💬｜general-chat

[pearl ocean]

apparently some people purely use SD for the babes

[surreal agate]

hihi

how can i get started with stable diffusion?

[karmic cedar]

I want to know how I can stop 🥁

[shell tendon]

all i can say is the answer is definitely not known by anyone here

[pearl ocean]

hammer

[split kestrel]

why do you need video?

for folks like me

[coral needle]

hi

[wet dawn]

hi

[jade bear]

Has anyone done like nes pixel art style ai art?

(NES sega genesis)

[shell tendon]

Yeah

Search civitai

[pearl ocean]

Civitai can change a life!

[grand dome]

hello,How to use this to generate images?

[echo aurora]

Hi evening. Anyone know how to make override settings appear on the main UI? thanks in advance🙏

[nova zodiac]

override settings appears in a1111 when you use png info on an image and then send it to txt2img or img2img and there are settings that differ from your setup.. If you don't get a override settings, then your settings match and it should be fine

[nova zodiac]

You don't, the bot is down temporarily

[echo aurora]

then mean we could not find any custom override settings in setting tab to turn it on?

[nova zodiac]

to turn what on?

the only reason the override appears is because the image in the png info tab was made with slightly different settings to the ones you currently have - if you send an image and the override doesn't appear, then your settings are the same and no changes required

[echo aurora]

i mean is there any setting to turn the "override setting" on at the first place? I could not find any place to turn it on. What you said it only appear if we use png and send it to txt2img?

[nova zodiac]

that is correct.. the only reason an override would be required is from the PNG info screen, otherwise you just change the setting for yourself in the settings tab

the override settings provide a temporary adjustment so you can recreate an image

[plush mica]

Im trying to find a good guide for finetuning, if anyone has a source or care to educate me please dm 🙏🙏

[echo aurora]

hmmm, then there will be a problem. without "downcast alphas_cumprod:true" and "face restoration:true" in the override settings, the face of the model smeared. What I did is use png info send to txt2img, after i got the override setting, i change the prompts, its troublsome🤔

[nova zodiac]

thats easy then - head to the settings tab and search first for "downcast" to change that setting, and then for face restoration you can enable that for any given image

to me that sounds like the face restoration was doing the heavy lifting though - instead of using face restoration, I would recommend the adetailer extension (made by bingsu)

does a way better job

[echo aurora]

ok, thanks for the advice, only found there is a tick beside "downcast model_cumprod to fp16 before sampling. but face restoration no where to be found. I am using SD 1.8.0_RC

[nova zodiac]

might be hiding under GFPGAN

[marsh steeple]

New text2image model by Stability AI dropped (v-pred SDXL, as far as my tiny brain understands)

https://huggingface.co/stabilityai/cosxl

[pearl ocean]

lol, I remember when people said, A.I will replace Developers XD

[nova zodiac]

Yeah Its the sd2.x stuff iirc

[marsh steeple]

Also I later re-read and there is another model in the same repo that should be for img2img (likely similar to UnClip I guess) but I'm not sure how's supposed to work since the text encoder with projection was already embedded in the base SDXL

[bleak matrix]

Good morning, everyone! How are we all this beautiful morning?

[finite cloak]

the Edit model is ip2p

you eg give it an image of a person and prompt it "give him a mustache" and it automatically figures out how to mustache him and leave the rest intact

[wraith notch]

i want to know how to use

is there anybody to teach

[thorny bloom]

What prompt should I put so that a little skin can be seen between the shirt and between the skirt or pants, a small line?

[trail lion]

crop top?

or photoshop a small area with a clone brush where you paint skin color and then put it back in img2img and inpaint it to make it look more natural

some things you cant just prompt for, you have to help it see what you want

[shell tendon]

it's really impressive.

[jade mason]

SD3 when?

[charred mesa]

oh my god its mcmonkey

[charred mesa]

2-4 weeks (CTO promised 4-6 weeks 2 weeks ago)

I wish it was at the end of April but by then my exams will come so I won't even get to use it that much lmao

[sweet flint]

can anyone help

https://tenor.com/view/frieren-wow-elf-peek-a-boo-gif-12265100463579712545

oh. why do I keep forgeting

no gif permssssssss

damnnnn

[static cape]

The link is dead.

[finite cloak]

(please wait, will be fixed soon)

it's back

[ornate flame]

Was it a promise? Or did they just give a rough estimate?

[trim magnet]

pinky promise so u know its real

[charred mesa]

rough ETA, that's why it's a range of weeks instead of a "uhh like 17.216 days guys"

but we can all be pessimistic about the release

the smaller models are done training, they are only going through 1024px fine tuning atm

[astral goblet]

about 3

maybe 5

[charred mesa]

I wish it was the end of April, but it will probably will probably come Mid May cause of some random delay

[sleek briar]

/imag

[charred mesa]

idk about the state of controlnets and model training

[sterile torrent]

Can someone share simple video 2 video workflow for comfy pleasE?

[astral goblet]

lol simple

[trim nymph]

forgot the name of the website that: scraped way back alot of stable diffusion generations+prompts.

anyone can point the name out to me?

[sterile torrent]

i mean without iadapte controlls and other groups

[finite cloak]

there's tons of sites that do that, i think civitai has become the de-facto go-to for that nowadays

[trim nymph]

i found it: lexica

[finite cloak]

yeah lexica was one of the oldold ones

they had a heaaaaaavy style bias

i made an sd1 model on their images to replicate that style bias over a year ago https://civitai.com/models/3798/lexica-testica

[warm bane]

Hi guys, what is the most up to date guide on how to set up a LoRA nowadays?

[karmic cedar]

how is it that Danny Elfman and Carrot Top look so similar

[astral goblet]

autoregression is so hot right now

[stiff obsidian]

I find AI art very interesting. However I want to gain more insight into why people use AI to produce art especially from those who use it or advocate for it. I would really appreciate a quick response if possible as I am doing some research.
Here are my questions:
Do you think AI art is real art and why?
Does AI have any consequences or benefits for artists?
What is real art to you if you could define it?
Do you think the prompt maker of AI art is given credit for the creation or does it go to the original artists, the machine or the company / developer that owns the machine?
Does AI art have copyright or is it public domain (can anyone own what is generated)?

[rugged mirage]

Yes, yes, creative activity, joint credit, depends but it can have copyright in many cases

[stiff obsidian]

If you use AI to make art is their a reason why?

[rugged mirage]

I can create things quickly that otherwise I can't create at all or will take me years to create

[stiff obsidian]

Do you think AI will be the future and why?

[rugged mirage]

is photoshop the future? part of the future at minimum

[oblique elk]

Well look at the roles distribution on the server (at least from those who set up their roles). About half of the people does not have a classical or digital art background. That does not mean that we aren't creative but until now we were not able to create art thru technology.
It is like in early days when digital cameras become main stream. It was easy to take pictures without the need to develop then. People with technical background startet with photography because it became easier.

[stiff obsidian]

thank you

[stiff obsidian]

Thanks for this thought I never really thought about this progression

[karmic cedar]

Creativity is also a matter of energy economy. This goes right down to the metabolism of your own mind.

To externalize it is to outsource

[stiff obsidian]

Could you explain this a little deeper. Do you mean that through AI you are able to spend more time being creative.

[forest trout]

To answer your questions:

Do you think AI art is real art and why?

What is real art? Does it require effort? Does it require human involvement? If your definition is that effort is required, we inevitably reject postmodern art. I personally reject the idea that effort = art. To me the only requirement is "do I like it?"

[karmic cedar]

Not more time, but rather you amplify your own inclinations through a medium that requires far more power than your own body can generate.

[stiff obsidian]

Thank you @forest trout so the pleasure the observer and artists gain is more important than the effort put in.

So would you say AI allows you to share your creativity like for writing prompts etc.

From a tecniquel stand point would you also say that making AI art you enjoy takes skill to polish it etc and the joy comes from here too.

[forest trout]

Contrary to what the anti AI art crowd claims, there's plenty of ability for personal expression that you can implement in generative art. There's a user here called Clownshark Batwing who adds clowns, sharks, spooky stuff into almost all of their images and every time I see their images I think: "yep, that's a Clownshark Batwing image."

[stiff obsidian]

Oh that's cool so people implement their own personal touches

[forest trout]

There's also a lot of different techniques that you can pick up that isn't just typing text into a box. You can provide starting images for partial diffusions.

I've been helping host daily img2img challenges in #🔆｜dailies if you want to see what we've been doing there.

[stiff obsidian]

So people also edit maybe their own work or images using AI

[forest trout]

Yeah there's a technique called in painting that let's you diffuse small parts of the image to give you more precision.

Sort of like Photoshop.

In fact adobe added generative inpainting to their software that basically does the same thing.

[stiff obsidian]

I've tested AI a couple times but I didn't know that it can be used in many ways. Thank you for discussing this topic with me it really opened my eyes.

[forest trout]

I just want to point out that effort absolutely should be valued too. I just don't think that it should be the only metric. I still love stop motion and I know that takes tons of effort.

[stiff obsidian]

Ahh that's a good point. I hope you have a good day!

[worn aspen]

Is all the energy going into NSFW models, etc., adding to speed improvement in this space above normal, in comparison to something just as interesting but not as tantalizing in the same way?

