✨｜sdxl

[zinc cargo]

first, username checks out 🙂

do you orgnize them? bcz i feel that my folder thumbnail file might be like 500mb or something 😛

[native knot]

All of this is why I keep working on automation and scripting to get everything tagged and hosted on my lychee server.

Shameless url drop: https://lychee.soulctcher.net
😄

[hardy cipher]

actually working on it right now. have a program to name them all according to date and then put them in folders according to date

[zinc cargo]

windows?

do you have somethign that can strip their meta data while at it?

[hardy cipher]

yeah

that's some of them

[zinc cargo]

i always get back to it, that i want to get a list of all my prompts ever

[hardy cipher]

putting them all in a pile before I sort them

[native knot]

You can save the images from comfy without the metadata. You can also use nodes that save the metadata into text file.

[hardy cipher]

yeah, I've thought about that. but I never do it

[native knot]

You could setup a workflow that will do it all for you for existing images as well.

[zinc cargo]

it sounds above my abilities

[native knot]

Not at all.

[zinc cargo]

not that it's somethign that ever stopped from trying...

🙂

guessing this the thing, right?

can i append it to a single file though?

[uncut fiber]

to me helped or actualy made it Aimingfail
@zinc cargo what all do you want to have saved?

[zinc cargo]

i want to keep a list of all my positive prompts

aiii, kid needs me. bbl 🙂 thanks Bernix

[uncut fiber]

save image file with prompt seems handy

[uncut fiber]

@zinc cargo i have it this way. All credits to Aimingfail

[indigo vine]

Which model did you use for this one? IPAdapter plus is indeed really good.

[eternal fog]

The input image is someone else's so I'm not sure. The output was done in counterfeitXL

[fading rock]

whats a good model for anime? also, how do i make img2img better?

[uncut fiber]

@fading rock try ask in #🍥｜anime and get better in anything key is practicing 🙂

[eternal fog]

CounterfeitXL is the best one I've seen for now.

[fading rock]

i tried counterfeit but its not the best

i use counterfeit v3

v30

idk

[somber hill]

https://twitter.com/GozukaraFurkan/status/1707782208952406222

Furkan Gözükara (@GozukaraFurkan)

SDXL DreamBooth vs LoRA

It is commonly asked to me that is Stable Diffusion XL (SDXL) DreamBooth better than SDXL LoRA?

Here same prompt comparisons. Not cherry picked. Look prompts and see how well each one following

1st DreamBooth vs 2nd LoRA
3rd DreamBooth vs 3th LoRA

Raw…

Twitter · Sep 29, 2023, 3:39 PM

[crisp owl]

Some luscious locks in that first block there

ah I see it was prompted for, and that followed better from dreambooth

[eternal fog]

We are so close to not even needing Dreambooth or LoRA anymore for somethings with IPAdapter Plus. I took one of your images from that post and put it through IPAdapter, it changes the face a bit too much, espeically when you lower the weight enough to change the background, but it's close.

They are apparently working on a version for faces on SDXL

[crisp owl]

Robot chicken

[ripe holly]

what kind of eggs does it lay

[vital ermine]

Grenades

[eternal fog]

Now you have to generate it laying a grenade

[vital ermine]

Hahahaha

[spring fulcrum]

lol

[uncut fiber]

@zinc cargo does it work for you?

[eternal fog]

The weight was too high so it put the Sofa outside lol

[vital ermine]

Always been hard to get SD to do more than one thing

[eternal fog]

I tried to combine the Grenade and the Chicken

[vital ermine]

yeah, I can't get it

[eternal fog]

Could try conditional prompting

I've not really bothered with that

[vital ermine]

Nor I

Yeah, without CN or something SD isn't going to do it

[uncut fiber]

weird is when bassethound run, his ears are still down along his body

[crisp owl]

so that's what they look like defeathered. I knew birds weren't real

[vital ermine]

The Man of Steel

[crisp owl]

Hulk lost some weight

Or it's just a super dense material 🤔 smart

[vital ermine]

Yeah, some space alloy, probably

[upbeat summit]

cool!

[crisp owl]

you see someone was just taken in as a suspect in Tupac's murder?

[vital ermine]

For real?

[crisp owl]

Yeah literally just got the notification like 20 minutes ago

[vital ermine]

Damn, that is what like 25 years ago now?

[crisp owl]

Yeah something liek that

1996

[vital ermine]

No statute of limitations for murder

damn, 27

Seems so happy to kil, err, meet you

[crisp owl]

He just wants to play a game

[half ivy]

So nice to not see the bokeh 😻

[ripe holly]

lol nice try, i believe it should work with gligen, but havent used it myself before https://comfyanonymous.github.io/ComfyUI_examples/gligen/

[static berry]

was this made with local version or the discord bot

[pure crystal]

https://www.mystic.ai/paulh/open-journey-xl:latest/play

[static berry]

ok

[pure crystal]

I've been trying to squeeze the most out of this, like prompt golf

(Three quarter portrait, vermeer lighting:1.3), black space mage, dark floral throne, (millennium falcon cockpit:1.3), 35mm, ( POV, cinematic afrofuturism:1.3),(adult, ghastly dark, no sky, horror), 1983, (extremely detailed depth), style of Antoine Blanchard, highly detailed
neg: group shot, text, watermark, signature, camera, infographic, art nouveau, blur, low contrast, cowboy, comic books,bad hands, bad anatomy, (bad quality, worst quality,sketch,incomplete,cropped, cut off:1.3)

[static berry]

actually nailed the tacoma

[half ivy]

Smells adverty

[pure crystal]

specific year + models has worked on a lot of cars for me across SDXL models. Very fun

[eternal fog]

Not sure that would work in SDXL. But you can use positional conditioning, but again, not sure how that would work with IPAdapter. I'll need to have a look into it.

