make a super cool editor for the prompt boxes. Like, you highlight some text and can increment or decrement the weight of that text using buttons.

I also want to break the box into several sections with pre-canned text like the Styles drop down. And then it would concatenate them all together when you hit Generate.

So, if your making, say, a person, have one box for subject descriptors, another for pose, another for location, activity. Whatever. And then you can mix and match things really easy. I'm always going on tangents and trying new things, and the saved Styles are great for holding successful prompts to reference, but having more granularity would be great. I've been putting bite-sized things in the saved Styles with "+" or "-" at the start of the name to indicate whether it's a positive or negative prompt, and then I use the paste button to build up a prompt from the pieces, but it's really not ideal. I want categories and sorting, and such. I sort the CSV file by hand. 😆

I wonder if anything like that already exists

[warm junco]

Yea it exists, mark a word, then hold Ctrl and press Up or Down

[hardy zinc]

Strg?

[main sparrow]

ctrl

[warm junco]

Haha yea xD Ctrl sry

[hardy zinc]

I'm learning all kinds of things tonight!

But the second part would be the huge feature. Having saved pieces of text that are easy to access.

[warm junco]

I think there is one extension that give little World bubbles suggestions to click

[hardy zinc]

It might be as simple as a category drop down that populates a list box and you multi-select items and hit the paste button to add them to your prompt.

Word bubbles might be good.

[warm junco]

With categorizations like Pose, camera angle, colors etc ?

Would be really cool

[hardy zinc]

Yeah, exactly

You could even publish JSON files that have good, curated prompts for popular models.

And mix them in with your own. Then you could load one of these files to populate the UI which then makes it easy to pick individual poses, camera angles, etc., with just some clicks.

It could be kind of like wild cards, except you would pick and choose from each category instead of it being random.

[warm junco]

Yea that sounds great! Also with the modular part of insert custom json

But how to display so much tags?

Dropdown, bubbles, search Function?

[hardy zinc]

I'm sure there's some reasonable way to do it. I was thinking of a master category drop down, and when you choose a category it populates a list box with choices, and you can multi-select items from that list box and add them to your prompt with one click.

It could even integrate with the wild cards extension to paste in your wild card file names.

[warm junco]

Imagine an alias function so when you type close to .. It displays close up or macro

[hardy zinc]

Sort of like an auto-complete?

Or are you thinking more like a search?

Closeness can also be things like "headshot" vs "portrait" vs "full body". There really are so many possibilities.

Or lens type like wide angle vs telephoto

[warm junco]

Yea thats what i mean. If the user dont know the right words it give him the similar ones

[hardy zinc]

That's why I was thinking categories. They pick a category and see all kinds of things they've never thought of. Hell that happens to me every time I open somebody's wildcard file. It really could be strongly linked to that feature. Just put all these categories in wildcard files, and show the contents when they choose one.

[near falcon]

Guys why do you sometimes get black image especially when using img to img ??

[warm junco]

Yea categories make the most sense. Maybe with autocorrect and when it detects a word you can press shift and it get added to the prompt

[hardy zinc]

Integrating it with typing in the prompt box would be pretty clever. If you enter a trigger word, it highlights it and you can shift-click it to open a something to make selections from.

What we need is a text AI that understands how to write stable diffusion prompts based on what users need. You type in "close up" and it suggests all the various ways to make a close up.

But based on it's own understanding of each model and not from user trial and error. "My AI needs AI."

I really need to go to bed. Stayed up all night playing with SD, and now I'm staying up even later talking about it. 😅

[warm junco]

Its 2:40pm here idk what you mean 😛 but good night!

[stray dawn]

We are looking for highly intelligent individuals. DM me for more information.

[waxen igloo]

Hi i am looking for documentation about Stable Diffusion 2.0 i dont know if i am blind but i didn find any thing about prompt parameters , file structure, where to put the model file and like 100 another totaly normal questions about setting up and runnung this software 🙂

[warm junco]

@waxen igloo i would suggest to watch this easy Tutorial to get startet:
https://m.youtube.com/watch?v=VXEyhM3Djqg

[waxen igloo]

i mean some original info direct from SD developers

i can/like to read

is there a manual or something

[warm junco]

There are different ways to use stable Diffusion, it depends on you graphics card mostly

There is no official Programm to use SD but multiple open source programns

[waxen igloo]

this would be a important chapter in that manual for sure

[radiant linden]

https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

[warm junco]

Someone works on a guide to read:
https://docs.google.com/document/d/1aHJ9RBt_vlCwJQBVUUsb7VghKB-wynv7WGTxm9ozL1k/mobilebasic

[warm junco]

Thanks i will try that later!

[cold siren]

but is there a way to extract known tags from a particular model?

[warm junco]

Since there is no official Software or guide i dont recommend written tutorials because most of them are outdated. The Video link is the newest one.

[waxen igloo]

from where got the maker of the video his info?

[warm junco]

He uses the Automatic1111 Webui. The info to get it running is here:
https://github.com/AUTOMATIC1111/stable-diffusion-webui

[radiant linden]

No. Iy's dome for booru because tags are known

for non-booru models, you could try to achieve this by going through the training dataset

but whether you'll succeed is unknown

[waxen igloo]

this looks like a tutorial about a gui not a tutorial about SD, its nice but very limited

[radiant linden]

there is SD2.0 repo

it has instructions for how to run it, where to put models

[waxen igloo]

thank you! i could not find the syntax and all about the prompt parameters

[radiant linden]

there is no syntax in SD2

or in SD1

[waxen igloo]

there is like -seed and so on the minus is syntax

but there is more

[radiant linden]

this is not part of SD but a detail of a whatever software you are using to run SD

look for explanation in the documentation for your software

or do you mean txt2img.py script bundled with SD2 repo?

https://github.com/Stability-AI/stablediffusion/blob/main/scripts/txt2img.py

this fine has explanations for parameters

[waxen igloo]

yes thank you

oh just the source code but ok parameters question solved 🙂

so this is how it works looking through code by myself?

[warm junco]

If youre interest in the "how ai image Generation works" you should look at blogs from stability.ai or Laion.ai

[ebon pewter]

hi..i try to install ultimate sd scaler but cant find it under script

[waxen igloo]

@warm junco @radiant linden ok thank you!

[silent fern]

hi

[bleak matrix]

hello! 🙂

[wide juniper]

Hi, I've been trying to solve my inpainting causing desaturation problem, but placing the vae doesn't fix it. Help please! https://www.reddit.com/r/StableDiffusion/comments/zk9zpy/img2img_resulting_in_low_saturation_faded_output/
My files:
Anything-V3.0-pruned-fp16.ckpt
Anything-V3.0.vae.pt

[wide juniper]

Fixed it. file name must match exactly

[tender herald]

How can i add dream bot to my own server

[naive aspen]

how do i know if preprocessing images does something? it just says its loading

[naive aspen]

i dont think its doing anything

[native sundial]

@radiant linden hiya : you may want to look into this issue with the webui: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/6676

[radiant linden]

commit from 3 days ago

[peak orchid]

Why does this guy keep trying to list the AI model as the copyright owner of the output? https://www.reuters.com/legal/litigation/computer-scientist-says-ai-artist-deserves-its-own-copyrights-2023-01-11/ As advanced as current AI models are, they are nowhere near sentient and just advanced tools for humans to use

[radiant linden]

I believe this part of the article is going to be a full and complete answer to your question

his

[river maple]

wait, AUTOMATIC is back now?

[soft basalt]

Why do people on this server hate midjourney?

[forest aspen]

anyone willing to help a fellow artist with some setup of custom diffusion model from adobe for testing?

[sonic pasture]

Any1 got a model that does graffitti art?

[steep copper]

Hey guys, I am working on an open-source photoshop plugin for Automatic1111 (locally or with Google Colab) with the integration of Lexica.art prompts. Does anyone want to beta-test it? Any feedback is appreciated
https://github.com/isekaidev/stable.art

[fervent thunder]

Hello guys, i have a question, can i install and run InvokeAI with a Nvidia 3050 RTX, 4GB (Vram), AMD Ryzen 7 5800HS and 16GB?

[swift snow]

Hey my drive is running out of space. Can I just move the Sdiff folder to my other drive

[still remnant]

I run SD from a 2nd hard drive and it saves all images to a 3rd, and it's working for me!

[river maple]

it's not open source

[soft basalt]

you hate anything that's not open source?

[river maple]

a lot of people do, and at this point, i kinda do too -- not anything but a lot of things

[soft basalt]

why does it have to be open source?

[river maple]

because freedom

[swift snow]

the problem is the models

[swift snow]

I think the main reason is to help the tech to advance but since it's not open source it doesn't help at all.

[soft basalt]

ok but aren't the midjourney devs paying a lot of money to train the models? why give it away for free?

I understand appreciating an app being open source

[river maple]

ok but aren't the stability ai devs paying a lot of money to train the models? why give it away for free?

[soft basalt]

but hating something because it isn't I don't understand

[river maple]

advancing humanity is more important than your profit

that said, i don't hate MJ

[soft basalt]

because they want to?

what i'm saying is why hate MJ for not being opensource for that reason

[still remnant]

That's true, you could try using the --ckpt-dir <URI> argument to your launch script and store only those on a separate disk?

[cunning slate]

For anyone that has used AUTOMATIC1111 and img to img, does anyone know what the denoising value translates to in api terms, is it the scheduler and if so which one

[swift snow]

Hmm I\lll give it atry

[still remnant]

I'm running Automatic1111 on windows, but this page was a big help for me setting up my webui-users files: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

[unkempt slate]

in MidJourney it's possible to input several images for a single generation and it seems to work a bit different than the the img2img in SD, is it possible to replicate the "several input images" from MidJourney in SD?