In other words, are all the horny f@@@ers speeding this tech along faster than it would move without?

Nothing against horny f@@@ers, and my input is in good humor.

Basing my questions on Civitai as a defacto resource and showcase.

In other thoughts: masking images for use with inpainting, etc.. I use Affinity photo, and it seems I cannot delete or erase to alpha. Does that matter?

[finite cloak]

on the copyright question, that's a legal question moreso than an opinion one, and the US Copyright Office has released official statements on the matter: https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence
tldr from memory: the direct output image from an AI image generator cannot be copyrighted by a person, however any modicum of effort applied on top of that can be

[stiff obsidian]

Thank you for your reply

[finite cloak]

also: "ai art" imo is a misnomer -- ai image generation is a tool. That's like saying "photoshop art"... nah, photoshop didn't make the art, a human did, they just used a tool

to be "art" rather than "an image" is a subjective definition, but usually some amount of creativity and expressing a message and etc. are required - so any random image out of a model isn't "art", but an ai model can be used as part of a creative expression which then is art

[chilly tangle]

Any news on SD3 release date yet?

[finite cloak]

nsfw creators do tend to push technology along, yes, as has historically happened in almost all tech circles, notably examples like online video software being pushed well ahead by nsfw needs for the tech (that one hub website you've heard of tends to have new platform features a few years before youtube adopts them). I've seen so many AI research publications that are definitely testing with pictures of underdressed anime women and the like

[finite cloak]

entirely depends on what AI software you're working with, some let you select a mask manually in-UI so you can just do that for the mask

[finite cloak]

the CTO tweeted that it might happen in a month or so. No promises yet

[chilly tangle]

Thank you for the reply man!

[finite cloak]

for training one? this UI looks pretty nice https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

[charred mesa]

oohh I used that

It's very nice and easy to use

[sweet flint]

guys

[charred mesa]

so it is still far away I guess

[sweet flint]

if im writing a prompt. can I write. wear tshirt or full sleves shirt or coat or jacket and then I do 100 batch count and the clothes being worn are random b/w these 4.

[finite cloak]

actually on this note ^ as well, cc @stiff obsidian -- In my time in the SD space I have met so many people that are/were artists, but were effectively unable to do their art (some that had grown too old, some that had hand injuries, etc.) for which AI image generation was exactly what they needed to get back into doing what they love.

[sweet flint]

or how to do something similar like this.

[charred mesa]

I remembered "4-6 weeks ETA" 2 weeks ago

2-4 weeks doesn't sound like "month or so"

and I know its just a rough ETA though

[finite cloak]

Yes, depends on what UI you're using. If using #🐝｜swarm-ui you can write that as <random:tshirt, full sleeve shirt, coat, jacket>

[sweet flint]

should i write

wear shirt, coat, jacket, full sleves t-shirt

or

wear shirt, wear coat, wear jacket, wear full sleves t-shirt

or

wear shirt or coat or jacket or full sleves t-shirt

what should I write If want batch count to randomly swap b/w clothes.

[charred mesa]

and Emad keep promising earlier release dates

[finite cloak]

2-4 weeks is "month or so" to me

[charred mesa]

I see

[finite cloak]

4 weeks would be a month, 2 would be "or so"

but also that's very rough, could be 6 weeks still

[sweet flint]

how do i check what i'm using right now?

[finite cloak]

i mean emad released himself from the company so...

[charred mesa]

oh I thought "or so" means "or more/two" like "a month or two"

[finite cloak]

means "or something in that range"

[sweet flint]

im using automatic 1.1.1.1 thingy. does that count

[charred mesa]

ah thank you

[finite cloak]

that's the correct name of the UI and the thing to specify when asking questions like that ye

[sweet flint]

is it the same as stableswarm UI?

[finite cloak]

no

[sweet flint]

i see

[finite cloak]

Swarm is swarm, auto is auto

[sweet flint]

but its a good one right? or should I switch

[finite cloak]

comfy is comfy but swarm is also comfy, lol

Swarm is the best

[sweet flint]

damn. I don't see automate 1.1.1.1 among channels lol

[finite cloak]

(I am biased in saying this but I am also right)

auto was big in the early days but hasn't kept up since XL era

[sweet flint]

guess. I'll switch once I get enough hang of stable diffusion in general

[charred mesa]

swarm is comfyui but with a simpler gui right?

[sweet flint]

right now. I don't wanna mess up my stable diffusion my switching UI

[charred mesa]

I saw it and it seems like it

[stiff obsidian]

Thank you again for this perspective. I hope you have a good day!

[sweet flint]

so what should I do in automate 1.1.1.1

should i write

wear shirt, coat, jacket, full sleves t-shirt

or

wear shirt, wear coat, wear jacket, wear full sleves t-shirt

or

wear shirt or coat or jacket or full sleves t-shirt

or should I switch to swarm

xd

[charred mesa]

try #📝｜prompting-help

I haven't used auto for a LONG time

I only used it up until 2.1

then with SDXL I switched to comfy and never came back lol

[sweet flint]

damn. ur making me feel like a caveman by saying that

[finite cloak]

lol

[charred mesa]

naah

I'm not a higher being just cause I use one with nodes

[finite cloak]

if you're used to comfy, then swarm can be thought of "comfy but with added quality-of-life/convenience/ease" yes

[charred mesa]

I just got used to stuff like blender and olive where nodes are king

epic

might be a really good ui for people wanting to try out SD3 but are too shy to try Comfyui

I might switch sometime in the future since its just comfyui backend anyway

[finite cloak]

you should just go for it

[sweet flint]

ok. so if im looking for a switch from automate 1.1.1.1 to swarm. how should I do that? also im a newbie. is this a good thing for me to do?

[charred mesa]

well

[sweet flint]

its not more complex or something.

right?

[finite cloak]

don't you want a workflow browser+quickswapper in ui owo

[charred mesa]

yeah that is very awesome

OH WORKFLOW BROWSER

YESSSSS

[sweet flint]

lol. is that swarm. that seem scary.

[charred mesa]

does it work with comfyui manager?

🙏

[finite cloak]

that's the comfy tab for crazy people like dark

[finite cloak]

yes

[charred mesa]

YESSSS

this is so epic

[finite cloak]

for normal people you'd want to the main Generate tab and not the comfy tab -- you can just follow the install info in the readme https://github.com/Stability-AI/StableSwarmUI?tab=readme-ov-file#stableswarmui

[charred mesa]

I suppose comfyui also updates separately from stableswarm and everything

[finite cloak]

for the most part it sets itself up. You can even point it to your existing auto webui models dir if you have a lot of models already

[charred mesa]

I mean its really just a git pull I guess

[finite cloak]

Yes. Swarm has an option to autoupdate comfy (but you can just uncheck that if you want to manage it yourself)

[charred mesa]

oh very nice

guess I'm switching

[worn aspen]

Thank you Alex. from a photoediting perspective, or also in the case of depth maps, I wanted to do precise mapping as opposed to scibbling in comfy. Maybe precision doesn't really matter? I have to sort out some memory issues before I can dig into that more.

[fervent thunder]

wasn't SD 3 supposed to get released today

[charred mesa]

what

where did you get that info

was it from a guy here who said it was "the next solar eclipse" or wahtever

cause that was a joke

[fervent thunder]

emad i think

[charred mesa]

emad?

[fervent thunder]

https://cdn.discordapp.com/attachments/193278651020738561/1170472696866414602/bruh.gif?ex=65592aa5&is=6546b5a5&hm=ea7e0796a99c51bbffdb83779ead4f6f287d1403baafc0a8e2ccbf9c2836d99b&

if i was sure i wouldn't be asking

[charred mesa]

no its like he always promises stuff earlier

and can you link it or something

I only remember him saying stuff like "soon"

"couple of days"

[fervent thunder]

no

[warm bane]

thanks!

[warm bane]

is there a manual on how to install that exactly 😂

just click the install bat file?

[finite cloak]

ima be real with ya chief, if you can't figure out the answer to this question yourself idk if you're ready to train loras

(hint; open the page, and scroll down approximately two millimeters to the giant "Installation" heading)

[trail lion]

tough love, we need more of that around here

[karmic cedar]

Tough is good, petty isn’t

[astral goblet]

tom petty was a rock dream. rip !

[karmic cedar]

He hit on my friend’s mom at a doctor’s office once

[astral goblet]

she had it coming. he's a rockstar

[karmic cedar]

True.

Though he didn’t come

Twas a doctors office you see

[astral goblet]

didn't come round ther no more

[karmic cedar]

Last dance with Mary sue

[warm bane]

I did everything and after I innput the last line I'm getting this error

oh I can't send screenshots here

[karmic cedar]

Images channel

[warm bane]

but basically "install bat the term install bat is not recognized as the name of a cmdlet blabla

[karmic cedar]

Have you watched any YouTube videos on LORA making?

I highly recommend getting a thorough primer on this process.

[pearl ocean]

Have you even met Lora?