[static berry]

really? any time i try to do most recent year/gen it does the previous one

[pure crystal]

def gets model years messed up a bit, but will often have both and they're pretty specific

[ripe holly]

let me know, im curious now ^^

[half cedar]

1991 buick century is the one on the left. You often get the one on the right

[static berry]

first pic looks nice

2nd pic reminds me of that one gta sa remaster video

was by 12th hour i think

[eternal fog]

looking at the examples I don't think it will work, as it's controlling the text conditioning, but IPAdapter changes the actual model. So I think at higher weights it 100% wouldn't work. Maybe it would at lower weights. It looks like a lot of effort though, so I might look at it a bit later. If I do and it's decent I'll post it in here.

[upbeat summit]

greetings, Arron 🙂

[pure crystal]

(vermeer lighting)

[static berry]

i really wish ai was better with the close up details

[eternal fog]

Hello 🙂

[pure crystal]

(close up, POV, intricate details:1.3)

[static berry]

why do i like that

[static berry]

nailed this one

[native knot]

I'm impressed...it knew the 2013 model year above from the 2003 model year:

Not perfect, but the main differences are known.

[pure crystal]

This is a mixed up w140

[native knot]

Dude. It knew the Volkswagen Thing.

That's just great. Impressive.

[zinc cargo]

is this soley txt2img?

[zinc cargo]

what is? Aimfailing's solution? we'll work on something for this purpes

[static berry]

which one are you using

[native knot]

Which what? Model?

[pure crystal]

@native knot

[shy kelp]

is there sdxl for lama cleaner?
can lama cleaner remove rain from the whole image?

[pure crystal]

I think this sounds like an IP Adapter task

[static berry]

ye

[native knot]

Current one I'm using is bluePencilXL

[static berry]

is that one local

?

like do you have to run it on your own machine

[pure crystal]

https://civitai.com/models/119012/bluepencil-xl ^

[static berry]

ive had a shit ton of issues when running local models

ill try this one later

[pure crystal]

what GPU do you have?

[static berry]

4080

why

[pure crystal]

ruling things out re: issues

what's going wrong?

[static berry]

ill have to redownload to show you

so lemme do that

[crisp owl]

Yup, just prompted. No ipa/controlnet/etc.

[native knot]

If you have a 4080 you shouldn't have ANY problem running SD 1.5 or SDXL on your machine...that is, unless you are woefully low on system RAM for some reason.

[zinc cargo]

i love it

[static berry]

64gb

[native knot]

Shouldn't be a problem, then.

[static berry]

ill just have to show you the issue

once it installs

[static berry]

could "no module 'xformers'" be the issue

[native knot]

could be, yeah...are you running comfy or a1111?

[static berry]

a1111

seems to be fine when i use pruned

[native knot]

I'm not too up on the latest for resolving a1111 xformers issues, but it sounds like you don't even have it installed at all. Typically if you just add --xformers to your command in the config, it'll install it for you.

[uncut gull]

These turned out so good! The icons are clearly readable and the colors and brightnesses are perfectly balanced. Very good hitrate. Generated 160, and easily over half look usable. 🙂

[pure crystal]

they look great, art style is perfect.

[uncut gull]

Thanks! Been slowly working on them for over a week now.

[static berry]

this is the result when i use sdxl

[pure crystal]

I've been 👀 ing. The dark fantasy sparkle is super nostalgic

[native knot]

What VAE do you have selected?

[pure crystal]

idk looks good to me

[uncut gull]

You're using the wrong VAE. Either using SDXL VAE on 1.5, or vice-versa.

[static berry]

none
is that where i made a fucky wucky

[native knot]

You need the SDXL VAE

[static berry]

alr where do i select vae

[native knot]

I don't know where they tucked it away on 1.6...if you don't see it on your screen, it's probably in a Settings menu.

I pretty much use Comfy nowadays.

[static berry]

maybe ill try that if i cant get this to work

[native knot]

You should be able to set the VAE and go...it's not tough.

[static berry]

its still not working

wait

hang on

i think i figured it out

caveman discovers fire moment

that is MUCH better

[native knot]

There you go.

[static berry]

i feel like i just discovered the existence of cooked meat lmao

[strange mist]

😱

[static berry]

what

walmart with volkswagen logo 🤣

[native knot]

Bolt nips.

[vital ermine]

works for me, lol

[static berry]

i really wish sdxl was good at current gen vehicles

[pure crystal]

https://civitai.com/models?tag=vehicle

look for vehicle loras. train your own LORAs with kohya

[static berry]

idk what that is lmao

[pure crystal]

haha, little models you can reference in your prompt. can be a person, a style, a thing. usually under 200mb. In general, if you want to prompt something not in a model, this is the path. not hard, a little fiddly and time consuming though https://www.youtube.com/watch?v=d4QJg4YPm1c

[static berry]

alright

ill probably try that later or tomorrow

whenever i have time

[eternal fog]

I had a look into it. It was a pain, you can sort of get it to make an image. But it doesn't really work properly because it needs text prompts in the final image anyway. I tried to combine unclip conditioning from them all and it just created a mess.

Maybe if you combined the 2 IPAdapter modified models together and ran that through it might work, but I don't have the VRAM for that.

[pure crystal]

is this a gen or a screenshot from a simpsons episode?

[opal blade]

gen

[pure crystal]

wow I literally thought it was a meme

it's so clean

[opal blade]

lol

yeah

[static berry]

the other foot is kinda wonky

[marble kindle]

who wants me to run a dalle 3 prompt for them?

[cyan crown]

I would like to try by myself. How?

[marble kindle]

not availible yet

i have early access

[pure crystal]

jealous. I have thousands of old dalle gens.

