#๐Ÿ’ฌ๏ฝœgeneral-chat

1 messages ยท Page 124 of 1

pearl ocean
#

Iโ€™m pretty hyped for the fixed text

nova zodiac
charred mesa
#

Well hopefully 3-5 weeks now that it's been a week since the CTO said 4-6 weeks sadcat

#

We could be playing around with the smaller models by now

nova zodiac
charred mesa
#

Especially with how much cheaper it would be to train on

nova zodiac
#

Thats why sd1.5 is arguably more used than sdxl

#

Better ecosystem, more time with it

bleak matrix
#

Good morning, everyone!

opal hedge
fervent thunder
#

Same

#

I have 1660ti, quite an old card, and forge is the only viable option for me to run something

#

I have a 3090 coming soon, but it was nice to be able to give SD a try even with low end pc

nova zodiac
#

I like forge, good inbuilt extensions and good model management under the hood giving good speed. They just need to get compatibility with supermerger sorted

fervent thunder
#

I will be moving into comfy with the 3090. Its complicated but also necessary for real world use cases.

rocky island
#

there is rtx 3090 for 670$ and 3080ti for 500$.I think it's better to buy a 3090 for sd 3. Otherwise, the 3080 might not pull it

nova zodiac
nova zodiac
fervent thunder
# nova zodiac Whats your use case??

Im visualizing menus. You pick style or just improve asthetic of existing menu photos. Its good for new restaurants, more engaging menus, and advertising

#

So text to image for all menu items

charred mesa
#

absolutely 3090

#

even if it ends up running on 12GB you will still benefit from being able to do higher resolutions and more tools running alongside SD3

#

like controlnets and such

#

im gonna a get a 3090 as well thomas

shy lotus
#

hi

regal wren
#

Will version 3.0 require a large onboard GPU memory?

#

Where can I test the model now?

#

I don't need a commercial licence

#

(yes I've signed up for EA)

trim magnet
#

its not out yet
its not out yet
its not out yet
its not out yet
(its not out yet)

regal wren
#

ugh, the pace of development in AI is SO SLOW >.<

#

I need instant gratification

opal hedge
#

How grateful our great grandchildren will be when SD3 releases

urban osprey
#

ham pagal nhi h bhaiya hamra dimag kharb h

#

๐Ÿคฃ

#

๐Ÿคฃ

trim magnet
urban osprey
#

u mean i am mad

#

?

#

๐Ÿ˜ก

#

i am indain

#

proud indian

#

JAI HIND!

trim magnet
urban osprey
#

why

fervent thunder
kindred folio
#

Does anyone know of any daily challenges anywhere? Iโ€™m looking to learn and broaden my horizons and a daily challenge would be a fun way to do it.

forest trout
foggy halo
#

sd3 release when

broken smelt
#

Could be tomorrow, could be next June idk

foggy halo
#

๐Ÿ˜ญ

opal hedge
#

(or maybe not)

worn aspen
#

Is there a reasonably accurate guide on chkpt models, as in which ones are good for what? I'm getting model fatigue as there seem to be too many. I know storage space is cheap, but I'm doing all of this on a notebook.

Or, can I run chkpt's from an external usb drive in ComfyUI?

honest spear
thorny osprey
#

I have some questions that hopefully someone here can answer as I have only been tinkering with ai for about a month now. I have used both Automatic 1111 and Forge. Forge is definitely a bit faster on my 3080ti. I am generating batches of images and I am coming across some issues. First the faces need to be fixed and I am using Adetailer and that seems to work just fine. A detailer adds about 1.5-2 seconds of extra generation time. This seems to be acceptable but I am curious if anyone knows any other fast fixes for faces at the moment. Next sometimes my fingers come out strange. I have tried embeddings such as badhandv4, negative_hand-neg, and deepnegative. They all seems to give better yet different results. I do not know what the best embeddings are for hands. I am curious if anyone has more info on this. Finally I use ChaiNNer to batch upscale my images by 3x. ChaiNNer honestly seems to be like a godsend piece of software. Very fast and very reliable.

left fulcrum
#

Hello, Where can I genara a Poster?

opal hedge
#

That's the stable diffusion checkpoint starter pack

loud solar
#

Oh yes! Ride me like a pony! ๐Ÿ˜„

ornate blade
#

Any news or hints recently on when SD3 will drop?

loud solar
worn aspen
#

@honest spear , @opal hedge Thank you both!

karmic cedar
worn aspen
#

@alex v 'preciate it.

loud solar
pearl ocean
#

Epicrealism is pretty good, it actually does what I say.
But I do have type the prompt more like a sentence

karmic cedar
loud solar
#

Better change your supermarket ^^

viscid stag
#

anyone?

#

or is there any article that explains those files?

fervent thunder
#

can someone help me

#

i downloaded a style LORA and idk where to put the file, in lora folder?

amber bloom
proven tendon
#

Is there any way to create videos

#

I heard somewhere there is a model

#

Not like Sora videos

split kestrel
#

New control net was released at 2pm today

#

(Fine tune)

rich kestrel
#

what does it do

split kestrel
#

Check general chat with images

rich kestrel
#

oh its comfy... ๐Ÿคข

keen rose
#

Is there like an img2img sort of thing where you put in two images and it generates something that averages the two pictures?

split kestrel
#

Why the barfy face?

rich kestrel
#

A1111 for life

split kestrel
#

Whatโ€™s your reasoning ?

rich kestrel
#

UX

split kestrel
#

Fair. I would like the interface to be better..

rugged mirage
pearl ocean
fervent thunder
#

i just downloaded forge

#

what is new abt it

#

noticing its faster

#

theres a model that used to take like 10 mins for an image on auto1111

#

but its taking way less on forge

shell tendon
#

it's much faster on my 4090 and my 3080 12gb, and every other card i've heard of other ppl trying

#

and uses less vram too

granite solar
#

hi

pearl ocean
thorny osprey
pearl ocean
uncut mason
#

How long should StableDiffusionPipeline.from_pretrained() take for a local model?

fiery mountain
#

hello

pearl ocean
#

remember to always take you're @fiery mountain

blissful finch
#

How does Imagine.art have SD3?

trail crag
#

good day

grim parcel
pearl ocean
#

will SD3 be better than SD-XL?

cosmic marten
#

hi

#

what is

twilit solar
#

Is SD3 going to be on civitai

broken smelt
#

will SD3 change the world

crimson zealot
heavy lark
#

will sd3 cure furries?

pearl ocean
crimson zealot
pearl ocean
vague pond
pearl ocean
low moon
#

and waifus

gaunt pulsar
low moon
verbal osprey
pearl ocean
full lark
verbal osprey
pearl ocean
full lark
verbal osprey
#

Ahhh. Do you have to mess with the backend, or can you just mess with it if you want to get your hands dirty? I like the simplicity of A111

full lark
#

Its like a1111, you dont have to use nodes or things like that

pearl ocean
#

You can even make the images animated with SVD

verbal osprey
#

Ok now you are selling me.

full lark
verbal osprey
#

Forge is available right now? I tried one a111 clone, it was called vlads or something like that. It didn't really impress me too much

full lark
verbal osprey
#

Is this agreed upon as the best successor to A111

pearl ocean
#

Forge can change a life!!

verbal osprey
#

Also, can I keep my models in my A111 folder and point to them in the forge UI? I have a lot of stuff organized and I really don't want to have to copy large model files to another directory

karmic cedar
#

anyone messed with this?

full lark
#

Not yet, i have to try it ๐Ÿค—

shell tendon
#

here's what's weird though.