[karmic cedar]

That’s the name of the secret anime big tee tee goth gf that hides in the latent spaces of all these gigantoid models and she only becomes more secret and more amazing and more big tee tee goth gf with each advancement in model fidelity

You can summon her only with a single precise token

She has hooks for feet. HOOKS FOR FEET

She’s the noob saibot of diffusion

Boobs SAI bot

(I’m done now, thank you for reading my ridiculous idea train)

[sweet flint]

what should I use for higher no of generations if I don't wanna lose quality. batch count or batch size?

[trail lion]

"which key is the "Any" key?"

[outer crescent]

the one that has a light on it on your PC. hold it in a few seconds

[timid fern]

Is anyone familiar with DreamBooth for SDXL? I want to train my custom checkpoint with ~1000 images

I would just use the default DreamBooth training script, but looking at the diffusers documentation for training on huggingFace, it seems like there is another option? Dreambooth: https://huggingface.co/docs/diffusers/en/training/dreambooth?gpu-select=16GB

and then just SDXL directly? https://huggingface.co/docs/diffusers/en/training/sdxl

[trail lion]

you can do dreambooth,most of us use kohya_ss. if you go much above 1000, then you may want to look at other options more suited for full finetuning

[timid fern]

do you know what the second option is? I got the vibe that the second option was for "more full" finetuning, but I have no evidence for that

[trail lion]

yah, it's just called 'finetuning' in kohya, there are other options such as onetrainer

I have no experience with it, but I know people use it

I'm more in the 100 image dataset lora camp, personally, if you want to know about that, I can talk all day 😄

[timid fern]

yeah I've trained a bunch of loras with kohya, but more looking to make my own checkpoint now

[trail lion]

training a dreambooth checkpoint on a subject/style is the exact same process as a lora, but if you want something on par with like juggernaut or something, that's way out of my league

and I also suspect, but have no way to know, that it takes way more that 1k images to do that kind of finetuning...like 200k maybe

[timid fern]

That's what I assumed too, but then I saw what the creator of RealVisXL posted on his model page. Check it out: https://civitai.com/models/139562?modelVersionId=361593

These are the change logs:
"""
RealVisXL V5.0 Status (Updated: Feb 15, 2024):

• Training Images: +0 (V4.0: 3340)
• Training Steps: +0k (V4.0: 672k)
• Approximate percentage of completion: ~0%
  """

so seems like it's "just" 3340

[pearl ocean]

DallE just can’t compete with SD

[trail lion]

DallE isnt open, so I agree

I would use a lesser product just for being open

[craggy arch]

On a spring morning, birds chirping and flowers fragrant, the sun shining brightly

[trail lion]

wow, that was relaxing

that would make a good prompt

[karmic cedar]

I think Midjourney needs a new model called Highadventure that just does 90s CGI

But that’s just like, my opinion, man

[pearl ocean]

I never used this, probs coz it wants you to pay 😦

[astral goblet]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/release_candidate/CHANGELOG.md new version

[pearl ocean]

SD3?

[astral goblet]

automatic1111 1.9

[stone glacier]

Apologies if this is answered somewhere else but I wasn't able to find this answer.... How would one go about training a SD model from scratch?

I understand that this is fairly expensive and difficult to do, but is there an explainer that shows how it is accomplished?

[pearl ocean]

Do you guys prefer ComfyUI over Forge??

[grizzled harness]

Team Forge.

[spark lichen]

Hello
isn't there anyone who wants to build web3 project intergrated with AI?
nowadays, BTC ordinals and upcoming Runes Protocol is trending.
if you want to build your own web3 project, as a web3 full stack developer, i can contribute my skills.
plz feel free to ping me.
thanks

[turbid walrus]

How to use SDXL on my pc?

I'm just a nerd 😄

[foggy halo]

guys when sd3

[pearl ocean]

maybe GTA6

but im told 6 weeks

[karmic cedar]

i really just want a gamecube or even an n64 to crunch some big data and give us some large-ish language model action

[narrow kernel]

Bro the N64 had 8mb of ram... With the expansion pack... Though you probably could do some hacky stuff and interface in a big GPU through the dev port/bus thing. I think someone hacked ray tracing into an SNES doing something similar I wanna say?

[karmic cedar]

i mean generating 16x16 pixel art

yeah it could potentially be done…with enough brainpower

definitely not mine

[warm bane]

Hi guys, can you lead me to a simple beginner's guide to using the
LoRA_Easy_Training_Scripts app ?

I have it set up and all but idk exactly what to do 😄 I'm looking for step by step guide if theres one for it

[narrow kernel]

Right but the point is you'd need a model that can store enough information to be useful and it would have to fit within <8mb. So yeah, you'd have to offload the model onto some other hardware. Maybe with some raspberry pi built into a cartridge?

[karmic cedar]

it could be done…with the EXPANSION PACK

…i don’t know

[unkempt hatch]

years ago, before AI, I remember watching a video about using multiple images of the same thing to create a "super resultion" image

it made me wonder, could I take 3 video frames, next to eachother, and combine them to create a higher quality image?

because I'm building datasets, and wonder if I could take video screenshots, but using such a technique to improve their quality

[frigid falcon]

hello everyone. I'm new here and i want to know how to generate images with Stable Diffusion.

[karmic cedar]

https://github.com/ckkelvinchan/RealBasicVSR

this code works similarly, i think?

[pearl ocean]

I downloaded "pixel-art-xl-v1.1" lora for pixel art

[karmic cedar]

i’m going backwards and visiting 1.x models

[coral needle]

hh

[blazing falcon]

where do i download the software?

[pearl quiver]

Hi everyone, how can I generate images using stable diffusion??

[rugged mirage]

I'm scared of the kind of person who'd go out of their way to find and join a discord to write a question before they'd consider writing that question in google

[nova zodiac]

Bots, bots everywhere

[unkempt hatch]

interesting... but not sure how I can use that

[livid sundial]

When I use SD Forge, after selecting the model in the controlnet, an error is displayed. Why?

[rugged mirage]

1. it obviously depends on the error. 2. that question is for #🤝｜tech-support

[livid sundial]

thanks

[karmic cedar]

It’s a bit outdated at this point anyway. I would expect more time series-based video upscalers at the rate everything else has been progressing…

But alas

[unkempt hatch]

ya, like I feel a computer should be able to take 3 sequential frames and intelligently combine them into a single higher quality image

I don't want to just AI upscale, because I don't want to change the image at all

[fervent thunder]

So i work in sales doing construction projects on houses.
i want to learn photoshop or some form of digital design to quickly do mockups of the windows and painting services we do on houses and how their house will look like if they pick our service.
i was wondering what would be a good way to do this esp on a lot of houses?
would there be a way to do this with AI or would it need to be made manually per house? if manually that's fine i would prefer it i just need to do a lot of houses.
Would i need a 3d software or can i just photoshop our windows and paint on the house?
also in the future i would like to have a 3d visualization of rain/environmental damages their house can have
best way to learn how to do this?

[vocal leaf]

whats the best model for Frienen Anime style

[pearl ocean]

I wonder if there is a Pepe Model

[vocal leaf]

I saw a Horny Jail model

close enough

[nova zodiac]

I'm pretty sure there was a Pepe lora floating around at one point

[vocal leaf]

@pearl ocean its not that bad compared to what I download on Nexus

using Vortex

[pearl ocean]

😮

[vocal leaf]

THERES SO MANY MODELS OMG

[pearl ocean]

Hot models?

[vocal leaf]

howd you know

YO I FOUND FRIENEN FINALLY

[fervent thunder]

wheres the bot ?

[cinder magnet]

Hi,
Anyone familiar with a model that would help qualifying scraped companies as good or bad fit for my B2B SaaS product?

[wet dawn]

hi

[vocal leaf]

grass jelly wintermelon goated

[pearl ocean]

Generate an image of it

[still glacier]

#1047610792226340935

[tiny adder]

？

[feral stone]

Hello, can anyone tell me if it is possible to run stable dream in local mode and make images in it? if yes how?

[final heron]

Is sd a1111 save for work? I generate a pickture with a prompt "1man, full body" and it generate a man without shirts. I just afraid if i prompt 1girl,full body it will generate shirtless woman. is there a way to make sure it will not generate nsfw?

[fervent thunder]

Hey all 🙂

[foggy halo]

6 weeks wait like 6 years 😔

[karmic cedar]

https://var.vision/demo

whoa

fast and accurate

[low moon]

They like Discord mroe than google.

[fervent thunder]

im scared of the kind of person who goes out of their way to be a snotty asshole when they could just ignore a question if they dont feel like answering it.

You could have told him the answer with the same amount of effort

im sorry i dont buy into the RTFM attitude

[full lark]

Anyone else can´t view live previews in sd forge? It worked last week but somehow withouth touching anything its broken now (and I haven´t updated)

[full lark]

nevermind looks like Live preview display period got changed to -1 instead of 1

[rich kestrel]

or maybe some ppl realize google is a joke these days

and chatgpt isnt versed on newest tech

[fervent thunder]

or maybe theyd just rather get help from a person. and if you dont feel like they deserve an answer, dont answer

'people come in here all the time asking dumb questions' well they're going to continue to do that. its not off-topic!

[fallen gale]

@stablediffusion3

[spiral inlet]

any sign of SD3 yet?