[wary hazel]

Anyone know how to keep the same faces throughout the video in duefourm

[peak orchid]

There was a paper published a while back that seemed to be really similar. I am trying to find it, but none of my searches are turning up anything. Basically in the paper (with supplimentary github code), they trained an image to text model using the model's outputs. It worked well enough that one could insert images into the text prompt like MJ does

[narrow sluice]

Filtered, not like just NSFW, but anything that's legitimately unChristian (I couldn't do horror or anything kinda flirty since stuff like parasitic and sensual was blocked, and they blocked big black if you have it in a sentence, mods have actually fight with users over taste), no privacy (even when you're paying an extra 50 bucks for privacy, it actually still forces you to use their Discord filtered system), a very specific style for images (can't do realism, can't do other specific styles well), closed source (if they banned the color red tomorrow, you would have to suck it up since they're calling the shots, instead of saying that's a bad idea and making them not do that.)

[prisma lake]

does anyone know if :: or : : can be used in automatic1111's version of stable diffusion to separate prompts with weights?

[outer crescent]

so basic decency?

[narrow sluice]

Honestly, it's beyond the regular Discord level

I like stuff like horror or other stuff, it just won't let me make it.

Not basic decency, I always maintain that, it's just...distasteful.

I don't have to block someone from eating cake if I don't like it, what goes around comes around.

It's using your opinion and hiding it as "it's for your own good" that ruins it.

[outer crescent]

well im seeing "big black" "sensual" "unchristian" in your post so... yeah they just dont want grubby edgy stuff crapping up their servers most likely

[narrow sluice]

No, they legitimately said it's for Christian values

I am not making it up

There was a screenshot that popped up

[wary hazel]

Anyone know how to keep the same faces throughout the video in duefourm @narrow sluice

[old marten]

Daaamn these protogen models just keep getting cranked out dont they

[narrow sluice]

What if you want a big black rock

[old marten]

I fuckin luuuuv how fast people are advancing this technology

[narrow sluice]

Again, lazy programming

There's legitimate reasons to use filters, SD probably does it better

With blocking the outputs instead of stuff

[wary hazel]

@narrow sluice any help

[narrow sluice]

Did I mention how frustrating it is when I get filtered out puts and wasting my money?

[outer crescent]

then dont put money into something that doesnt match your purile values?

[old marten]

Yeah I never paid a dime for MJ and their gay ass censorship

[narrow sluice]

I'm not an expert in this stuff, you're better off asking someone else. 🙂
I'm no help.

[wary hazel]

O ok Would you know who can help that’s active @narrow sluice

[old marten]

Honestly MJ is to computers the same way Macs are

[outer crescent]

thats true

[old marten]

Yeah they work yeah some people like them but serious users never even consider it as a real platform.

[narrow sluice]

Don't need to be a politician to say that if you don't like something, just avoid it.
That's after spending 15 bucks since I only used MJ at the time since any competitor didn't exist.

It was that or DALL-E 2

[outer crescent]

https://discord.com/channels/1002292111942635562/1025266140445933648 this is the animations channel. they likely would 🙂

[narrow sluice]

It was for a project.
These days, I have much better options

[narrow sluice]

Also, MJ got boring after awhile.
It just sucked with other genres.
Essentially, it became less valuable since I have much more better choices in terms of AI generation and not worry about hitting the pretty bad word filter.

Like we're on the SD server, there's so much out there to try.

If you like the MJ style, whatever floats your boat

And SD is free, which I can't say with MJ

[outer crescent]

yes. but its lol to cry censorship

[bleak wolf]

this is the point

SD is already better than MJ

MJ is just more "end user friendly" and faster when it comes to total time required for an artwork.

SD requires more time and experience, but once you have it, it makes truly wonders

[narrow sluice]

Tbf, SD has Dreambot, which does pretty close to MJ these days.

A few more extra options, but can definitely give MJ a run for its money

[young ivy]

why? you can complain about censorship and then stop paying for the product. MJ might have a greater dataset that can generate out of a wide range of styles...but it gets annoying when you have no idea if the word you typed in is banned. Then having to pay for privacy

MJ`s biggest issue is how expensive it is

for features I would take for granted

[bleak wolf]

That is the opposite, My assumptions are MJ has a smaller, cleaner dataset

[young ivy]

they've talked about the millions upon millions of images theyve parsed

[river maple]

That doesn't negative cleaner

[young ivy]

in that voice chat that you can listen to their office hours

[river maple]

You can have a cleaner but still large dataset

I have a mostly well-tagged dataset I'm building, and I've hit more than a hundred million images already

If you know the right places to source images from, it's not too terribly hard

[young ivy]

yeh

[bleak wolf]

100 million? Damnnnn

[river maple]

technically 140 million+

and still growing

[bleak wolf]

and im thinking the 26K dataset i have is a lot

[young ivy]

my biggest gripe with MJ is still that I have to actually PAY for privacy

[bleak wolf]

Is your dataset public or private?

[young ivy]

something that imo should be free, since it doesn't cost them a thing

[bleak wolf]

Capitalism

[outer crescent]

^

[river maple]

It will be public when it's complete (as close to 1 billion as possible!)

[outer crescent]

its something people want but dont need so $

[young ivy]

sure capitalism. Also free with dream, novel ai etc

[river maple]

Had NovelAI ever fully recovered from the leak/hack?

They had the easiest to use anime model back in early October

[outer crescent]

wasnt sure if this was something you could talk about actually

but im interested too

[river maple]

the dataset or the aftermath?

[outer crescent]

the whole drama

[river maple]

it's just that... haven't heard anything about NAI imagegen in a while

it's clearly OK™️ now since AUTOMATIC is back here

[outer crescent]

oh didnt even notice

i never know whos chatting or on servers tbh. just like the pretty pictures and occasionally lolling at people getting angry at AI generated stuff

[narrow sluice]

They're doing fine. They're working on beta furry stuff

The only things that has springed up is people using, cough their datasets/models illegally

But well, cat out of the bag

[river maple]

Are machine learning models actually copyrightable?

[narrow sluice]

It is implementation theft, from what I heard from a programmer friend.
Hard to prove though.

[river maple]

They had the option to send a DMCA to certain hosters, but never did. This is beginning to make me believe that model weights aren't actually copyrightable anyway (at least in the US)

[narrow sluice]

They can DMCA but mirrors exist

Plus 4chan

[river maple]

"Implementation theft" isn't a legal concept I've heard of. Anyway, I have a feeling their lawyers advised them not to or something.

[narrow sluice]

Good luck tracking everyone

[river maple]

They claimed to have sought legal counsel

[narrow sluice]

Well, they did

But again, too late

Once it's on the internet, you beeped

Even the best folks in the world can't stop the spread

It's going to be going after shadow puppets

[river maple]

Anyway, once I train my new large open models

all of proprietary models will be 100% obsolete

[narrow sluice]

Yeah, I mean, I would love for NAI to get new toys

They are looking out there for open source commercial stuff

Not just for images, but text as well

[river maple]

There are new toys (BLOOM), but text is really hard

nothing compares to GPT-3.5/ChatGPT

[narrow sluice]

Ironically, they won't touch most open source stuff if it's bad or non-commercial

[river maple]

well, nobody touches bad stuff

and non-commercial violates the definition of free software and open source

[narrow sluice]

And they wouldn't have gotten big if OpenAI didn't eff up, so they're not ever going to use that.

[river maple]

OpenAI wouldn't allow them

They'd get banned for content policy violations

[narrow sluice]

Yeah, no kidding

Best for everyone 🙂

20B is good, but I am curious what an uncensored ChatGPT would be like. For science ofc.

[outer crescent]

like twitter

[river maple]

like 4chan

[narrow sluice]

Well, OpenAI admitted to scraping the entirety of Twitter when training ChatGPT.

Even when it's...against ToS

[river maple]

You don't need to comply with ToS for anonymous scraping

LinkedIn vs HiQ

[narrow sluice]

Wouldn't that fall under serious privacy concerns?

[outer crescent]

no

[river maple]

No, because it's public data

Tweets are public. Courts have decided that public data is obviously public and can be used publicly.

[narrow sluice]

What other stuff is technically public data?

[outer crescent]

twitter has to be the best source for training language though, questions answers, small to the point, countless variations on replies etc.

[narrow sluice]

Out of curiosity.

[river maple]

I think it's fun to bring up home address databases. "Doxxing", while controversial, isn't actually illegal unless you intend to cause harm (and intent is hard to prove).

LAION consists of public data.

[narrow sluice]

Once you mentioned public data, it's now very interesting frontiers.

[outer crescent]

but for factual stuff, gpt supposedly weighs sources credibilities, like if they've published papers on something or if theyre an anime avatar oti 😄

[river maple]

You can look at Common Crawl to see what else is considered public data.

[narrow sluice]

What about if you disclose military secrets in a tweet?

How would that work lol

[old marten]

How much storage do they take

Thats gotta be petabytes and petabytes I call bs

[river maple]

Whoever first disclosed the secrets would be in trouble, but I think subsequent sharers would be fine.

[river maple]

metadata alone is 120+ GB

However, that's all the dataset is: metadata

[old marten]

Oh

[river maple]

LAION is just metadata

nobody's carting around 1 billion images - that's insane

[narrow sluice]

Actually, is there any weird cases with copyright in a tweet, in the cases of people sharing say, a book that has gone to public domain in one country, but not in another?

[river maple]

I don't think that'd work. Twitter would follow US copyright law, and everyone else would be dead out of luck.

[narrow sluice]

Yeah, there's some difference in UK/US law, and the former is earlier, if I remember correctly.