#

idk why.

#

forge is actually significantly faster than comfyui...

pearl ocean
shell tendon
#

for dpmpp_2m_ancestral with karras, 1024x1024 on a 4090...

#

a1111: 1.83 it/s
comfyui: 2.96 it/s
forge: 3.71 it/s

pearl ocean
shell tendon
# pearl ocean I never understood these it/s, I always assumed it was mb a second or something

100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 60/60 [00:16<00:00, 3.62it/s]
Requested to load SDXL
Loading 1 new model
100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 60/60 [00:27<00:00, 2.16it/s]
Prompt executed in 55.23 seconds
got prompt
[rgthree] Using rgthree's optimized recursive execution.
Requested to load CLIPVisionModelProjection
Loading 1 new model
Requested to load SDXL
Loading 1 new model
100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 60/60 [00:30<00:00, 1.98it/s]

#

take the steps and divide by the it/s and then look at the runtime

pearl ocean
# shell tendon 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ...

๐Ÿ˜ฎ

trail lion
shell tendon
#

lol

#

good to hear ppl are trying it

#

it really is like a1111 minus the whole being a memory hog slug

verbal osprey
shell tendon
#

more stuff is ready to go out of the box

pearl ocean
#

Do you guys normally do 1024x1024???

verbal osprey
grizzled harness
#

It generally supports them all.

shell tendon
grizzled harness
#

Some extensions have specific Forge versions.

shell tendon
#

but eveyrtihng i've tried has been no prob

#

my advice for the last month or so to everyone has been...

#

use forge, comfy, or ideally, both.

pearl ocean
shell tendon
#

literally the only thing i use a1111 for now is lycoris-ia3 as neither forge nor comfyui seem to support it (at least, that was the case a few weeks ago)

trail lion
# verbal osprey what are some things about it that you like better than a111?

for one, I never get OOM anymore, even after switching models multiple times. it does seem faster, but that really wasnt what I was looking for. the built-in extensions are are well thought out and at first I though, why all this extra crap, but it starts up pretty fast, so it's not actually a burden and I appreciate most of them being there

shell tendon
#

yep those extensions are really useful ones you def want

#

i was happy to see them there out of the box

verbal osprey
#

You guys are selling me. I'm going to try it.

shell tendon
#

you're going to laugh your ass off if youre anything like me

#

at how much faster it is than a1111

#

and again, as a primarily comfyui user (btw, i hate that shit ppl sometimes do on here... Team This, Team That)

#

forge is definitely faster. don't know why, but it is.

#

you're not gonna get the control and power user capabilities you'll get with comfyui, but you'll be able to do most things still just fine

#

i use comfyui mostly because i like to make really, really wild images

grizzled harness
#

I haven't noticed it being THAT much faster. I ran some benchmarks and it had some mild differences. Maybe because I have 12GB of VRAM...

shell tendon
#

version of windows vs linux, etc

#

but on my 3080 12gb, forge was vastly faster than a1111

#

i couldn't run batch sizes greater than 3 on a1111 with sdxl without OOM

#

forge, i was running batches of 8 without it even blinking

brazen forge
#

anyone know how to do text pictures like SD4 localy?

shell tendon
#

and yeah, on my 4090, forge is about 2-3x faster depending on the sampler

grizzled harness
#

batch size != speed

shell tendon
#

yeah, i know

#

commenting on vram use

#

obviously

pearl ocean
#

Since I discovered Forge, A1111 is dead to me

grizzled harness
#

Need more controlnet stuff for sdxl ๐Ÿ˜
Last I recall there were still limitations.

brazen forge
#

what is forge?

shell tendon
#

yeah controlnet for sdxl isn't great

#

but there's other things now that are even better imo

pearl ocean
grizzled harness
#

like, I would like to create images of the same character

#

character consistency or whatever

shell tendon
#

right now, afaik it is comfyui only, if there's a forge/a1111 equiv i'd def like to know so i don't misinform ppl... but ipadapter plus has new nodes (these are NOT new models) that can do stuff that's sdxl-only like style transfer and composition transfer

#

the style one in particular is spectacular

full lark
trail lion
pearl ocean
shell tendon
#

i'll drop a piece of software in 2 seconds if i find something better

#

it's just a tool

grizzled harness
#

Comfy drives me absolutely batty, so I only fire that up if absolutely necessary.

shell tendon
#

#TeamNothing

brazen forge
verbal osprey
#

hahaha. A111 definitely feels slow......I'm excited.

shell tendon
#

a1111 is slow as fuck

full lark
#

Its faster than comfyui on my PC ๐Ÿ˜

shell tendon
#

the reason i tried comfyui in the first place was because a1111 performance was so god awful on my 3080 12gb

shell tendon
full lark
shell tendon
#

what's your gpu and what it/s for dpmpp_2m_ancestral with karras as the scheduler, 1024x1024?

full lark
#

Im using a rtx 3060, 1024x1024, comfyui up to 22 seconds, a1111 up to 19 seconds, forge up to 14-15 and sometimes 17 in some models

shell tendon
#

oh, well that's really weird, i wonder if your comfyui install is wack

verbal osprey
#

I'm mostly concerned about when stable diffusion 3 comes out. A111 is always slow to incorporate new SD releases.

trail lion
#

my only concern with comfy has nothing to do with performance, its simply that I would spend more time tweaking workflows, which has nothing to do with why I'm actually in it in the first place

brazen forge
#

can you still expose the api to share with ollama?

verbal osprey
shell tendon
#

i don't worry about that shit, i have lots of functions hotkeyed on my g502 gamer mouse

#

i just fire up a workflow, copy paste stuff from old ones and go

#

no organizing or anything unless i'm sharing it on request

trail lion
#

haha, nice, but what would make that first image killer is if all those tiles made a stick figure or something, that would be meme-worthy

shell tendon
#

i find if i start carefully organizing anything it turns into the rabbit hole you described... hours wasted, nothing made

#

ha yeah

pearl ocean
shell tendon
#

ahh nice

#

here's the issues with comfyui as it stands imo

#
  1. documentation for most nodes either is shitty and incomplete, or flat out doesnt exist.
  2. a lot of times there's unnecessary redundancy (why can't there be a preview or save image node that has a vae and latent input?!)
  3. needs better node grouping support... in particular the ability to group stuff like a function, where it could be expanded and collapsed, would be hugely useful, or if you double clicked on one, it'd open a new window with the "subworkflow" in it for that grouped node
  4. chaos with node packs where people just throw in the kitchen sink with everything they've made that again, is often undocumented
#

great things about it: you can do just about anything, and if you stop worrying about making it look nice, you can work just as fast as in webui

trail lion
#

I think it shines with automated workflows, so maybe when video workflows are less crappy