[shell tendon]

Yeah it's virtually useless now

I don't really fault ppl anymore for just asking for help instead of throwing hours into searches

In the last 3 years in particular it's swirled down the toilet

[pine fiber]

https://github.com/TencentQQGYLab/ELLA

weights for 1.5 released

[charred mesa]

WOW CODe

and weights

holy crap

is cogxl only trash for me?

just tried it out with comfyui

[trim nymph]

thanks for pointing out

[pine fiber]

how did you test it?

[trim nymph]

in auto111

[pine fiber]

oh its just a text encoder so u can drop it in/

?

I was expecting it to be bigger

[trim nymph]

its an model actually

[pine fiber]

🤔

post results

[olive zealot]

Impressive!

[trim nymph]

one moment

gonna ping you once ive uploaded images, may take some 10ish minutes.

[vocal leaf]

whats a good model for Valorant artstyle

[static cape]

That's great on the one hand... but horrible that they didn't release their (existing) SDXL weights.

[pine fiber]

yea not their fault tho :p chinese board

[fervent thunder]

https://soundcloud.com/4dreamsy/blondies-and-weed my 3rd music track

[loud solar]

how is the selection of the model type in top of UI SD Forge called? Like to add it.

[fervent thunder]

Sdxl v1.0

[trim nymph]

https://github.com/TencentQQGYLab/ELLA

in which folder should the model be put?

auto111

[static cape]

I actually couldn't care less about the reason. It sucks. Also funk that board.

[trim nymph]

do you know maybe? Fawks

[pine fiber]

so you dont care but also fuck them?

[trim nymph]

[static cape]

#Oxymoron4Life

[arctic sedge]

Do you guys know where i should put the Ella text encoder?
It appears to be just a stand alone file. So i assume it's a text encoder model file?

[nova zodiac]

Id guess models/CLIP

[sweet flint]

ok so I messing up with batch count and batch size. trying to figure them out, except I'm getting the same times it takes to generate images

like its taking the same time to generate images whether I use

batch count 1
batch size 4

batch count 4
batch size 1

wassup with that?

[sinful tartan]

hi

[loud solar]

count = images created one after the other, batch=images created at the same time

[sweet flint]

yes ik.

but why is it taking the same amount of time. shouldn't batch size be faster?

[loud solar]

It's the same rendering time ...

4 pictures are 4 pictures ...

[strong cradle]

Hello guys, I am a beginner at Stable diffusion i am trying desperately to change my photo into digital drawing style, but even with many tutorials I couldn't get a great result...

Which way is the better ? Because I tried with img2img, txt2img with ControlNet but nothing get a great result as I see in youtube tutorials

[sinful tartan]

Have you made friends?

[loud solar]

Hard to say what your problem is. Maybe IMG2IMG play around with CFG Scale?

[strong cradle]

Yes i tried too, but it still too much different from my original photo

[arctic sedge]

I'm on Forge and i don't have a Clip folder. At least not in a place where i can see it.
But it all else fails. I guess i could try putting it into the Clip folder on Comfy with a corresponding workflow, and see what happens. Haha
Thx btw.

Maybe it needs a extention. (For Forge.)

[loud solar]

Still no solution here 🙂

[arctic sedge]

Yeah. I think it may need a extension to work on Forge. People really want to use it, so i doubt it will take too long for one to be developed.

[loud solar]

That's the problem working with early beta versions ... 🙂

You start from 0

[arctic sedge]

Yeah haha.

[loud solar]

But Comfy could be an idea ...

[strong cradle]

But do you think it's better to do it with txt2images with ControlNet or is it only when you want to keep the pose of your character etc. and not to keep the whole image ?

[loud solar]

To be honest that's not really the stuff I am doing. But you could also try to combine different ControlNets for a better result. For sure the model you are using is pretty important and your prompt, too.

I'm an old fashioned prompter. Don't wanna change existing stuff ^^

[strong cradle]

Ahah ok thanks a lot man !

[trim nymph]

how to change clip model?

once web ui started

[nova zodiac]

Yeah needs extension for a1111/forge and the custom node for comfy

[nova zodiac]

You dont - a1111 only able to load the clip from the selected checkpoint

[trim nymph]

i think this ella model cant be run yet inside auto111

[nova zodiac]

Unless there is an extension allowing a selection

[nova zodiac]

It cant

[trim nymph]

a friend of mine suggested its like an adapter

so guess we have to wait a few days/weeks till its implemented

[nova zodiac]

I still havent seen x-adapter for a1111 yet so not sure which will come first

[loud solar]

Yeah ... take over ... you have a better base of knowledge 🙂

[nova zodiac]

The wisdom comes from the collective, not the individual

[full lark]

Anyone knows how to transform a sdxl model into an inpainting one? I tried merging sdxl_inpaint (A), custom model_sdxl (B) and sdxl_base (C) in A + (B - C) * M using add difference but I keep having errors and the merge doesnt happen

[pearl ocean]

You’d need a pretty fast GPU for Batch then

[sage reef]

@woven panther hey i tried your ella wrapper for comfyui, it works fantastic ❤️ there is just one small problem. if i close comfy and reopen the workflow for ella, it redownloads the t5 sharded models again... even tho they are downloaded already... is there a way to make a simple fix so that it bypasses downloading by detecting the models already downloaded?

[loud solar]

I have ... but I only use batch count. Also see no benefit for the batch size ...

[woven panther]

huh, that should not happen... I dont like using hugginface cache but didn't have time to figure out better way yet

it's definitely not happening for me

[sage reef]

can you make it so that the shards can be loaded from the same folder as the ella 132mb checkpoint?

i can then just move the shards there

[woven panther]

yeah I do want to do that, just never really used these google t5 models before

[sage reef]

thx ❤️

[pearl ocean]

lol, when people put t then the number in, I keep thinking banking stocks when they have the T3 system

[sage reef]

ella generates so quick and it works under decent hardware too which is awesome

[loud solar]

Wasn't T5 a text input mode for mobiles?

[honest mica]

That was T9

[pearl ocean]

T900

[loud solar]

Ahhh... my old brain 😄

So still 4 generations to wait ... ^^

[sage reef]

sd3 hopefully by end of April right?

[trim nymph]

[sage reef]

@woven panther just a quick question, the ella sampler node doesnt have a previewer? i mean it generates fast anyway, but just wondering

[pearl ocean]

[woven panther]

nah it's just using diffusers

[sage reef]

kk np

[woven panther]

can you try now if the model loading works

[sage reef]

kk let me check

[woven panther]

actually worked nicely, it copied the model from cache to the folder I designated without redownloading

[pearl ocean]

AI models training other AI models when

[trim nymph]

you got ella working? if so in which folder did you put it

[sage reef]

it downloads automatically, thanks to kijai custom nodes 🙂

ok hold on

im gonna actually first delete the models to see if it downloads correctly and then see if it doesnt redownload

[woven panther]

I can't tell if this would be hard or easy to implement natively in ComfyUI, for diffusers they used some workarounds that I can't make sense of with my limited experience 😛

[sage reef]

ah now it works perfectly

but there is one more thing to fix kijai :3

[woven panther]

just one? 😄

[sage reef]

in comfy nodes menu, your ella nodes are for some reason split under 2 categories, why not one?

[woven panther]

because I rushed it 😄

[sage reef]

haha

i mean im happy it works ❤️

thx for all the hard work

[woven panther]

fixed, thanks for letting me know

[sage reef]

np

@woven panther sorry, just forgot to ask something :3 what is the max length and the fixed negative params?

[gentle dome]

How to access SD3

[charred mesa]

no

enter the waitlist, that's all I can suggest

otherwise, wait for the weights

[woven panther]

It should be the maximum tokens the t5 model can generate, in the context of ELLA, I don't know how it would affect things, 128 was the default. I dont know what the fixed negative does exactly myself, just decided to expose it to see 😄

[woven panther]

the max lenght wasn't actually used, now it is, affects results a lot, as does the fixed negative, not always worse either

[rich kestrel]

sd3 weights may never become public

too dangerous for the masses. like giving people an atom bomb. gens indistinguishable from real images. dogs and cats living together, mass hysteria

[thorny orchid]

hey

wich chat do i go to for support?

[brittle maple]

tech support?)

[thorny orchid]

ty

[sage reef]

@woven panther hmm is it possible you didnt implement ella correctly? i mean aside from the default workflow prompt with the red book and the flowers, everything else, even with Photon just seems so bad... like any human i try to generate outputs some cartoony looking pic.. i mean even testing all the other schedulers doesnt seem to give good outputs... i guess il experiment more, but so far it's not great quality wise. i was also looking at their research paper on arxiv, on page 12, they show a comparison between ella and some 1.5 models. one of them is realistic vision and they generated a human without ella, looking good, the ella version looks weird even looking at it in the paper... so this technology is perhaps still in it's infancy stage, needs a lot of improvements..