Same with EU, I believe

[bleak wolf]

Artists when another artists learns from their image to make/influence their own:

"Yes! Im so pRoUd"

Artist when AI does the same that the human artists did:

"nO!1!!1! tHeiF, sTeaAliNg WaAaAaA"

Its just coping

[narrow sluice]

Artist when you spend 6 hours on one hand and it still sucks.

[bleak wolf]

most of them are genuienly scared that all the effort they placed in for years is in risk because technology advances

Again, when collecting images for my datasets, I ask myself:

Did this artist pay/get permission for every image they have seen with their eyes which may have conciously/subconciously affected them?

The answer is no.

Thus im relaxed, ethics wise.

[narrow sluice]

Gives me huge food fusion, cultural appropriation vibes

[bleak wolf]

This topic is something that cannot be met in the midfield logically, you either allow it, or copyright everything, delete memories of every artists, and tell all of them to start from scratch (a.k.a caveman paintings)

It can be met in the midfield in law tho

[narrow sluice]

Does the Irish have to pay Mexico because they used potatoes?
Does the rest of the world have to give credit to places that gave them non-native spices?
Does anime/manga artists need to pay Disney cause the guy used Bambi's eyes for inspiration?

[bleak wolf]

Does Corruptlake have to pay Potato for this inspirational quote?

[still remnant]

don't even get me started on Italian food and tomatoes!

[narrow sluice]

Does the British Museum have to give back everything they stole, no cap?

[bleak wolf]

uhhh

that is another topic

but uh

mostly, eh?

idk

[narrow sluice]

Exactly 🤣

I just put that joke in

[bleak wolf]

oh

sorry for being dense fr

[narrow sluice]

NP

I was moderately dense as a kid. I get it.

[bleak wolf]

also would you like to try my model?

[narrow sluice]

Probably not now, I have to move my hard drive to a bigger one

I'll let you know if it's done, but yeah

[bleak wolf]

Does keeping models in HDD instead of SSD have a big impact on gen speed?

[narrow sluice]

Not sure, I have been using SSDs mostly

[bleak wolf]

cause my models file is 81 GB lol

[narrow sluice]

OH LOL

Definitely not a small download

So yeah, later on

I was on a 1050 Ti, so I didn't do much generating

Locally

[bleak wolf]

ic

[bleak wolf]

https://aqualxx.github.io/stable-ui/ you can try my model here

[narrow sluice]

Ahh, cool

Thanks for letting me know

[rotund umbra]

Hi everyone, I'm fairly new to stable Diffusion, so I want to ask some things - what are generally good negative-promt words to use in order to create good anime-character image ? What are good keywords for this ?

In 1.4 model

( And I know that I should download community model for better res, buuuut I want to get around first and try to get smth from basic model )

[still remnant]

honestly I just start by adding Anime:1 to the end, and vary the number up and down to get a feel for how much it affects a given model

on Automatic1111 I would use the X/Y plot script to get side by side comparisons and use S/R to increment from like "Anime:.25" to "Anime:2" by increments of .25

If it's looking too shaded for me I'll often then start adding "cartoon" or "cel-shaded"

to flatten it out

[rotund umbra]

Yea, I use Automatic 1111

And, what about negative promts?

[still remnant]

Aside from my standard ones, I rarely use them in 1.4, i might put in like "3D render" if it's always looking overly CG, i usually aim to keep my prompts simple and use weights to make changes

[rotund umbra]

What are your standart ones ?

I am looking for such

[still remnant]

I am running some tests now to see with the standard 1.5 checkpoint, I haven't used that model in so long!

[sharp falcon]

how to generate textures in dream studio website for dream texture to blender?

[icy pendant]

What’s the general consensus on Waifu diffusion 1.4 vs anything v3? Because for me, anything seems to completely clobber it

[warm junco]

WD1.4 ist not fully trained (1 epoch), but it can generate good images if you know how. I think it will be better than AnythingV3 when its fully trained

Tags work different for both models so you cant compare that easily

[rotund umbra]

Well, How to train AI by yourself?

[warm junco]

With Dreambooth and enough Vram

[rotund umbra]

Well, with my 6GB VRAM this is not an option xd)

( I have Geforce 940-mx in my notebook , that is my only PC )(

[warm junco]

Even with my 8gb its not an Option xD but there is a new thing called Lora training that only require like 7,6gb

[rotund umbra]

Still not an option )

[warm junco]

Yea but i want to try it when there is a good Tutorial out. Hope Aitrepreneur will make one. Can recommend his Videos.

[rotund umbra]

From his videos I have heared about community modules )

[warm junco]

Ahh nice

Yea he also compares them very detailed with lot of effort (like NovelAi vs WaifuDiffusion vs AnythingV3)

[rotund umbra]

BTW, is protogen actually that good ? And does it need training ?

[warm junco]

The models you Download are already trained and wont learn new stuff.
Protogen is pretty good. Not for anime but for photorealistic and fantasy stuff

[still remnant]

I kind of like the modern-anime look I get with jamix

[vapid tangle]

anyone know of analog diffusion colab out there?

[rotund umbra]

Does 1.5 module understand keyword " Neko " ?

))

[still remnant]

I feel like I get some differences with "Kawaii" and the -chan suffix

ran a couple test prompts and I don't feel like neko is doing anything.

[icy pendant]

Ahhh ok I was using the same prompts and comparing and it wasn’t even close but that explains it

[short tapir]

What's a good model for doing stuff with pokémon?

[latent flower]

check #1047197565365538826 for https://huggingface.co/Timmahw/SD2.1_Pokemon2DSugimori

[still remnant]

Which model? I wanna catch em all!

[latent flower]

heheh

[short tapir]

Well like anime styled. Basically want to like like Raiachu wearing a tuxedo

[short tapir]

um.. how do i download that..?

[jade meadow]

💵Cha-Ching! Up to $200.00 CASH for you!💰️Click and accept my invitation on Temu！ https://temu.com/s/nN34q42ydgLeOpX

[river maple]

Warning: a scam and/or hacked account has been detected by Discord Security Suite 2023

[latent flower]

click the ckpt file then "download"

under the files and versions tab

[still remnant]

for training can you split the job across multiple GPUs?

[fervent thunder]

how do you report posts?

like that spam above

i cant believe an ai discord doesnt have a better auto mod haha

actually i never see mods here but idk

[still remnant]

It has blocked me before 🤷

I guess I am pretty spammy - lol

[umbral geyser]

Is there a deviantart-equivalent for AI art, poetry, etc.?

[astral goblet]

my favorite negative is "jpeg artifacts" as it completely destroys any effort of SD to create authentic compression artifacts. But i tend to add it to a seed after i see it containing them

[still remnant]

thats a goo idea

[astral goblet]

i got a couple negative embeds i've been playing with, but i don't add them to prompts right away. i tend to not use negatives unless needed

[still remnant]

I always add "watermark, signature, logo" to the negatives, since the training set seems to contain a lot of stock photos tahat have those "features"

[astral goblet]

A fun one i needed recently and it took me abit to figure it out, i was trying to generate doom guy with better ultra doom armor and he kept coming out hooded. Well, what could that be confused for? Doctor Doom has a hood always. So i negative'd "Doctor" and finally got what I was after

Yeah watermark, signature, logo, name are quickly becoming default embeds actually. i'm always just throwing those in after 3 generations

[latent halo]

Do embeddings change the image even if you don't use the keyword for the embedding?

Just by having embeddings in the folder, will it change the output of all generations?

[astral goblet]

sometimes i think a lot of the voodoo like "too many fingers" doesn't work because it's not like the model understands such things. It's hard to convey what i mean. The clip model could probably understand too many fingers means polydactilism through relations, but that's still such an abstract connection. But then the trained data might have any images to know what that would look like

[astral goblet]

no. They're just small files loaded and ready to go

[latent halo]

Ah, so they're only activated and used in the prompt if the keyword is detected in the prompt?

[astral goblet]

yup. the generation details in a1111 webui tell you which ones were active

[latent halo]

Very cool

[astral goblet]

use more than one even

[latent halo]

That's much better than having 1000 models.

Can embeddings be used to add new objects?

[astral goblet]

embeddings are infinitely better for playing with, but sometimes a refined model is needed for a goal

embeddings sort of pull what the model knows out in a very targeted way

[timber silo]

can someone clear my doubts please? can i generate an image based on some public image link? like in midjourney

[signal nest]

any idea how many people run Stable Diffusion locally? (ballpark obviously)

[wary hazel]

@latent flower is there a way to have my own picture throughout a defourm video generation

[wary hazel]

This guy said he did this on stable dis and defourm using a trained molde of him self any one know how to do this

https://www.instagram.com/reel/Cml9degou6E/?igshid=YmMyMTA2M2Y=

@latent flower

[outer crescent]

well theres 28k stars on automatics' repo, and as a rule of thumb only about 1 in 10 people thumb up/like/comment/interact on stuff online they use/download/view. ive no idea, havent thought about it before but that's how i'd go about eyeballing for a number

[humble ledge]

Good day fellow prompters

[outer crescent]

hello fellow skellington profile pic friend

[signal nest]

Perhaps this gives us a reasonable lower limit of 10,000. I hope the actual number is a lot higher. Maybe the number can be discovered from huggingface. What i'd like to know is does the world have enough people that are seriously interested in running generative AI locally to keep it going. (I'm also guessing there are people that would run it but lack a big enough GPU )

[outer crescent]

10k? nah. easily 100k

also theres going to be an exponential uptick in AI interest this year, this is still at the early adoption stage

[still remnant]

I mean, I'm running 3 local instances all by myself!

[hardy zinc]

I'm no dummy, but I just don't get Dreambooth. Anyone have a really detailed tutorial on what class to train against, and what regularization images to use, and why you would chose that class and set of images? If training a face that happens to be a woman, why would I train "person"? Why not "woman"? Why not "beautiful woman"? Why not "face of a beautiful woman"? What's the theory behind choosing just "person"? No tutorial I've found explains any of this. It just feels like people copying and pasting instructions from someone else who did the same from someone else, and no one really knows what they're doing or why they're doing it. In programming we call this "cargo cult programming". If it works, it works, but it's better to understand exactly why something works.