#

but by then someone would probably make a cool video workflow for forge

shell tendon
#

i find working in webui to be more tedious and repetitive because a lot of tasks i just group up some nodes for in comfyui i end up having to do manually

#

but yeah, again, i'm doing some weird shit

#

if you want to make images that make ppl think your acid use has expanded into the workweek, comfyui is the shit

verbal osprey
#

So out of the box forge allows you to make video out of images? Or do I need an extension for that?

harsh crescent
#

Why canโ€™t I find the discord creation channel, which is a channel where I can generate pictures online? Can any expert explain it?

dry trellis
harsh crescent
#

Hey hey hey, no wonder they are all deployed locally, I see they are all comparing the differences

harsh crescent
shell tendon
trail lion
pearl ocean
#

I bet some people use SD just to generate babe's

shell tendon
#

i think that's most ppl ^^

#

boring imo

frank rampart
#

Anyone familiar with using KOYA to train your own models?

harsh crescent
shell tendon
trail lion
shell tendon
#

oh i mean if you type a prompt in there i might copy it verbatim into comfyui and see what comes out

pearl ocean
# shell tendon boring imo

I perfer to generate things, I would also like to generate charaters from games like mario and sonic, etc etc, but I don't think SD is built for game characters like that, coz its more of a realistic model im guessing

shell tendon
#

oh you can def make those

pearl ocean
#

๐Ÿ˜ฎ

verbal osprey
frank rampart
trail lion
verbal osprey
#

I have a 4090 but I am salivating at the power that a 5000 series card might have. I need more power!!!

trail lion
#

you have the best consumer on the market, chill ๐Ÿ˜‰

#

I built a PC from scratch, but didnt want to cheap on those parts to get a good card (or pay the nvidia tax), so I just got a decent GPU at the time and splurged a bit more on the board and other components knowing I'd eventually upgrade the GPU anyway

frank rampart
trail lion
#

bf16 seems like the safe choice, it's about the mixed precision

shell tendon
#

yeah bf16 is pretty good

trail lion
pearl ocean
#

I think some people underestimate the power of SD-XL

karmic cedar
trail lion
#

not that I know of

karmic cedar
#

I wonder if 32-bit precision will eventually become a thing

#

โ€œOn the full SWE-bench test set, SWE-agent resolves 12.29% of issues, achieving the state-of-the-art performance on the full test set.โ€

#

So itโ€™s a robot code janitor that can hypothetically go into the entire GitHub repository database and clean up 12% or so of all lingering issues?

haughty saddle
#

xin chร o cแบฃ nhร 

autumn grove
#

Does anyone know when SD 3.0 will be released to the public?

trim magnet
#

soon

autumn grove
#

lol I hope so

pearl ocean
autumn grove
#

Ok thanks!

shell tendon
karmic cedar
pearl ocean
#

People who have 32gb of RAM have never experienced 64gb of RAM!

shell tendon
oblique willow
pearl ocean
lucid oxide
#

Hello, Everyone.
I want your help.
I am going to convert any dog image into sympsonized one using stable diffusion project.

oblique willow
#

what is sympsonized

rich kestrel
#

looool

fervent thunder
#

high soggy camel variable

#

the taco way shrimp cattle way

#

upper echelon tacorex-PYrex

#

kawasaki cheese rex

#

airbag jennie chewables

#

WHY WONT THE BOT GENERATE

lusty beacon
#

SD3 where are you? ๐Ÿค”๐Ÿฅน

loud solar
#

No bot here

mild hedge
#

@lavish lake @wintry stream Sorry for the ping, you both were involved in a support ticket I had open last September. I'm hoping you may be able to help push along my new ticket. I opened it Mar 21st, it was then claimed by someone on March 22nd, and I have not seen them online since then. If you were able to take a look I would really appreciate it!

slow aspen
#

1324

hoary ridge
#

hello

forest wyvern
#

hello

abstract stratus
#

1

spice crypt
#

1

tepid vortex
#

1

cyan quarry
#

1

loud solar
#

11 00 101

fleet pelican
#

heelo

tidal tangle
#

hello

rocky island
#

what about RAM for sd xl? I have 16 GB of ddr3. is this enough if the graphics card is rtx3090? will there be any problems due to lack of RAM?

bleak matrix
#

Good morning, everyone!

#

How are we all this morning?

loud solar
#

I would try Stable Diffusion Forge ...

rocky island
#

okay

loud solar
# rocky island okay

Easy to start and uses less resources ... advanced might wanna check ComfyUI ...

buoyant hornet
#

https://forms.gle/9i4jM9BQu9bVVAAF6

As a young professional just a few years into the workforce, there is a constant, low-humming anxiety about proving yourself and finding that mythical work-life balance everyone talks about.
Sometimes you can't help but wonder - is this really what you signed up for? Or is there a better way to approach this whole "work" thing as a young professional?ย 
We at 5day.io have the same questions as you.

Our goal for this survey is to understand your work habits and present you with a work management ecosystem so good that it brings your spark back.ย ย 
Do you have a minute to simply tap on 7 answers?

It will not take you more than 1 minute 49 seconds. We checked.

austere pollen
#

how fast u guys generating a 1024x1024 XL 20 steps on A1111? Idk if im going slow or what

full lark
austere pollen
full lark
#

๐Ÿค”

austere pollen
#

will keep trying

full lark
#

It should be way faster, what settings are you using? I use dpm++2M karras and about 5 to 7 cfg scale

#

I hope you can fix it ๐Ÿค—

austere pollen
#

Euler a, 7 cfg, dynavisionXL checkpoint.

still new to all this so hopefully figure it out

rugged mirage
#

you can always try forge instead of a111 or lightning models to speed up a lot but yeh that sounds too long

#

check if your drivers and whatnot are up to date

#

also idk maybe you are using a vae and everyone else isnt or something

austere pollen
#

yeee so many variables to compare i suppose, everything upto date so prolly smthn dumb on my end

full lark
#

Are you using xformers or other optimizations in the webui.bat COMMANDLINE_ARGS? @austere pollen

austere pollen
full lark
#

Aaaah thatยดs the problem

#

you donยดt need medvram with that gpu

austere pollen
#

oooh lemme try

#

actually took 2.1 mins now agony

full lark
#

ร‘ooooo ๐Ÿ˜ญ

austere pollen
#

dooomed

karmic cedar
#

Stable Audio 2!

#

๐Ÿ‘

full lark
#

Maybe you gotta use medvram and remove xformers but im not sure, here are all the optimizations you can use

austere pollen
#

ill keep trying, ty tho โค๏ธ

full lark
#

๐Ÿค—

trim magnet
#

forge is faster with XL

austere pollen
#

does Forge have dynamic prompting n stuff? might have to try

trim magnet
austere pollen
last wasp
#

hello

full lark
#

Hi!

patent prawn
#

dasgvs

#

hi

honest apex
#

hello.

sand verge
#

hi

modest otter
#

ไธญๅ›ฝไบบ

shell tendon
karmic cedar
#

people are getting mad because SWE-agent is already about as good as Devin

#

lol

shy wolf
#

Hi everyone i hope I'm not bothering you. I tried to install the program but couldn't. I really need an AI drawing. Could anyone help me out with that? sadcat

loud gale
#

hello

trim magnet
#

some of the sfuff in supermerger works but not all of it

karmic cedar
shy wolf
#

I wrote A1111 and open Github and followed the rules, but there was always a problem (sorry for my english btw)

karmic cedar
#

Itโ€™s okay, I DMed as well

graceful wren
#

nihao

worn aspen
#

What is the functional difference between models and their XL counterpart? I see that image generation size is a thing, but from what I read it comes a price in variety and ease of getting the image to match your prompt if you're generating anyting other than people.

halcyon cloak
#

helo

lavish lake
forest trout
#

Guys, stable audio released their 2.0 model.

lavish lake
forest trout
#

It's an audiosparx model so it's still web only.

