✨｜sdxl

[nimble heart]

you could 100% make a python script that i2i's a masked out face and emails you the result every hour.

[sharp robin]

Caith and Dio img so similar i forget whos who. They same person.

[paper phoenix]

yep. i have win 10 and got annoyed that they moved the different wallpapers on virtual desktops feature to win 11 so hobbled a silent python script that does it for me in the background 😄

[boreal bough]

look. using my typical test seed of "2", to prove first attempt
a black cat sitting inside a washing machine saying "if it fits - it sits"

it even did the hands right

no need for the oldest trick

[paper phoenix]

😄

please dont put cats in washing machines. even ai cats 😄

[boreal bough]

XD

[upbeat summit]

1920x1024 native

[boreal bough]

noice c:

[paper phoenix]

this is really nice!

you told me it was key art from a disney pixar i'd believe you

[nimble heart]

they can pry 10 from my cold dead 8-fingered hands. I work in PC repair and the amount of absurd issues 11 manages to somehow produce reminds me of when 8 first came out.

[glad fulcrum]

Is sdxl made with a new dataset?

[paper phoenix]

i thought we werent allowed to talk about 8. 🙂

[static prawn]

this is exactly what i mean , its an awesome picture, super stylized, great art -> but its blurry . I never struggled with that on 1.5 , even on lower res.

[amber fulcrum]

Heavy metal style this time: "discord user in form of zombie sitting next to computer with glass of wine with text saying "dissapointed"

[nimble heart]

I'd love that but unfortunately I get lots of old people bringing in desktops that have the original 8 installed like "pls make work" and I have to bully them into letting me install 10 on top of it

[paper phoenix]

try running it through controlnet & tiled upscale but with some 1.5 model on super low denoise, see what happens.

[trail bay]

he does seem quite unhappy

[inner ruin]

do y'all get grainy results with diffusers?

[paper phoenix]

you're doing gods work!🫡

[boreal bough]

...close enough

sticking to seed 2, wont always get you what you want, but it is amusing XD

[autumn forum]

Nub hand

[static prawn]

this one has way more sharpness, it just seems to be super random if u get a blurry or sharp result 😄

[boreal bough]

its a prompt thing really. some words are biased to be blurry

[paper phoenix]

this is what i imagine the bottom of the irish sea looks like. all that waste from sellafield had to go somewhere.

[boreal bough]

either avoid those words - or find a set of counter blurry words to put in negative, that wont mess up your picture

[upbeat summit]

yes, you are right. I overall like the composition - no further processing or in-painting done. it tells a story, but it's true that it's easy to get a super soft depth of field or make everything just a bit out of focus.

I think with further workflow improvements and prompt engineering we might be able to modify it.

[boreal bough]

also, weighting works. so you can do this in negative
(blurry:1.3)

depends on artstyle/photostyle if it works well or not

[paper phoenix]

I'm still being a boyscout and waiting for 1.0 to drop before using it. y'all are the canaries down the mineshaft

[amber fulcrum]

It has been happening in 1.5 too

[primal nimbus]

when is sd xl dropping?

18 or 19th

[inner ruin]

tuesday

[urban fjord]

Yes depending on your time-zone

[primal nimbus]

people talking already about finetuning sd xl, whats that about?

leak?

[primal nimbus]

im in germany and its 2 am almost

when it will release for me then?

[boreal bough]

1920x1024 works oddly well

"young woman summons fire magic that goes brrrrrr, in the style of (disney:0.7)"

[paper phoenix]

looks like the most cheerful frankenstein scene. just before they storm the castle....with a musical number.

[upbeat summit]

really cool. the expressions especially!

[amber fulcrum]

It is quite good considering so many characters

[upbeat summit]

that's a real warm character portrait. great atmosphere!

[boreal bough]

Cursed, cursed ♪ be the fiend
We'll no longer live ♪ under the horror of his

[compact flax]

So, I modified KTiledDiffusion to pass the new conditioning parameters relating to cropping for each tile.

It doesn't work so well with simple text to image, but works extremely good for img2img upscale

Base 1024x1024 txt2img:

4096x4096 tiled upscale:

0.5 denoising btw

[glad fulcrum]

Can I do img 2 img with sdxl in COMFYUI?

and how...?

[upbeat summit]

the upscale looks good!

[eternal fog]

Can you do it on a person and see if it washes out any skin detail?

[compact flax]

Pass an image node into a VAE Encode node and pass the output into a sampler

[compact flax]

Hmm, would you be able to provide a sample and prompts for me?
Only asking because I'm exhausted for the day

All I need is an image and some text describing it and what it isn't, standard stuff lol

But yeah I am curious how it handles that

[amber fulcrum]

Is this space chuck?!

Imagine future space exploration games looking like that ^^ my dream

[compact flax]

Hyped for Starfield?