I also don't understand at all when or why you would use filewords in Dreambooth and what effect they have on training. When training an embedding, you use BLIP to caption your images. I imagine it's better that the training mechanism understand that if your subject is wearing "a blue sweater", than it learns that that's not intrinsic to the image you're trying to train it for. It instead already knows what that is, and can focus on the things that make the image unique. I imagine this helps your subject editability. And you can also use filewords which just a list of prompts like "a cool photo of [subject]". How does that help training at all?

In Dreambooth, there doesn't seem to be a preprocessing mechanism. You don't generate text prompts for each image, but you do have a chance to specify filewords in the webui. Why would I do that? What's the value? Also, why are images not preprocessed like in an embedding? And again, what subject should I choose and why?

Any help on truly understanding this would be greatly appreciated! Just pointing me to a really solid tutorial with reasoning behind the choices would be sufficient.

[astral goblet]

choose any class you want and experiment with it. I made one of myself on "Dude" and whenever i prompted myself i got dude versions

the goal is to contexalize the concept you're training

[hardy zinc]

Like, you're in a coyboy hat and such?

[astral goblet]

no, more like good burger

or big lebowski

surfers

[hardy zinc]

So if you chose a class of "cat" and trained a person, the result would come out as cats when you used your subject in a prompt?

[astral goblet]

maybe? try it

[hardy zinc]

And "person" keeps it more generic?

I'd rather read the theory behind it than just experiment.

[astral goblet]

i typically do the gendered generic term, man, woman, but i wonder if it matters much

thats the theory. you're contextualizing the concept into the model based on what it knows

[dense torrent]

I haven't dug in, but some on Reddit pointed to this video tutorial as a great explainer for Dreambooth https://www.youtube.com/watch?v=mfaqqL5yOO4

[hardy zinc]

Woah, an hour long. Excellent!

[old marten]

the class prompt is what generates the regularization images, iirc if you use ur own reggies then theres no point to that prompt

so yes, in theory you would end up with cat human hybrids but more realistically just really deformed cat-human abominations.

[hardy zinc]

Alright, I see that in the tooltip now, but I guess that still leaves me asking why I would choose "person" vs "man" vs "portrait of a man" (which is what the video linked up above is using.)

He actually used "face photo of a man, 8k, hd, smooth, sharp focus, Cinematography"

So it only generated good images of a man

[old marten]

Correct, and that's what I usually do now. For example, my reggies are "photo of a person in high definition"

[hardy zinc]

Meanwhile, I downloaded a whole folder of "person" regularization images, and some aren't even persons

[old marten]

I have experimented a LOT with reggies. I've used female models, women in bikinis, etc.

[old marten]

That's what you want. Because if the subjects look too much like your subject, you will overtrain your subject

Think of the training as the machine getting to know ur subject. If u have 20 images and 1000 steps, ur bot will study each image 50 times. If all it does is study the same image 50 times, it will learn it "verbatim"

If u throw in some reggies, it will "blend" the images together to a certain degree, learning patterns that you want it to follow. The end result is a more variable and flexible model that works better in different positions and poses, scenarios, etc

[hardy zinc]

I did have the experience once of a model literally outputting one of the input images almost pixel for pixel

[old marten]

that's overtraining

[hardy zinc]

Indeed

[old marten]

The reggies help battle that, in fact if u have a ton of images you will have a more flexible model bc it got to use such a wide variety of reggies

thats why some of my best models are almost 500 sample images long, I have some cool high-res reggiest that I have yet to test in 512, I trained a 1024 with...some...success with these reggies

[hardy zinc]

So why wouldn't you always just use "person"? Why use "female model" or "women in bikini" or whatever?

[old marten]

Because you want the reggies to be very broad and generalized. Thats why the person reggies u saw are so deformed. I trained several models with women in bikinies and supermodels and the result ended up looking like ass

Unless you specifically want to train a woman that's already in a bikini, or a model, etc. otherwise the machine will have a very hard time shoehorning ur person into those bikini models and supermodels

[hardy zinc]

Oh, when you said, "I've used female models, women in bikinis, etc." I assumed you meant you tried it and had success

[old marten]

No they suck although I haven't trained specifically bikini women, I think however, I did train a dude using the buff dudes folder

Let me check actually lol I'm curious how that model looks

[still remnant]

those sound the same to me. lol

[old marten]

Nah I just checked and the buff dude ended up also looking like ass with those reggies. Because those reggies focus a lot on nude bodies, they will generally work better for training bodies, not faces. I recommend creating ur own reggies with something that resembles what you want ur model to look like, and set the CFG really low, to like 4

That way u have a very wide variety of images to mix in with ur model, all while keeping faces as the main subject

[still remnant]

anyone else have strange dreams when they generate images for like 2 hours before bed?

If I could record my dreams after playing with SD, I wouldn't need SD

[fervent thunder]

good to see you

[junior juniper]

Hello. Quick question, anyone have a link for pre-made classifier images (512 or above) so I don't have to generate my own?

[charred otter]

What are some of the latest notebooks you people use for img2img?

[quick tapir]

How do people make thos pictures on tiktok like: Economy as a human. Or, countries as a human.

[sour marsh]

this server is nice

[fervent thunder]

Guys what is the difference with pruned or unpruned

[tawny quail]

Pruned model which is modified to make it faster, good for inference
Other is slower but good for training/fine-tuning

[fervent thunder]

what is fine/tuning

I generally use for inpainting so i should use pruned version

so it uses less ram and vram

[tawny quail]

Yes, use pruned

[fervent thunder]

Oh and do you know why i need to use vae ?

what is used for ?

[sleek rain]

I've got a problem where the photos I save are being put into the RAM instead of the hard drive, pls help

[bleak matrix]

Good morning, everyone!! How are we all this fine day?

[tiny finch]

A text to video version of SD is supposed to be released this year?

[shy hornet]

Hi to all (don't know is this is the correct chat)
I'm new and trying to use it to generate art for TTRPG(DND) games. Is there a model or enbedings focus on this topic?
This is because I haven been able to create non-human base portraits.

[main sparrow]

yep. check out this guy's work. I can't remember what model these embeddings are trained off of, but he did beholders, dragonborn, and a bunch of other stuff. https://huggingface.co/VTTRPGResources

I think he even has his own discord server

Actually, I think they might be all trained models, not embeddings. Either way. good place to start.

[latent flower]

yep, train a model on yourself, install the video extension... Play with the parameters... Thats it

[latent flower]

hello! Its been nice... I hope to gen lots of cool images!

[burnt inlet]

Hello !
I make op-art / abstract / geometric / paterns art

[bleak matrix]

Good to hear! What kind of stuff are you prompting today?

[frosty schooner]

Hey, I've made too many images and searching between them is a hassle now. How are you guys doing this?

[latent flower]

I was think about space pirate ships! Those are cool...

[latent flower]

well I just use an extension to view the images and save the best I gen in my pics folder

[frosty schooner]

I want to search and filter by prompts keywords. Any way I can do that?

[bleak matrix]

You mean locally, or on Discord?

[frosty schooner]

locally, on the images I've made

[latent flower]

are you using the WebUI? The images are saved using your prompt as file name...

[main sparrow]

I use ACDSee. Unfortunately their Digital Asset Management is only with their paid versions, But Adobe Bridge is free, and is close in features for just digital asset management.

Might want to give them a shot.

Once you start using software to manage your inages you need to be careful to only move/delete/copy images in that software of the database can get trashed. So once you start using one, you sort of get locked in, but it's invaluable when you start having to manage tens or hundreds of thousands of images

[frosty schooner]

If I want to search two tags or more together. It's not possible with prompt as file name.

[main sparrow]

saving data in filenames works short term, but it's a horrible way to manage images.

the only thing I care about is to make sure that auto1111 generates completely unique names, so if folders get merged, nothing gets skipped/overwritten. I use [datetime] in filename generation settings.

[frosty schooner]

Are there online versions of these as well. I don't want to redo everything if a hard-drive gets corrupted/ anything else happens. 🙂

[bleak matrix]

I typically use the prompt itself to search. But I also just move the images I really like into a folder.

Then I organize my images.

And sort them by the things I'm working on, etc.

[frosty schooner]

You cannot multi search on them. For example if you want to search tags "beach" and "gun" together for a previous prompt, u can't do that.

[bleak matrix]

all of my files have the name written into the folder/image, so just searching for anything with those tags will come up

But I have all of my folders organized anyway (into people, projects, etc)

[frosty schooner]

Let's say my prompt is "4k, digital art, concept art, anatomical accuracy, hyper-realism, chiaroscuro, ray-tracing, caustics, cinematic, full body, A middle aged mexican construction worker taking a break, [sitting|eating], [respectable|hard working], [friendly|good natured], awe inspiring, absurdres, masterpiece, sublime, character sheet, art by Michelangelo, greg rutkowski". This will also be the filename.

I want to search "awe inspiring" and "greg rutkowski" together. Simple windows search will not pick this file.

Adding some generation data like this: Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 872579092, Size: 1024x1280, Model hash: 16e33692. And then searching becomes impossible.

[latent flower]

maybe theres a program for that, on linux theres the "find" command line program maybe theres something like that on windows

i think git bash has the coreutils programs

"find" is one of them

[main sparrow]

the one thing you can do is have auto generate text files for generation parameters, and use soething like notepad++ to search through those to find the corresponding images.

[latent flower]

then id "find . -name awegreg"

ops

i cant send *

[frosty schooner]

I've written a python script for this but I was wondering how non programming folks do this.