[woven panther]

native implementation btw: https://github.com/ExponentialML/ComfyUI_ELLA

but I can't get same results out of it yet

[woven panther]

yeah I had one mistake, fixed it though

had wrong beta schedule as I always mix them up

[sage reef]

well looking at that other implementation you linked, they seem to be doing extra stuff you are not doing i guess?

il try it out and compare

[woven panther]

I'm still getting better prompt adherence with the diffusers version

and definitely the real ELLA is the sdxl version sadly I think :/

but this does give huge boost for 1.5 prompt adherence still

[sage reef]

also their implementation is using a different t5, not bf16 rip 😦
but yea seems like the real thing is the sdxl version perhaps, but sd3 will have t5 capabilities anyway, so i guess we are waiting for that anyway haha

[woven panther]

the t5 dtype shouldn't matter

it works with the one you downloaded, just put it in the folder he designated

[sage reef]

but for example their implementation dont tell me it's gonna download literally the whole huggingface repo, cause thats a lot of GBs because t5 comes in different versions... or are you saying i can just dump the t5 i got from your version and it should work with their version? cause that would be awesome

il try it

[woven panther]

yes that's what I did

I have both running side by side

[sage reef]

nice

[woven panther]

I just put all the files in ComfyUI/modles/t5_model/flant-t5-bf16

and then you can select it from his node too

dunno if I should change the autodownload folder... again 😄

[sage reef]

ah nice, also what is the get sigma node, is that from blender neko?

[woven panther]

or just forget about this as native is obviously better in the long run

it's not like I have lack of things to do 😄

[woven panther]

it is

[sage reef]

you have a lot of repos on github aha 🙂

kk

ah nice it works with bf16 shards and also their version uses ksampler, so we have a preview too :3

now to see if the results are actually better with this version

hmm ok i think im starting to get maybe a bit better results with this implementation, just need to figure out better params for sigma and the ksampler i guess

ugh... idk... i give up for now LOL, results are still bad.. maybe il wait until someone shares a workflow that generates a photorealistic human and then use that workflow cause i have no idea 😦

actually now im getting good output quality, but it's not listening to my prompt, i want to generate a male, it gives me a female for example.. the hell

i guess il wait for sd3 lol

[young lintel]

Hello! I have a question, please, if someone knows the answer. If I pay the 20$ for the membership, can I have the video generations private? I couldn't find the answer

[heavy lark]

You can use the stability ai api to generate images for yourself.

[young lintel]

Thank you! It's a bit more technical, I think, than I was into until now, but I'll try to understand it 🙂

[heavy lark]

https://platform.stability.ai/docs/api-reference#tag/Generate/paths/~1v2beta~1stable-image~1generate~1core/post

[heavy lark]

If you go to that page and scroll down, they have a curl example so you can just generate using that on the command line along with your api key. No programming required.

[young lintel]

Thank you! Perfect

[nova zodiac]

So just wondering - thats taking the prompt, turning it into an embedding with t5, then using the ella model to convert the embedding to feed the conditioning/unconditioning??

[woven panther]

something like that yes

[pearl ocean]

what would you with $1,500?

[forest trout]

Probably use it to pay down a loan or something.

[shell tendon]

buy another 4090

[pearl ocean]

4090 loan

[nova zodiac]

you can get a 4090 for $1500??

They all over $2200 retail here 😦

[pearl ocean]

they are $3200 AUD here

[nova zodiac]

Yeah.. bout the same on this side of the Tasman

almost to the dollar

[pearl ocean]

a few years ago I went with the 4080, didn't want to pay an extra like $1,200 AUD at the time for a 4090

[nova zodiac]

Yeah.. my car is worth less than either of those 😛

[pearl ocean]

same

[nova zodiac]

#priorities 😛

[trail lion]

even 3090 ti still goes for $!500 - $1800 US. I just refuse to spend over a 1k for a computer part, and 30% of that is the Nvidia we're the best and we know it tax

[shell tendon]

just checked... trade in for my car is the same as what i paid for my 4090 ($1800)

it drives. it has airbags. it always starts. no mechanical problems. so who cares

[shell tendon]

they're $1600 at best, and that's only from nvidia direct and only if you're really, really on top of your "add to cart" game

plus tax, so really $1700 usd

[nova zodiac]

Has anyone tried the $200USD nvidia tesla p40 from aliexpress?

[shell tendon]

mine was $1800 usd (gigabyte oc), about $1900 after tax

[nova zodiac]

https://www.aliexpress.com/item/1005006419681213.html?spm=a2g0o.productlist.main.21.96a83a95GmpZVk&algo_pvid=76b78d8e-5a0c-4793-9b83-6449d5a3b323&algo_exp_id=76b78d8e-5a0c-4793-9b83-6449d5a3b323-10&pdp_npi=4%40dis!NZD!393.90!389.96!!!1690.55!1673.64!%402103200617127111681816263e5d62!12000037099677061!sea!NZ!199233445!&curPageLogUid=T0Cmz9z7WGkV&utparam-url=scene%3Asearch|query_from%3A

$250USD... my bad

but a 24gb gfx card with more cores than any of the 20xx series bar the 2080Ti seems like a bargain at that price

[pearl ocean]

I think I got that 4080, it was $1700 AUD for me

[trail lion]

this is a fun page if you want to see where your card ranks in terms of SD https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

[shell tendon]

So about $1100 usd? Sounds similar to here

[trail lion]

so a 7900xtx is sub 1k and lands in the middle

[shell tendon]

The gap is even bigger with SDXL

[trail lion]

dont doubt it

[shell tendon]

4090 was so worth it

I figured if it just starts on fire in two years, that's $2.45/day

Equiv to a coffee and a take out sandwich once a week

[nova zodiac]

Doesn't have the workstation cards on it though sadly

k80, m40, p40, p100, and t4 would be a great comparison set

[shell tendon]

3090 new is a waste of money right now

Might as well just get a 4090

Used 3090 might be worth it if you're cool with the risk of used

[pearl ocean]

Isn’t there a 4090 Super coming?

[shell tendon]

nope

[trail lion]

used or refurb, I've done refurb on a few things, but I know that doesnt always work out

problem with used is you could get a crypto mining card

[shell tendon]

yeah it makes more sense to just get the 4090 unless you're really cash limited imo

it's a big purchase at the outset, but if you use it a lot, and aren't just gonna toss it, it's not actually bad at all

goes back to the same thing as the takeout coffee math

easy to bleed thousands a year on dumb crap you don't really care about that much

[trail lion]

I see some 4090's new on amazon for $1800, so I agree, why buy a 3090, it's rediculous that they're not discounted at all based on being last gen

pulling an apple

[shell tendon]

yeah i just wont pay scalpres

that's my line

is nvidia ripping us off? hell yes they are, and i think they should be subject to some regulatory action for abusing their monopoly status

but that's just where we are right now

[trail lion]

they just have to point to amd and intel to prove they arent a monopoly, do they have pricing power? yah no doubt

[shell tendon]

for AI compute they have a monopoly

where's the alternative to the A100/H100? there isn't one

[pearl ocean]

I just have a simple 4080 😦

[shell tendon]

and so a card with about 3x the vram of a 4090 that is actually a bit slower goes for 10x the price

[trail lion]

4080 is 16G?

I have a 16G card, if I bought a new card, I'd get a 24

[pearl ocean]

The real question is, how many weeks of work will it take to buy

[nova zodiac]

how many hours of runpod could you get instead

[shell tendon]

think of it this way

what percent of your workday would be dedicated to that card over the lifetime of its use

[shell tendon]

not that much really compared to what you can do with a 4090 slotted into your machine

[nova zodiac]

I can get nearly 4000 hours of runpod for the same price as the card - would need to be 2.5 hours a day for 6 years lol

[shell tendon]

yeah i def run mine a lot more than that

https://www.runpod.io/gpu-instance/pricing says 0.49 spot 0.74 on demand

tbh not familiar with renting cards, whats spot mean?

buy a chunk of time that's contiguous?

[nova zodiac]

I have a 5 year old... unless it was a work machine, I ain't got that much spare time

[nova zodiac]

spot means they can kick you off

[shell tendon]

ahh, so... not great

[trail lion]

vast.ai is way cheaper than runpod

[shell tendon]

so it's really 0.74/hr if you don't want jobs randomly interrupted

[trail lion]

like half as much

[shell tendon]

i see 40 or 50 cents per hour

[nova zodiac]

correct.. so 3000 hours for a 2200 card

but you can't game on it

[shell tendon]

i'd imagine it's something you're going to keep around for at least 3 years or recoup much of its value with a resale

[trail lion]

I rent cards for training ,but I want to be able to inference local, too much hassle to spin up cloud resources to generate images

[shell tendon]

yeah i can def see the logic in renting stuff where you don't have to worry about 24gb or 16gb limits

24gb is juuuust enough to finetune a checkpoint which is partly why i said f it and went for the 4090 when i already had a 3080 12gb

[karmic cedar]

my gochujang tolerance is shit 😦

[cosmic drift]

You working on a personal projects or something

Or just to generate some images?

[shell tendon]

just fuckin around really

but been training loras

24gb is really nice for that... no gradient checkpointing

[karmic cedar]

damn this tteokbokki

[cosmic drift]

Ok cool i thought your are making something

[pearl ocean]

I just have 12 😦

[shell tendon]

i thought the 4080 had 16

[pearl ocean]

I mena 16 😦

[shell tendon]

what would making something mean?