[amber fulcrum]

yeah! I know will be buggy but probably closest to what I want in space games

[compact flax]

Same here, I'm like extremely hyped, I'm expecting a buggy launch but I'm still debating on preordering it or not.

I mean, the Direct spoke for itself to me at least

[eternal fog]

I've managed to get Img2img and upscaling to create more detail, but it also introduces background noise

If I try stop the noise, it goes smooth again 😦

[compact flax]

Is this the image you want me to test on?

[eternal fog]

Use the original gen, to see if it adds detail

[compact flax]

Mind providing the prompts?

[eternal fog]

It's from an img2img I'm playing with so it doesn't actually really match the prompt I'm using

So I think these prompts might make it weird, because they don't really make sense

[boreal bough]

sry frankenstein

tracer from overwatch summons a fireball with castle frankenstein in the background, in the style of dnd artwork

[eternal fog]

@compact flax Try something like Movie Still of a young Woman

And then in the 2nd encoder I have Close-up Upper body.

[compact flax]

I'm haven't gotten the chance to use the second encoder with this particular workflow yet, let me see...

[eternal fog]

I'm so close to getting something decent, but when it adds detail it seems to leave that noise over in other places and looks a bit crap

And when you remove that noise, it removes all the detail

[lament rune]

some illustration styles are waaaaay overfitted yes

maybe rlhf issue

[civic sigil]

I worry 1.0 will be even more overfitted

[compact flax]

Alright, it's queued

I'm gonna queue one exactly with those parameters, and other with some detail prompt fluff

[lament rune]

anyone notice if you make someones hair purple.... everything in the background gets purple too!?

[uneven dove]

[compact flax]

I saw an addon called "cutoff" for both Comfy and Auto, not sure if it works with SDXL on either, but seems to be a solution to colors affecting things they should not

Haven't tried it yet, full disclosure

[boreal bough]

it will definitely be T.T I just really hope they dont put more weights into anatomy

else it will genuinely become impossible to train

like yes, its nice to have almost no deformities. but at what cost?

its like hardcoding loras into the base model

[boreal bough]

we doing this again? xD

cause I'm down

[lament rune]

what if you have rlhf.... but then everyone had bad taste 😄

[compact flax]

First one done baking, here you go.
This follows your simple prompts to the letter with a negative of text, watermark downscaled by 0.5, upresed with RealESRGAN-4x, fed through with 0.5 denoising, can try with higher and different settings.

Hair looks detailed

Can't say the same of the skin

[eternal fog]

Yeah it seems to do the same as what I've been having

Although I think RealESRGAN has added even more artifacts in the background

[compact flax]

Second one with slight prompt engineering almost done

[eternal fog]

img2img definately works differently in SDXL than SD 1.5

[eternal fog]

I think it's the way it does the denoising that's causing it

[civic sigil]

How so

[upbeat summit]

wide shoulders and stretched heads. a phenomena I'm familiar with these days - at least in some aspect ratios.

[compact flax]

What the fuck?

[eternal fog]

Well with SD 1.5 (Although maybe I'm being unfair comparing to finetunes)

When you use img2img with a low denoise, you can slightly change things or add a little detail to things like eyes, by upscaling and then doing img2img.

With SDXL, it really feels like it doesn't do small changes, and when you upscale and do img2img, it washes out detail. Unless you add so much noise it can't remove it all.

[grizzled warren]

absolutely cursed

[eternal fog]

Now that's deep fried

[uneven dove]

when the tiled VAE kicks in

[compact flax]

I've BEEN using Tiled VAE lol

I ask again, wtf?

[uneven dove]

cursed seeden?>

[upbeat summit]

I want to watch something from this universe. great images!

[compact flax]

Same seed as the last but I am using a specially modified tiled difusion method so it might be a little weird

Well, let's hope my GPU isn't suddenly dying or something lol

[uneven dove]

mansplaining tiled diffusion prompt

[compact flax]

This one is special

[upbeat summit]

Greetings @uneven dove

[uneven dove]

that million yard stare

[compact flax]

I'm serious, I modified KTiledDiffusion to pass the new conditioning parameters for crops that SDXL added

[upbeat summit]

I really like the "cleanness" of this

[boreal bough]

real damn pretty!

[compact flax]

Very nice

[uneven dove]

can't tell if it knows what mansplaining is or if i need to explain

[boreal bough]

cat staring at you from deep with the depths of the washing machine
style: (horror:1.2) still, grunge, vignette, chromatic aberration, dark, lovecraftian

[uneven dove]

brother, mine was a kitchen sink

[boreal bough]

well, it fit the horror theme better

what horror movie happens in the sink

[uneven dove]

ಠ_ಠ

[uneven dove]

@boreal bough It Came From Under The Sink

was that a serious question

what about The Blob?