[bleak matrix]

(Besides parameters) like I said, organize your files. Be proactive in knowing where your things are. Be aware of what you create. Make a people folder. Make a science folder. Put those images in that folder. Then, you know those images are there for when you want them, you know? Or, when you create them, create folders/use folders for them.

[warm junco]

I made folders for my most used Models, then sorted the generated images based on my prompts in this folders with more folders xD

[latent flower]

hey guys, an attack is happening rn in the bluewillow server. Bots are spamming @ everyone and using the imagine bot a lot. I think its an anti ai croud, so this server is prob also a target

#🏞｜general-with-images

[young spruce]

damn, i wanted to try their service

[latent flower]

yeah Im afraid they are gonna attack this server but I also dont wanna @ the mods

[fervent thunder]

Is there any model that makes pixel art for games?

[latent flower]

yep, wait a sec gonna send the pink

link

ok found it https://huggingface.co/PublicPrompts/All-In-One-Pixel-Model

I have the all in one pixel v1 but I think the v2 is out there also

[hallow karma]

how do i change the image bit depth of the output image?

it seems to be 8 bits/channel

[karmic brook]

yeah, I just had to leave the server after all the ping flood

we will keep an eye out

not sure if that was a raid or triggered by someone using the @ in their prompt and then spamming it

[hallow karma]

i counted 20 different users

the way they did it was to tell the bot to imagine something with @ everyone in it and the bot repeats the prompt back to you in the same way midjourney does it

so the bot essentially spam pinged the whole server hundreds of times for a couple of minutes

[old marten]

I, for one, am ready to move to the future where txt2video doesn't look like a really bad acid trip and looks like real video lmao

[karmic brook]

yeah, figured it might've been something like this
in that case, not sure why server staff haven't turned off the ability for @/everyone pings?

unfortunate

[hallow karma]

i think they just overlooked that it would be an issue like that

it's a smart way to ping a whole server but really annoying 😂

[karmic brook]

well, if anything...i guess now they know lollll

[hallow karma]

yeah hahaha

wait can the bot in here do the same thing?

[latent flower]

wow its fruit! Thx for answering and being alert! I like here and dont wanna see ppl raiding the server

[karmic brook]

you're welcome!

i was wondering if anyone else here saw that since it happened so quickly haha

ty for caring

[latent flower]

👍

[karmic brook]

https://tenor.com/view/nichijou-anime-handshake-bros-shake-hands-gif-12270042

[plush sky]

i need a little help with something..were do i ask for it?

[karmic brook]

depends on what kinda help! if you've reviewed #1014939219904450590 , and need more technical help #🤝｜tech-support is a good place to start. We have a lot of different channels for different topics/help ^-^ so it depends !

[plush sky]

thank you..

[wary hazel]

We will together

[wheat hazel]

Somewhere I've seen a list of prompt styles for portraits, landscapes, etc... you know the url ? can't remember it

[austere wharf]

Hi, does anyone have a good colab for text to speech?

[sleek osprey]

is there any work being done to improve long/complex prompt interpretation

[rapid bobcat]

Any news on generated video? Are there any models out right now capable of it?

[fervent thunder]

Hi, is there a way to get the bot to generate 2:1 ratio images?

[fervent thunder]

does anyone know what the --nohalf does ? is it for speed? i have a new install/computer and am wondering if its worth adding to my command line

[warm junco]

dont use it if you dont get black images or other problems at generating. i think its for half precision generating, dont think it performs better

for better performance and low vram usage you should use --xformers. if you have 3-4gb vram you should also use --medvram

[fervent thunder]

yes i have the xformers, i just remembered i had that nohalf too before but didnt know if it was really worth adding, i guess it isnt. ive got a 4090 so plenty of vram but i guess it doesnt do much in stable. hella fast in games thats for sure

[warm junco]

haha its also the fastest gpu for generating i think

[fervent thunder]

it is pretty quick, but someone said something isnt optimized yet for it. torch i think but idk haha

[warm junco]

if you want to battle your vram you should use Batch Size and set it to max! xD

[fervent thunder]

let 'er rip!

[rocky beacon]

Hey, can someone explain to me what a model hash is? Can i change it?

[warm junco]

the model hash is a unique id for every model, you cant change it, but you can merge models it will generate a new hash.

[jovial tinsel]

https://www.youtube.com/watch?v=P10HOvTNPbk&t=3s

no way omg

[fervent thunder]

https://preview.redd.it/g8ezvmb98tba1.png?width=1278&format=png&auto=webp&v=enabled&s=df361f01a867dfc9832a4c9a6e700d87e1c4b93f

new model or maybe video 🧐

[sleek stratus]

Hi, does anyone know if there are any legal issues with selling images generated by Stable diffusion or derivative products?

[signal nest]

yes^^ grey area

i see some claims that if you do significant work eg img2img you can claim its yours,but also there's hard rulings like "you can't copyright AI generated images"

i get the imperssion we are still in uncharted territory

fair use seems like a valid defense of training on scrapes to me, but when you actually try to sell things.. fair use only goes so far

[brazen fox]

I'd argue the legal issues would be the same if you were to sell other types of artwork- as in, draw Mickey and Disney will sue your ass regardless of the methodology

[magic widget]

US copyright law has notable exceptions for personal, educational, or satire/parody purposes

[sleek stratus]

Thank you for your answer, I will take it into account!

[sleek stratus]

Of course, if the concept already has a copyright, it doesn't matter if it was made through AI. But I'm referring to own ideas.

[magic widget]

“This action from the USCO may serve as an early warning that anyone filing works that contain any portions generated by artificial intelligence must disclose such portions and be prepared to support their registration and prove a degree of human authorship.” https://ipwatchdog.com/2022/11/01/us-copyright-office-backtracks-registration-partially-ai-generated-work/id=152451/

Basically, it's legal gray area right now, but it's a safe bet that large corporations will still use the law to protect their property.

[fervent thunder]

my iq test came back negative today.. grateful

[magic widget]

Our best hope is that such a demand is shown infeasible, or at least that all works generated with large models be made public domain.

[fervent thunder]

This whole ai controversy is making me scared to share my generated stuff

[magic widget]

The biggest problem is all the people who think it's their chance to be holier than everyone else.

[fervent thunder]

Yeah

Strange how ai got attacked month ago ( i think) by artist's cuz its "steals"
Why didn't they do anything in the past?
I bet its all their strategy to ban ai art cuz "new technology scary, it will replace us"

[rotund umbra]

BTW, how to generate , like, 3 images per 1 generation ?

[magic widget]

If you're just using stock SD and you're not going out of your way to violate trademarks, you're probably fine.

I think the real problem will be for fine tuning.

[fervent thunder]

Im starting to believe that some artists dont care about the "stealing" stuff. Only to ban the ai

[brazen fox]

if you're using a local install or DS, adjust your batch size

[signal nest]

i can only hope common sense will prevail. corporations could shut stable diffusion down with the arguments about scrapes. I have artists friends and dont want to use AI in a way that shits on them. I'm using 'img2img', personal use. i might be interested in trying to sell things in future

[brazen fox]

you are correct

[magic widget]

It's a huge bowl of tangled emotional noodles

[fervent thunder]

No matter what im still gonna use ai

[signal nest]

i see people calling ALL sd users "petty criminals" for example, which i think is rather unfair

[magic widget]

The underlying problem is that our society doesn't give a shit about artists or labor.

[fervent thunder]

Im not against artists at all.
But some of them need to chill out and stop crying about ai

[magic widget]

Most people don't realize that intellectual property isn't our friend. It doesn't work for us.

[fervent thunder]

Its 2023.. isn't it already a great achievement we created such technology

[magic widget]

We also get an inflated sense of self -- I CREATED THAT! Well, you didn't create it in a vacuum.

[brazen fox]

it's just an argument as opposed to the main issue- which is fear of the loss of work

[fervent thunder]

True

I mainly use SD by drawing, then i use img2img, and edit my stuff

[magic widget]

https://pbs.twimg.com/media/FkCQHBHacAAcn73?format=png&name=small I made this a while ago.

It sums up my feelings on it.

"Twitter artists after completing a tutorial."

[fervent thunder]

I saw some death threats by artist's

[magic widget]

Yeah, that's fucked up.

[fervent thunder]

All because someone used ai

[magic widget]

We need to vociferously fight that shit.

[fervent thunder]

Yes

[magic widget]

At the same time, I think people should be mad about all the closed models.

[fervent thunder]

Yes!

[sleek stratus]

Correct, that's what seems most reasonable to me

[magic widget]

Corporations shouldn't be allowed to suck up every available public byte as they please and keep the result private, especially when it's all built on research done at public institutions

[signal nest]

opensource seems win win to me. value creation, posative sum. "ok we used your images to make this, in return we give the model back to you"

[fervent thunder]

It's crazy how google also doesn't release their text to images cuz they are "too dangerous for the public"

[magic widget]

It's bullshit.

[fervent thunder]

If they are, why announce it?

[magic widget]

It's about control.

All the huge corporations are maneuvering in this space, trying to figure out how they're going to try and dominate it.

[rotund umbra]

well, IMAO I just will stay tuned for corp not to take out AI art from commuity

[magic widget]

In that sense, I think the achievement SD represents cannot be overstated. It's a model the public has free access to, can tinker with, can extend. It's incredible.

[rotund umbra]

Will It be salleble or not - this is another discussion

[fervent thunder]

I lost faith in openai long time ago

Im glad SD is public

Im not glad that the ai war is going on

[warm junco]

Batch Count 3, never use Batch Size with your vram

[fervent thunder]

Why do artists even care.

It doesn't harm you

Just draw and ignore it

[signal nest]

jobloss

[magic widget]

I really hope they do well economically with it -- I hope they choke out the competition and force others to go open.

