#💬｜general-chat

1 messages · Page 114 of 1

opal hedge
#

512 token context window goes brrrr

ember hill
#

gm

shell tendon
#

unfortunately, SD3 will be released during WWIII

#

i looked that up in a book

astral goblet
#

they won't call it ww3 until its done so we could be in that time right now

opal hedge
#

The good news is that we'll be able to generate booba while the nuclear echoes are ringing

honest spear
#

literally metro 2033 but with an ai waifu chatbot

pearl ocean
#

imagine if SD3 came with video like Sora

opal hedge
#

Gotta start saving up quarters in the change jar for 3 B100s

astral goblet
#

next stop is the blockchain. stabilitys partnered up with nft networks

opal hedge
#

In 10 years people (or robots) will look back at us and go like, not even 1 Terabyte Vram per card?

astral goblet
#

we'll never need more than 40gb of vram

trim magnet
astral goblet
#

no but like, imagine if like, sd3 could do anything

opal hedge
#

Imagine all the people (that sd3 can generate)

pearl ocean
#

impossible

trim magnet
#

ok but when sd3 🗿

astral goblet
opal hedge
#

Imagine no VRAM limits or inference times.

#

Wonder if you can

pearl ocean
#

imagine if SD3 ran on 1GB VRAM

opal hedge
astral goblet
#

yoko's blockchain idea is gonna fuck it all up i swear

opal hedge
#

Though I imagine you need at least 4GB not to rely on off-loading

astral goblet
#

i dont see anything in the white paper about how they're going to take 8b down to 800m . wonder how they'll do it. blockchains?

pearl ocean
gaunt pulsar
#

The SD3 turbo looks very good too. Maybe it can be used for real time generation in games together with an LLM and audio gen for fully custom games

shell tendon
#

i'd imagine it'll prolly be more useful for non-realtime applications like video generation

gaunt pulsar
#

Yeah it looks promising for video too

opal hedge
#

turbo seems really good for generation but it doesn't do well in editing

#

I mean, 1 step inference looks insane

gaunt pulsar
#

It will be really good for experimenting with ideas given how fast it is

opal hedge
#

Yep

#

100 batch image generations will be pretty lit

astral goblet
gaunt pulsar
#

I'm sure countless types of games will come up once the tech improves. I gave it a try for fun, but I think local LLMs for 8gb vram are still not there yet to make entertaining characters + custom world

pseudo nacelle
#

Is there any verified article that talks about why Artists are against AI Img Generator?

What I know from it is because the AI databases are from the unpermitted arts by the artist themselves.. and when some companies use the AI img generator as a commercial use like in the online paid AI img generator, the artists didn't get any affiliation to it (or something like that idk..). (tho I think the company just want to pay their electrical bills and upgrade GPUs)

gaunt pulsar
#

I think the true reason most people are against AI is because it will take a lot of jobs. Once the AI that does X gets close to human level, some people will go against it, some fields more than others

shell tendon
#

ultimately AI is the most potentially beneficial and most potentially thing we've ever created IMO

#

all comes down to labor protections

fallen gale
#

does anyone know what version of clip to use with sv3d

shell tendon
#

the two obvious extremes: unprecedented prosperity, or basically everyone reverts to meat bags roaming the countryside while a handful live it up in castles

pseudo nacelle
karmic cedar
#

Bunkers seem to be the new castles.

verbal osprey
#

ai will create new jobs. There will be new film makers

karmic cedar
#

Bigger room, lower ceiling

foggy halo
#

guys sd3 release when?

verbal osprey
#

I would love to make a fully featured film if AI gets good enough where that is possible for me to do it all myself

gray fern
#

sora will be nutz

astral goblet
#

what i noticed is nvida's new initiavie. ACE. Avatar Cloud Engine. So npc ai's will be software as a service

#

single player games that are always online work right guys?

#

how long till MMO's have no players and it's all bots?

trim magnet
#

if mmo's have a gold trading market it will have like 70% bots,if u cant trade then it will be like 10 to 20%

verbal osprey
trim magnet
#

yes it is,its like letting ppl print their own money,they flood the market with cheep stuff and devalue the in game coins

astral goblet
#

just an always online single player game then. not massively multiplayer

verbal osprey
#

If we live in a star trek society we won't need money anymore

#

It would make a game like hitman interesting if you could talk to the different NPCs and they would react differently to certain situations

astral goblet
#

lol startrek still has replicator rations that crew have to save up. biggest reason their future economy is so efficient is transporter tech

#

the more we understand special relitivity though, more it seems we can't go ftl

honest spear
# gray fern sora will be nutz

I think it really broke other people's plans to keep tech releases at a certain pace. I just hope Runway and Stability will relase something similar soon

astral goblet
#

its a deluge of resaerch unfolding right now. anyone in this space isn't counting on slow releases anymore.

hidden dagger
#

hello, guys

#

what kinda dataset do ya use for running ur diffusion model

trim magnet
#

the thick one

astral goblet
#

i train loras with like, 10-20 images per

#

once i did one with a couple hundred

wet dust
#

Is SD3 trained on 1024x1024?

pearl ocean
#

is there anypoint in going higher than 1024?

wet dust
#

1024 is already rough when it comes to Deforum unless you have a beefy machine

crimson pilot
#

Hi all. I'm new here. I read the Pricing for Inpaint, it is shown as $3.00 for credit. Is that 1 instance or for 100 instance? Thanks!

pearl ocean
opal hedge
pearl ocean
#

if you set it very thin, it will be more of a standing shot

opal hedge
#

Yeah, I think image resolution and ratios play a large role in parameters

pearl ocean
light crystal
pearl ocean
light crystal
#

no worries about mutilated stuff now, hehe

wary mantle
#

Sorry to bother you, I'm trying my hardest to figure out how to upscale and enhance a very small image, but nothing I've come up with has been suitable for my needs so far. Maybe it's just not possible with such a small image and I shouldn't bother?

pearl ocean
light crystal
pearl ocean
gaunt pulsar
pearl ocean
light crystal
astral goblet
#

deforum is an animation thing yeah

pearl ocean
light crystal
pearl ocean
#

i see

opal hedge
#

These should all work unless your image is absolutely tiny, at which point it's worthless and you can just generate a new one entirely

astral goblet
shrewd jasper
#

noob question: when i pass a txt2img image into img2img for slight adjustment, do i need to include the original prompt?

opal hedge
pseudo nacelle
#

since, he said he pass the text2img result into img2img for a slight adjustment

opal hedge
#

img2img is very vague. It could mean inpainting, upscaling, downscaling, style change

shrewd jasper
#

yeah i meant inpainting. sry for the confusion

opal hedge
#

For most cases yes you need the original prompt, but some methods of inpainting for example with the krita plugin all you need to tell it is what you want to put in the inpainted area

shrewd jasper
#

thanks!

pseudo nacelle
opal hedge
grizzled void
regal glacier
#

yooo
I've been install the run.bat
after it's doing installation and download some stuff
and press any key to continue

#

nothing happens it's just shut, any hint ?

last kraken
#

u got an error

#

can u open webui-user.bat again and send the error?

regal glacier
#

once I run 'webui-user.bat' it's said Python was not found

#

launch unsuccessful

regal glacier
last kraken
warm junco
opal hedge
#

Panhandling on discord? That's a first for me

#

We truly live in troubled times

regal glacier
jovial wraith
#

morning all!

gray wolf
#

For what

stone latch
#

I have an itch for $2000 for an RTX 4090 if someone could send me the money 🥺

#

I won't be giving my username for privacy reasons so just give me your card details and I'll handle the rest!