[upbeat summit]

prompt:
highly detailed pencil and watercolor, looking out of the window seeing a huge alien spaceship ready to board, dim and dark, awkward anxious butch brilliant blonde tomboy teenage girl engineering student with short messy hair looking out the windows from behind, in flight suit, modern children's book, cinematic, muted colors, faded, dynamic lighting, art design by horizon zero dawn

[uneven dove]

Attack! Of the Killer Tomatoes?

[compact flax]

My CFG was set to 100

HAHAHAHAHAH

[uneven dove]

what about that scene where the garbage disposal...

[compact flax]

How the hell

[boreal bough]

you just pointed out what I'm avoiding XD I dont want a dying cat, to be the horror aspect

[compact flax]

I swear ComfyUI has a ghost that changes random values from time to time

[uneven dove]

@compact flax i had a typo in my code that tried to get the CFG from the wrong value and overwrote it woth 10.0 or greater and then, when i fixed it, EVERYTHING CHANGED

[compact flax]

I see lol

[boreal bough]

there XD

[compact flax]

ComfyUI interrupt/cancel button when

(Unless I'm missing it lol)

[eternal fog]

There is one

Under Queue Prompt

See Queue

And then you can click cancel on running jobs

[compact flax]

Bruh, that's kinda non-user-friendly

But it is a node based software

[eternal fog]

Yeah it took me a while to find it

[compact flax]

Thanks

[boreal bough]

one man open source, for the longest time unpaid, development team
he's already working in overdrive right now XD

[uneven dove]

ooookay feline pope francis calm down

[compact flax]

ComfyUI misnomer correction update when?

lol

[boreal bough]

THE PLOT THICKENS
cat staring at you from deep with the depths of the kitchen sink, in the style of Attack! Of the Killer Tomatoes

[upbeat summit]

Amazing coherence

[upbeat summit]

was he cooking?

[compact flax]

"Y'all need Meowsus"

[boreal bough]

he was waiting

he has a tomato for you. just for you

[upbeat summit]

looks like the sauce is ready

[compact flax]

What was he cooking?

[paper phoenix]

I feel like cat's wouldn't really like the whole baptism/holy water aspect. maybe the incense

[upbeat summit]

you tell me

[upbeat summit]

maybe he was getting the groceries

[compact flax]

Does he know?

(Not AI)

(In case he doesn't know)

[uneven dove]

maybe they're born with it... maybe it's maple beans

love teh use of goggles inside

that's just smart thinking

[boreal bough]

cat playing with the infinity stones, infinity gauntlet in background, in the style of marvel movie

[compact flax]

My God...

Lol

[paper phoenix]

storebrand Rocket from GotG still kinda rules.

[boreal bough]

"mrrrrhhhghhh" he says, pointing out that with unlimited power, anything can be turned to tomatoes, and you have seen what I can do to tomatoes

[uneven dove]

wants to gwow up to be just like daddy, a pilot

[upbeat summit]

yeah 🙂

[compact flax]

Marvel universe question: Can you use the Infinity Stones to render the Infinity Stones powerless? Say turning them into tomotoes?

[uneven dove]

can you smell what the rock is cooking? loss of smell impacts one in four americans each year

[boreal bough]

instructions unclear. stones gone.

[compact flax]

[boreal bough]

nvmd. infinity tomatoes confirmed

[upbeat summit]

great mood!

[compact flax]

Nuh uh, I have obtained all 5 infinite rocks for my Thanos Glove

Say goodbye to half of your floating point precision

[boreal bough]

hahahahaha

fp16 infinity tomatoes XD

[compact flax]

Damn, you run on FP32?

[glad dock]

what is the current model used in #1100170312106127410 those channels? is it SD XL 1.0?

[compact flax]

I was hoping you were on FP16 so I could turn you into INT8

[eternal fog]

They all have different possible release candidates of 1.0

[glad dock]

Thanks~

[boreal bough]

infinity tomato running in fp16 only

[compact flax]

int8s your fp16

Haha

[compact flax]

Has anyone managed to quantize SD models?

[boreal bough]

does vae count?

[uneven dove]

an obese feline airline pilot named bob, overworked, never sees his family

[boreal bough]

int8 infinity tomato caused NAN. world gone.

[compact flax]

It's a start I suppose

[boreal bough]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[compact flax]

"Not a Number"
Then what the fuck is it? Hmm? A string?

[boreal bough]

just dont go int8. I kinda still like our world

[compact flax]

That's FP16, not int8

Is converting FP32 to FP16 considered quantizing still?

[boreal bough]

for sdxl it is, cause we're all using bf16 to run it - which uses 3gb more vram than fp16 vae

[compact flax]

I see lol

[boreal bough]

best optimization we have so far ^^'

[compact flax]

Nice, now how do I bake this into an SDXL model so that I don't have to concern myself if Auto/Comfy/Whatever can load it separated or not?