16 is pretty good though

12 is where you really hit your head on the ceiling now and then

drove me nuts having 12

[pearl ocean]

imagine having 8

[cosmic drift]

No i mean training your own model

What model is best for text based images. like making memes and stuff

[pearl ocean]

imagine training on 4gb

[cosmic drift]

You can't carry a car in your pocket

[pearl ocean]

unless its hot wheels

[woven charm]

unless your doing it in stable diffusion

[cosmic drift]

What model is best for text based images. like making memes and stuff?

[floral nimbus]

anyone know why when i try to use 2 loras or 2 chracters in pony SDXL why it doesn't work well

or what i need for 2 character or more generations

[cosmic drift]

Use cloud if you cant run it on your pc

[karmic cedar]

cascade’s pretty nifty for text

[cosmic drift]

Oh thanks

[karmic cedar]

tends to be the most stable

pun intended

[trail lion]

2 character loras at once will always be an issue, it doesnt know where to apply the look so it gets blended

[pearl ocean]

imagine training an LLM on a Smart Watch

[cosmic drift]

What is the difference between sdxl and sdxl turbo

Quality is worst in turbo?

[karmic cedar]

less details, sometimes less vibrancy.

also turbo is designed to generate 512px if i’m not mistaken?

[woven charm]

Do you guys know if if there is a way to make stable diffusion prompting link words together?

take the promt "a big truck next to a green tree"

[karmic cedar]

the way you order your words in your prompt is essentially just that

[pearl ocean]

Depends on the model.
Epicrealism has this

[cosmic drift]

Low res but faster right ?

[woven charm]

How does one make sure that stable diffusion understands that its not meant to be "a green truck next to a big tree"

[karmic cedar]

essentially yeah—less complexity.

[woven charm]

or other vartiations 😛

[karmic cedar]

it really depends on the model’s architecture and text processing capabilities

[cosmic drift]

Dall e 3 is good for that

[pearl ocean]

Epicrealism

[karmic cedar]

i do like that one

[cosmic drift]

Is it open source ?

[pearl ocean]

¯_(ツ)_/¯

You can download it

[karmic cedar]

but if you have a model that has a ton of well-structured data, as in it’s got a ton of photos of X with lots of details about the contents of each image of X, etc. then you’re going to have a more robust prompting experience when you’re calling up those particular tokens

etc.

[cosmic drift]

It is open source

[pearl ocean]

¯_(ツ)_/¯

[shell tendon]

loras are basically models

[karmic cedar]

^

[pearl ocean]

Hot models

[karmic cedar]

hot pockets

BBQ sometimes

[shell tendon]

you don't need to train a full checkpoint to get shit you want

[pearl ocean]

Imagine training Sora

[karmic cedar]

remember when loras first hit the mainstream and there were shit tons of hour long YouTube training courses about how to do it

[pearl ocean]

Have you even met Lora?

[karmic cedar]

someone is going to move their business to the moon (read: microsoft, openai, newman’s own) and then train their models using entirely moon rock helium as their energy source

because they’re gonna have to

[shell tendon]

or they'll just burn more coal

[fervent thunder]

Does anyone know how to change stable diffusion relying on cpu?

I have an installation problem saying that torch cannont use gpu

[trail lion]

head over to the tech-support room, they can help you

[karmic cedar]

i’m just waiting for someone to build a rudimentary tensor-based architecture within Minecraft

it’s going to happen

i feel it in my bones

https://github.com/dendenxu/fast-gaussian-rasterization

insanity

[pearl ocean]

maybe triple monitors is the answer 🙂

[karmic cedar]

Okay so all this talk about Lora, but what about Ella? https://github.com/TencentQQGYLab/ELLA

Enhanced semantic alignment is what we’re all after, after all.

[rich kestrel]

seems like a gimmick tbh

[karmic cedar]

https://ella-diffusion.github.io

Seems really consistent compared to the other examples they give

[shell tendon]

anyone use that yet with comfyui?

or anything whatsoever

[karmic cedar]

I dunno but with the right base model, that thing plus an upscaler could make someone’s creativity spree a lot more efficient / effective

[shell tendon]

no doubt

i'd really like to see a sdxl version

[pearl ocean]

what about an XXL version

[karmic cedar]

I want a Championship Edition

With hidden bosses

[nova fable]

If you had 2.5 million images and you wanted to automatically remove watermarks/logos from the corners of them, what tools would you use? (I know how to do by hand in GIMP/Photoshop, but need something that can do efficiently for huge image set)

[pearl ocean]

needs Microtransactions

[pearl ocean]

MS Paint

[nova fable]

Took you 15 minutes to think of that joke? Pitiful.

Now, back to the adults in the room: I would even be satisfied with a model that just accurately highlights the watermarks, and could then write inpainting code to remove myself.

[trail lion]

there's an online service that does it, search the web for watermarkremover. It's paid, but for as many as you're talking, maybe that makes sense

[shell tendon]

comfyui

[pearl ocean]

Does it make you feel comfy?

[nova fable]

Are you referring to just using inpainting by hand in comfyui, or does it have batch editing tools to automatically ID watermarks and remove?

[shell tendon]

use segmentation to detect watermarks and/or text, mask that area, inpaint it, next image

you'd have to learn how to actually use it but yeah

i've automated processing of hundreds of images

inpainting certain objects out, replacing them, etc

[trail lion]

is the watermark always in the same place?

[rich kestrel]

im pretty sure there is software out there dedicated to removing watermarks

[nova fable]

yeah, I know how to use inpainting on individual images, but the problem is that this set contains different watermarks, of different sizes, located in different corners

[rich kestrel]

yeah the technology isnt there yet for SD

[trail lion]

ah, nevermind, you answered

if it was always in the same place, or always in the 4 corners, you can create a mask image, and use batch processing

[shell tendon]

what tech isn't there

[nova fable]

yeah, it would be easy if it was the same image, in same location

[trail lion]

I tend to just crop those out, like in an image editor, but I'm dealing with dozens, not million

[nova fable]

I'm feeling like I might just have to train a custom segmentation model ... and then write something that does the inpainting/removal automatically ... was just hoping somebody had already done this to save myself a few weeks of work 🙂

[trail lion]

well, that would be cool as hell, so if you do it, I think it would be popular

[shell tendon]

yeah, if there isn't one, you prolly will have to do that

[nova fable]

Yeah, I'll definitely publish to github if I end up doing it 😉

[trail lion]

I get that it's one of those things where "....surely someone has done this"

but I myself have searched and ended up just doing it by hand

[shell tendon]

i think the biggest issue is the diversity of designs

[nova fable]

yeah, if doing it by hand was an option, I would, but I'd literally be spending the rest of my life if I did 2.5 million of these one by one lol

[shell tendon]

getting one to recognize text i bet would be fairly easy

all the other weird ones are another story... is it part of the composition or not?

[trail lion]

that watermarkremover site does work, however that thing does it...even those crazy watermarks that stream diagonally across the image

[shell tendon]

espec if it doesn't repeat

[nova fable]

the nice thing about this image set is that ALL of the watermarks are in the corners - I don't have to deal with the ones that cover the whole image and repeat

[narrow kernel]

Then don't go around stealing a bunch of watermarked images to train loras? Seems pretty simple... The watermarks are there for a reason lol

[nova fable]

yawn ... who says I stole them?

they were freely available

and the original owner still has them

[narrow kernel]

If they are watermarked, it's to show their source. Have you asked the owners of them if you're allowed to use them?

[shell tendon]

are you using stable diffusion?

[nova fable]

so nothing was stolen. maybe go look up the definition of the word "steal" in the dictionary

[trail lion]

sheesh, the clowns are here, cool conversation is over, I'm out

[nova fable]

lol for real - literally EVERY major LLM / diffusion model was trained on copyrighted data ... I don't understand why someone would be in this chat if they were a copyright troll

[shell tendon]

def train something on watermark removal if you're able to though, it'll help get us better finetunes from more sources

[pearl ocean]

do you guys have iPads and MacBooks?

[nova fable]

I prefer high performance hardware, at a lower cost, so no

[shell tendon]

Generally with ya on that - however the newest ones are apparently amazing for LLMs

Have never bought a Mac before but yeah that's a thing now

[pearl ocean]

😮

[narrow kernel]

They didn't attempt to remove the watermarking or signatures. That's why you still see the models try to recreate them. Also, newer models are being a lot more strict about it and there are opt out lists that they have to legally abide

[shell tendon]

Their training sets aren't publicly available so you don't know that

Maybe they got some, and not all

Regardless, you can prompt it out

Spirit of the law and all applies here

Fact is if you're using SD you're using models trained on copyrighted data without owner permission

[trim magnet]

you should also try with this one,it was called lama cleaner and it removes/replaces objects in imgs https://github.com/Sanster/IOPaint

[nova fable]

thanks, I'll look into it - someone in another chat had recommended lama cleaner, and I kinda discounted it because I was looking for something that didn't require me to mask by hand ... but I'm curious to see how they implemented single click masking like that ... seems like it would be very useful for what I'm trying to do

[narrow kernel]

Yep, the difference is that I don't have hundreds of millions of dollars to handle potential lawsuits and they do. Point is, dude here is complaining about actively stealing people's watermarked images that are watermarked for a reason: to help prevent unauthorized usage of their images.