#

But you can prompt for 3 minutes for free now.

#

Stable Radio is back.

karmic cedar
#

itโ€™s okayโ€”seems to be a little repetitive for the most part

bleak matrix
austere pollen
shell tendon
#

so glad to hear you're getting such a huge improvement!

#

moving from a1111 to forge is as good as moving from a 3080 to a 4090

rotund meadow
#

Hello friends, would anyone be able to provide aid to someone who's a little bit stupid

shell tendon
#

just ask

rotund meadow
#

I'd like to make some means of converting an image into a set style, consistently

#

I tried to make my own model but the results are all over the place

sudden ruin
shell tendon
rotund meadow
shell tendon
#

what gpu do you have

pine fiber
rotund meadow
#

No just a bit dim

rotund meadow
pine fiber
#

you wanna make a lora

#

what did you try already

rotund meadow
#

I used DreamBooth and fed it a bunch of images

shell tendon
trim magnet
shell tendon
rotund meadow
#

Much appreciated, shall I DM you?

shell tendon
rotund meadow
#

Alrighty, shall do

austere pollen
#

got another Q if anyone knows:
im making stuff with dynamic prompts and it keeps putting text from one prompt onto other prompts.

i.e. wearing an adidas tshirt also makes the hat say the words "adidas" even though the hat is defined to say 420

#

and its split up like {adidas tshirt}{hat with "420" text}

arctic sedge
#

I hope everyone at SAI are doing ok. catlurk

shell tendon
arctic sedge
trail lion
#

maybe not today, but that forbes article a couple days ago sort of painted a gloomy picture

rocky island
#

are there still no anime styles in stable diffusion like in midjourney? for example, if I want the Osamu Tezuka style, then I will only have to train my own?

honest mica
hybrid ivy
#

So, is the bot ever coming back online?

rocky island
#

is there like all in one?

mortal delta
#

What do you all use stable diffusion for? if i might ask?

rich kestrel
#

porn

mortal delta
eternal ledge
#

Otherwise, it's just for the fun of art

mortal delta
eternal ledge
#

Also I have used it to create a logo before, but I don't need to create logos very often

#

My PFP on here is also made with Stable Diffusion

mortal delta
#

Im just trying to get ideas of things i can do with stable diffusion besides just for fun.

mortal delta
eternal ledge
mortal delta
eternal ledge
#

The subtle-est, but most important tip I can give for getting good results though is this:
Be intentional about your aspect ratio. You're not gonna get a good side view of a horse with a tall skinny aspect ratio

#

That and just use the finetunes, DreamShaper Lightning (SDXL) really lowers the bar for what makes a good prompt in my experience

mortal delta
#

you know ive been searching for a creative outlet and i think ai might be able to help, but at the same time im lost.

mortal delta
eternal ledge
coral aspen
#

is stable audio going to be open source?

mortal delta
eternal ledge
eternal ledge
mortal delta
eternal ledge
mortal delta
#

3 frames is not alot to work with.

#

but as seen on the subreddit it possible.

eternal ledge
#

Constraints breed creativity
-- I don't remember

mortal delta
#

ive also seen some use only images to create a movie, thats commonly done with ai versions of music videos.

eternal ledge
#

Yeah, and AnimateDiff can get you pretty far if you have some source video you want to reskin or you're okay with your base images not being faithfully recreated as frames

#

(When doing more advanced stuff for myself, I use ComfyUI, and I haven't gotten deep into animation yet)

astral goblet
#

sparse control net works really great with animatediff v3 modules. but it's not supported on forge last i checked. just base a1111

charred mesa
#

3090 acquired

astral goblet
#

i use it with my 4080 16gb with fp8 enabled

charred mesa
#

now I can use SegMOE

nova zodiac
charred mesa
#

not officially, but there's an addon for comfyui

#

its super easy

nova zodiac
#

Cool!!

#

Im surprised it never took off tbh

charred mesa
#

its smarter but not a massive massive boost

rugged mirage
charred mesa
#

thank you

sharp socket
#

Hello, I don't see bot channels anymore... Will they come back?

tepid finch
#

Hi, does anybody no when stable diffusion 3 is coming out?

charred mesa
#

3-5 weeks ETA

#

end of this month if we're lucky

tepid finch
#

ty

amber bloom
plush pilot
#

anyone knows where can I try latest models for free?

tepid finch
#

im compelety new to this. Is it anything like midjourney? I know we can install on the computer but is it pretty much the same concept where we prompt on discord?

plush pilot
#

or dreambooth

fervent thunder
#

decrepit smooth chili rice

tepid finch
#

i cant wait for 3 because im tired of ai mispelling stuff lol

crude notch
#

it can do text though

tepid finch
pearl ocean
charred mesa
#

12 on the 4080?!

crude notch
#

its 8b

#

and it really good for 8b

honest spear
crude notch
#

:3

#

(with offloading)

charred mesa
#

8B or 6B

crude notch
#

8b with offloading

#

6b too iirc

charred mesa
#

really? T5 at what quantization

crude notch
#

4bit

charred mesa
#

wow

crude notch
#

clips and NotUNet?! at fp16

honest spear
#

impressive, but it'll take more time to generate, or not?

crude notch
charred mesa
#

NotUNet is the best architecture ๐Ÿ”ฅ

crude notch
#

but i also guess that most people will use clip only

#

since t5 only good for text

tepid finch
#

would my system be able to run it? Alienware laptop 13th Gen Intel(R) Core(TM) i9-13980HX
NVIDIA GeForce RTX 4090 Laptop GPU
64gb memory

charred mesa
crude notch
#

yea

crude notch
#

for other things?

#

not so much

charred mesa
#

I know the 46% win rate when it comes to prompt adherence sounded good, but I don't know

crude notch
charred mesa
#

like could it do my prompt with 2 different fighters

charred mesa
#

w/o T5 and with T5

crude notch
#

t5 being there was a last second decision if i had to guess

charred mesa
#

wow

#

then what did they make the MultiModal part of MMDiT for exatly?

#

could we use other LLMs?