#

(joke)

#

Does JuggernautXL have an inpainting model 🤔 I should check

fallow veldt
#

But sadly really I didn't experience any improve in the inpainting process, sometimes just worst results than just inpainting with the original model I used for the generation, so I stopped using it. But maybe I was missing something.

stone latch
#

🤔 I read something about Soft Inpainting in A1111 on Reddit a few minutes ago

fallow veldt
#

It requires making some changes in Automatic1111 to make it work.

stone latch
#

Maybe that could make inpainting better?

fallow veldt
stone latch
#

I tried using a ControlNet + inpainting with hephaistos nextgen dpo (a model) and oftentimes it refused to make changes to hair color specifically 🤔 It worked only a few times. I doubt this is a Canny ControlNet issue, as I can change features if I don't inpaint, but say I don't want the rest of the image changed for example.

fallow veldt
#

but the inpainting model is the same

stone latch
fallow veldt
#

Also if you want big changes you need to choose "fill" in Masked Content, and kind of high denoising (0.75). If you want to make a minor change over something that is already there, "original".

#

But fill is a gamble, it is like generating from scratch. So sometimes it is better to make the change manually and then inpaint with original. But for certain things it works fine.

#

It depends on the denoising strengh and the context SD has (if you use "only masked", instead of "whole picture")

stone latch
#

Plenty of tips I will try! Thanks!

fallow veldt
#

with "only masked" you change the context "Only masked padding, pixels"

#

no problem!

violet ledge
#

HOLA!

unkempt hatch
#

any way I could like, volunteer to help SD3 somehow?

#

I'm just so anxious for it to come out

fallow veldt
#

I don't understand why people do prompts like a story, as I understand it, words in prompts just trigger stuff, so that just add random stuff. Sometimes it do the images as certain words trigger what is prompted, but I believe the better to prompt is some kind of tag system, deep boru style with some short, concise ideas

#

I don't know how much grammar SD understands

#

Also I see huge prompts with a a lot of stuff, much of it doesn't appear in the image, it is funny

still glacier
#

with sd1.5 (which is using CLIP) it's usually better to use "word salad" prompts

#

with SDXL meeeeeeh not so much

fallow veldt
# still glacier with SDXL meeeeeeh not so much

yes I use SD XL and most of the time I do very succinct prompts, I don't even use negatives as I believe that may affect the image good or bad, like if you use "bad hands" it will try to not draw hands at all, or it won't really matter

unborn nexus
#

i've been trying to do style transfer with SDXL but all the turtorials seem outdated, does anyone have any recent tutorials on style transfer?

trail lion
#

style transfer like from an image? probably better to train a lora

trail lion
#

style is harder than subjects, imo

honest spear
unborn nexus
unborn nexus
honest spear
teal pagoda
trim nymph
#

anyone can tell me what ''basicsr'' is doing? trying to use softedge hed in auto111, and it says this module is missing..appreciate

#

No module named 'basicsr'

honest spear
abstract hawk
#

i long for sd3 so bad

#

gimme gimme gimme

charred mesa
#

apparently they will be inviting more people this week

#

its already Wednesday and nothing so far

abstract hawk
#

i really mean the model release beacuse im renting a machine on lambda for ollama and sdxl rn

#

after dealing with the openai api i never want to deal with an ai api again lol

charred mesa
#

ohh so model weights

#

that may come in april

#

or that april come in may

karmic cedar
#

ahhh ipad pros are going to get a matte option

#

if only i had money

trail lion
karmic cedar
#

regardless of the timing of SD3–every day with open source is a day with gratitude.

trim nymph
#

never mentioned which april, april 2025?

charred mesa
#

lmfao

karmic cedar
#

hey i mean pony diffusion is a thing, let’s see someone start building horse diffusion lol

#

at this rate someone else could have bronco diffusion by the time april 2025 rolls around

charred mesa
#

SD3 pony will be interesting

karmic cedar
#

oh definitely

charred mesa
#

Pony3? lol

#

I wonder if they will have to recaption the dataset, idk what their caption types are

teal pagoda
charred mesa
#

if its booru-like where its a bunch of adjectives and nouns separated by commas then it will need recaptioning

karmic cedar
#

this is true!

#

i’d think dataset recaptioning would be a bigger deal for a lot of point releases.

charred mesa
#

its unfortunate that a lot of pony finetunes and loras are mostly uhhhh, nasty

#

For photos i dont even know if we'll need finetunes

#

they look as good as sdxl photo finetunes like dreamshaperXL and juggernautXL

karmic cedar
#

juggernaut v9 is pretty slick

charred mesa
#

I wonder if more actions and expression swill be in massive finetunes or just separate loras 🤔

#

because it would make a lot of sense to finetune that

teal pagoda
#

or logos or illustrations

karmic cedar
#

it shows community bias in a whole new light, don’t it

charred mesa
#

heh

teal pagoda
#

but guess that the NSFW generators can make some money now with pony

karmic cedar
#

i mean the adult industry is probably already churning out fake websites complete with 100% fake content. it’s a win-win for those folks as they get to start laying claim to being more relatively moralistic in principle.

teal pagoda
#

:)))

#

kinda

charred mesa
#

if they manage to train SD3 with NSFW successfully with detailed and obedient prompts then imagine the possibilities catwhaaa

#

I personally don't use AI for corn

teal pagoda
#

Imagine the possibilities when a NSFW model like SORA will be out

#

:))))))

charred mesa
#

eeishh

#

hmm im wondering that

karmic cedar
#

with the context that Sora seems to demonstrate, I would imagine their NSFW controls are very robust.

charred mesa
#

if it will turn out that 8B will only run with 12GB if its like super slow CPU RAM offload even with comfyui (so no matter what it won't fit), then we may still rely on SD3 8B Turbo though as it only needs 4 steps

#

imagine FastSD or whatever with openvino, generating 4 step super intelligent images

#

(provided you have the RAM)

charred mesa
karmic cedar
#

well, it’s probably like how Fooocus utilizes GPT-2 with its prompts

#

the prompt has to be “baked”

charred mesa
#

unfortunately fooocus mostly adds tags at the end of the prompt

#

so its for aesthetics

karmic cedar
#

it is, and that’s okay by me as long as I’m using DPM 3m

#

and juggernaut v9

#

and a high guidance scale

charred mesa
#

but guys.. imagine the amazing possibilities once Fooocus gets SD3 support, then we get Fooocus V3 prompt enhancement (SuperPrompt-V1) running on CPU (super fast, don't worry), which MODIFIES your prompt entirely like DALLE3 and Ideogram's Magic Prompt

#

you basically get DALLE3 offline

karmic cedar
#

it will be amazing.

#

lol

charred mesa
#

im running superprompt on my cpu right now and its very very fast yet still coherent

#

its just a small 77M model

#

Brian (stability dev) made it and he might make a comfyui node and a newer model

teal pagoda
charred mesa
#

so far it has some issues like similar prompts ('gazing at the skyline' appearing for a bunch of portrait shots when rain is in it)

charred mesa
karmic cedar
#

has anyone tried to make “cyber woman with corn” yet using AI?

charred mesa
#

I use SD for entertainment

#

to entertain myself or others

#

and with SD3 having a bunch of artist opt-outs I will feel even happier

karmic cedar
#

me too.

charred mesa
#

especially how the quality hasn't degraded despite it

karmic cedar
#

absolutely that yeah

charred mesa
#

I mean for photos its to be expected to remain nearly the same quality

karmic cedar
#

it’s what we’re wanting from AI

teal pagoda
charred mesa
#

ah yeah I get it

karmic cedar
#

I want the economy to compatibilize with AI

#

not the other way around.

charred mesa
#

I mean there were a bunch of AI Art competitions where you compete with AI Art only

#

once there was one where you could win a 4090!