[cyan crown]

how to gain early access?

[marble kindle]

its random

[marble kindle]

just a random image i made

its very good

[pure crystal]

this would sell as a all over print garment 100%.

[pure crystal]

(extremely detailed depth:1.3), intricate detail, natural land forms,Futuristic landscape with some spacecraft flying over the planets, in the style of mind - bending murals, floating structures, futuristic landscapes, (style of El Greco, style of Leonardo di Vinci:1.3)

[marble kindle]

i can run a prompt if u have anything in mind

[stone fossil]

Very glitchy I like it.

[eternal fog]

I put this into IPAdapter and tried to get it to put a wing on it, and it just turned it into a 3 Series

[static berry]

whats ipadapter

im curious

[eternal fog]

https://github.com/tencent-ailab/IP-Adapter

[eternal fog]

https://github.com/cubiq/ComfyUI_IPAdapter_plus

And for comfyUI nodes

[marble kindle]

prompting for dalle 3 is more natural language

[eternal fog]

Basically you give it an image or images and it can create new images out of it that look similar, or you can add text prompts to change the images.

[eternal fog]

For example I took your car and used the prompt "Desert Dunes"

Unfortunately it's turned it into a Merc

[eternal fog]

Alternatively I can use the same sort of background and change the car

[pure crystal]

ick, not my favorite. "Alien landscape vista with flying space vehicles leaving chemtrails. Ringed planets fill the red sunset horizon. Lighting of Vermeer, a mind bending oil painting mural by El Greco and Da Vinci"

[woeful patio]

did you all see the post about dall-e 3, pretty incredible stuff:
https://www.reddit.com/r/StableDiffusion/comments/16uynpp/dalle_3_blue_ball_on_a_red_cube_on_a_wooden_table/

[eternal fog]

We can just steal their images to use in SD

[marble kindle]

lmao

[pure crystal]

cool

[marble kindle]

dalle 3 is insane

[pure crystal]

Christmas card illustration of a mouse relaxing in a mug of hot chocolate with marshmallows. Beautiful, nostalgic

[pure crystal]

lol so cute

3 is my fave, all perfect

[marble kindle]

and text

nice prompt 🙂

[woeful patio]

this dall-e 3?

[marble kindle]

yes

[woeful patio]

damn thats crazy

[marble kindle]

i can run some prompts for u

[static berry]

try 2020 bmw x5

on dalle 3

[pure crystal]

adorable aware fictional pygmy marmoset cuttlefish chimera, beautiful studio photograph. calm try a hard one lol

[woeful patio]

try this one i made in sd

Horror-themed , on the moon, dark side of the moon, space suit, dying , dark, creepy, lying down, dead, . Eerie, unsettling, dark, spooky, suspenseful, grim, highly detailed

[static berry]

nice

other than the license plate it looks great

and it nailed the rims

[pure crystal]

omg lol

[static berry]

also how did you get access

[woeful patio]

holy shit I'm done, that's amazing. The other image took me an hour to get right lol

[static berry]

to dalle 3

[marble kindle]

is random on bing image generator for now

some people have temp access

[static berry]

so you just randomly get it

[marble kindle]

yes

[eternal fog]

Dpes Bing tell you if it's used Dalle 3?

[static berry]

welp time to use bing till i get it

[marble kindle]

no but try use complex text in dalle if it works then its dalle 3

[static berry]

can you give an example

idk if this is dalle 3 but it looks very good

[slender coral]

Suggestions on removing these splits?

[pure crystal]

Sega Hologram Time Traveler

[upbeat summit]

yes, exactly 🙂 nice catch

[static berry]

i meant a prompt example

[marble kindle]

that is quite good yeah

[pure crystal]

major nostalgia

[marble kindle]

A woman walking in a street with a sign that says "DALLE 3 can do text", in cartoon style

[eternal fog]

Is this complex enough lol

[woeful patio]

this is insane stable diffusion needs to catch up

[static berry]

it will

[woeful patio]

at least we still have anime tiddies

[marble kindle]

then u have access

[static berry]

W

[marble kindle]

lol

[pure crystal]

how do I even get to bing he wondered

[marble kindle]

i only have access on chrome not edge lmao

[static berry]

im using edge

lol

LMAO I LOVE WHAT IT DID

[marble kindle]

nice

[static berry]

it slightly messes up the text

but other then that

its very VERY good

[marble kindle]

yeah its a massive leap

[static berry]

it might make me use dalle more

[eternal fog]

Problem with Dalle is they are never going to open source it

So you'll either have to pay or have limited generations

[marble kindle]

its free on bing unlimeted

[static berry]

yeah

but its slower

when u run out

[eternal fog]

Is it actually unlimited, last time they said that it still stopped you using it after a while.

[marble kindle]

yeah i think its just a bit slower

[static berry]

even dalle struggles with current gen vehicles

i wonder why 🤔

[static berry]

LMAOOO

THAT ACTUALLY GOT ME

[pure crystal]

no pictures of the new model in the training data

[static berry]

they need to add them fr

[pure crystal]

wobailondy means family, and family means sharing prompts

[eternal fog]

Is that just from putting in "xbox 360 game cover" or something?

[marble kindle]

yep

[static berry]

dude wtf dalle actually nailed the walmart logo

[marble kindle]

with dalle 3 u can be very spesific and still get all the details

[woeful patio]

the amount of detail you can make in a prompt with this is insane

[static berry]

i hope dalle 3 will get img2img

[woeful patio]

context is understood so well

[static berry]

just imagine what dalle 4 will be like

[woeful patio]

makes sd prompting look like oonga boonga language

[crisp owl]

What's with all the promoting talk?