#

cause iirc they just simply replace the T5 weights with zeros to NOT load T5

tepid finch
#

Alienware laptop 13th Gen Intel(R) Core(TM) i9-13980HX
NVIDIA GeForce RTX 4090 Laptop GPU
64gb memory would my system be able to run it?

crude notch
charred mesa
#

(rough memory from the paper)

crude notch
charred mesa
#

oh

#

so only T5-XXL can be used or..?

crude notch
#

the ram will help a lot for offload

tepid finch
#

how long is the setup process from start to finish?

crude notch
charred mesa
#

ahh

tepid finch
#

can you point me to a site where it shows me step by step on how to get up and running?

charred mesa
#

thanks

crude notch
charred mesa
#

Also how does 1 step of SD3 MMDiT work? Doesn't it do like 6 forward whatevers per step

tepid finch
#

you guys sound like how I sound when I talk about crypto lmfao

crude notch
charred mesa
#

does that mean that it does 6 steps in one step or does it not matter to us

crude notch
#

it doesnt matter much

#

flow is faster iirc

honest spear
charred mesa
#

๐Ÿค”

#

I mean I will probably daily drive Turbo anyway

crude notch
charred mesa
#

SD3 Turbo is basically SDXL Lightning, which is just so epic habby

pearl ocean
#

To think, with SD, you can download Models, and make Hot Models with them

tepid finch
#

what does this mean? Prompt outputs failed validation
CheckpointLoaderSimple:

  • Value not in list: ckpt_name: 'v1-5-pruned-emaonly.ckpt' not in []
#

i guess i need to install a checkpoint?

#

oh nice i got it working!

fervent thunder
#

Hey guys im ๐”ญ๐”ฆ๐” ๐”จ๐”ฌ๐”ฉ๐”ฆ๐”ท๐”ž๐” ,im a certified and a walking W for work so im a pretty big fucking deal i'll tell you that,i used to be a part time W but recently i became full time so yeah i js wanted to introduce myself to y'all

tepid finch
#

ok so whats the best way to prompt. I want a rocket going to the moon and I want the text $Print written on the rocket. Is that possible?

tribal dagger
#

hey guys, idk if this is the proper channel to ask but i hope so.
what's the difference between sdxl and sd 1.5, at first I thought that they were just checkpoints but then I was investigating and found that they have like completely different architectures i guess, that's why they have different LoRas and need to be implemented, etc. but i suppose that they are also checkpoints, like there is the architecture and then the model, and the model only works on that architecture, and then the checkpoints that take the sd 1.5 or sdxl as base model take the model, adds its own data and publish the new checkpoint that runs on the architecture of the base model, am I correct? Also, why is the sd 1.5 model used as a base and not the sd 2.1?

trail lion
#

1.5 was an architecture, 2.1 was the next architecture, but not well adopted, SDXL is most recent architecture, and SD3 will be next

crude notch
#

cascade is the other one we dont talk about

#

mainly since its just StableWurstchen

pearl ocean
trail lion
#

mostly none of them are compatible from a trained model POV

charred mesa
#

me when imagen without T5

#

nah idk the exact idea of wurstchen

crude notch
pearl ocean
charred mesa
#

well finetuned models don't make those

trail lion
#

yah, but every time I go back to 1.5 I'm horrifed with all the deformations

charred mesa
#

but then again most finetuned models give you portraits

trail lion
#

like it was sooo bad, that I forget sometimes

pearl ocean
charred mesa
#

are you sure you are not talking about the base model lol

pearl ocean
#

Gotta go XL mate

tepid finch
#

im amazed on how fast this shit is

pearl ocean
tepid finch
#

4090 on alienware laptop

#

just doing 512*512

pearl ocean
tepid finch
pearl ocean
tepid finch
#

give me a prompt and i'll do it at 1024

tribal dagger
shell tendon
pearl ocean
#

XL can change a life!

shell tendon
#

major downgrade from a desktop 4090

#

different bus, vram, etc

crude notch
shell tendon
#

less than 60% the cuda cores of the desktop 4090

crude notch
tepid finch
#

i got this laptop for free from work so I aint complaining lol . They gave me the option for the alienware desktop but then I woulnd't be able to game on it at home lool

charred mesa
#

SEGMoE is kinda fun

#

yoinked have you heard of it

pearl ocean
#

SD3-XL when

tepid finch
#

yah at 1024 it does it in like 5 seconds

trail lion
#

prob never

crude notch
#

theres 4 sizes

tribal dagger
crude notch
pearl ocean
#

Arch Linux

crude notch
tepid finch
#

whats the best sampler?

crude notch
trail lion
#

arc de triumph

crude notch
#

everyone else will probably say dpmpp_2m

#

and for scheduler, use karras

shell tendon
#

reality is though that there is no best sampler

tepid finch
#

man you guys got me up and running in no time. Thank you!

shell tendon
#

there are some cases where dpmpp_2m is better, the sde samplers are better, the gpu versions are better, res_momentumized is better (and worse!) along with a litany of others

tepid finch
#

i thought it was like this 5 hr long process getting it all set up

shell tendon
#

the easier thing to do is identify which ones are BAD

crude notch
crude notch
shell tendon
#

but... that doesn't mean it's the best sampler

pearl ocean
#

Forge is best

trail lion
#

agree with the comment there's no best, some of them take less steps, and there's ancestral vs non ancestral, which I seem to recal the non-ancestral ones are better if you're trying to reproduce an image

crude notch
shell tendon
#

that means it may be best via a somewhat subjective metric (image quality) with regard to only runtime

shell tendon
pearl ocean
shell tendon
#

ancestral injects a bit of noise with each step i think

crude notch
#

i prefer comfy

shell tendon
#

the best sampler is the one that does the job you have at hand

tepid finch
#

give me a prompt i can test

pearl ocean
shell tendon
#

want a really crazy image with a lot of wild reimagining with each step? res_momentumized is king

crude notch
shell tendon
#

want stability with an image when upscaling? dpmpp_3m_sde_gpu with exponential, after unsampling with dpmpp_2m

pearl ocean
charred mesa
#

lol

tribal dagger
crude notch
#

yep

charred mesa
#

to think that SD3 could be named like SD3-S, SD3-M, SD3-L, SD3-XL

shell tendon
#

the easier thing to understand when you're starting out... is which schedulers you should be using

tepid finch
#

yah karras is making up some mangled images lol

charred mesa
#

interesting

shell tendon
#

rule of thumb there: karras should be your default, exponential if you want more stability in the image with each denoising step (so upscaling...), sgm_uniform is kinda in the middle in terms of behavior and has some special uses too

crude notch
charred mesa
#

ah

shell tendon
#

the others are handy too but if you just stick with karras and exponential in the beginning you'll rarely have problems that trace back to the scheduler

shell tendon
#

the other thing is... keep that step count at 25 or higher even if your computer sucks ass

#

my recommended default is 35

crude notch
#

uhhh

rugged mirage
#

unless you are say using lightning and 25 will fry it really bad

tepid finch
#

schedular is karras, cfg 8, steps 20, control randomize, denoise 1

crude notch
#

20-25 is good enough

pearl ocean
shell tendon
rugged mirage
#

I usually just check what the cool looking images on civtai with a given model used and use that

shell tendon
#

^^good advice too

tepid finch
#

this is awesome! im tired of paying midjourney

shell tendon
pearl ocean
tepid finch
#

sampler is dpmpp 2s ancestral

shell tendon
#

yeah try bumping your steps up to 35, drop your cfg to 6

#

see if that cleans things up

rugged mirage
#

idk, out of the box midjourney looks way more amazing at way less effort if you just want images, tho I dont use it

pearl ocean
#

I wonder what 100 steps would do

shell tendon
rugged mirage
#

usually 100 steps will make it overprocessed

tepid finch
#

someone give me a prompt, my prompts suck lol

shell tendon
#

my advice is cap it at 60, with the SDE samplers and the ancestral ones you can get really good results with the higher step counts

#

res_momentumized is the exception... especially with cascade, 100+ steps can do some amazing shit

#

much more intricate compositions

sleek scarab
#

when is this coming back, i swear its been down for ages

pearl ocean
#

Imagine making an image 2048x2048

tepid finch
#

check chat with images

#

lol i made arnold from hr giger style lol

shell tendon
tepid finch
#

pruned emaonly

shell tendon
#

and is that with dpmpp_2m_ancestral, karras, 35 steps, cfg = 6?