#

on civitai I think

teal pagoda
#

Yea, I remember those

charred mesa
#

but yeah SD3 is going to be great if we can run it on 8-12GB (the 8B model)

#

even if not the 8B Turbo model is there or the 2B model will get heavily finetuned to look way better

teal pagoda
#

but isn't really any other way for now to win money with AI art? I mean etsy is oversaturated...

charred mesa
#

yeah

#

idk I don't want to open that can of worms here

#

I just want to make either stupid funny imagery or nice looking portaits and masterpieces

teal pagoda
#

Alright

karmic cedar
#

“the board is set, the pieces are moving”

charred mesa
teal pagoda
#

I would really like to make my own comics, but SD is not really realiable for that atm. I mean it's really hard to maintain the same character (with the same clothes, same hair) and it's even harder to make him perform certain actions (like landing a punch into someone).

#

Guess you need 4-5+ ControlNets at the same time to make something "good".

#

And I don't have a RTX 4090, so yea

charred mesa
#

hopefully IPAdapter comes to SD3 so we can get consistent characters

#

or something similar

#

and detailed prompting will also help a lot with consistent characters

trail lion
#

consistent got a little better with ipadapter + instantid

#

that right there is 2 controlnets

#

I've found it's somewhat better to not focus as much on the look during txt2img and get the pose right, fix it in img2img

#

having said that, yah it's extremely difficult

teal pagoda
trail lion
#

yup, time consuming, you'll never get it all in 1 gen

#

I imagine probably 20 gens + inpainting

#

that's one of the things MJ has been touting, I havent tried it, never used MJ

#

but for me it's like Dark was saying ,hobby, not trying to monetize

broken cave
teal pagoda
#

And then you have to sell the comics, so that you can get your investment back (2000$ for RTX 4090).

cold fog
#

hello, can i get some good generic tags for image generation?

teal pagoda
trail lion
#

more eyeballs than the superbowl 😄

forest trout
#

Eh, I see trying to monetize art generation similar to drop shipping. Yeah there's a market for people to make money off of commissions just like how there's a market for dropshippers, but it's pretty easy to get lost in a sea of people doing the same thing since the bar of entry is pretty low.

unborn nexus
#

is it possible for automatic1111 to run without internet? tried doing this but it keeps saying it needs to install requirements(ive ran and installed the requirements multiple times)

trail lion
#

yes you can

#

if your script is trying to pull from git, that's why

forest trout
#

Once you have properly installed everything it is able to be run offline.

unborn nexus
trail lion
#

your startup might literally have "git pull" in it, check it out

forest trout
#

Because your program is set to update when you open it.

teal pagoda
forest trout
#

Yeah artists will always have more control over the concept. They have their limitations too, but honestly I think artists should pivot to the performance of creating art, similar to how musicians perform songs.

trail lion
#

yah, create a base and train their own loras for example

forest trout
#

Did you know that during the time period impressionism was kicking off, that the final stages of creating a painting were performed in front of a crowd? The act of varnishing the painting was a spectacle because it was the point where the color exploded on the painting and was fun to watch.

#

Fun little art history tidbit.

unborn nexus
#

@forest trout @trail lion do you guys know what config file i have to edit to fix this?

teal pagoda
#

right click on it and press on Edit file with notepad

forest trout
#

Could be in webui-user.bat

#

Check both.

trail lion
#

yah might just have to comment it out in the -user one. I'm in linux, but same concept

forest trout
#

Both are in the top level of the parent folder

pallid marsh
#

hello

fallow veldt
#

inpaint should have a negative mask that prevent SD to get the context from that area.

So when two objects or subjects are really close and you want to define one, it doesn't try to extend the other one.

bleak matrix
#

Good morning, everyone! How are we all today?

trail lion
#

if you use too much denoise you'll start to destroy the coherence, so it's a game, you have to play with it

fallow veldt
fallow veldt
#

for example I was inpainting some railing, there was a hand near, so it draw another hand

#

Also when trying to add another character, if it sees another character in the same context, SD will think it is about that character

trail lion
#

sometimes you can also help it along with inpaint sketch

fallow veldt
trail lion
#

many of these tasks are, and if they arent worthwhile then giving up is an option

fallow veldt
trail lion
#

just venting, got it

fallow veldt
#

SD gets easily "distracted"

charred mesa
#

how much does the 16 channel VAE improve our images

#

4x more detail for each pixel whilst the VAE is decoding?

#

or better colour accuracy?

#

I need to see the paper again

vernal dawn
#

Hey guys, I've created an open-source prompt engineering & prompt chaining toolkit in Javascript - it's kinda like the OpenAI playground but you script in Javascript. I'm looking for a couple people to take through version 0.0.1, get some feedback. Please DM me or @ me if you're interested.

slender field
dusk canopy
#

Controlnet better

#

💀

visual torrent
#

hello everyone I am new and I would like to know how to generate images on this server

hidden dagger
#

ofc, GAN is previously popular several yrs ago

#

especially a lota art students use StyleGAN for their graduation thesis

#

currently, I am still learning a previous Kaggle competition abt stable diffusion

#

and maybe need to read some paper for further comprehension on it

#

I hope I can develop an algorithm abt diffusion model in medical imaging

#

and wish I can contribute sth to my github soon

inner lintel
#

Stable Video 3D looks so promising, guess the entire gamedev and CG industry will be changed soon. agony

hidden dagger
visual torrent
#

thanks anyway

hidden dagger
#

and not so many people apply this model to medical imaging atm

#

so I wanna try it

#

by learning from a professor

inner lintel
hidden dagger
#

btw, I just lost my job

#

so I need to publish a SCI paper or a good conference paper for my next job

#

I don't like teaching, that's why I am fired

visual torrent
#

good luck with your project

hidden dagger
#

I want to do research so I need my own paper first, then I may have my second, third one

#

and I need new job that can be helpful for my research

#

Losing my job is very terrible

#

I was angry, drunk

#

but Ik I still have to figure out how to continue my plan and my research

inner lintel
#

Good luck, anyway

hidden dagger
charred mesa
#

its quite unfortunate though

primal star
primal star
iron pendant
hidden dagger
primal star
charred mesa
#

ah hell nah wtf

true canopy
#

that..changes things

primal star
#

so whatever forbes etc are writing don't trust it blindly

astral goblet
#

the stable diffusion research team all leaving before its even finished.. weird.

iron pendant
astral goblet
#

it aligns with the time frame of pivoting to crypto partnerships too

iron pendant
charred mesa
#

its just not the most brightest future

astral goblet
#

two partnerships announced this week

primal star
astral goblet
#

to their credit, otoy makes octane, but show me a company that has ever succesfully pivoted into nft/crypto

iron pendant
#

did we not* all learn our lesson from the Folding Ideas video on nfts and daos?

astral goblet
#

never saw a video. i just paid attention over the last 10 years

iron pendant
astral goblet
charred mesa
iron pendant
#

I guess i'll try not to jump to conclusions yet there

charred mesa
#

I just shrugged it off until flowwolf started talking about this

#

my only conclusion is "does SD3 come out either way or not"

full lark
#

😭

astral goblet
#

i think sd3 will make it out the gate, but the stable is on fire (see what i did there?)

iron pendant
#

@coral yew 🤗 🫡

grizzled void
astral goblet
iron pendant
#

it goes deep into DAOs and how they're used to grift poor people for spec work

astral goblet
#

seems that way. hard to find these gems on youtube since the thumbs down button was killed and it's turned into a pure hype/engagement machine

primal star
astral goblet
#

not for the whole though

iron pendant
#

he only puts out a video every few months, but they're bangers, and I feel that NFT/DAO video was basically the final nail in the coffin for the collective consciousness after the (actively dumb) hype over NFTs, not that it was necessarily news, but deep diving into some of the more ridiculous acts is something really everyone should watch

astral goblet
#

i could invidiually use thumbs down buttons but nobody else does. just us few users with an extension. so it doesn't matter towards any of the content creaiton and algorithm boosts

primal star
#

just another one of those circus decisions from google

#

next one is the chrome manifest change which kills adblock

iron pendant
#

I mean, it has 14 millionviews, its made rounds

astral goblet
#

i still think vitalik buterin is an animal. love that guy. but the legacy of the network is a whole lot of crime and fraud

iron pendant
#

yeah, some good technical ideas that unfortunately, when reality of application commences, is generally not good, largely there to enable money laundering and abuse people who cannot understand the complex economics of what they're doing and like to play the lotto

astral goblet
#

-cough paidoff cough-

primal star
#

firefox will maintain v2 while chrome discards it

astral goblet
#

i have no doubt that large sums of cash change hands behind the scenes, because why wouldn't it happen?