[boreal bough]

comfy -> load vae
is the easiest solution for now

cant help you on any other implementations

[compact flax]

[eternal fog]

I used that and it made basically no noticable difference for me

[inner ruin]

the image quality kind of tanks when I add a LoRA to SD XL, even if the LoRA just started training. I think either I'm doing something wrong or the diffusers script is broken? 😄

[uneven dove]

making bullies is hard

[glad dock]

is it possible to bring StableDreamer bot to my own channel?

[compact flax]

Difference in VRAM usage or...?
Because if it didn't make a perceptual difference that's good lol

[eternal fog]

Didn't make a difference in VRAM usage or visuals

[compact flax]

Damn

Maybe it still loaded and coverted to BF16? Is that possible?

[uneven dove]

didn't change my vram use either

[boreal bough]

in that case rip

[uneven dove]

but i'm tiling and slicing the VAE anyway

[compact flax]

Very nice, Tiled VAE is the greatest invention here

Aside from Stable Diffusion itself lol

[upbeat summit]

SD 2.1 / SDXL 0.9
miniature sailing ship sailing in a heavy storm inside of a horizontal glass globe inside on a window ledge golden hour, home photography, 50mm, Sony Alpha a7

[uneven dove]

steve irwin holding a stingray

[uneven dove]

try pseudo-flex-base or pseudo-journey-v2 (or both, since one is photoreal and the other is adventure)

[upbeat summit]

Already sorted into my model library. I will definitely check it out and maybe try including it in a SDXL workflow 🙂

I already recommended pseudo-flex-base to a couple of my community friends

[boreal bough]

"a male angler fish, also known as the saddest animal in the animal kingdom"

its almost 100% accurate too XD

[uneven dove]

interesting, i forgot to change my cfg

[uneven dove]

lol pseudo-journey-v2 and its lack of aspect bucketing

flex-base has an interesting interpretation

[upbeat summit]

it got the mood right 🙂

[boreal bough]

no no, that's a mirror, and a vampire glass stand

[sharp robin]

ok im tired but im happy w workflow now

ready for 1.0

[wicked frigate]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[boreal bough]

in 3 days, after 1.0, maybe possibly vielleicht release of griddy thing?

[upbeat summit]

inception globe

[uneven dove]

globeception

[wicked frigate]

i don't know what the word in the middle of that sentence is

but yes griddy thingy soon

[elfin cobalt]

https://brage.info/GAN/ganbot2/8cca726e-d06e-43f9-92ac-b0c1bd8f0cfc.1.jpg

https://brage.info/GAN/ganbot2/8cca726e-d06e-43f9-92ac-b0c1bd8f0cfc.0.jpg

I think someone said XL is bad at anime?

[eternal fog]

You need to remove the width and height off the links or they won't embed

[sharp robin]

Aight. I think I changed my mind. Refiner is a useful tool and am warming up to it being a thing. One caveat is it’s going to have to have variations. Does add substantial details but it needs to be a tool. Because it will turn anime into almost real.

[elfin cobalt]

They don't embed anyway, but that's fine I guess.

[boreal bough]

it is, if you prompt wrong
its amazing if you prompt right - no lora needed

[sharp robin]

Does amime ok.

[boreal bough]

though lora will remove the need to prompt properly - reducing it to a single trigger word

[elfin cobalt]

Heck, it's amazing if you just get an AI to do the prompting.

[urban fjord]

WaifuXL is supposed to be released the 18th too so it doesn't matter what base can do.

[eternal fog]

You wot m8

[boreal bough]

basically no too short prompts

[sharp robin]

How they have 1.0 already?

[elfin cobalt]

I would call this 'great'. It's not as good as 1.5, for pure aesthetics scores, but it's got 500% more personality.

[sharp robin]

Did they train on .9?

[eternal fog]

They don't

[urban fjord]

Emad mentioned something about certain finetuners getting early access.

[boreal bough]

trained on 0.9

[eternal fog]

They just got 0.9 the same as everyone else

[urban fjord]

But I don't know what's the case here.

[sharp robin]

That’s horrible

[urban fjord]

It doesn't matter really.

[boreal bough]

i mean not like there's any advantage to waiting for 1.0, if you only want anime. the current model is only lacking in certain real photography aspects (which are in the refiner)
base can easily be fully converted into an anime model - before its overfitted with face data 🥲

[eternal fog]

They are going to re train on 1.0

[elfin cobalt]

So they just wasted money?

[eternal fog]

Not really, testing out what works before doing a full one

[boreal bough]

I genuinely wonder though - since now that refiner data is going into base, they may genuinely be better off not using 1.0
while niche, for them it feels like the right choice

[boreal bough]

some people own an A100 XD

stares at pseudo

[sharp robin]

It’s still not confirmed that they are doing away w refiner no?