[marble kindle]

cursed

[static berry]

NO

NO

DELET

PLEASER

I WISH TO NEVER SEE THAT

AGAIN

[static berry]

im getting snowrunner vibes from this

[eternal fog]

I think sometimes it goes back to Dalle 2

Because that is scuffed

[marble kindle]

yeah

[static berry]

it does everything decent except the actual person

[marble kindle]

what was the prompt

[woeful patio]

crazy stuff that controller is perfect

even the letters are in there

[static berry]

thank you dalle team 🙏

this is truly insane

[static berry]

same prompt on sdxl

[woeful patio]

lol

[floral island]

the image generation wars are real xD

[marble kindle]

we thought sdxl was good 😦

[woeful patio]

its like going from a flip phone to an iphone

[floral island]

for real, dalle-3 IS extremelyh good

i really hope anything "self-run" will be available soon

[static berry]

going from a note 3 to a samsung s23

[marble kindle]

yeah i dont think stability ai will be able to compete

cuz dalle 3 is built on chatgpt

[woeful patio]

I dont think so, I think you need a huge language model to pull something like this off

[floral island]

cries in nokia 2210 age

[woeful patio]

that just doesnt excist yet

[floral island]

for sure -> it's multiple models (llm/diffusion) models

[marble kindle]

they also just a bigger company

[floral island]

i'm pretty sure openai used a lot of model tech to make sure each part goes to the right agent (ai model)

[woeful patio]

art is dead?

[marble kindle]

dalle 3 is the first model that is actually good in hindsight

[cyan crown]

well Can you try some prompt for me?

[marble kindle]

yes

[static berry]

stole prompt from #🏅｜pantheon

[woeful patio]

is this the biggest leap in image generation yet?

[static berry]

might be

[cyan crown]

by Caravaggio, impactful color paint of Xenomorph alien with fruits in Venice, highly detailed, vibrant colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

This is SDXL

[woeful patio]

how about realism tho?

[cyan crown]

well I prefer al lot SDXL result

[marble kindle]

dalle 3 follows prompt better

[noble shoal]

Wait until it get neutered to death 🙂

[woeful patio]

yup. like chatgpt

[pure crystal]

I saw you post this earlier, it's so good

[woeful patio]

used to be good

[cyan crown]

thanks

[marble kindle]

its didint get the Caravaggio part

[noble shoal]

Is that blood on the aliens face? Open AI: Yes! Banned.

[cyan crown]

another try

cinematic photo Charlize Theron in sexyiest red Star Trek TOS uniform sitting in Enterprise , in the style of 1950 vintage sci-fi movie, Star Trek TOS, lomo color 100 35mm photograph, film, bokeh, professional, 4k, highly detailed

[woeful patio]

hmmmm

[cyan crown]

again this is SDXL

[eternal fog]

It seems pretty bad at people

[eternal fog]

The Dalle 3 ones

[eternal fog]

Does she like Cocaine a bit too much

[woeful patio]

sdxl makes great images, it's just that when the prompt gets complex it fails where dalle 3 succeeds

[cyan crown]

because of chatgpt

[woeful patio]

yup sadly

[cyan crown]

it also have big servers

[woeful patio]

the fuck, I asked for varations and it gave me this lol

[marble kindle]

again i think sdxl might have a better image but dalle follows the prompt better and more accurate

[crisp owl]

what's more likely, dalle adds to make better, or puts in more restrictions (like ChatGPT) and gets worse

[marble kindle]

i dont think their gonna change it that much more

[static berry]

2020

[marble kindle]

but if its free and very good can you really complain

[crisp owl]

I've got SDXL locally, I'm good

[woeful patio]

crazy stuff

[marble kindle]

ill use both

[pure crystal]

(close up, POV:1.3), cinematic oil painting by Caravaggio, impactful color paint of Xenomorph alien with fruits in Venice, highly detailed, vibrant colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

crystal clear is a good helper, ty

[cyan crown]

another : cinematic photo alien car, 3/4 view, in a cyberpunk downtown , futuristic 35mm photograph, film, bokeh, professional, 4k, highly detailed

SDXL

[cyan crown]

this is good

[marble kindle]

my favorite game

[static berry]

my 2nd favorite

[static berry]

new lego set dropped 🔥

[marble kindle]

this one is insane

[static berry]

how

[marble kindle]

a robot painting itself painting itself painting itself painting itself

[woeful patio]

perfect text

nice

[static berry]

the fact that all of this is now possible is crazy

[crisp owl]

I did another variation

[static berry]

i did the same prompt but with a hand

[marble kindle]

wow

[woeful patio]

pretty close but not perfect

[marble kindle]

lmaoooo

OK THIS IS GOOD

[pure crystal]

haha

[woeful patio]

lol

[glad grove]

but dalle3 cant make porn ( oh wait sdxl cant either )

[marble kindle]

there is plenty of porn already

[pure crystal]

it's rated teen for the text options when they hand you the bill

[cyan crown]

is this a prompt?