There's a BOOMING ecosystem going up around it, so it very well may happen

as long as a bunch of fucking politicians don't screw it up

[signal nest]

this is obvious , fear of jobloss. they need to focus on 3D IMO

concept art is a role in gamedev.. that is transitory, between ideas and the final assets

[brazen fox]

yeah people are complaining it's not as 'good' as MJ but they don't seem to see how incredible it is to have models like these open source that can be endlessly tinkered with on consumer hardware

[quasi yarrow]

Anyone have any luck using stable diffusion support making simple shapes like logos?

[brazen fox]

batch size is fine if you have enough vram, idk what you're on about

[magic widget]

For average users, I get it. But for power users, Midjourney is a joke.

[rotund umbra]

Well, I would add that AI art is progressing

[fervent thunder]

MidJourney is a joke

[rotund umbra]

I have 6GB VRAM

[signal nest]

right a model in the public doimain that can be endlessly finetuned

this is just awesome

[magic widget]

Truly. There is no parallel.

[brazen fox]

It's better at some things, less good at others- I sometimes use it as a base to train embeds

[warm junco]

yea ik, but i tought she had like 4gb, so i never recomend batch size for low vram

[brazen fox]

6 gb is still on the low end- you were right at that

I tend to forget that I'm spoiled with my 12

[outer crescent]

can confirm, 6gigs makes SD cry

[magic widget]

There are some really impressive private models that can do incredibly detailed prompts, but I still think they're junk and I won't use 'em until they're public.

We can build new products with Stable Diffusion.

Fuck all those other toys

[brazen fox]

I just wish DS supported embeddings

[brazen fox]

The tinkering and customisation is what makes SD so much fun- I honestly grew bored of DALL-E and MJ pretty quickly when they were made accessible to the public, yet I'm still on the embedding grind after like 2 months

[signal nest]

its writing your own scripts and tooling for me

and the volume you can run locally

[magic widget]

And knowing some ass hat suit can't snatch it out of my hands because quarterly profits need to go up

[signal nest]

but as soon as i get a new PSU i can start on finetunes

[magic widget]

We are met with a tremendous opportunity by AI: it shows all the contradictions of our society.

People have to twist themselves up in pretzels to explain why one thing is acceptable and another isn't -- until they're forced to think hard about it and realize that the story isn't all that it seems

Intellectual property and corporate control are a total scam.

The mouse always wins.

[bleak wolf]

Is it fine if i share my patreon here? Sci-Fi Diff already cost me a lot and I dont think 2.0 will cost me any less 😅
Or would that be too much to ask?

[brazen fox]

No but you see, this time they're teaming up with Disney and the Copyright alliance for the artists' benefit! /s

[outer crescent]

see you coulda went for The Mouse always wins. and got a good pun in there.

[magic widget]

I totally empathize with artists who feel something terrible about all of this, because they should. Society isn't built for us, it's built to coerce labor out of us

But we have to recognize the problem. Taking more tools out of our own hands is the wrong move.

[bleak wolf]

I dont think you guys would mind, so if anyone wants to contribute to the development of v2.0 and future projects, please consider supporting me:

https://www.patreon.com/user?u=86594740

[outer crescent]

i feel there should be a shameless self promotion community hub somewhere, to post like latest vids, things like this. i dont mind personally no but i like when things have a place on a server

[candid oriole]

Where do I see my prompts?

[bleak wolf]

I absolutely agree, feel bad posting it in here

[outer crescent]

nah you're actively contributing and doing something that takes time and is used by others

youre not shilling elsa from frozen with an ol meat an two veg or whatever 😄

[bleak wolf]

😄

[outer crescent]

heres a thing though, you wouldnt want this server and by extension stable diffusion to give the nod to all fundraiser/shilling because by not deleting someones stuff theres almost a tacit endorsement

[narrow sluice]

And also to not turn into one of those sad messes of spammy Discords where everyone is promoting stuff, and nobody is giving back.

Cough, whenever artists share art Twitter links on Discord.

Spam makes it so everyone's stuff just gets lowered by proximity

[outer crescent]

its probably almost better not having a fixed rule and just quietly letting through productive things and putting awful shills on blast, otherwise someones entire day would be filled with deciding what is and isnt legitimately worth promotion and all the glorious drama that would come with it. that's why i say a little self promotion thing, but maybe only for channel boosters. that'd at least filter out people who join, spam and leave.

anyway, not my monkeys not my zoo, good luck with your stuff though @bleak wolf

[narrow sluice]

There's folks that don't talk at all and just use servers as an ad zone

Admittedly, I'm not a fan of that approach, since sometimes these folks do put out cool stuff but I don't know who the beep they are and why they're doing this

It's kinda cold, I would love if people would put in a wee bit more effort than a simple link unless that's what people are requesting.

But to one's own.

[outer crescent]

yeah like i do like seeing showcases and love when people put up videos and tips and tricks.

[spare moon]

they really used a ai rendered music video lmao
https://www.youtube.com/watch?v=P10HOvTNPbk

[outer crescent]

i really like the non professional AI art video stuff. it reminds me of early 90s janky greenscreen vids in a good way

in time it might actually let older musicians remaster or reenvisage music vids from pre 3d stuff

[spare moon]

also really cheap to produce, only 1 dancer in front of a greenscreen and ofc ai lmao

[outer crescent]

yes!

https://www.youtube.com/watch?v=87nkJquHnAU

(mark gormley, an old meme greenscreen vid)

[spare moon]

The body of a science teacher, the voice of an angel

love this comment

[outer crescent]

saw that and was about to post! lol hivemind

just thumbed it up

some pretty good covers ended up getting done of his song. but yeah this is the stuff i get a nostalgic feel for with AI vids at the minute. much more charm than polished stuff

[spare moon]

when i could time travel to that time, all id do is to wait and buy btc

as much as i could

[dawn yarrow]

hinsight is 2020

[outer crescent]

@fervent thunder you're into music and stuff. you should make your own AI The Uncharted Zone 😄

[spare moon]

but i dont look nearly as cool as the science teacher

[outer crescent]

no but you cool be cool as the space happy phil thomas cat

[spare moon]

promt: science teacher
negative: not science teacher

[outer crescent]

i forgot! i have the poolsuite model. its 90s-tastic

i think gormley was made in 2000s but definitely has the same 90s qualities

[spare moon]

lol

basically if it looks trash it is from the 90s

[outer crescent]

https://huggingface.co/prompthero/poolsuite-diffusion trash as hell but totally confident and chic expressions

[rocky beacon]

any tips for merging models?

[fervent thunder]

i wonder if outpainting will progress to the point where we can turn movies and television recorded in 4:3 into 16:9 (other than cropping, obviously)

[fervent thunder]

it's really just trial and error until you get the results you want

[cursive gate]

We released a new model and feedback is welcome -> https://www.fyf.com/

[wheat hazel]

regarding training and batch size... I'm wondering suddenly, how does the batch is treated in the backprop ? is it the mean of the batch or is some linear transformation of the batch is done before handling it to the backprop ?

because if it's the mean, then a batch size of 1 is almost always better

[past vortex]

with embeddings, you can mix them with other models to get different results, right?

[sick wing]

how can I create my own ckpt model? I don't care if its going to take hundreds of hours

[lavish zinc]

hey guys, I create an add-on for Blender for text/image to 3D. Would love to hear your feedback! https://github.com/Firework-Games-AI-Division/dmt-meshes

[past vortex]

where is the link for the SD 2.0 model?

searches seem to just come up with the full install and whatnot, I just want the ckpt file

[brisk bloom]

https://huggingface.co/stabilityai/stable-diffusion-2

[past vortex]

thanks

hmmm, can't find the yaml file

nm got it

[sick wing]

@keen citrus you made the anything models right?

[severe field]

Alright youtube its Shout out time

[astral goblet]

albeit, for an extension for a1111's webui, I found an easy to diagnose bug! steps to recreate. select the "create inspiration images" script provided by the inspiration tab extension. Load a txt file that has a double quotation mark in it, ". The script then tries to escape all the slashes in destination folder's path leading to c:\\stable-diffusion-webui\\extensions\\scripts\\ etc.. etc..

Fails on file save. damn it's exciting to discover a bug

[outer crescent]

ooh i have that installed

[astral goblet]

i solved it by just deleting all the " in the txt file i was using

apparantly there's another bug where if a file has a slash in it, it messes wiht everything too

[outer crescent]

just putting a # in the prompt box shits it up a bit when it comes to save the filename. it changes all other characters you'd sanitize to an underscore. nothing nasty about it, just a thing

i do kind of worry about possible injection of stuff in png info, but seems fine

[astral goblet]

yeah i feel a little woozy running my instance natively. i might move it to a docker

[manic mulch]

hi

[astral goblet]

as far as i can tell, nobody has generated a downloadable image set for 2.1 yet. I think that would go a long way to shutting the debate about artists being removed from 2x up. There are a ton of names to prompt that get radically different styles out of the model. watching it generate has been fun times

[tame bloom]

are there any @stone charmds av?

mods I mean stupid discord

well if there is can a mod msg @frosty turret , he's having some issues that need to be helped with please. i also submitted a ticket

[peak orchid]

If mean the training dataset, they used LAION Aesthetics. If you have many terabytes of space, then you can download it yourself

Or you can dl smaller subsections

[astral goblet]

I mean the "Create inspiration images" script i'm running is catching a ton of names that have very unique and fantastic styles inside the model

[peak orchid]

Ah, I see

[astral goblet]

it goes through a list of names, or any text file, and plants them into a prompt. Uses those images to drive an inspiration browser

[astral goblet]

https://twitter.com/DrJimFan/status/1613918800444899328 badasss thread. if i'm understanding this right, you can encode logic and code into models

step towards self improving models?