[uneven dove]

"refiner data is going into base"?

[boreal bough]

the bot does faces now - so clearly the faces got more training now - is what I meant

[uneven dove]

base did faces before

[boreal bough]

but it felt significantly undertrained

[uneven dove]

i was just hoping you had more info because they never tell us anything lol

[boreal bough]

sadly not T.T

[uneven dove]

"i'm undertrained?! "

that's absolutely John Candy

[slender coral]

Thanks mate, quick question with WAS, do I just drag drop into the \custom_nodes dir?

[uneven dove]

you're going to make him cry.

[upbeat summit]

vhs quality effects, one of the ninja turtles working at a pizza place, new york, from a 1980 tv series, analog film, analog distortion, cinematic, tape glitch effect

vhs effects leave something to be desired about, but at least it has some analog footage look going

[upbeat summit]

you're welcome. you git clone or copy the WAS nodes folder into your custom_nodes folder. try starting comfyui. it should download the requirements. if not you need to manually install the requirements.txt with pip install -i requirements.txtin the WAS nodes directory

[slender coral]

Got it, I don't like making submodules so just urared the zip, going to explore, if I make that node I'll shoot it off to you as you mentioned you didn't have one yet 👍🏿

[upbeat summit]

I don't want to overload everything as well but WAS is currently essential for me. lots of great helper nodes.

[slender coral]

Last question, is there currently any repo or docs around it, specifically for community tools?

[upbeat summit]

I mostly getting all my info from the community here or github commits

[autumn forum]

i get this when trying to installl requirements ERROR: You must give at least one requirement to install (see "pip help install")

[upbeat summit]

pip install -r requirements.txt

[slender coral]

I would suggest also being in a virtual env venv don't just install requirements on your system, makes a mess for any future use of AI modules.

[glad dock]

for clip drop function, once I clip the image from my phone, how do I drop it to my PC? is there a tutorial on this?

[autumn forum]

yeah im not smart enough to figure that out

[upbeat summit]

since I'm on windows I'm using a portable WinPython installation. this way I can easily switch between different versions and make backups.

[paper phoenix]

it'd actually be worth making a spoonfeed video or something for people to teach them how to use git and conda to create a little AI directory to keep all that good stuff in.

[urban fjord]

python -m venv venv
Use this command to create a venv

[paper phoenix]

if you know symlinks, have git and conda installed you can pretty much futureproof nearly anything AI related

[upbeat summit]

I guess most were waiting when a1111 was going to update, even though you could use it since days with sd.next

[urban fjord]

Funny, Conda always break on me.

[uneven dove]

the pope eating a miniature pope

[paper phoenix]

funny thing is, they do sell tat like that around the vatican

[uneven dove]

the pope eating a stack of money?

[nimble heart]

to add to the venv thing if you create the venv with something like python3 -m venv --system-site-packages venv you can re-use system-wide dependencies while keeping anything new in the venv, so you can say have torch 2.0.1 installed system-wide but still have local copies of pinned gradio versions for whatever UIs you use

[paper phoenix]

edible papal effigies. dudes 90% bling anyway.

[paper phoenix]

exactly the beauty of it.

[paper phoenix]

wait, why is a chef dog doing barista work? This is totally unrealistic.

[glad fulcrum]

jaja

[nimble heart]

labor shortages

or maybe labrador shortages

[paper phoenix]

HAH

shouldnt have, but did laugh

[uneven dove]

a rural kentucky farmer

[glad fulcrum]

maybe try barista, as I didn't know that word existed

[paper phoenix]

it exists in countries that take coffee seriously 🙂

[glad fulcrum]

well the dressing that comes out of it is not so charming

I like the chef better

[uneven dove]

Jorge Mario Bergoglio apparently works

helps to bypass the 'pope'ness

[upbeat summit]

nice gloves

[uneven dove]

he's a plumber so i guess plumber = super mario?

[upbeat summit]

in the data center?

[uneven dove]

computer tech shop 1995

[upbeat summit]

nice

[paper phoenix]

clearly fake. a pope would never wear a silly hat and outfit like that.

[upbeat summit]

really like the high quality cinematic look that you add to your images

[uneven dove]

💀

[upbeat summit]

it's like one of those 1980 IBM ads

[uneven dove]

so u liek moose, eh

[dense chasm]

anybody install the https://github.com/OpenTalker/SadTalker img2video application?

[urban fjord]

I've tried it, it works but I've not had too much use of it.

[boreal bough]

wheeeee. overwatch properly works now. no more needing to train it XD

[paper phoenix]

train it to do clothes. 😛

[dense chasm]

i wanna to make talk show short videos anyway

[boreal bough]

it can. the point was if it's able to be adapted to other things, or if the prompt learned weird clothing details like in 1.5

[paper phoenix]

It takes a long time to render but once installed its pretty painless. pick a photorealistic image to work from, give it a long audio file of speech and go make lunch or something while it renders.