[marble kindle]

Squid ward from sponge bob digging a hole with a shovel, a tombstone with the word Midjourney is next to it

[marble kindle]

i love these game ones lol

[static berry]

fr

[woeful patio]

damn

[marble kindle]

didnt know he was chill like that

[static berry]

ultra rare gta v copy for the gameboy color!!!!!@!!@!@!@ (only one was made fr)

[eternal fog]

😢

[woeful patio]

the text is so much better then sdxl, and that was the main selling point of sdxl lol

[glad grove]

damn he solved the worlds #1 problem

https://tenor.com/view/calculation-math-hangover-allen-zach-galifianakis-gif-6219070

[pure crystal]

in all honesty it's pretty damn easy to add text to a photo in photoshop. it always blends well

[marble kindle]

i just dont get why people make porn with sdxl lol, like i just rather watch real porn

[static berry]

catching fortnite dubs in the 2000s 🥶

[marble kindle]

coldddd

[glad grove]

maybe they like to generate their squidward lewds that u cant get anywhere

[pure crystal]

This looks like a 90s playstation magazine ad

[glad grove]

[half cedar]

I was trying to make fake sega peripherals

[static berry]

guys gta 5 launched on android!?!??#@$

[crisp owl]

Oh look an SDXL picture in the SDXL chat channel

[marble kindle]

lies

[marble kindle]

jk its very good

[pure crystal]

the sequel had better mini games

[marble kindle]

nice comfy workflow

clean

[woeful patio]

https://tenor.com/view/we-dont-do-that-here-black-panther-tchalla-bruce-gif-16558003

[crisp owl]

peripherals out of shot
here's the whole thing

[eternal fog]

Is mine better or worse lol

[crisp owl]

https://tenor.com/view/ron-swanson-parks-and-rec-its-so-beautiful-gif-15644547

[eternal fog]

it took me a very long time to split it up into sections

Before it was just everywhere

[marble kindle]

the best workflow is the one you made

[crisp owl]

Yeah the organizing part takes a shit ton of time

[marble kindle]

lowkey fun tho

[glad grove]

this one is better and smaller

[eternal fog]

What are you even doing with all that

[marble kindle]

what in the mother of god

bro made an image with the nodes

[pure crystal]

whats going on down in australia on the bottom right

[marble kindle]

is this the world map

[static berry]

part of it

[marble kindle]

half the fun of sdxl is playing with settings

[woeful patio]

lol

[marble kindle]

how many of you use Foocus?\

[woeful patio]

turtle duck

[static berry]

why is the carpet texture in 8k lmao

[glad grove]

ok but how do i get that car out of the room

[static berry]

i did not think about that

uhhhh

just rip it through the roof

[eternal fog]

Window

[static berry]

ill buy you a new one fr

[static berry]

what if its a cliff house

[eternal fog]

big crane

[static berry]

what if its a flying house

[eternal fog]

People have done this before, they normally take the side of the house off and pull it out

[native knot]

It's Digiorno.

[glad grove]

🍕

[marble kindle]

these are really good

[static berry]

so good they are making me hungry

[cyan crown]

Harry Potter - the dark side

[sage sand]

Hi, how use xl model?

[slender coral]

Ya'll going to be pissed at me asking this but is that the paid version of GPT?

[eternal fog]

It's Dalle 3

[crisp owl]

nosejob voldemort

[slender coral]

For sure, but how are you guys accessing it?

[static berry]

bing create\

[eternal fog]

It only uses it sometimes it seems

Because it's not using it for me

[static berry]

it is for me

use a prompt like "a man holding a sign that says "dall-e is amazing"

and if you can read the text mostly

its dalle 3

[slender coral]

Thanks lads, do boosts reset daily?

[static berry]

weekly

but you can still use it without them

itll just be slower

[upbeat summit]

love the 80s cinematic look

[noble shoal]

Clearly very close relatives. I guess triplets.

[vital ermine]

Yep

I am just looking at how well all these hands are

[noble shoal]

For this distance, definitely very good

[static berry]

EW WTF

[native knot]

SDXL wants to play some complicated af Monopoly.

[static berry]

thats normal

[trim vigil]

brad pitt x trump 😄

[glad grove]

brad trimp

[dim steeple]

Does anyone know what SD is gonna release after XL?

Have they said anything publically

[glad grove]

XXXL

[static berry]

do settings save

like width and stuff

[half cedar]

Embedded in the image EXIF

Or in a1111?

[static berry]

wym

[half cedar]

Which settings?

[static berry]

these

[half cedar]

You can upload an image that was generated and send it to the txt2img page. In png info tab

[static berry]

so they dont save

?

[shy kelp]

SD N

[half cedar]

Yeah it's blank. Was weird to me too at first

But saving useful setting pngs in a folder works fine

[static berry]

i just took a picture of them

lol

[half cedar]

Haha

[glad grove]

hello ladies

[half cedar]

Cornpus

[hoary saddle]

anyone know what the SD Doodle drawing box was made with? language wise...

assume javascript maybe?

[half cedar]

IPA workflow? Bing?

[opal blade]

bing

[slender coral]

Frontend is almost always JS,

[nimble heart]

be the change you want to see in the world

write your frontend entirely in Rust

[hoary saddle]

getting there, not sure how to display the image when complete but api is good (gotta tweak control net obviously)

replit with ghostwriter is pretty dope

[half cedar]

Are you making calls to a local a1111/comfy?

[hoary saddle]

local comfy

[restive elm]

Is there equivalent of a1111 where SDXL models can be used?

[hardy cipher]

you mean a1111?

[uncut fiber]

@restive elm probably you mean comfyUI. Or you can use sdxl in A1111 itself

[hoary saddle]

hey @visual glade, dont suppose you have a free sec to look at something do ya?

[spring fulcrum]

https://tenor.com/view/tyson-mike-punch-gif-18463668

Dang @soft zealot you came out with the boxing gloves on with that heavyweight workflow release on Civit... Nice job... RTFM lmao Boom!

[supple knot]

https://media.giphy.com/media/C76XNCsfMpGJq/giphy-downsized-large.gif

[ionic gulch]

ip adapter described in one picture (not by me):

[slender coral]

Lads, I want dallie3 how do?

Will it ever be open source?

[supple knot]

its here to test, you gotta sign in with outlook microsoft https://www.bing.com/create

[slender coral]

For sure got it running, do you know what they are doing with it?