#

oooo

tepid finch
#

yup

shell tendon
#

you have a 4090 mobile right?

tepid finch
#

right

shell tendon
#

first thing is that's the base SD1.5 checkpoint and base SD1.5 is pretty damn bad

rugged mirage
#

looks good, I think you can decrease cfg and increase steps based on me eyeballing it

shell tendon
#

second is, with a 4090 mobile you can absolutely run SDXL no problem

#

and i recommend doing so

tepid finch
#

how do i do that?

shell tendon
#

the tiling is probably because you generated at 1024x1024

pearl ocean
rugged mirage
#

download an sdxl checkpoint

shell tendon
#

looks like civitais down right now

tepid finch
#

are those links for me?

rugged mirage
#

a proper 4090 is the size of like 20 phones, and uses the energy of a phone per 15m or something when running at max

tepid finch
rugged mirage
#

just play around in those directions

shell tendon
#

don't bother doing anything till you download that link

#

are you on a1111 or forge or...?

rugged mirage
#

though if you are changing checkpoint it doesnt really matter how it was on your 1.5 one since the difference between checkpoints and sweetpoint for settings is big

tepid finch
#

the juggernaut one?

shell tendon
#

yeah, juggernaut from huggingface

#

civitai has tons but it's down right now

rugged mirage
#

juggernaut is the most popular sdxl one though

tepid finch
#

damn these files are huge lolo

shell tendon
#

yup, it is worth it though

tepid finch
#

gonna run out of sdd space talking to yall lol

shell tendon
#

you def want at least one sdxl checkpoint

rugged mirage
#

yeh you need a lot of space, especially with controlnets and upscalers and shit

tepid finch
#

how about steps 50 cfg 4?

rugged mirage
#

try it

shell tendon
rugged mirage
#

and change the batch so you generate like 2/4 at once to choose the best one

shell tendon
#

best default is around 5.5-6.5 for most checkpoints

#

below that can give you a nice grainy photographic look, but tends to lose detail, and get washed out, faded

#

higher than that can be nice for illustrative styles and means it'll follow the prompt slightly better

#

but can qucikly start looking blown out - insane contrast, hypersaturated, just bad looking

#

i rarely go outside 5-8

tepid finch
#

ok i dropped the check point in the folder but its not showing up

shell tendon
#

if i do, there's a reason

#

which folder?

#

you have to hit the refresh button usually

rugged mirage
#

click the refresh button next to the models

shell tendon
#

if you're in webui it's a lil blue arrow

rugged mirage
#

and when you load the new model change the resolution to 1024/1024

tepid finch
#

got it

shell tendon
#

k now i'll give you something to test

#

dpmpp_2m_ancestral, karras, 35 steps, 1024x1024 resolution, CFG = 6... negative prompt = "low quality, low detail, bokeh, blurry, text, watermark, signature"... positive prompt = "a freak cannibal clown shark ninja with a pickle sword leaping sideways with a metallic giraffe out of an airplane inside a scuba pool in a rooftop barn in a cyberpunk city"

#

i'm sure it'll be interesting

tepid finch
#

alright here goes nothing

#

about 16 secs

#

i dropped it in chat with img

pearl ocean
shell tendon
#

hugely better than what you saw a min ago! ๐Ÿ™‚

tepid finch
shell tendon
#

are you using forge or automatic 1111

#

cuz if you're using automatic 1111, use forge, it's much faster

pearl ocean
#

Forge

#

No point in 1111 is forge exists

shell tendon
#

what about you kill krill?

tepid finch
#

im using comfy

shell tendon
#

k cool that's good

tepid finch
#

not sure where the giraffe came from lol

shell tendon
#

ever need any workflows, i always leave mine embedded in my images... lots of ppl here do

#

the prompt ๐Ÿ˜„

tepid finch
#

let me rerun it actually time it

pearl ocean
tepid finch
#

lol

shell tendon
#

comfy waits until you run the workflow the first time to load it

tepid finch
#

yah 18 seconds

#

let me put alienware in performance mode

pearl ocean
tepid finch
#

damn same 18 sec

shell tendon
#

that's not too far off from where it usually is with that sampler

#

dpmpp_2m_ancestral is about half as fast as a number of the others

#

dpmpp_2m_sde is a lot faster and also pretty damn good

#

same with dpmpp_3m_sde and dpmpp_3m_sde_gpu

#

but def recommend sticking with the slightly higher stepcounts not 20 cuz otherwise you end up wasting time rerunning shit when it looks weird or crappy

tepid finch
#

i have a m2 mac book at home i wonder if it would be this or not

shell tendon
#

another 3-4 seconds per gen sure isn't gonna kill ya

tepid finch
#

im sure the I9 is better than m2

shell tendon
#

yeah you're def better off with the 4090 mobile

pearl ocean
#

I have noticed 512 images are extremely fast to do, but they are also a more blurry and low quality, which makes sense

tepid finch
#

i have gigapixal ai which is awesome

shell tendon
pearl ocean
shell tendon
#

yeah that's the wrong resolution for sdxl

pearl ocean
#

๐Ÿ˜ฎ

shell tendon
#

you wanna stick with 1024x1024 and other resolutions with similar total pixels

tepid finch
#

thanks again guys, you are awesome!

shell tendon
#

1024 x 1024

1152 x 896

896 x 1152

1216 x 832

832 x 1216

1344 x 768

768 x 1344

1536 x 640

640 x 1536
pearl ocean
#

does any models work well with 2048? I've noticed it tends to just repeat a person, Im told you need to turn tiling off, but I never see a titling option lol

charred mesa
#

probably controlnet tile

pearl ocean
#

๐Ÿ˜ฎ

topaz plover
shell tendon
#

you gotta upscale if you want that

#

natively generating at that resolution is pretty hopeless atm unless you're using cascade

tepid finch
#

sometimes faces come out perfect sometimes not lol

topaz plover
#

crisis ัˆะพ

pearl ocean
topaz plover
#

ะฏ ะฝะธั„ะธะณะฐ ะฝะฐ ะฐะฝะณะปะธะนัะบะพะผ ะฝะต ะฟะพะฝะธะผะฐัŽ)

tepid finch
#

I did taylor swift came out perfect, but kylie jenner and it was botched

pearl ocean
tepid finch
pearl ocean
pearl ocean
#

To think @shell tendon could possibly be the great AI artist in existence

rocky island
#

how do I generate the same thing? promt gives out different options every time, is it possible to regenerate the option you like and add details? change the character and so on

civic sphinx
#

If any teams are looking to locally train/develop their own AI text-to-image or text chat models, but need help securing the right on-premise hardware, we have liquid cooled RTX 4090 24GB, RTX 6000 Ada 48GB, Nvidia L40S 48GB, as well as the just recently released, proprietary Phison AI100 SSDs that are compatible with aiDAPTIV+ middleware for 70 billion+ parameter, large languge models all in stock. DM for more info, happy developing

pearl ocean
willow cave
#

WEED AND CIGARETTES THE TWO BEST THINGS IN LIFE. PERIOD.