[dense chasm]

where is Rudolph's red nose huh

[autumn forum]

okay i ran that code, honestly i am a noob at python and coding and stuff so dont know exactly how that helps but a venv is there now. i was also able to install was correctly.

[paper phoenix]

I was just making a joke about ashe's lack of clothing in the image 🙂

[uneven dove]

do Darkwatch now

[dense chasm]

yep,videos take much time than pics and have to edit frames with some software

[boreal bough]

the amount of hours I spent in this game XD

[paper phoenix]

are those handles attached to the bear or his groin?

[uneven dove]

yes

[boreal bough]

its all fun and games, until the monster hunting you starts being hunted by an even bigger monster

[dense chasm]

The Call of the Wild film seemd render with a lot of AI effects

[uneven dove]

i don't get it, 'riding' adds handlebears to them

[boreal bough]

I've seen this. and I regret nothing. Dont take my motorbike pidgeons away from me ❤️

[boreal bough]

Tala from Darkwatch

closeup portrait only, since... reasons

[uneven dove]

i dont think im doing this right

they look confused like they don't know what to do

[boreal bough]

Jericho Cross (pre vamp issues)

[uneven dove]

i can remember how great that combat in that game was

lol

[boreal bough]

they also put a huge effort into VS mode. which most people never got to experience T.T

[uneven dove]

i got to

[upbeat summit]

hybrids are pretty difficult to make. but I've seen some good ones with SDXL

played my fair share of OW with a group but at some point you need either invest more time to get better or get a bit annoyed by the toxic community

[steady chasm]

All the cool hybrids I've seen so far on sdxl are in this server, and met with disappointment as they're always accidental X3

A lot of them look so sick

[boreal bough]

the trick is to be greatful for losses, for keeping you low and away from toxic people XD

[paper phoenix]

i moved from european to us servers to avoid the awful community. its like the online version of the stanford prison experiment.

[upbeat summit]

haha yeah

[paper phoenix]

ended up just maining pharah because i could lead projectiles with the latency

[upbeat summit]

I did that back playing Quake 2. Played on US servers. The lag pretty much made me invincible sometimes 😄

[paper phoenix]

i did just give up on OW though, was either playing sweaty or being tilted by absolute strangers making strange commands about what i have to do. not fun at all, it was a game i bought, i wasn't trafficked into playing it lol.

[boreal bough]

"Winston from Overwatch"
yeah... I give up

[paper phoenix]

He could be nemo's monkey butler in the league of extraordinary gentlemen though

"I'm not a monkey....I'm a steampunk etsy store owner!"

[upbeat summit]

it was nice while it lasted

[uneven dove]

@dapper current look at this display of defiance we can do locally

[boreal bough]

no brain. all floof.

[autumn forum]

so was node suite has sooo many nodes, im gunna have a fun time with that

birb

[true wing]

SDXL really is a lot of trial and error sometimes

[autumn forum]

its the 3 blind mice!

[true wing]

well that's bad news for these guys then

[upbeat summit]

yeah. I really like it. I've build some quality of life features for my workflow just using WAS. I should instead be making my own nodes, but it's real fun to create components and helper tools just with ComfyUI and WAS

[boreal bough]

because I can

[upbeat summit]

this image is fantastic - the version you posted in #🌠｜show-and-tell message 👌

[upbeat summit]

are you doing training tests already?

[eternal fog]

No

That's base model

[upbeat summit]

yeah - I mean in general

you made some pretty good ones during the SD 2.1 launch

[eternal fog]

No because I can't train because the trainer is busted

[upbeat summit]

oh okay. I didn't know that :/

[eternal fog]

Yeah it worked a couple of times for me fine

Now it just OOM every time

Even if I try put the resolution at like 256x256

[upbeat summit]

strange

[true wing]

We should be just a few days from the 1.0 launch anyways, I can't wait to see what people do with it. I believe emad said there was some sort of optimization done on it which I think was post 0.9

[boreal bough]

Watercolor painting, delicate washes, translucent layers, soft and ethereal, fluidity and movement, luminous colors, atmospheric effects, spontaneity, versatility, dreamlike quality, expressive brushwork, subtle textures, natural and organic, evocative, soothing
nice for it you want to print some text, then these images can be overlayed with some white, and used as a pretty background!
printed text instantly 90% better

[true wing]

Honestly that's just a great style prompt in general.

[slender coral]

https://github.com/comfyanonymous/ComfyUI#i-already-have-another-ui-for-stable-diffusion-installed-do-i-really-have-to-install-all-of-these-dependencies

[boreal bough]

omg the toast XD

love it

[true wing]

yeah it's really good.