[supple knot]

heres a video i found today

https://www.youtube.com/watch?v=C4RU5KPM9to&t

looks like its good at text and real characters like simpsons

[stone fossil]

StarCraft: https://civitai.com/models/154238?modelVersionId=172817

StarCraft - Transforming Images into Celestial Constellations

Introducing StarCraft, a revolutionary LoRa (Text to Image) model designed to reimagine ordinary images as celestial constellations adorned with connecting stars. With StarCraft, you can effortlessly usher your visuals into the cosmos, creating a mesmerizing celestial experience that ignites the imagination and curiosity.

Use trigger word: c0nst3llation.
Trained on 3000 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @masslevel
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)

[woeful patio]

what he gonna do?

[woeful patio]

you wot m8

[ionic dragon]

is dalle-3 only accessible for gpt4 users?

[west breach]

you can use bing create to use it

[kindred plinth]

So dalle-3 actually seems to be WORSE than SDXL when it comes to fine details. Anybody got a good comfyui workflow that might be able to take dalle3 images and fix the fine details

[cursive hearth]

Hello, can I find out when the new SDXL will be released? :). Let me guess by about 2024 or summer 2023?

[zinc cargo]

what's the new sdxl?

[ionic dragon]

bing now uses dalle3?

[ripe holly]

interesting, but you said it doesnt work in SDXL so would it work in SD1.5? or is it just because of the combination you used with the IPAdapter, wouldnt it be possible then if you load the model again without the adapter influencing it and/or having its own CLIP text prompt like in the example? i dont know why it wouldnt work looking at the example, but then again i still havent tried myself ^^

[molten breach]

Yeah, I"ve been testing it and its pretty amazing.

Been really impressed with its text and hands vs anything else I've tried

[kindred plinth]

So applying SDXL refiner to Dalle-3 images seems to work quite well

[ionic dragon]

its kinda confusing

[kindred plinth]

Before and after:

[molten breach]

This is where I"ve been using it. I think you have to use Edge browser
https://www.bing.com/create

you get 100 free generations per day I think

[kindred plinth]

100 boosts, you can still generate when they run out, its just slower

[molten breach]

oh ok thanks

[soft zealot]

Appreciated. Most of what’s in heavyweight is now in cruiserw eight in v4.2 which is tidier/cleaner to look at

Cruiserweight being my personal daily driver

[ionic dragon]

dalle text seems great

[spring fulcrum]

my money is on elon

he would probably math him to death

[cursive hearth]

Well the next generation of SDXL after version 1.0

[spring fulcrum]

2.0

lol

[cursive hearth]

Really?

[spring fulcrum]

idk

but it sounds logical

SDXXL

[cursive hearth]

Are there any rumors about this news?

[spring fulcrum]

not that i'm aware of

I'm sure they are always working on something though... these things take a lot of time to really flesh out into something amazing like sdxl

the devs are amazing

[cursive hearth]

It's true

[spring fulcrum]

but then you have the folks that are working really hard in the after products of developement stage... such as comfy... then further down the line you have people that make amazing things with that framework like searge, and winston woof and sytan and tdg8 that understand how to really streamline it to be even better.

its a long and winding road

then you have people that use what they put together and make amazing images, gifs, videos, etc... it just keeps on going

[cursive hearth]

How can I contact SD developers personally?

[spring fulcrum]

a lot of them float in and out of this channel. stick around long enough and you get to know who the players are around here.

[cursive hearth]

Isn’t it an idea to contact the SD developer personally, like Emad?

[spring fulcrum]

Even Emad chats in this channel once in a while

[spring fulcrum]

Just take a look over to the right in the discord chat you can see who the devs are that are in this channel... You could @ one of the devs and start a conversation with one of them if you really want to talk to a dev

[dapper dragon]

Thanks thaevilone for making this model and share it with us 👍

[ionic dragon]

no, you shouldnt do that

[glad grove]

you should bother them with dumb questions to get a free ban

[spring fulcrum]

most of them are pretty nice. but I agree. don't ask them a bunch of stupid questions... however if you are also a developer looking to contribute in some way you could try to talk to them to see what they develop and see if they would be willing to collaborate with you or maybe answer some of the questions about how to make something you are developing work with something they have already developed.

[glad grove]

https://tenor.com/view/developers-gif-13292051

[spring fulcrum]

I need a counter to see how many times I said the word develop... lol

[ionic dragon]

SAI doesnt just creat diffusion model, they have other projects so I dont think they are working on SDXL 2.0 atm

[spring fulcrum]

that would be why I said... something

[zinc cargo]

which model is that? can we have the prompt for this?

[stone fossil]

https://civitai.com/models/154238/starcraft-sd-xl-10

[dapper dragon]

StarCraft SD XL 1.0 lora and you can get all the prompts there

[zinc cargo]

thanks

[upbeat summit]

also made with StarCraft SD XL

[ionic dragon]

@stone fossil train another lora, with the best images, so you can get the best images

[glad grove]

ty for the multi channel spam

[wicked frigate]

multiple different teams working on multiple different projects at any given time

[ionic dragon]

so, SDXL 2.0 is under development?

[wicked frigate]

there are definitely SD-related things being built, can't say what exactly tho, you'll find out specifics when they're announced

also language model things being built and audio things being built and other stuff too~

[ionic dragon]

can we expect a gui where can use all SAI products in one app?

any plans to add RLHF into comfy(different release build)?

[dapper dragon]

Illuminatis told me no SD XL 2.0 to expect for few months 🙂

[molten breach]

I wonder if SDXL 2 will come out before SD 3.0?

or they may be same thing...I dunno

[molten breach]

Hopefully video 🤞

[native knot]

Don't mess with a space train.

Space train is best train.

[stone fossil]

Ok then. 🙂

[native knot]

How are you doing today, @stone fossil?

[stone fossil]

Fine u lol and u? 🙂

[native knot]

Kinda terrible.