iron pendant
#

I mean, git source control uses successive hashes to help provide provenance, closely adjacent to blockchain (but without the eco-destroying hash mining), its a good technical idea, but nft/daos are next level gross bastardization

astral goblet
iron pendant
#

right, an algorithm is just an algorithm until a person uses it for bad things

astral goblet
#

algorithms don't kill people. I kill people! (good tshirt idea?)

peak orchid
#

But the crypto pivot isn't really a good look

full lark
full lark
#

I tried searching but i got nothing

astral goblet
autumn egret
#

Any updates on SD3 or a final release of Stable Cascade...?

peak orchid
#

Open source community projects might be able to produce better models on longer timelines, but I'm not sure that they can do it at the same speed as Stability AI can. The EU wouldn't have been able to place restrictions on GPT4 if it had been open source and the weights easily accessible

full lark
#

F*ck EU restrictions, too bad that even in the US and Russia/China there is censorship/restrictions too

astral goblet
#

Because a site asking you if it can use cookies is really going to stop big data. smart move EU. Thanks for doing that and adding an annoyance layer of privacy theater to the whole web

peak orchid
astral goblet
#

realtime automated surveillance of all online communications. yup

#

in order to protect the kids

#

they didn't actually knock that off and have a v2 proposal in the mix now

full lark
iron pendant
#

backdoors are how adversaries hack you

astral goblet
#

effectively making it moot

full lark
#

if its "scanned" before is encrypted what is the point of encryption

astral goblet
#

i dont plan on abusing children ever but false positives do exist

iron pendant
#

who and what does the scanning

#

there's your backdoor

full lark
#

It would be better if parents raised their children instead of giving them the damn cell phone to talk to strangers so the government has to do something

astral goblet
#

here kids, go play in this asylum lounge

crude garden
#

when will the bot come back

#

i can't wait it

astral goblet
#

Might not.

crude garden
#

i don't believe that

#

I think they are using the bot's computing power to calculate new video models.

iron pendant
vapid tangle
#

SD3 is still not released, right? no way to try it?

astral goblet
#

can post prompts in the #sd3 hashtag on twitter and maybe they might do them. People have been posting dall-e images with prompts and someone with access posts a comparison

#

soorry i mean the website formally known as twitter

meager galleon
#

Hello, inquiry:
Is it possible to feed SD with different photos to get one with different characteristics? I mean, I want to create a warrior and I have a photo of a person with a hat, another with a weapon in his hand and another with a specific gesture and I would like to gather those characteristics in a single photo.

charred mesa
#

I'd be surprised if we got access this week, they said they're givin out more invites

astral goblet
#

invites seem to be picked and chosen and signing up for the list doesn't matter. its not a spool.

willow crescent
meager galleon
charred mesa
astral goblet
#

us PC gamers may reign over the console plebs, but the ML phds and candidates reign over us agony

charred mesa
#

true..

willow crescent
astral goblet
meager galleon
shy eagle
#

remove if not allowed ofc, but im looking for someone who is capable in faceswapping onto videos - DM for more info.

astral goblet
#

once you learn nodegraphs in comfy, you can move onto things like touchdesigner

willow crescent
astral goblet
#

node graph for creating visual stuff

willow crescent
#

ah, I see , thank you

astral goblet
willow crescent
broken smelt
#

researchers have left before, SAI will be fine lmao

charred mesa
#

well I don't think its going to be that terrible

astral goblet
willow crescent
#

do you read it? - no, I reddit

charred mesa
#

how

willow crescent
#

MJ is based on SD

clever heart
#

which upscaler do you use to make 8k instead of 4x-Ultrasharp for anime when you use ultimate SD upscale ?

karmic cedar
static cape
clever heart
#

oh ok

karmic cedar
clever heart
#

Or is there a simpler way to make loras ?

regal snow
#

For anyone using A1111 / Forge who is familiar with the extensions with AR / static res buttons, I did not like any of the ones available. They all output bad resolutions with no control over rounding precision, while having between 1px - 4px while 64px is ideal in most cases.
I forked the pick of the litter and spent a long time enhancing it.
https://github.com/altoiddealer/--sd-webui-ar-plusplus

karmic cedar
#

holy work

astral goblet
#

^ this kind of spam is only going to deluge more and more. now that stability is partnering with blockchain companies, we're going to see a faux gold rush to capitalize on hype and naivity. sucks.

iron pendant
#

it's going to the moon!

karmic cedar
#

Strapped to, by then, “ancient” A100s

astral goblet
karmic cedar
#

disgusting thought, sort of like how bitcoin mining really is basically digital coal

#

too bad Nano never took off as an altcoin—the entire network could be powered by a single wind turbine

astral goblet
#

bitcoin is hitting some kind of halfening. where the rewards for each block are cut in half. this is driving the whole market into a frenzy and ponzi , i mean investors are going full salesguy mode

karmic cedar
#

BTC is driving carbon tho

#

it’s ugly

karmic cedar
mortal delta
#

I wonder if ai has gotten to a level where it can create whole projects like games, software, and such, i feel like we are close but im not sure.

astral goblet
#

i miss the days of the internet where the worst thing was an irc split

iron pendant
astral goblet
mortal delta
iron pendant
mortal delta
iron pendant
#

half joking, but people already have langchain, you can build "project managers" and "coders" in llms and string them together with image models, etc.

karmic cedar
#

I think we’re going to see a more distinct correlation between creativity and energy efficiency in future MoE subagent workflows

#

so creativity in models will be more prized as subagent designs popularize more

iron pendant
#

ask an llm to come up with a product idea, ask an llm to list features it should have, ask an llm to configure a project that will have large epics, then break it down into smaller features, etc, etc., and down to creation of art assets, its just a big complex graph

astral goblet
#

flava flav has a point... well probably a lot more than that. (heyohhhh)

karmic cedar
#

also a strained neck

mortal delta
#

I must say i feel like im too invested into ai, but is that a bad thing?

karmic cedar
#

No, because AI is a blob of future potential right now.

#

It is the future.

iron pendant
#

I think we're very close to someone publishing a (simple) game completely created by AI, not just like "hey Llama write Pong in python for me" but like "come up with a game idea that is simple" ... "write down the major features..." "write tasks for creation of all the art assets" and soforth

mortal delta
iron pendant
#

give your llms different system prompts/prompts for specific tasks, let it run the code and fix its own errors, and soon

karmic cedar
#

You can take pride in knowing what others don’t. People are extremely stubborn and for the most part deny a very basic reality: that they need to see to believe.