[autumn forum]

thanks for the link. i just ended up creating a new venv. lol guess i should have read the readme better when i was first installing comfy

[boreal bough]

toast intensifies

[slender coral]

bruh, can you dm me the prompt if you can. 👌🏿

[autumn forum]

second pic dont look right..

[upbeat summit]

tried some style mixing:
highly detailed pencil and watercolor, aerial shot time travel cyborg fighting against a rogue AI that has taken control of all military drones on Earth, modern children's book, cinematic, muted colors, faded, dynamic lighting, (art design by Zdzislaw Beksinski:1.15) and Remedy Control and (pixar:0.85), directed by Steven Spielberg

[boreal bough]

when prompts go horribly wrong, and right at the same time

I mean so much for her reading a book

[true wing]

It looks like she's the cover of the book

[upbeat summit]

so much detail

[west breach]

loving this style

[true wing]

Anyone else getting hungry?

just me?

[upbeat summit]

the style blends really well

[true wing]

Yeah Caith's prompt is amazing for this type of thing.

[west breach]

I just had curry for lunch, very full right now!

[boreal bough]

not gonna spoiler this one
try this prompt - then anything with lithograph, risograph at the end

cat jumping on a box, lithograph, risograph

N: text, background, signature
negative needs to be like this, for maximum effect XD

it works for every damn thing

[boreal bough]

that steak is 10/10

[true wing]

I think his kitchen might be on fire but it is a great looking steak

[upbeat summit]

uh that is nice

[autumn forum]

he was right risograph works on everythiung haha

[upbeat summit]

Thanks for sharing, Caith! Really cool 🙂
delorean from the movie back to the future driving through a tunnel, synthwave vibes, lithograph, risograph

[vast narwhal]

4320x4320 SDXL + Juggernaut

[boreal bough]

the detail is real! ❤️

[true wing]

isn't juggernaut a 1.5 model?

[vast narwhal]

Yes. My full SDXL isn't ready yet

[upbeat summit]

Impressive - great work!

[true wing]

So how does it work? Are you using XL to gen the base image and then img2imging with juggernaut? vice versa? something else?

[upbeat summit]

Yeah I would buy that workflow 😉

[autumn forum]

Shrimp on the Barbie

[upbeat summit]

how is duplication treating you in this aspect ratio?

[boreal bough]

I didn't even know that :: is a thing when prompting x_x

[west breach]

it's doing really well, i'm using 2.39:1

[glad fulcrum]

what's the workflow to use the refiner in auto1111? It says the refiner has initial support already

definetely something wrong with my comfyUI since in auto1111 the image starts to diffuse as soon as I click generate, in comfyUI I have to wait like 20 to 30 seconds beore it starts to iterate.

[west breach]

auto1111 loads the model on start up

first time you run in comfy is when it loads the model

[glad fulcrum]

yes I have been running comfyUI for a few hours

every time I click on queue it spends that time

doing no diffusion

[uneven dove]

i find the dupes aren't an issue with super wide it's just that the quality drops for a lot of prompts

[lament rune]

Yes 100%

Weirdly swampy untrained latent zones

[vast narwhal]

I need more tests, give me a good prompt for testing

[upbeat summit]

yeah. Im experimenting with many differences sizes. currently I'm mostly using 1920x1024 (~16:9) and 1444x1080 (4:3). of course the ratio goes down but compared to previous models it's a joy to have so many interesting images to look at.

[autumn forum]

huh? whats it do?

[upbeat summit]

high definition photograph shot with canon ef - s 5 5 - 2 5 0 mm f / 4 - 5. 6, full body, low focal point, narrow depth of field, bokeh : ( subject = robot + subject detail = futuristic, running stance, dark metallic, glossy, led), lit cyberpunk city by night, heavy rain, heavy fog

one of the crazier prompt constructs 😄

[boreal bough]

https://aidailies.com/midjourney/what-are-double-colons-weights-iw-and-multiprompts-how-do-i-use-them
or an image example showing the insane blending

A detailed portrait of Amy Quesada :: City center dripping with black ink and black slime in the background, lights reflecting gasoline colors:: Bojan Jevtic + Ashley Wood :: maximalist intricate detailed :: ray tracing :: hyperdetailed, maximalist, psychedelic, post-apocalyptic, photorealistic, 64k resolution concept art, dynamic lighting, trending on Artstation

[upbeat summit]

portrait of a dark sorceress, sharp focus, extremely detailed, photorealistic, RAW image, 8k, RAW candid cinema, 16mm, color graded Portra 400 film, ultra realistic, cinematic film still, subsurface scattering, ray tracing, volumetric lighting

[autumn forum]

so instead of each word having a weight of 1 its giving a whole phrase a weight? and if you put a number after say ::3 itll give the phrase a weight of 3?