[stone fossil]

Oh thats a fair awnser I like that.

[native knot]

I'll get through it. Art helps.

[stone fossil]

I do not call it art, I call it image output.

But that might depress you even more forget that hehe.

[native knot]

It's all art in my eyes, regardless.

[native knot]

Nice warhammer there!

[native knot]

Now do a warhammer space train? 😄

[half cedar]

Lol

[stone fossil]

Like u siad.

No fucking with the space train.

[half cedar]

Do a locomotive train and try to make the smoke starry

[half cedar]

How showering with a bad sun burn feels (bad)

[stone fossil]

This is what is whats to be not sure or its good not into war hammer lol.

[native knot]

Interesting mix. Not bad, tho.

[half cedar]

"Too much soup, I made"

[floral island]

heard you felt bad, i hope my spiced up car makes you feel better

[zinc cargo]

@upbeat summit as usual, your prompts are always fun to play with 🙂

[native knot]

That's one spicy wall of peppers.

[floral island]

spicevalanche!

it's regional prompter + mmy songmaster, and it makes the most whacky stuff 😮

[zinc cargo]

regional prompter for sdxl?

also, where's my commuinity regional meme?

[floral island]

yeah 🙂

[native knot]

lol that one on the right

[floral island]

can't an ogre enjoy his drink in peace?!

[eternal fog]

I don't think Gligen would work as it uses sd1.5 model for the text which wouldn't work with the clip_g text that SDXL uses. The positioning you can do with conditioning will work fine.

[half cedar]

Oops

"Dinosaur hunter"

[native knot]

Well, it's not wrong.

[ionic dragon]

txt2vid, need a huge dataset, and SD only uses public datasets, so it would be difficult to build one atm

[ionic gulch]

Version 20 of my workflow is online - better cropping options (center, top, bottom, left, right + offset) for IP-Adapter and Revision, new layout (with previews for cropping), better documentation
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[icy brook]

I can see an image I made in there 😶‍🌫️

🙂

[ionic gulch]

which one? seems i liked it :)

[icy brook]

look at my comment again and make a guess 😄

[ionic gulch]

the cloud couple

[icy brook]

yup

I made Aether Cloud

[ionic gulch]

it's nice as input image for ip adapter / revision to get a cloudy look

for example the hair in this one is done by mixing the cloud input:

[icy brook]

that's really cool

where can I watch more on how to use ipadapter and what version of it is it?

very helpful I bet for making datasets

I made the whole of my Cloud LoRA from ControlNet images

[vital ermine]

ipadapter isn't that Controlnet?

[icy brook]

that I made myself that is

[ionic gulch]

you could download my workflow or nodes. or search for ip adpater in comfyui manager to get the nodes needed for ip adapter

[icy brook]

I am still on Auto11

🙂

I haven't had time to learn comfy

I finetune all the time I have

besides family and such

[ionic gulch]

it's controlnet depth for the pose, but the first two ip adapter images for the style

[vital ermine]

I just haven't played around with ipadapter as I was digging realvision

[ionic gulch]

i think there is an ip adapter extension for a1111/sd.next too

[vital ermine]

there is

[ionic gulch]

or if you just want some technical information you can look here: https://github.com/tencent-ailab/IP-Adapter

[icy brook]

thanks!

[cursive hearth]

HOW?!?!

[icy brook]

New LoRA I’m uploading soon

Aether Fire.

[ionic gulch]

example:

[eternal fog]

Anyone not using IPAdapter is missing out, especially the new Plus model

@icy brook Put one of your fire images in and just used London Bus as the prompt

[ionic gulch]

or with different weights:

[icy brook]

But can it make the texture take over the subject/object altogether? Bc what I see in these images is mostly how it becomes a feature.

[eternal fog]

Depends on the model,weights and input images really

[zinc cargo]

i didnt get that lucky with the clouds

[eternal fog]

I think part of the reason JPS's works so well is because the position of the clouds matches where the character is

[icy brook]

Try this instead

[eternal fog]

I'm interested how JPS is doing the mixing, because usually it tends to do the clouds as a background like this

[icy brook]

maybe it tries to find commonalities in the images based on composition?

[eternal fog]

There's a few different ways you can mix the images together

[icy brook]

ok

[eternal fog]

I'm feeding multiple into 1 IPAdapter, which can reduce the control somewhat, but with IPAdapter Plus and 2 IPAdapters it uses over 10GB of VRAM, so it's super slow

[zinc cargo]

is that why with your great workflow, we can't control each picture's str?

[eternal fog]

You can control the conditioning of it via UnClip, but you can't control each images individual IPAdapter Weight

You can chain IPAdapters together with a single image in each, but each one you add uses more and more VRAM

[ionic gulch]

i use the chain mode of my workflow (instead of loading all in one batch), so i can adjust the weights for each input image. and after that it's some fine tuning of the weights.

[eternal fog]

Yeah I can't do that with IPAdapter Plus Model as I run out of VRAM after more than 1

Need to get a 4090 or something lol

[ionic gulch]

yes, i have 24gb vram - even that can be too little if you want to upscale and use more than 4 ip adapter images

a used 3090 should be enough if you don't also want to use the card for gaming

[zinc cargo]

4 IPa!!!

[south horizon]

Gonna flip over to comfy for a little bit to try some things, though I generally use A1111. Is there a new place folks are sharing workflows?

oh, looks like JPS posted one up there...

[zinc cargo]

pepole here share and you can ask for stuff for specifics, people here are very helpful

[south horizon]

Thanks @zinc cargo 🙂

I wanted to switch over for a minute to see if I could get longer Animatediff outputs. Apparently comfy supports it quite well.