#

Seeing doesn’t just mean with the eyes—it also means understanding.

mortal delta
karmic cedar
#

I like the idea of basing a game concept on a console format, like a preconfig

iron pendant
#

a couple weeks ago one of the big boys released a paper I think on having an AI pilot your desktop, like, click on windows, move the mouse around and soforth

#

but most of this can be done programmatically without a UI at all

#

humans are very visual, but its extra cruft to actually click on a UI

karmic cedar
#

I would love to see some kind of supergame where you start off playing in 4-bit, OG gameboy style, then as you progress through the game it jumps generation after generation

mortal delta
iron pendant
#

the VLM models already have grounding, like Kosmos2 or Cog, they can literally identify where specific objects are in frame

iron pendant
#

i mean you dont want it clicking on your banking website link and logging itself in with 1Password and sending your money off to some dude in Tunisia

mortal delta
astral goblet
#

i was on board with the internet long before people thought it was going to be a big deal. once it was, most of that old experience was worthless. /shrug

iron pendant
#

you mean learning how to setup my modem init and SLIP/PPP isn't applicable anymore?

karmic cedar
#

I want to make Devin create a Virtua Fighter clone, like straight out of 1994 or whatever and then I would make all of the playable characters choo choo train faces

iron pendant
karmic cedar
#

yep and then the show me what you got head will descend from the sky and ask me what i got, and I’ll show him and he’ll like it and float awayyyyyy

nova zodiac
iron pendant
#

I don't think that is what I was remembering seeing

#

but it looks like the same sort of idea

mortal delta
#

i wonder if there a tool that can help makes games because ive been wanting to make video games but im to dumb to understand code and such, but i also dont have money to throw at api's.

iron pendant
#

I think you could optentially fine tune Cog or Kosmos2 to learn how to identify buttons from screenshots, then just write a program to click on the buttons, for instance, and then chain that into a larger program using LLMs and such

astral goblet
iron pendant
#

you ask the VLM "where is the close button on photoshop?" and it will spit out a bounding box for where the X button is, in theory, and you built that out behind the scenes, and you could pilot your computer that way with AI, or voice recognition, and soforth

#

not quite there, but close

astral goblet
#

now ask it how to close vi

iron pendant
#

you'd need to tune it for desktop screenshots, but that is not a big ask

opal hedge
#

Devin in 2024: It works, but it's nowhere near as good as a software engineer
Devin in 2026: You won't get the same quality as a team of software engineers.
Devin in 2028: You still need a software engineer to look over it and give his approval.
Devin in 2029: We need legislation or else new CS majors have nowhere to work!

karmic cedar
#

Folks are seeing the consequences of applying human precedent to a decidedly nonhuman era

#

cough yann lecunn cough

astral goblet
karmic cedar
#

you know there’s some DARPA-level clubhouse shit going on that’s possibly years older than what we’re seeing now.

opal hedge
astral goblet
#

i spoke in future because there isn't one yet. it's claims are a lie

opal hedge
#

Hasn't he been proven wrong in his predictions by large margins several times already?

karmic cedar
#

I have this hunch that he’s a spiritualist.

#

yes

opal hedge
#

I need to watch some of his past interviews and see what went different

#

Because I watched a few newer ones and they just seem... pessimistic about the rate of progress

#

Like pessimistic in an unrealistic sense

astral goblet
karmic cedar
#

My theory is that his idea of intelligence is too narrow—as in he’s got some religious thinking possibly affecting his imagination.

astral goblet
#

he presents this problem like language can't possible solve it. but then people with inner monologs show up

opal hedge
#

How does someone not have an internal dialogue? Is he a buddhist monk lol

karmic cedar
#

He keeps painting the notion that X or Y “isn’t possible” yet, who is he to declare what is and isn’t possible in an age where neuroscience is fueling the most advanced technology developed yet?

astral goblet
#

i mostly dont have inner vision. just flashes of images here and there

#

mostly a chorus of me like a price is right audience

opal hedge
astral goblet
#

consider deaf people

opal hedge
#

I suppose you're right

#

So they just walk around with silence in their mind like enlightened beings? Isn't that the point of meditation

karmic cedar
#

it’s precision moment-living.

#

fueled by disassociation.

astral goblet
opal hedge
#

I guess they could technically envision the world in images and sensations

astral goblet
#

ton of other senses and concepts that could steal attention

karmic cedar
#

drugs provide more context, but i’m not going to advocate for them.

opal hedge
#

I meditate regularly, and I have experienced true thoughtlessness for seconds at a time, but I envy people who can do that all the time. I can't get my inner monologue to shut up lol

astral goblet
#

"is it quiet now? fuk i just thought!"

#

Its like playing the game (you just lost)

stone latch
#

Vivid imagination constant music and voiced thoughts

honest spear
karmic cedar
#

synesthesia is what we’re getting at here

stone latch
#

I guess it's like aphantasia where people can't see faces in their heads but can recognize people anyway

karmic cedar
#

there is a lot of sensory overlap, because the pathways in our brains that process our senses decussate and multiplex

#

totally

opal hedge
#

Back to the topic of yann lecun, if he has no inner monologue, does that mean he can't really ponder his ideas?

#

At least on the spot

karmic cedar
#

ehh it means he suppresses his ideas which tells me they manifest as subconscious abstract and he probably reveres them more meaning he probably leans more into his egocentrism

#

aka stubbornness

#

(i might be projecting a little there)

honest spear
#

It's probably more difficult without an inner voice to think something like "no this idea is trash" or "I shouldn't post this"

stone latch
#

I'm a little out of the loop, what are we talking about? 😅

karmic cedar
#

we’re talking about how yann lecunn is an enlightened guy and all, but his predictions on AI have been pretty far-off

valid carbon
#

wich generator can i use to make pixel art? without credits

karmic cedar
#

ummm

#

idk anymore 😕

opal hedge
#

Especially when he's put on the spot in an interview since he can't deliberate his thoughts

karmic cedar
#

I think if he took other disciplines and perspectives into account with his thinking, he might have a different approach to the future—i.e. neuroscience

#

and hell, philosophy

astral goblet
#

if i have vision its for a flash. like a strobe light hit it

stone latch
#

I'm not d'

karmic cedar
#

aphantasia happens on account of a specific region in the visiotemporal cortical area going offline

astral goblet
#

meanwhile, my eyesight and acuity is on point. 41 don't need glasses and can see better than 20/20

opal hedge
#

This may sound strange, but I think I have aphantasia for cars. I can't remember what my family members' cars look like

stone latch
#

Oops, my goodness I keep sending my messages prematurely 😂

karmic cedar
#

where visual qualia is mapped to facial attributes

stone latch
#

I'm not 'diagnosed' with anything but I may have hyperphantasia, took a quick test once.

astral goblet
#

while i don't see stuff, in a moment i hear an indepth description of what i remember

karmic cedar
#

it can happen with any region of the brain encoded to a specific pattern

#

if it goes off, then your brain loses the ability to process.

#

simple as that.and sometimes your brain can compensate for it in other ways

astral goblet
#

my going theory is my visual cortex works on real vision more than inner vision.

karmic cedar
#

there are different layers of visual acuity, with distinct regions where the “math” is computed

#

all towards the very back of your head

astral goblet
#

yea we also know the brain is very plastik

#

there likely was some damage at some point

karmic cedar
#

and the amount of wiring going back and forth is prone to a lot of dynamic growth

#

could be! probably true for most folks.

opal hedge
astral goblet
#

/thinks back to all the hang overs and knocks ive taken/ hmm could be indeed

opal hedge
#

We only have inner vision, there is no such thing as real vision

astral goblet
karmic cedar
#

there’s the visual sense, and then the imagination.

stone latch
karmic cedar
#

lol

#

it’s very confusing because…reality, man

astral goblet
#

at some point the visual cortex was like "fuck all that i'm gonna focus on the eyeballs"

karmic cedar
#

phylogenetically speaking, eyesight is the most recently developed sense

#

and the frontal lobe is the most recently developed (evolved) region in the brain as well.

opal hedge
karmic cedar
#

Clint—very poignant point

astral goblet
karmic cedar
#

also—we technically view the world upside down, but our brains actually convert the signal right-side up. and each eye decussates (crosses over) to the other side of the brain.

astral goblet
#

lowkey we're all australian

karmic cedar
#

like an upside-down X.

#

lol

astral goblet
#

ever found your blind spot? remember that xfiles where this assasin used the blindspot?

karmic cedar
#

yeeeea!