[hardy zinc]

An tips on producing a group of people posing for a photo who aren't all clones of each other?

[astral goblet]

generate what the model knows and build on that image. 768x or 512x generally

[short tapir]

My god.. why do I get shit on by artists when I show an image generated? Like they foam at the mouth and have a full on tantrum

[astral goblet]

vocal minority

they represent goon squad more than they rep the craft

[short tapir]

Like. I play a web based game. I run a business of sorts. I make profile pictures for people in game for in game currency. (Costs nothing irl).

I generate images based on what they want roughly. Then I edit slightly, change minor details using Photoshop and stuff. I really like it. But I got shit on a lot for not making the art myself

Like I add personal touches. Stuff to make it incorporate there faction, there username, and other stuff they want.

Then part of the issue I don't understand is there are lots of other players that do similar but instead of AI generating the image they just use Google and find an image they like and then edit it.

[astral goblet]

Good hobby good craft. awesome energy

Photobashing, yeah. Same thing really. These are just new tools for it

[short tapir]

Yeah. And like. It's just a fun side hobby to earn me in game money. And it lets me make cool images for people

But everyone keeps getting so mad about it.

[astral goblet]

Yeah. It's probably 1 or 2 people running 5 accounts each, acting like a bigger group than they really are

and then 10miillion other people who arent

the silent majority are the real everyone

[short tapir]

Just sucks

[astral goblet]

toxicity always does

[short tapir]

I quite like being able to make cool characters and stuff. But then anytime I'm proud of what's made and try to share it with people I generally get shit on for it.

Just hope eventually people grow up eventually and stop throwing a tantrum over ai art

[astral goblet]

Yeah i wouldn't fight any of those manufactured battles. They're just inventing drama for theater purposes. don't participate

[cedar plank]

Any news on newly trained models for 2.1? Everything still seems to be made on 1.5

[ebon pewter]

any usable prompt that i can use to allow SD to ONLY COLOR my black and white lineart?

[astral goblet]

1.5 just has tons of momentum and a lot of people have been deceived into not making the upgrade. the tools for 2.x also haven't matured as well either.

Most of what comes out for 1x is just merges. I'm more interested iin refined 1x models that are more unique than the 1000 very similar merges that are coming out now a days

[astral goblet]

No but you can use the lines as a mask to tell SD where not to work

[ebon pewter]

😅 i try that suggestion

[astral goblet]

good luck

[viscid vector]

Can i have your guidance, regarding starting with machine learning and deep learning for GAN ?

I don’t have any experience with ML. But i am a React Native developer so i at least know how things will work, somewhat.

I want to learn more about this and help community build better AI and also implement some IMG-2-3D AI’s

So that building a 3D based on a top view or mix of other views like side/top/isometric view and create a 3D for it.

Please help if possible.

🙏🙏🙏🙏🙏🙏🙏🙏

[ebon pewter]

🥲 AI just ignored my mask and create better and prettier lineart than mine..

SD like "ur drawing is ugly this is how its done" 🥹

[outer crescent]

have you a good clip description of the piece?

if so, you could add "wet on wet watercolor" to the prompt. as a style/medium wet on wet doesn't really use lines at all and might give you the colouring in effect you want

another thing i guess you could try, assuming its black ink on paper, put white in as a strong negative prompt

[astral goblet]

oof. butttt.. sometimes it "blurs" the mask by default. i haven't figured it out lately, but when i was first beginning, i was sure i just loaded an alpha mask layer and it took it exactly how it was instead of bluring any of the edges

i'll look into it an @ you if i ever come across something specific

[ebon pewter]

well..mostly it is my fault...i dunno what these settings do and i just moving the sliders left and right without knowing what it does 😅

[chrome spade]

Happy Friday all!!

Saturday**

[hollow mortar]

Hey, does anyone know if Point E can be fine tuned?

Or, are there any 3d model/stl file generators out there, that can be fine tuned?

[elder cradle]

What are you guys' thoughts on this: https://stablediffusionlitigation.com

[signal nest]

"Hav­ing copied the five bil­lion images" if they managed to fit 5billion images into 5gb, they deserve a medal not a lawsuit.

[brisk bloom]

consent to copy and save images...lol

that case will fall flat

[lofty zephyr]

I'm happy they finally went and did it

this needs to be settled

in regards to their case, after reading it, best of luck

[fervent thunder]

Bloody hell

[fervent thunder]

why do people doesn't like 2.1 that much?

[neat condor]

i personally don't like it for compatibility reasons since large amount of community is creating their embeddings and safetensors on 1.5 not 2.1 and i hate to deal with all that conversions, giving it a year or so to become standard.

[lilac reef]

I went to 2.0, then 2.1, awaiting 2.2 to never look back, BUT I am in the minority according to the stats.

[neat condor]

yeah i did the same, went to 2.1 and when i heard cricket sounds there, rolled back to 1.5 👀

[lilac reef]

Not me. I mean everyone can stay with 1.5 and I will continue to release my embeddings (etc...) for 2.x.

[bleak wolf]

https://www.reddit.com/r/StableDiffusion/comments/10bj8jm/class_action_lawsuit_filed_against_stable/

[lilac reef]

"a collage tool"? LOL

tbh, I seriously know AI is going to grow in all fields (heck, there is now an AI robot lawyer who was just assigned its first case so no one is safe), but I think the luddites might win this one. Even if they don't they will never cease, and my faith in SAI is waning.

[bleak wolf]

SAI?

StablitiyAI? why

[lilac reef]

Stability AI

Just not liking the movements I see right now from them (no I don't mean 1.5 vs 2.x either). Not going to get into it but it is a gut feeling I have about them. I see things behind the scenes I just don't care for. They may do whatever they wish I am just saying even if they were victorious over the luddites I need to see some serious shit from them than I have. Anyway, we shall see where all this goes, but for me I lived before all this and will live if it fails as well.

[bleak wolf]

i see, respect your opinion

do not know myself honestly, just having a tool that allowed me to visualize my childhood imagination is fun and dont want to see it go because some coping artists will not be so "special" anymore

[lilac reef]

We agree as I can't even draw a stick figure but so many visions in my head. This tool was going in the right direction, but I am finding limitations that in time would be removed. The thing is I refuse anything that I can't do locally. If for pay, or only on the cloud then I excuse myself from it.

[bleak wolf]

SAI is going to remove the local/open source aspect?

[neat condor]

I think best thing any of us can do in this situation is just keep creating and not get dragged into this shouting match. I almost did a big heresy and expressed my opinion about all this drama, then i thought that someone with a knack for arguement would answer with something bitter, then i would get bitter and answer the same way and so on and it would just take away from the time i have for creating cool stuff.

[lilac reef]

I didn't say that I just see them chasing stuff, like Lexica now, as they see the money MJ makes. I think they may let it sort of wither on the vine, but never just remove it all.

[lilac reef]

Oh, yeah, I saw a couple of replies I wanted to jump into but refrained.

[lilac reef]

Internet and special olympics come to mind with that group.

[neat condor]

Yeah, unless someone is literally threatening to take away my PC where i have automatic1111 installed, i really see no point in arguing or "defending" something. Some people will hate it, some will love it, that's basically same with everything else so why would this be any different.

[fervent thunder]

This whole ai situation makes me crazy

[lilac reef]

In the end the courts will decide and one side is going to get plastered, butt hurt, and wake up to how the world works. No need to add to the noise.

[bleak wolf]

I am so done with arguing too, because all the people i encounter are emotionally driven air raid sirens.

[lilac reef]

Yes, precisely.

[neat condor]

don't fall into it rlly, this situation is very closely like to politics, once u get into it, your mood swings with each news and basically they got their claws in you which leaves u paralyzed and takes oh so much time. we have amazing tool we can use for free, there's huge artist community who creates something amazing everyday so all u do by following all this "non-technical" part of the ai debate is lose time u could've spent to just have fun with it.

[lilac reef]

You can't win a logical discussion when one side is pure emotion.

[fervent thunder]

I use ai and draw.
I mostly use ai for reference or just for creativity, (or just for fun)
I don't know why can't artist just accept this tool.
It's very useful

[neat condor]

is there even point in winning? i asked myself same question when i was arguing with my friend about religion and my conclusion was that they couldn't accept my atheism and i couldn't accept their faith without loosing who we are and why we were friends in the first place so we just kinda vettoed religion in our discussions.

[bleak wolf]

Mad respect

[neat condor]

i also come from drawing background, had to do it the hard way for last 10 years, now i do it the easy way. no sweat and i'm loving it.

[lilac reef]

Winning is that they can concede that they are thinking about this the wrong way. If you even attempt to shake their beliefs (akin to religion and politics) they get emotional and irl violent. You can't change them so let them eventually wither.

[neat condor]

best way to fight any kind of "oldschool" ideas is just to leave them be and let them wither indeed

[lilac reef]

You can't, no matter how many there are, stop progress. You may slow it down but it will eventually roll over all.

[fervent thunder]

Im glad Stable diffusion exists.
So much creativity

[neat condor]

I already started working on my game, as a coder with somewhat okay art skills, making card game solo was just out of my reach but now i already have all the artwork done and ready. Sometimes i still think i'm in a very fancy dream and i'll wake up from it.

[lilac reef]

See, a digital artist could load up SD and train on their style. Pop in some text, or do img2img using their style and save them countless man hours. Go in and correct what you like which would take far less time than the ground up. NOPE, they can't see the forest for the trees.

[wheat hazel]

Sounds like a very good friend of mine, we sometimes have discussions on topics we used to strongly disagree. And for a long time, these topics became more or less taboo. But as time passed, we found ways to discuss about it and still disagreeing but with so much respect that we often accept that there may be other ways to think and view these topics

Really glad to have someone like that around me

[fervent thunder]

I used stablediffusion since the beginning and i JUST can't see myself leaving it behind

No matter what i will use this technology

[neat condor]

same, if you have people who think like you around you it just gets... boring.