All so he can obviously train a lora or model with them

[trim magnet]

fact is this is still unregulated thats why lawsuits still get thrown around

[karmic cedar]

Soon steel-eyed agentic scrubbers will turn loose on the internet

It’ll suck

[nova fable]

I'm not "complaining" about anything. I'm celebrating, because I have an awesome image dataset with 2.5 million photos, and I'm super excited to make a badass LORA with it. Sorry you're jealous. Now run along and keep using software that was created with copyrighted data, while lecturing people on the immorality of violating copyright

[narrow kernel]

Well you guys can drink whatever cope coolaid you want, but the field is going to be locked down a ton within a year or two worldwide. Pretty much every major "first world" country is scrambling to come up with legislation to combat and contain AI. I guess enjoy it for now while you can

[nova fable]

If they don't want people to use their images, they can stop sharing them for free on the internet

[trim magnet]

evey first world country except japan right

oh wait now they gonna claim japan is not first world

[nova fable]

You all think LORA creators are why AI censorship is happening? Get a grip. This isn't about people making Waifu LORAs. They are trying to monopolize AI to maintain political power, and they are going to do it whether you use copyrighted images or not

[karmic cedar]

The argument may yet boil down to outright banning the training of models or worse yet, the outlawing of specific types of mathematics or code structures

[nova fable]

The "we're protecting artist" bullshit is just what they are pushing in the media to distract people from what they are doing.

They are busy making mass surveillance AI, automated propaganda tools, killer robots, and you are here bitching about people making waifu loras

"steeeealing"

[karmic cedar]

Well, it’s a slippery slope

Not so stable

[shell tendon]

Which will never happen

And will become totally unenforceable once nvidias monopoly breaks and we have customer gpus with 80gb or more of vram

[karmic cedar]

Unenforceable by a human

[trail lion]

aliens can do it

[karmic cedar]

Probably

[shell tendon]

Think of LORAs

What would a Lora ban do? Nothing

Except make the places you download from them bounce from domain to domain like pirate sports video feed sites

That's cuz tons of us have cards in our systems that allow us to do it without permission or knowledge of any authority of any kind

It's only a matter of time before those cards are better than H100s

[frank wedge]

especially becuase you can train one in a couple of minutes locally now(SD 1.5)

[shell tendon]

And we can train our own base models at home

[shell tendon]

Exactly

I grew up with MJ illegal and it was everywhere

If it's popular and ppl can afford what is needed to use it, it will be everywhere

Ban or no ban

[trail lion]

I'm just thankful we have the ethics police here to keep our conscience in check

[trim magnet]

yea when japan passed that law saying tha copyright doesnt apply to ai training these kind of ppl just moved the goalposts now they claim other countries will create a better law

[shell tendon]

And it's not like packets have to pass through customs when crossing borders on the Internet

If it's legal anywhere it's accessible everywhere

[shell tendon]

Well, I've heard that most artists have no problem with you training on their paintings so long as you don't inpaint out their signature 🙄

[karmic cedar]

I think the sharing of AI content will eventually become damn near impossible to proliferate on account of powerful, fast AI that can audit content in a microsecond

[shell tendon]

encryption

tor

etc

[karmic cedar]

Ehhhh

[shell tendon]

they'll never stop anything

[karmic cedar]

We shall see

[shell tendon]

it's also error prone

i think it's a lot like two different situations

[karmic cedar]

Information control is at the root of it all, as always

[shell tendon]

one is the development of weapons and defenses over the last century. offensive innovations have outpaced defensive ones, and there's no way around that. the only defense against nuclear war is the ability to start one.

the other is cryptography: it's a lot easier to increase the strength of a cryptographic algorithm than to break it. much, much easier.

right here, it's much easier to generate an impressive image than to determine its origin.

how easy is it to make a great image from noise? how easy is it to unsample back to the exact same noise that was used to generate it?

i have no doubt surveillance will be a problem, and it'll be used and do some really bad shit

but i don't think they'll ever be able to even come close to stopping it unless the entire world turns into a totalitarian dictatorship

[nova fable]

I think that fueled by exabyte scale AI, pervasive video/audio/GPS surveillance in the form of cell phones, and hordes of killer robots, a global totalitarian dictatorship might be exactly where we're headed

[shell tendon]

it's certainly possible if not likely

that we'll have the first

[karmic cedar]

People in power have new resources to manifest their versions of reality with

And that is too bad

[shell tendon]

global totalitarian dictatorship... i think if AI gets smart enough for that to happen, we're toast anyway, life moving one row down the periodic table

[trail lion]

will our new ai overlords care about copyright? 😄

[shell tendon]

only if you inpaint their watermarks

[echo aurora]

Hi mate, I have found the the location to turn face restoration and downcast back, I just go to setting/user interface/user interface/quicksetting list to add it back to the Main UI, cheers!

[dense sequoia]

Assalam o Alaikum.
𝐸𝑖𝑑 𝑀𝑢𝑏𝑎𝑟𝑎𝑘🌙
𝑇𝒉𝑒 𝑚𝑎𝑔𝑖𝑐 𝑜𝑓 𝑡𝒉𝑖𝑠 𝑑𝑎𝑦 𝑏𝑟𝑖𝑛𝑔 𝑎 𝑙𝑜𝑡 𝑜𝑓 𝒉𝑎𝑝𝑝𝑖𝑛𝑒𝑠𝑠 𝑎𝑛𝑑 𝑏𝑙𝑒𝑠𝑠𝑖𝑛𝑔𝑠 𝑖𝑛 𝑦𝑜𝑢𝑟 𝑙𝑖𝑓𝑒 𝑡𝒉𝑎𝑡 𝑤𝑖𝑙𝑙 𝑓𝑖𝑙𝑙 𝑦𝑜𝑢𝑟 𝒉𝑒𝑎𝑟𝑡 𝑤𝑖𝑡𝒉 𝑗𝑜𝑦. 𝑊𝑖𝑠𝒉𝑖𝑛𝑔 𝑦𝑜𝑢 𝑎𝑛𝑑 𝑦𝑜𝑢𝑟 𝑓𝑎𝑚𝑖𝑙𝑦 𝑎 𝐻𝑎𝑝𝑝𝑦 𝐸𝑖𝑑.
┅┄•※✤͜ ⃟✤͜※•┄┅

[narrow kernel]

Yes, actually, it's one of the exact main reasons why they are going to be cracking down on AI generated image creation... Two words: Deep Fakes. Random weirdos will train loras on thousands of images of a famous person(or some person they know) they are obsessed with and then create content with them. Some of those weirdos then go on to make things like NSFW content and potentially politically damaging content that can dupe your average social media user. LLMs are their own beast, but overall, ALL AI is going to be cracked down on within a couple years and will be governmentally regulated. Believe me, I'm a huge fan of AI, LLMs, stable diffusion, etc, but the few will always ruin it for the many and in the case of AI image generation, it's the many ruining it for the many.

[nova zodiac]

cunts ruining it for the masses as per normal

they can't put it back in the box though

[shell tendon]

so how are they going to stop it?

just like they did with illegal music downloads? live pirate video feeds?

illegal drugs? illegal guns?

all still here

[trim magnet]

they will send soldiers thru your pc obviously,i heard biden's gonna leak through my router anytime

[narrow kernel]

And if you get caught, you get in trouble.

[trim magnet]

oooh im shaking

[narrow kernel]

We're less than five years out from a "DRMed" internet anyways where everything you do is thumb-printed by your CPU+OS anyways. The IME and secure platforms in your CPU are ring -3, so you can't try to VPN or VMware your way around them either.

[trail lion]

dmca happened, worst legislation ever. and yah, it's still pirated, but the music industry gets to sue everybody, so it's still a win

the lawyers win at least

[shell tendon]

they won the battle and got to pick off randos but they lost the war

yeah, lawyers won bigtime

[trim magnet]

when you copy paste jargon from reddit to try to scare people

[shell tendon]

anything and everything is freely available illegal or not

[narrow kernel]

it's not jargon and not from reddit, it's going to happen whether you like it or not. especially now with all the botting on the internet. man, the whole dead internet thing really is becoming a reality lol

[shell tendon]

some things really shouldn't be freely available... and are anyway

[trim magnet]

yes its going to happen i heard biden is a repltilian and has a deal with aliens and they will poison the water with nanomachines

[trail lion]

I agree with Ruiner, it's probably gonna happen. this is literally dmca 2.0 in progress

liberals always find a way

[shell tendon]

i'd argue dmca was really ineffective in the end

[turbid bay]

A ban like that would be so unenforceable that I think it will never happen

[shell tendon]

all it did is make things more inconvenient

[trail lion]

you could have made those arguments back in the bittorent days though. it's decenteralized, how can you enforce it, etc etc

[shell tendon]

and it remains decentralized, and the enforcement remains a joke

[narrow kernel]

microsoft and intel have been hinting about it for a couple years now. and if the cpu makers get behind it, you can't tinfoil hat linux your way around it either since again, IME and SP are ring -3 and every single thing any OS tries to do is passed through them. Those threads you think your OS has physical access too? virtualized thread workers spawned by the IME/SP