karmic cedar
#

weed is nice

woeful quail
pearl ocean
karmic cedar
#

i wanted that upgraded Sony ToF sensor.

civic sphinx
# pearl ocean Iโ€™d recommend AWS Sage tbh

SageMaker's Nvidia V100 Clusters are quite capable for those with solid internet connectivity, however the TCO over the next 3 years doesn't scale well compared to the shear volume of add'l AI users that will be active by 2027. Per SageMaker, "The compute cost for Amazon SageMaker is $2,549 in year 1, $3,059 in year 2, and
$3,569 in year 3, totaling $9,177 over a three-year period."

A lot of us in this group will still be using AI models in 2027, so why pay $9k+ when we can just build our own machines for less? Training takes a bit longer, of course, but a lot of users back in the day with 56k dial-up connections left their machines on overnight to download 480p movies. This same multi-tasking methodology is still feasible today for training large models, especially for those living in disadvantaged cities with poor high-speed internet availability. Starlink isn't really that cheap for AI devs that want to live out in nature, but still work remotely.

slender lily
#

Which channel do we use the dream command in again?

shell tendon
slender lily
shell tendon
#

try typing a prompt at me in there as plain text

fervent thunder
#

ant arctic sour choosings of the chooser family

#

and yes all of that chewy rice

woven charm
#

I am using forge, but I cant seem to get openpose to recognize models. Any ideas what folder models are supposed to go in for forge's inbuilt openpose?

tepid finch
tepid finch
sterile raven
#

SD3 is never coming out :,(((((

tepid finch
#

Y because ceo left?

pearl ocean
woeful quail
dry trellis
pearl ocean
thick kite
#

hello

edgy tundra
#

hi

novel blade
#

hoew to use

karmic cedar
#

Stability is probably sitting on SD3 hoping investors will come around, but our current global economy would serve SD3 up as a poison apple.

oblique quartz
#

Hello is Stable Audio 2 available via API?

karmic cedar
#

Not yet, just the site widget

obsidian bane
#

hello

dapper ocean
#

good

ripe plaza
#

hi

candid juniper
#

Nihao1

karmic cedar
#

spaghet

trim magnet
#

ball (singular)

karmic cedar
#

,

#

^ oxford comma out of place

young compass
#

hello

karmic cedar
#

gesundheit

dull heath
#

hello

neat finch
#

ไฝ ๅฅฝ

marsh haven
#

hello

trim magnet
#

yellow

shell tendon
#

thing

shell jay
#

hi

plush gazelle
#

123

trim magnet
#

tunak

forest trout
#

Kamusta

trail lion
pearl ocean
#

I feel like these SD models tends like nude, I tend to have to put it in the negative prompt sometimes to prevent it lol

fallen spoke
#

hello

jagged mantle
#

hello

native siren
#

hi

whole talon
#

hi

near vapor
#

you see see you

hexed kraken
#

1

meager anvil
#

hi

deep basalt
#

nice

sturdy sierra
#

nihao

trim path
#

้š”ๅผ€ยท

#

Stability is probably sitting on SD3 hoping investors will come around, but our current global economy would serve SD3 up as a poison apple.

pearl ocean
#

I bet people would use SD3 for furries

austere pollen
#

Could anyone help? Using Forge XL webui with dynamic prompts.
My prompts are leaking into eachother i.e. text designated solely for a tshirt also shows on the hat - I have it seperated with { } so not sure if its just something that happens always.

forest trout
#

Pretty sure a few of the stable diffusion breakthroughs came from them.

trim magnet
austere pollen
#

aaah - bummer

#

ty habby

trim magnet
#

try with "cutoff" extension it helps with bleedin

austere pollen
red canopy
#

ไธ‹ๅˆๅฅฝ

crude egret
#

.

edgy tundra
#

1

pearl ocean
#

2

woeful anchor
#

3

fervent thunder
#

Hey everyone, i needed help with something. I want to generate photographic images of people that look as realistic as possible. I used realistic vision v6 for this and it worked fine for lower resolution. But for higher resolution like 2k it fails miserably and takes lot of time to generate on local system and even Google Collab GPU. Any suggestions to speed this up and create realistic looking 2k images ?

crude notch
fervent thunder
#

Reduced steps to 25, still no good

crude notch
#

use fewer for hiresfix

fervent thunder
#

And the image generated even after taking lot of time is distorted

crude notch
#

do you have hiresfix on

fervent thunder
#

I don't think so

#

I like started 2 days ago, reading about stable diffusion and trying to impleme the APIs

raven agate
#

Daz AI finally ๐Ÿ˜…, you know how long I've been trying to get Daz and SD to team up on some level lol.

shell tendon
bleak matrix
#

Good morning, everyone! How are we all today?

charred mesa
#

Very nice

#

I am going to make a sick abandoned building lora, I'm making the images for the dataset rn

#

How about you

wispy cloak
#

why SD3 is delayed?

trim magnet
#

why it wouldnt?

arctic sedge
fervent thunder
trail lion
#

yah there has to be a communicated release date for it to be delayed

gaunt pulsar
#

I can't wait for any update at all about Stable Diffusion 3, maybe just a few new invites so we know it's progressing

reef shard
#

I've been waiting for research access for a month... I've just accepted that the company is tanking and it's not gonna happen anytime soon

hot wraith
#

I have a question, I'm trying to make clothing Loras on SDXL, but whenever I do and apply the Lora, the image quality of the generated images severly gets degraded, like if there was a filter over the image and it doesn't look sharp anymore.
What am I doing wrong in the training for it to look like this?

wet willow
#

Even at lower values @hot wraith ?

hot wraith
# wet willow Even at lower values <@399693786491387935> ?

So at lower values that goes away, definitely, but then the problem is the item of clothing just becomes something general. I want it to reproduce that exact piece of clothing. Like say i'm doing a jacket I made in real life, I want it to learn that jacket exactly ๐Ÿ˜ฆ

trim magnet
#

you overcooked the lora thats why

hot wraith
#

I guess that's it, yeah. How do I avoid that? I followed some youtube guide for clothing, and i did 15 epochs at 20 repeats, had 31 training images.

trail lion
#

20 * 31 * 15 - that's like 9300 steps

#

and with only 31 images each image was repeated 300 times

#

what I typically do is take 500 / (num images) and round down to get a decent starting number for base repeats, save every 1 epoch, and maybe do 8 epochs, test each one until you are satisfied

hot wraith
acoustic forge
#

Hello, I haven't used this bot for months and I see that it changes a lot, how do I create images now?

trail lion
hot wraith
#

One outfit per lora yes, definitely

trail lion
#

just dont describe the outfit would be my approach, describe everything else, it's basically a subject training, but not a character

amber bloom
wet willow
eager tendon
#

we got a number 1 victory royald

trail lion
light marlin
#

hey guys, quick question - If we put a second graphics card will it help to generate quicker? Does itย workย withย sli?