[wheat hazel]

same. Except I'm not using it to create game, but to illustrate my novel (and I'm thinking about making a comic, maybe)

[fervent thunder]

That's nice!

[wheat hazel]

yep. And tbh, that friend is really the best ally I can get to brainstorm ideas about my novel

[fervent thunder]

I just create characters with stable diffusion then give them lore and descriptions

A backstory sort of

[wheat hazel]

He won't hesitate to tell me "that's stupid" when he think it is, because he knows I can handle that kind of comment. And he knows it will lead to a much better version

[neat condor]

there's no possible way it will go away, thinking from a technical point, there's no possibility to ban ai art everywhere because there's no way to differ ai art from standard art that is upscaled via gigapixel ai for example. even if by some magic that happens and good recognition tool will be created though, people will create whole communities around ai art like civitai.com

[fervent thunder]

True

[wanton swift]

it’s just so good

[fervent thunder]

Yes!

[bleak wolf]

The court cases, if the judge has logic, and is aware of the tech, it will be over before it starts. If the judge doesnt know, its going to be that who can teach the judge better

The liars or us

[wheat hazel]

The issue I have with SD right now, regarding that novel, is that creating the images I really have in mind is sometimes a bit difficult, and also.... Well... The best models I've found to draw high quality characters are usually nsfw models. And I'm a little bit annoyed that the increased resolution on nsfw parts "ate" the resolution and creativity on other parts of the image (like, landscape, sci-fi, fantasy elements...)

[wanton swift]

sci-fi is my favourite and I really have a hard time finding a model on my linking

I usually like space related sci-fi

[bleak wolf]

uhm i made a model called sci-fi diffusion

LOL

[wheat hazel]

the code and models are on everyone computer. You can't remove them.

[neat condor]

I am also working on comics, it's kinda long shot project and slowly figuring out ins and outs. one thing that helped me alot is embeddings, embeddings, embeddings!

[wanton swift]

that is so cool

where can I find it?

[wheat hazel]

same for me, and same with space related sci-fi

[bleak wolf]

https://huggingface.co/Corruptlake/Sci-Fi-Diffusion
https://civitai.com/models/4404

[wanton swift]

It’s just something special for an ai to try and understand concepts that humans really don’t

[bleak wolf]

come to #🏞｜general-with-images for samples

i have got some better ones

[wheat hazel]

not sure if it can also help, but have you tried chatgpt (or better: gpt-neox-20b with nlpcloud for example, or sudowrite) ?

[wanton swift]

looks great! I am going to try it

[neat condor]

i constantly show my lore-book to chat gpt and ask to find plotholes, amazing tool indeed.

[wheat hazel]

I've seen it. But it doesn't work well to create high quality characters

[bleak wolf]

yeah the dataset is lacking on that side, you can try protogen infinity for that

[wheat hazel]

but I think it would be best to use different models for different areas of the image. But when it comes to character costumes, it's really hard

[wanton swift]

oh and futuristic city landscapes look so good

[wanton swift]

chatgpt is a tool great for anything

[neat condor]

i'll give it a try and leave a review aswell, looks cool.

[wanton swift]

I do some of my academic research using chatgpt

[wheat hazel]

in fact, for characters currently use a mix of uber realistic dreamy porn merge and anything v3. Until I find something better. Recently I've tried dreamshaper 2.52, and it's decent, but not as HQ as the urdpm+anyv3 merge

[neat condor]

so far best thing i did with chatgpt was writing a python tool that generates images which then if used with imgtoimg generate amazing results. 😄 it feels like that movie "inception".

[wanton swift]

I wrote a live chess game with it

and still adding elements to it

[neat condor]

live chess?

[wheat hazel]

do you know you can do a lot more ? for example, you can ask him to rewrite full paragraphs or even fill blanks in a paragraph

[wanton swift]

live by two players can play

express js

[wheat hazel]

chatgpt can generate images ?

[wanton swift]

the thing is it made the server side

[neat condor]

sure but thing is i have very cryptic writing style and i tried hard to teach chat gpt that but it kinda goes reeeeeeeeeeeeeeeeee

[wheat hazel]

I'm mostly using chatgpt and other ai writing assistants to kick start my first drafts

then I rewrite it to my taste

[neat condor]

no i wrote code with the help of chat gpt for software that generates images that are good for using as base in stable diffusion. 😄

[lilac reef]

I am not fond of chat bots

[wanton swift]

do you usually use any other chat bot?

[wheat hazel]

but, for some tasks, I find gpt-neox-20b much better than chatgpt

[wanton swift]

For example I sometimes use the openao playground with davinci

openai*

[wheat hazel]

for example, if you ask chatgpt to suggest ideas for character names in a sci-fi fantasy novel, it will suggest things like : paul, giselle, beatrice, brian...

[neat condor]

nah not rlly

[wheat hazel]

whereas gpt-neox-20b will give you pretty wild ideas and would even invent particles of an imaginary language to explain the meaning of names

anyway... I gtg, I come back later probably. My sister's here today.

[neat condor]

see ya ppl gtg aswell, was cool chatting with yall

[oblique wigeon]

https://www.reddit.com/r/StableDiffusion/comments/10babza/protogen_photorealism_dreambooth_guide_link_in/

[signal nest]

I still wish they'd make a model trained purely on photos to avoid this accusation of copying artists. (I wonder if its prohibitively expensive to train 2, one with and one without?). I hope the arithmetic .. "<=2bytes per image => obviously it doesn't copy" is persuasive, but i'm reminded of the expression "hope is a shitty hedge"

[lilac reef]

The price tag to train just one is 6 figures.

[signal nest]

emad's answer to this in the AMA was (a) it only breaches copyright if you ask it to, and (b) you can finetune material out of it. Perhaps one workaround is that they could train the main model and release a bunch of obviously safe finetunes?

whats the worst case scenario here - what happens if the luddites win the case

would the courts basically sue Stability.AI out of existence

would it be impossible to continue to distribute the model - or would the finetunes be an acceptable defence (and people would jjust judge them on a case by case basis as with any other digital asset)

for reference , what was the timeline of Youtube like? That had literal copies of copyrighted material for a long time, right..

[neat condor]

nothing

[signal nest]

another idea - is there a well presented "case for the defence", explaining (a) the technical 2bytes per image defence, (b) the ways in which artists should feel empowered rather than threatened (addressing the concern "unfairly puts artists out of work")

[fervent thunder]

Hey guys

[wispy jasper]

@wise stratus Guessing you guys are financially equipped to handle the bogus law-suits coming your way. In case you guys need donations, just let us know I'm sure the community would be happy to support.

[lilac reef]

If they truly are bogus then the courts will dismiss for sure.

[wispy jasper]

From what I read initially re the statements that were made legally regarding the new tech; the 'system' thought it wasn't worth pursuing, some technical term for it. Alas you never know.

[lilac reef]

I suspect the ludites will get a very old, and sympathetic judge.

[peak orchid]

If this bullshit lawsuit delays upcoming model releases, then I'm going to be very angry

[lilac reef]

Get ready to turn green.

[wise stratus]

Thinking maybe community support of education and outreach efforts may be a nice thing to coordinate. Help folk create and understand

[wispy jasper]

Agreed. The lack of basic education on the tech is astounding and as usual where most of the misconceptions are coming from. (People fighting windmills). With that said, we need a single source (well) to draw from. Does Stability.ai wanna put together a simple white-paper or guide on communications?

[wise stratus]

Yeah we are making a docupiece and stuff too

TV starting soon think CBS tmrw

[wispy jasper]

I think part of the PR strategy ought to be messaging > Youtubers and podcasts.

It's probably worth sponsoring a few big podcasts ie) Jordan Harbinger, Joe Rogan? ie) Rogan would likely welcome you on the show to speak about the tech, I don't doubt it. Harbinger would too. They love to clear misconceptions (they are both very curious about the tech but have a very novice understanding)

Your team should reach out.

[fervent thunder]

I heard some Luddites are already attacking chatgpt

[wispy jasper]

The way get to people today is through the tubes.

[lilac reef]

Get on Rogan. YES!

[fervent thunder]

Proves they want ai technology gone

It will never be gone

Joe Rogan, nobody gets fooled by clickbait more than that guy

i always wonder if its ever gone dawn on him that no the libs havent gone crazy, hes just gullible as hell

like anyone on the right i guess

[wispy jasper]

You're joking right?

[fervent thunder]

https://www.yahoo.com/entertainment/joe-rogan-fooled-hoax-tweet-203311304.html?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAACyVuLzoolqpHIcQZWTOSuVmYEzQkeuV1nkR4o67noDXIOGUYxfMAcGyIW9K7pEzWKEswx-5sJONaLGf6QE2syrroUMfLL96ZPIUjTOHMWFTpWFIZml5VMXEqoJrhqpPiE5Jr7OxPg7VEW0bp8_oMqTb5HCZ9zKdkFTEGiq_VUoL

https://www.bloodyelbow.com/2023/1/6/23541566/ufc-commentator-joe-rogan-fooled-again-fake-tweet-doctor-covid-vaccine-opinion-news

your master communicator is a dumb ass yeah

but but hes just asking questions

https://www.youtube.com/watch?v=OFm9zM8nX7I

[wispy jasper]

Rogan has indeed been fooled, but he's also owned up to it and corrected it. He deals with a ton of info and guests.

As for libs going crazy, it's been shown through copious studies (left pole principal) that the left and MSM has drifted further and further left.

[fervent thunder]

can keep going

the guys a dumb ass

[wispy jasper]

Oh boy...