#

hey, if you want to read about something trippy—look up Charles Bonnet Syndrome

#

some folks can have blind spots that are much bigger, called scotomas

#

and your brain does the same thing with those as it does your blind spot—it pulls a photoshop-style interpolation effect.

astral goblet
#

scotomas deez nuts... /shrug

karmic cedar
#

Except with CBS, the scotomas are so large that the brain projects straight up crazy hallucinations

astral goblet
#

the brain hallucinates all the time

opal hedge
karmic cedar
#

yes it does, we are always trying to find homeostasis. stability.

astral goblet
#

most crashes happen in the km near people's homes. i heard that once. cause the world is a reconstruction based on memory, not realtime

karmic cedar
#

fascinating. and creepy.

astral goblet
#

given how unreliable memory is, yeh

karmic cedar
#

hence the race for stability, lol.

opal hedge
#

We don't really have a memory. We reconstruct it every time based on a few tidbits

#

Kind of like neural nets

karmic cedar
#

and whenever you recall a memory, you are degrading the proteins that have encoded the memory very slightly each time. so it’s never technically the same memory.

#

that one always bakes my noodle.

opal hedge
#

We're basically biological AI, using synapses and neurons to reconstruct an idea based on how it's wired

astral goblet
#

sounds like some never ending story shit

karmic cedar
#

the door is open.

grizzled void
#

hi

karmic cedar
#

hiiiiii

opal hedge
#

👋

full lark
# willow crescent MJ is based on SD

Yes, but I'm not sure if they're trained only on high quality data or how they accomplish what they did in v6, they surpassed even dalle3 by...alot

karmic cedar
#

it’s trained on everything in the world, probably.

#

or—they have a fundamentally different approach to the text encoding.

opal hedge
#

Does that mean midjourney has to pay SD royalties?

karmic cedar
#

no one is paying anyone anything…yet

#

lol

#

well, i mean….everyone is paying microsoft

#

$$$

opal hedge
#

the SD commercial license means anyone making money off of SD has to pay SAI

#

There's also the research license but that doesn't allow commercial usage

willow crescent
#

including a style prompt as well it seems

gaunt pulsar
#

I think ppl are underestimating Dalle-3 just because they censor literally everything

potent spire
#

Surpassed in what exactly?

ember hill
#

hello

#

?

karmic cedar
#

he mad

crimson beacon
#

when are we gonna get a sd3 discord channel where we can see community gens w new model

karmic cedar
#

i mean yeah talk normally

#

grifters have not changed in the decades that I’ve spent online

#

you guys are all the same

#

i feel sorry for you 😦

full lark
#

bruh what?

karmic cedar
#

this guy is trying to communicate. it’s not working very well. 😦

#

do you need an SD3 invite?

#

lol

full lark
#

Maybe he is a midjourney/dall-e spy 🕵️

karmic cedar
#

oh, those are eeeeeverywhere

icy quest
#

bro mov2mov extention is so gnarly. it halucinates stuff that you cant make with img2img

#

any alternatives to mov2mov?

karmic cedar
#

not yet

icy quest
#

anyway to run SD from a batch file?

karmic cedar
#

there are batch scripts out there, but temporal persistence is a general issue right now for what could be awhile.

icy quest
#

maybe we could split a video into images, then run each image through a bat file

#

well if it could use the previous image as a prompt, with the settings of the new image together

karmic cedar
#

every image represents a different input, and even with the same seed you’re going to get a different diffusion for each different input

icy quest
#

something else is, like, if the bat file only runs if there is a missing file, u could just delete bad frames and keep running it until you have better temporal persistance

#

yeah but some promts are very limiting, so it might narrow stuff down enough

karmic cedar
#

it’s noise

icy quest
#

i noticed using a bunch of specific negative promts severly lowers the diversity of the output

#

i think if u could just hand delete bad frames and regenerate them, it would probably make stuff a lot more feasible

karmic cedar
#

maybe for a 15-frame sequence

#

but a 200? 5000?

icy quest
#

beats doing it all by hand, its not 1000x productivity increase but even a 50x helps

#

well actually right now its not even dooable at all.

#

how do u run sd from a script anyway? because all the settings are temporary in the browser

karmic cedar
tough oak
#

Hey, i'm pretty stupid with this stuff, i purchased a subscription to midjourney specifically to use a stable diffusion generator to give me images of this building if it was rehabilitated; but since I am too dumb to figure it out i am blowing my fast GPU away

#

I put it in a server called Dr Toboggans Historical Rehabilitation

#

i'm totally lost in this server stuff; can someone help me? or tell me if it's even possible?

opal hedge
tough oak
#

yeah pretty dumb, i'm pretty computer illiterate

#

can it be done tho? i put the picture in the general-with-images hashtag below this one

opal hedge
#

If you have a fast gpu as you say you can do it on your computer. Look up image to image stable diffusion on youtube and start learning

#

Don't have to pay anything except your electric bill

karmic cedar
#

YouTube is a good resource for how-to’s on nearly all of the stuff everyone here is using.

tough oak
#

No, i have a macbook pro from 2015, it kind of sucks; i think there was like fast gpu points on midjourney; i didn't realize there were other good/free diffusion softwares

karmic cedar
#

But it starts with what you want to do—if you want to do an image to image, then search for that technique particularly.

tough oak
#

image to image, thank you, i did'nt even know the term to start

opal hedge
karmic cedar
#

AI is sort of a use-case thing right now for folks. People describe it in terms of what it does. i.e. txt2img, img2img, img2vid, vid2splat, etc.

tough oak
#

thank you, i'll see if midjourney can do image2image; also should focus on learning a little bit more of this stuff; thank you guys

opal hedge
#

Just to get you started

#

The second link is more recent for the current version of midjourney (v6)

karmic cedar
#

Also keep in mind that these models we’re talking about are free on their own—it’s how they get used that the money starts to factor in. Running it yourself? You’ll get pointed in the direction of the code repositories. Weak GPU or no GPU? Then you’re surfing the cloud. It becomes a matter of finding the cheapest way to get the code running virtually.

#

Approach it like that and you’ll save money

opal hedge
#

He already bought a midjourney sub so might as well use it

karmic cedar
#

That is true! and Midjourney is a powerful tool in its own right

tough oak
#

interesting, yeah i plan on building out a computer in the future; may be easier to run it on that in the future; for now, i have to figure out how to use midjourney to to image2image

karmic cedar
tough oak
#

thank you guys!

karmic cedar
#

yep…just remember, we are in an age of answers. there are so many out there. you just gotta know what you want to ask. and try using GPT for stuff too if you haven’t started yet, you’d be surprised at how helpful it can be whenever humans aren’t around to help 😛

coarse anchor
#

有没有Chinesecatsprout

tough oak
#

true, i'm so used to asking a person to "show me", been giving off boomer vibes for some time now

karmic cedar
#

boomer vibes are okay. boomer vibes need to happen in order to evolve into a zoomer (unless you have an evolution stone)

#

ahhh….pokemon jokes

#

anyway

crimson beacon
#

will the 800m sd3 loras be compatible with the 8b model?

opal hedge
#

Maybe eventually through new techniques like x-adapter, but the vanilla versions probably not

regal glacier
iron pendant
astral goblet
#

i've only heard one mention of the 800m param model in the original announcement. nothing in the paper

gaunt pulsar
#

I'm starting to lose the hype since it's taking so long for the preview invites, ngl sadcat

opal hedge
gaunt pulsar
#

Wait let me see if I can- nvm

opal hedge
#

The hype ain't for us sadly

gaunt pulsar
#

Yeah it seems that the hype was mostly for investors

karmic cedar
#

They need to do a better job of labeling the hype trains so us bottom feeders know what’s taking us where

astral goblet
#

you underestimate the bottom feeders ability to entitle themselves

#

i deserve sd3. i've worked hard for it!