#

currently I have 24GB nvidia 4090

karmic cedar
narrow kernel
#

Also, you describe everything you want to be able to change and the thing you want to "keep" is usually associated with a unique token

karmic cedar
#

sophistication of context will be what divides the public from the private sectors. I think. /talkingoutloud

narrow kernel
#

So it might be something like " a 36 year old woman with brown hair wearing a yfeh7sd, denim jeans, a blue hat in an empty stadium. She's posing with her arms in the air"

#

If the shirt is what you're trying to train into the lora

broken cave
#

also, what did you do with stable cascade? it is extremely powerful

#

they really had just 1 shot at capturing enough mindshare for interesting stuff to happen. it happened with stable diffusion 1.1, for the kind of idiosyncratic reason of generating porn. stable cascade is phenomenally capable. yet the manic individuals building stuff in the stable diffusion ecosystem haven't done anything with it

#

it is too bad really.

#

deepfloyd/if foreshadowed all of this

shell tendon
#

Yeah

#

Part of the problem is the announcement was followed almost immediately by the sd3 one

#

In retrospect they should've waited to announce sd3 for a bit

#

Obviously they weren't as ready as they thought

karmic cedar
#

I think theyโ€™re getting pressure

#

lots and lots

#

from many different places

broken cave
karmic cedar
#

heart of the matter is data ownership rn

broken cave
#

because SD3, when the "full weights" version of the model is hosted

#

and the lightweight model, that is of limited use, is released

#

it will be obsolete in a year. but their dataset will not be

shell tendon
karmic cedar
#

and with upscaling becoming more contextually powerful, the industry has already underestimated how useful lower resolution models have managed to remain

broken cave
#

it is easy to raise money to spend on compute, not because it makes sense (it doesn't), but because investors buy it

karmic cedar
#

at the moment

#

compute itself is in danger of becoming monopolized like every other commodity

broken cave
#

well i think they tried to convince people that compute isn't actually expensive (it's not) and that there are unlimited kinds of charts to be #1 on (there are)

broken cave
karmic cedar
#

compute isnโ€™t expensiveโ€”but neither is ketamine, neither are tomatoes.

broken cave
#

lol

karmic cedar
#

๐Ÿ™‚

#

itโ€™s the overhead

#

thatโ€™s what i worry about

#

the greed

#

one day, we will be mining helium in a sustainable way, and we may yet have zero point energy to actually offer all of this stuff as a basic free service for all humans

#

etc.

broken cave
#

okay well, well capitalized enterprises, they never pay someone else's monopoly pricing for a long time. anthropic is a great example of this: they have a $XX billion convertible investment from amazon, let's say $10b, which costs amazon only $100m to provide, but $10b in "investment from amazon" lets them raise whatever, $100m-750m in hard cash from investors.

karmic cedar
#

^^

#

but the system at the moment isโ€ฆ.so incredibly biased

broken cave
#

because "compute is expensive"

karmic cedar
#

itโ€™s investments for sure

broken cave
#

stability is going out there telling a much more complicated story

karmic cedar
#

ape no like complex

broken cave
#

yeah. and nobody wants another ideogram

#

i like ideogram

karmic cedar
#

but thatโ€™s what the system is instinctively trying to make for itself

broken cave
#

but i'm going to "just" "use the best thing."

karmic cedar
#

by design

#

๐Ÿ™‚ yeah

#

AI is beyond product-nessโ€ฆitโ€™s medial

broken cave
#

i mean there are people who watch this chat who work at stability

#

it's a simple yes or no question: do you want to make another ideogram?

#

they have all the evidence in the world right in front of them that the dataset is more valuable for the role of a private enterprise in an ecosystem where researchers give out brilliant ideas for free

#

but it still sounds so sexy, to be an "R&D" guy, and sexier than going to do research at the universities they are too non-traditional to be admitted to

broken cave
# shell tendon True that

personally i think they should directly monetize the dataset for enterprises. accept the most straightforward billing possible: a sort of DLC where each new part of the dataset is a small additional charge. they should be doing the kind of enterprise work that i am doing lol

#

this thing where they want to stick to a 0th or 1st level expanding brain meme playbook. it's over. they lost their chance for that. and anyway, i want them to thrive, and they will not thrive competing directly with ideogram and midjourney. they will only thrive coexisting with them. i'm sure emad said all the same exact things.

karmic cedar
#

models are amalgamations. we build things, products. AI has a medial nature that contradicts our own intellectual metabolism, reinforced by economic principles

karmic cedar
#

not to play the devilโ€™s advocate

broken cave
#

@shell tendon scale.ai already does this, and has been very successful, for a very long time

broken cave
#

generative AI is the best thing that ever happened to dead authors

#

it just goes to show that the real problem is that people are really fucking stupid

#

i don't mean you lol

karmic cedar
#

etc*

broken cave
#

but that there is a big, popular discourse about many of the issues stability faces, and it gets things 200% wrong

karmic cedar
#

itโ€™s likeโ€ฆa royalty at that point. isnโ€™t it

broken cave
# karmic cedar itโ€™s likeโ€ฆa royalty at that point. isnโ€™t it

okay but if it weren't for the goodness of a few people's hearts, there wouldn't be an Internet Archive. they had very humanistic ideas for what they want to do, and until generative AI, its greatest "real" value to ordinary people was as a paywall bypass for very, very recent news articles. kind of the opposite of what they want to be involved in. presumably if the people who put up random geocities crap thought it was actually valuable, they would store their twee little writing somewhere, but they don't. that suddenly your twee little writing about twee shit has ancillary value for training an AI: nobody could have anticipated that. but it fulfills all of internet archive's humanistic goals, that because they preserved this twee geocities internet culture stuff, it can "provide universal access to all knowledge"via chatgpt, that is used by millions of people every day

#

nobody was reading your geocities crap ever. it is only interesting in some big aggregate with everyone else's geocities crap

karmic cedar
#

Hm, but if large data represents structures of a whole new magnitude, then that data represents building blocks to pyramid-like vectors that enable the already powerful to be more powerful, etc.

broken cave
#

i don't know

karmic cedar
#

โ˜บ๏ธ

#

I get ya tho

broken cave
#

nobody is misusing geocities crap to get more power

karmic cedar
#

Thatโ€™s beside the point

#

But I know what you mean

#

When it comes to AI models, for every hypothesis you find in your output you are eliminating xxxxx amounts of null hypotheses in the form of those geocities rants, etc.

broken cave
#

it has similar energy to saying that mark zuckerberg will personally will himself to be president or whatever the fuck, because he controls all the most powerful communications platforms in the world right now. and yet. you talk to a real political analyst, and the most likeable thing about him, the real reason he could get elected, is because of his lovely family.

karmic cedar
#

The null data is perspective

karmic cedar
broken cave