[shell tendon]

you can still find pretty much anything on bittorrent

the pirate bay is still online

even still has its original domain

[trail lion]

even if they dont catch everyone, the threat is there

[trim magnet]

only way they could enforce it would be if they controlled the internet with massive surveilance/firewall like china and even in china i can still access torrent sites,to implement that kind of system in the u.s both congress and senate would have to pass a law (both with majority) and even them it can be challenged in the courts,so chances of something like that happening are basically less than 5%

[trail lion]

it's like speeding

you do it, everyone does it, but there's always that...what if I get caught

[shell tendon]

exactly this

i had chinese friends in grad school

they already knew what was going on that their great firewall had blocked

they'd all evaded it already anyway

and that's the most complete surveillance state in history

even north korea doesn't have total control

ppl smuggle in cell phones and pass around dvds and thumb drives

[narrow kernel]

oh no, it's even easier than that... that's literally what i'm talking about... everything you do will be thumbprinted and if you try to circumvent the security measures, you won't be able to use the internet since even the ISPs will be on board with it all. basically, they are going to remove anonymity from the internet and it will likely be within this decade

[shell tendon]

i've been hearing that for a loooong time

[narrow kernel]

it's for real this time, AI is big scawy to rich people

[shell tendon]

AI doesn't scare them at all

[trail lion]

yah I'm calling bs on the thumbprint thing, at least in the free world

[shell tendon]

it's opportunity

[turbid bay]

The pushback on such legislation will be massive though

[narrow kernel]

there is no free world

[shell tendon]

AI = lower labor costs, higher profits, higher productivity. rich ppl LOVE ai.

[narrow kernel]

there won't be, it will come from even higher up like HS level "in the name of national safety"

[turbid bay]

What is HS?

[karmic cedar]

Number games are a slippery slope—because in the end, all of this code represents products of simple data convolution algorithms

Numbers are an attempt to represent reality but are not reality itself

[trail lion]

there's possible and probable, they can barely pass measures in the states to extend the natinal debt cieling, I'm so not worried about them trying to pass this thumbprint thing

[narrow kernel]

AI is going to gut roughly 60% of middle level jobs by 2030, they will pass anti-ai bills effortlessly

i'm an engineer and even i'm not dumb enough to think i can't be replaced...

[turbid bay]

Doubt

[narrow kernel]

you're welcome to doubt it all you want. people said the same thing during the dot com boom as well

[turbid bay]

Yeah but we don't have a comparativs boom

[trail lion]

google laid off thousands of workers already, he kind of has a point

[narrow kernel]

but AI, as a whole, is going to be 10-100x as devastating to not one, not two, but almost every majory industry

[trim magnet]

cant really compare a boom to a full big brother dictadorship law

[narrow kernel]

major*

[turbid bay]

I'm not saying there won't be lay-offs. 60% is just a too high number

It's easy to forget that LLMs are really bad at anything that requires reasoning

[shell tendon]

and all of this could've been avoided if ppl merely didn't inpaint watermarks with their lora training sets

[turbid bay]

And there is no reason they believe they won't be in the future. It's possible that there will be some innovations, which would probably change my opinion, but I don't think it's likely.

[narrow kernel]

pretty much no government of a country with more than 50-100million people is ready to handle having to do UBI yet, in order to offset the loss in GDP from nobody having money to spend, due to middle/upperclass jobs being lost to AI

[restive herald]

Hello

[shell tendon]

AGI will come eventually and replace us all, we'll move a row down the periodic table to silicon based "life". but whatever really to that

[turbid bay]

UBI is nowhere near

[shell tendon]

all species end eventually

[narrow kernel]

that's what i just said guy with the asian name

chinese*

[shell tendon]

at least that's the one way we'll end ourselves that doesn't involve mass destruction

at least that creates something new

[turbid bay]

Maybe, or maybe not.

[shell tendon]

eveeeentually it'll probably get a lot smarter than any biological system.

when? who knows.

maybe we'll nuke ourselves first.

[turbid bay]

I don't see where you said that tbh

[narrow kernel]

AGI is still a couple decades away, but it will come quicker than contrarianists believe

[shell tendon]

there's no practical limit on the size of an AI brain, there def is for a biological one

[narrow kernel]

"no government of a country with more than 50-100million people is ready to handle having to do UBI yet"

[turbid bay]

That sentence is pretty ambiguous. I thought you meant handle as in being able to regulate

[narrow kernel]

wording could be a little better, english isn't my primary language

[turbid bay]

It's always been a couple of decades away

[trail lion]

ok but does that end look like the depiction in Wall-E where everyone is fat and happy and watching videos and drinking slurpies while being driven around on mobile self-driven cars

[turbid bay]

I think it's kind of a mistake to believe that LLMs are the stepping stone.

[narrow kernel]

Curve fitting models look like magic, but they really aren't. LLMs, diffusion, etc, you're just fitting curves. They aren't what is going to lead to AGI.

[turbid bay]

They sure are the best thing we have right now, but the difference to AGI is leaps and bounds

[karmic cedar]

Neuromorphic chips could power some very interesting inventions moving forward

[narrow kernel]

AGI is just going to come from shear computational power mixed with some shade of quantum voodoo magic. if not that, organoid cyborgs where it's a brain in a dish

we're already doing the latter

they trained some organoid brain in a petri dish to fly a plane in a flight simulator by feeding it sugar

[turbid bay]

Brain in a dish might be the most feasibile way lol

[karmic cedar]

I’d rather diffuse into the future, not have the future diffuse into me

[turbid bay]

Imagine having to buy a new brain for your computer every few years

The future

And feed it sugar and nutrients

[shell tendon]

that actually sounds awesome

imagine if you could just grow one in your bathtub

[karmic cedar]

“It’s Tuesday. Pizza gel day for brain.”

[shell tendon]

a bigass brain the size of big fat pig

[narrow kernel]

well picture having a specialized organoid brain the size of a football field

that's the kind of freaky stuff governments are going to be doing with them

[karmic cedar]

It would be an organoid with specialized heterogeneous gray matter

[narrow kernel]

granted, electrochemicular signals have a speed limit and all. think somewhere around 20m/s or something like that in the brain, but still

[karmic cedar]

Action potentials in such a thing could be measured by frame

Interval

[turbid bay]

It's creepy af

[narrow kernel]

absolutely is

[karmic cedar]

Yes it is

[narrow kernel]

and 100% not sci-fi...

[karmic cedar]

And fuck Minority Report

And Idiocracy

[turbid bay]

We're turning grandma into the AI compute of tomorrow

[nova zodiac]

idiocracy was a documentary from a time traveller, change my mind

[karmic cedar]

“BEEF SUPREME”

Okay where is our idiocracy finetune

[turbid bay]

Imagine you keep your consciousness

And your life is nothing but flight simulator

[karmic cedar]

Altered States

[turbid bay]

And joy for doing well

[narrow kernel]

i have enough existential dread to cope with, i don't want to think of being the brain in the vat until the heat death of the universe

[turbid bay]

Organ Mafia will turn people into bitcoin miners

[karmic cedar]

I mean personally I’d love to see an AI that behaves like an Animal Crossing game, where it helps you plant crops in a garden or yard space

[shell tendon]

isn't that what it already is

[nova zodiac]

noone clever enough to make it 😛

[shell tendon]

i think this actually makes sense vs the matrix

we're not useful as batteries, but maybe our brains could be divided up into computing cores and plugged into supercomputers to power LLMs

better watch out for openai

they're comin for your noggins

[narrow kernel]

that's how it actually was in the matrix btw

they just changed it to "human batteries" for the movie

because it was something simpler that people could understand without being confused

[shell tendon]

lame

should've left it like that

that actually is very cool

[narrow kernel]

they essentially used human brains as computational power to power the AGI and the matrix

like a giant renderfarm

[karmic cedar]

Extending consciousness means training your mind, a sensory jungle, on a new medium—like a deep sea.

Going from a jungle to the deep sea would be a shock.

Your brain would need training and exposure to make sense from the nonsense

So, you know, unless they come up with some magic hoodoo shit, that’s seizure land

I’m sure neuralink will usher in the Deus Ex era in no time, tho

😂

Hard to believe there could be a real Adam Jensen out there somewhere

(More or less)

[shell tendon]

thats the future

[floral nimbus]

is there anything other than like auto 1111 or comfy for vram usage? dosnt need to be image generation just wondering if i have the power is there anything i should look into

[karmic cedar]

Don’t you want to talk about the future

[nova zodiac]

Have you seen the global temperatures for sea surface temp and 2m air temp?? There is no future!!

The top 10 hottest months of all time were recorded in the last 10 months…

[pearl ocean]

I use Forge

[floral nimbus]

thats a1111

wondering if there is any other software or programs that require vram

[nova zodiac]

They generally all require vram

[floral nimbus]

bruh, i want to know whats "all"

[nova zodiac]

Other user interfaces for generating images - fooocus, easy diffusion, stable swarm, invokeai

There are mac specific ones as well

[floral nimbus]

is there anything other than image generation?

[nova zodiac]

How long is a piece of string - there are text to speech, speech to speech, speech to text, text to text, text to music, text to video and video to video models out there too

Go have a geeze at futuretools.io

[pearl ocean]

How long is a piece of VRAM

[rich kestrel]

69

thts exactly how long