honest spear
#

I think it didn't hype anyone, they said few dozen of invites have been sent and to look for #sd3 posts, but after a week looks like only one or two people got invited

pale latch
#

Most of the hashtag is other people spamming stuff unrelated to sd3. Just because so much thirsty attention is on it

ornate flame
#

SAI has to know that more people become impatient every day, so it probably won't be more than a few weeks before mass invites are sent out

crude notch
#

according to emad, limited invites rolling out this week

#

mass invites soon

ornate flame
#

nice, i was right lol

crude notch
#

every smaller model was an afterthought

ornate flame
#

are the parameter counts of the other sizes known? Or just 800m and 8b known so far?

astral goblet
#

the paper mentions a 2m param model i think

opal hedge
#

Not randoms who put clouds in a bottle or generate booba

karmic cedar
#

what about people who make hourlong youtube videos with hundreds of examples of them with the same pair of glasses wearing every kind of superhero costume you can imagine

astral goblet
#

and gingers

crude notch
sterile raven
#

Sd3 how long will u make my poor heart wait 😥

crude notch
sterile raven
#

😦

crude notch
#

or less

#

idk i dont release them

crude notch
trim magnet
#

will SD4 loras be compatible with SD3

crude notch
#

i hope not

pearl ocean
#

SD4 when

trim magnet
#

SD4 is old news

#

when SD5

warm junco
mystic rapids
#

hey mates

opal hedge
#

👋

mystic rapids
#

nice to meet you

unkempt hatch
#

I want to prepare datasets for SD3, but will there be anything different from SDXL?

mystic rapids
#

When preparing datasets for SD3 (Shoelace Data 3), compared to SDXL (Shoelace Data XL), there might be some differences based on the specific requirements or characteristics of each dataset.

#

Here's a general some differences you might encounter.

#

Data Size: SDXL typically refers to extra-large datasets, implying a larger volume of data compared to SD3. SD3 might be smaller in size, containing a subset of data or focusing on a specific aspect or subset of the larger dataset.

#

Scope and Coverage: SDXL datasets might cover a broader range of topics, domains, or categories, while SD3 might be more focused on a specific area, industry, or domain.

#

and so on...

zinc egret
#

hi ! i m using webui to run Stable diffusion , is there a pc app to run it local ? i didn't find one.

gritty dust
#

webui is local, it's only using your browser to display stuff

#

you can disconnect your internet and it will still work

zinc egret
#

i know ... i m looking for a interpreter that can run SD

#

??? there s none up to date as i know ... that s why i ask

#

using webui is very complex for normal users

gritty dust
#

you mean programs like comfy or Automatic?

#

have you tried fooocus, i've heard that's easy to use

zinc egret
#

i ll check 2 sec

#

it s not local

#

more like lm studio ... but it for chat bot

gritty dust
#

what do you mean? you can download fooocus and run it locally

zinc egret
#

i dont code python lol

gritty dust
#

you don't need any coding experience for foocus, A1111, or comfyUI

zinc egret
#

i m using a1111

gritty dust
#

and what do you dislike about it?

zinc egret
#

it s to complex for normal user ... not for me lol

gritty dust
#

oh you mean run an LLM locally to use SD?

#

what's a normal user?

#

do you mean you don't want to bother with step size and stuff, isn't fooocus then pretty easy to use?

#

A1111 you can also just copy settings from others and then go from there and only mess with stuff you know, like prompt

zinc egret
#

i got to run 3 prog for full ai power ... a chat ai , open interpreter , and webui ... why not make one app ?

gritty dust
#

oh i havn't heard about open interpreter

zinc egret
#

i test it ...very powerfull but but it very dangerous to lol

gritty dust
#

dangerous how?

zinc egret
#

you can arm your own pc ... lol ooops be carefull with it

#

open interpreter use python code to execute command in your pc ...it can do reseach on the web for you , it can start app ... event format a disk ooops

gritty dust
#

oh dang, yeah i wouldn't use that after my experiences with how bad language models can be lol

zinc egret
#

i didn t make it use webui yet ... is to complex

#

but i m sure it can

gritty dust
#

i guess you could run SD without a user interface

zinc egret
#

the probleme , is to run SD in a web page use ram for nothing

#

a browser is not the best interface ...

#

also a browser is to easy to hack

gritty dust
#

yes, you might want to look how to run stable diffusion through python code directly

zinc egret
#

so i ll keep looking but i think some interpreter will soon come wt options wt SD models

gritty dust
#

sure

zinc egret
#

like jan.ai , gpt4all , lm studio , ... someone will do it ...who will be first ??? lol

gritty dust
#

i'm surprised that those LLMs run locally on 6GB VRAM already, havn't kept up to date in a while, might check out some of it as well

zinc egret
#

i m running SD 1.5 wt controlnet on a gtx1060 ti 6g Vram laptop and it make 688x470 ( wt 3 ctrlnet running ) under 3 min lol

#

it very powerfull ... i m totaly amazed

#

it s a bomb lol you dont know what you created in the world of CGI lol

stone latch
zinc egret
#

a agree ... i tested alot of them ...lm studio simply rock lol

#

multi AI now lol

#

yes they did that ..you can run 2 or more AI in 1 pc

#

lol

stone latch
#

I wish I was a rich kid who could afford 2 4090s and 128 GB of RAM 😂

zinc egret
#

it s pure madness

gritty dust
#

i bought a new laptop last year with 8gb vram to be able to use SD and it's running quickly as well, before I had to use CPU and wait 50 minutes for every image lol

stone latch
#

Runs SD3 + a 30-billion parameter LLM on his PC at the same time

stone latch
#

Depending on the SDXL checkpoint 1024x1024 usually takes around 30-40 seconds

#

Base SDXL takes around 10-12

zinc egret
#

@gritty dust , well my friend got a gtx 1650 wt 12g of vram laptop ... i ll try to find a used one soon

#

i cant run SDXL in 6g of Vram

gritty dust
zinc egret
#

@stone latch i got a lap ... i cant change my GPU

stone latch
#

My little brain might not be able to handle it 😅

gritty dust
#

yeah, buying a desktop was also out of the question for me, so i got a laptop with a 4060

zinc egret
#

the idea of working wt laptop ... it cheaper than rtx 3060 lol ...no hope for a 4090 at all lol

gritty dust
stone latch
#

Now I'm compelled to try ComfyUI but I wonder if I have disk space...

zinc egret
#

@wally dm me a link plz

gritty dust
#

for comfy?

zinc egret
#

yess

gritty dust
zinc egret
#

i ll try it in the morning ... it s late

#

ok

stone latch
zinc egret
#

see runing a aI use all the ressource of your pc

stone latch
#

A1111 has way more capabilities and settings to tweak... That are easy to tweak

zinc egret
#

so having laptops let you work wt different ai and event share them with a local server ...yes

#

@stone latch yes i agree , but i m lookinf for a pc app that can run SD localy ... i dont tink someone did it yet

#

it s the concept of browser app that bug me

stone latch
zinc egret
#

me too

gritty dust
#

yeah i also don't know why it's always webui, browsers take up previous ram

stone latch
#

There might be a reason for this, but I'm unsure why someone hasn't sat down to write a C# or C++ application for Stable Diffusion 🤔

zinc egret
#

it leave trace in the cache ..

#

i m sure 2024 will be the year of SD

stone latch
#

Dall-E 3 for prompt adherence (currently) + SDXL ControlNet to enhance the image and do inpainting if desired 😌

gritty dust
#

you can't run Dall-E 3 for free, nor locally though

stone latch
#

I did this to generate portraits - Dall-E 3 because I often like how it does male faces, then SDXL to add realism

zinc egret
#

i got a issue with AI that are not local , i study local ai only lol

stone latch
gritty dust
#

i want to try some workflows where you multiprompt sspecific regions.