[icy brook]

but can this be done in auto11? I really can't move to comfy just yet

[eternal fog]

I think it has a basic IPAdapter mode, but you can't do chaining multiple I don't think

[south horizon]

Ooof, can't use JPS's... got like 50 red boxes. 😄

[ionic gulch]

@south horizon links for the required nodes are in the description :)

you can load them through comfy manager

[south horizon]

I think your workflow might actually do far more than I need 🙂

... famous last words I guess

[ionic gulch]

you won't find many workflows that don't use custom nodes, as many features like ip adapter require them. only most basic stuff is in the default nodes of comfy.

[eternal fog]

lmao, I tried blending the images together first and got this

Not what i had in mind

[south horizon]

What is "IP Adapter"? 😐

[icy brook]

what did the input images look like?

[zinc cargo]

yea, but that's not in SDXL, it's all 1.5 😦 at least afaik

[eternal fog]

Was just a simple blend so

Makes sense why it did it

[south horizon]

Yeah I think you're right, don't think it supports sdxl yet from what I'm seeing.

[ionic gulch]

use image prompts in addition or instead of text prompts

[south horizon]

interesting

[eternal fog]

Not sure why it's made this so blurry, but the dress is cool

These are one of Joachim's fire images, but with a prompt of "Woman wearing a long dress"

[native knot]

I feel like I need a meme template for this:
CFG < 3 = Jesus Take the Wheel!

[native knot]

So...which one should I upscale? 😄

[ionic gulch]

here is one that works more like a texture - text promt was "nuclear explosion in the distance" - so the clouds were not in the text prompt - guess with a little more fine tuning that result could be further improved

[fathom pasture]

Hey guys, I've trained a lora SDXL model on my dog and I cannot get quality pictures. I have tried many different combinations of steps and text encoder epochs and nothing works.. I must be mising something..

Any suggestions what I'm doing wrong?

[native knot]

What VAE do you have selected?

[icy brook]

would you mind sharing your Comfy Workflow for IPadapter over dm?

[fathom pasture]

I have not done anything with VAE

[native knot]

What is selected, though?

[fathom pasture]

I'm using comfyUI if I add the node load VAE there is no options. If ComfyUI have some default I'm using that. Cant figure out what though 🤷‍♂️

[native knot]

You have to have the SDXL VAE.

Unless the model has it baked in. Which it doesn't seem to.

[fathom pasture]

thank you, where do I find it?

[native knot]

https://huggingface.co/stabilityai/sdxl-vae

[fathom pasture]

Thanks, how would I use it ? I'm running on paperspace

I ahve not found any tutorial showing me

[native knot]

Needs to go into the VAE folder, then load it and decode the image with it.

[vital ermine]

In ComfyUI how do we make a new default for the load default? I can't seem to find it.

[fathom pasture]

Thank you very much brother

[eternal fog]

I'll post a simple one in here once I'm back at my PC. The main one I use has loads of other extra stuff mixed in.

[fathom pasture]

Where in file structure should I put the VAE folder?

[native knot]

There should already be a VAE folder under the models folder.

[fathom pasture]

Hmm, there was not

Oh, found it!

[vital ermine]

Thank you as I am trying to follow vids and yeah, not happening.

[zinc cargo]

make into a gif

[woeful patio]

has stable diffusion said anything yet on dall-e 3 and it's features?

[halcyon tusk]

this could be either the wrong VAE or a wrong combination of sampler and scheduler, DPMs with normal tend to do stuff like that, you have to do karras or another scheduler.

[woeful patio]

you still need help, shoot me a message

[south horizon]

hmm.. trying to figure out, is there a way to preview the first frame of an animatediff workflow? 😐

[eternal fog]

@vital ermine @icy brook Here's a simple, albeit messy workflow. It's got weighting on both Positive and Negative Text Prompts, so you can turn them off or on or anything in between.

It currently has 2 image slots that both have unCLIPConditioning that can help weighting, these are both fed into the IPAdapter. It also has a crop tool to help position the images. IPAdapter takes a small square of the centre of the image, so if it's not a square image and you don't crop it, you might not get the expected effect.

There's a few custom nodes you can get via the ComfyUI manager. And you'll need the Clip Vision H model for IPAdapter and the Clip Vision G model for the unCLIP.

[icy brook]

great. thanks!

[eternal fog]

And the IPAdapter models as well of course, if you aren't sure about anything just ask.

[woeful patio]

is lineart ever gonna work for sdxl a1111?

[vital ermine]

Thank you

[vital ermine]

ImpactMakeImageBatch is missing but it can't find.

[eternal fog]

It's part of the Impact Pack, but if you are only planning on using 2 images, you can use the built in "Image Batch" node instead

[vital ermine]

Weird as I have that pack

[eternal fog]

try deleting it and then adding it back, it might be acting weird as the node changes depending on how many images you have attached

[vital ermine]

Yeah, it is

says it is missing but look

[indigo carbon]

tf

idk what did I expect

[eternal fog]

just delete the node and manually add it back in

[vital ermine]

loaded now

[strange mist]

first one is sdxl, second one is Dall-e

[zinc cargo]

pixar 3d cartoon in the prompt also

[strange mist]

here's a better comparison. first sdxl then dall-e

I dunno, I think dall-e is better 🤔
for this specific art style at least

[woeful patio]

in realism sdxl rules tho

[ionic gulch]

i'm sure you can get the right one (or something close) from sdxl too

[indigo carbon]

I have a good theory why. It's pixel art, isn't it?

[strange mist]

yep 🙂

[indigo carbon]

Dalle is a pixel diffusion model, so all the pixels in the pixel art will be the same size

SDXL is latent diffusion, so it's expected that the pixel diffusion model will be better at diffusing pixels

[strange mist]

ah, I see 😅

[vale eagle]

Where could I know Dalle is a pixel diffusion model?