#

promt adherance is less important if you can specify a region in a picture where a thing is supposed to be

zinc egret
#

my freind pay 30$/month for gpt pro ... i show him the power of local ai ... he fall down lol (chat gpt pro got a limit of 40 question every 3 hours) ... make me laught again

stone latch
#

The subscription is $20 without VAT 😅

narrow kernel
gritty dust
#

you don't know which $

stone latch
#

AUD?

zinc egret
#

30$ can

#

lol

#

ooops

stone latch
#

Ah my bad

narrow kernel
#

All the python level code these generators use is just tapping into the API of torch

#

The actual hard calculations are done at a very low level

zinc egret
#

ok ... i m not that deep into coding ... i runit without any internet ... it work but contronet need the web to edit is control lol so it not perfect yet

#

i use 2 webui ...one for my chat AI and the other for SD ... why 2 webui ...i just dont get it

#

i can run lm studio as a server for my AI ... but i cant run it with sd

#

yes ... auto generation of original visual content ... lol that s the plan

#

that would be a bomb

#

lol

#

we can do it via open interpreter ..but it s alot of work

#

... a guy ! make open interpreter deliver him a pizza at home ... wt a txt line ... it s that powerfull ..

#

my problem a cant run both at the same time on my pc

#

you need 12g vram to try it

honest mica
#

With comfyui you could just integrate a LLM into the image generation process. It would offload the LLM model after the inference is done to free the resources for SD.

zinc egret
#

i need 2 laptop

#

ok good i didn t know , everyone talk about comfy ...i intall it tomorow ..

#

open interpreter do more then just answering question ... it can run your pc

#

i need it

narrow kernel
zinc egret
#

i use only local AI ,

narrow kernel
#

you just have to make sure to call it in the node when you send the request to llama.cpp/python or ollama or whatever local server backend you're using

honest mica
#

That might be a bit of a problem for a comfy beginner 😬

zinc egret
#

@narrow kernel i m new in this ..lol

#

@honest mica exactly

#

i just discover SD a week ago ... i m amazed

#

i m a cgi artist ...everyone is freaking out lol

honest mica
#

Prepare yourself for a steep learning curve with comfyui. But you will get rewarded with high flexibility.

zinc egret
#

good

#

then think sd will steal there job ...they are so wrong

#

it will make it more creative lol

honest mica
zinc egret
#

anyway i ll give you news tomorow about comfy

honest mica
zinc egret
#

@honest mica sincely to get excaly what you what from sd it will take you 3h of prompt testing , in 3 h i can do 10 images lol

#

so ...we are far away to be replace lol

#

but for creation it blow me off totaly

#

change backround in a second wow

#

it speed up the process alot

honest mica
zinc egret
#

... you all know the expression ...with me , without me ... well with AI , without AI ... lol the new version of it lol

honest mica
#

Well, you say you use it a week. Have you heard of controlnets already?

zinc egret
#

yeah

#

controlnet is impressive

#

i try many options but it still limited to SD 1.5

#

still hand and feet proplem

#

need to train it to get better result

#

there is so many extention ...to learn them all it will take a year lol

#

mastering controlnet il take me a month at least lol ...

#

ok comfy is nodal ui ...yes i read about it ... hummm need hd space for it

#

anyway thanks for your advice

limpid bluff
#

Hey all ... I've been off the server for a while so I might have missed it ... I can't find the bot channels to put in a prompt any more ... sorry if this is a noob question ... is that no longer available?

gritty dust
#

10 years ago, who would have thought that AI programs struggle the most when asked about basic maths and not creative prompts, and image generation can draw nice organic shapes, but can't do text, straight lines and proportionate repetive building elements.

zinc egret
#

@gritty dust .... i need to test lora ... if you give him building reference ... it can create a city

#

lora is a major part of SD

#

as a fact ... SD is the only AI that can learn localy ...that i know

young sinew
supple acorn
#

I have learnt that deep faking people into animated characters is incredibly difficult

#

Especially really stylistic charachters

wary belfry
# crude notch 800m, 2b, 6b and 8b

so then we will get loras for 4 different sizes and none are compatible with each other? Sounds horrible. You will basicly never be able to combine multiple loras of your choice because they probably are all on different sizes... or are they compatible with eeach other ?

fresh pond
#

guys, what is the exact commant to use that SDXL

#

command*

flint fox
shy lagoon
#

Is there a general online solution for an actually good quality upscaling of watercolor or oil paintings ?

pseudo nacelle
#

what does the controlnet do, and where to put the file?

trail lion
glossy igloo
#

hello guys, i'm wanting to instal the extension DreamArtist-sd-webui-extension, in fact i instal it like said in the github page, adding that command git clone "https://github.com/7eu7d7/DreamArtist-sd-webui-extension.git" in the folder of c:/stable difusion/extension, but i havent the tab to acces to the extension when i open stable difusion, someone could help me ?

trail lion
glossy igloo
trail lion
#

yes

glossy igloo
#

thanks a lot

glossy igloo
trail lion
glossy igloo
pseudo nacelle
trail lion
#

that was rude btw

trim magnet
pseudo nacelle
#

ok, thanks

trail lion
#

it's becoming clear to me now why ai bots will eventually replace search engines. the culure is changing, how people research

olive wave
#

Hello folks. I have a question. Does anyone happen to have a tutorial or how2 that explains sensibly how you can host a GPU or alternatives. I have to admit that I'm unfortunately too stupid to understand all the providers' own docs.

stiff valve
#

Option1: you give them money they let you borrow GPU
Option 2: you give store money, you take GPU home

#

but to answer your question, i dont know of anything other than youtube guides

trail lion
#

are you asking how to be paid to lend out your GPU or how to rent a cloud GPU?

#

when you say host a GPU, that normally means your GPU is being used by others

olive wave
trail lion
#

I've only rented them, the price/hr depends on the service and the hardware, like Vast is really cheap for like a 3090, I can get one normally for like $0.25/hr US

stiff valve
#

renting out your gpu is like the 2024 version of torrenting, lol

#

gotta have a good seed ratio 😄

olive wave
#

my intention is. I would like to host stable diffusion local on my PC. my own gpu is more than too weak for that...lol... I only have a GTX 970. That's why I wanted to get the necessary gpu online.

trim magnet
#

yea plus cant really make money with just a 4090 because clusters offer better hardware and prices

upbeat kestrel
#

Hello - is it publicly known what the enterprise level of SAI entails? (% of revenue? something else?). I emailed the address 7 times (SEVEN) and never got a reply. This is on behalf of investors who are hesitant to back entrepreneurs without first understanding how much SAI will take of the final revenue. I hope this is the right channel to post, thank you.

trail lion
stiff valve
#

tbh, no one is going to rent a 970

trim magnet
#

no i think he wants to rent a better gpu online because he has a weak one

trail lion
#

hah, I think they were saying they cant use the 970, so trying to setup a cloud resource for permanent use

stiff valve
#

he said both, as in renting his out, and renting a cloud one

trail lion
#

ok, I misunderstood then. if that's the case, that'll never work

olive wave
trail lion
#

best option is just to rent a cheap cloud gpu when you need to do something. If you're using it for both inference and training, over time that will get expensive, and you should consider buying a better GPU

stiff valve
trail lion
#

some people have made free services like kaggle or colab work, but it's definitely a lot of hassle

stiff valve
#

i need to order a new ssd, a 4090, or rocm needs to come to windows asap. the 7900xtx is great, but still kinda slow in windows, and my linux ssd is decaying

olive wave
#

oh man...lol...you confuse me just like the explanations on google cloud colab...lol...it gives me a headache reading it. 🤣

stiff valve
#

haha, i get too excited about tech

trail lion
#

I'm on a 6800xt in linux, works ok for most things, but I do use cloud for training

stiff valve
#

i'm signing off, going to spend the day in Akiba tomorrow so need to rest 😄