[boreal bough]

apparently - but it also merges them somehow?

[autumn forum]

wow okay. thats wild, ima have to test that out

[boreal bough]

same XD

[tender timber]

How does SDXL handle object repetition above 1024x1024? I am getting lots of it at 2400x1600, almost none at 1536x1024

[upbeat summit]

I settled on 1920x1024 and 1440x1080 for now. still getting repetition but enough good ones as well

[tender timber]

So we'll still need a "highres .fix" type thingie

[west breach]

I use the RealESRGAN_x2 model and then go through another base + refiner to get the final image

[boreal bough]

while it can do bigger sizes, you will have the best results if WxH = 1024x1024
the math needs to math out

[upbeat summit]

must have a very powerful engine

[boreal bough]

512x2048 is the max magic math number you can get

[tender timber]

Yeah, 1920x1024 seems OK most of the time too, but 1920x1280, not so much

[west breach]

looks like a kenworth

[tender timber]

Oh, no... I get that, same as SD 1.3 was best at 512x512, but highres.fix had some interesting tricks, so hopefully there's something similar to that for SDXL eventually 'cause sometimes you just want a bigger canvas to play with as opposed to scaling things up

[upbeat summit]

64,64,64,64,64

[boreal bough]

already exists. multiple people here use various version of it

[tender timber]

Cooooooll, will hunt it down

[west breach]

I made a lora with a 1.5 model, the used it on the same model and was able to get higher resolutions without duplication, but had some interesting results, e.g. much stronger bias towards portrait photo of a woman even when not prompted

[boreal bough]

https://civitai.com/models/81540?modelVersionId=115129
here's one of the many versions of it. most people make their own after lots of testing

[tender timber]

My Snowpiercer train didn't work out too well 😦

[civic sigil]

SAI did mention that it adapts better to wider aspect ratios than tall ones

[tender timber]

Thanks, ~~precipitate ~~appreciate t! Damn spell correct 🙂

[boreal bough]

'pierce' checks out XD

[tender timber]

Yeah, I noticed that back in the old days before SD when things were on Colab in late 2021 ... wider seems to work better than tall. I remember Katherine speaking to it once, but can't remember where

[west breach]

what's the prompt? I'll try it on my workflow

[tender timber]

Snowpiercer train engine with 996 cars, going through the mountains, professional, detailed, cinematic

[upbeat summit]

SDXL's CLIP_L uses the same interpreter as SD 1.4/1.5. I tested some SD 1.5 embeddings and they work in ComfyUI but since they are obviously trained on a different latent space it creates mostly unusable images.

[boreal bough]

human/person photos work really well in portrait though - just nothing else

[tender timber]

Gettin' there now that the res is down to 1536x896

[tender timber]

I'm surprised my ol' 3060 is doing well under Windows

[tender timber]

Wow

[upbeat summit]

no funny gifs

[tender timber]

Nope... they work in the show-n-tell bot though 🙂

[sour obsidian]

[tender timber]

So, if the model is 1024x1024, does it not like to dip below ... say 768? Like SD 1.5 was not good if the image was 512x320

[tender timber]

https://tenor.com/view/joey-batey-bow-little-bow-you-are-welcome-gif-25142469

[upbeat summit]

I'm not falling for that!

[tender timber]

So weird... works now 🙂

[west breach]

https://tenor.com/view/crossfit-barako-crossfit-keiana-gif-21943704

[upbeat summit]

https://tenor.com/view/tim-robinson-i-think-you-should-leave-you-sure-about-that-gif-22429959

[tender timber]

Dustin's mad 'cause my kraken beat up his whale yesterday

[sour obsidian]

it was a good fight my whale put up

[autumn forum]

https://tenor.com/view/aunty-donna-looking-for-cowdoy-instead-of-promoting-our-netflix-show-cowdoy-in-the-city-mark-mark-bonanno-gif-23259783

[tender timber]

OK, back to tech stuff... what are people finding are the min image dimensions before SDXL craps the bed?

[west breach]

what i've noticed with using the refiner is the base might make something intentionally blurry, like motion blur of a hand or something, then the refiner comes in and tries to sharpen it up and makes it look weird. not everything needs to be sharp

[tender timber]

1258x832 is good(ish) ... same prompt at 512x512

and... 768x768. Other than a bad case of Oligodactyly, it's OK

[boreal bough]

photo of the pope :: Magical embossed strips of tape, black and white, patterned, antique, photographic, 2d, rustic
vs
photo of the pope, Magical embossed strips of tape, black and white, patterned, antique, photographic, 2d, rustic

[autumn forum]

i accidentally generated something horrific.. Photograph of a man::1 eating a sandwich::3 , Seed: 70,

[boreal bough]

I name this, The Dunwich Horror

[autumn forum]

i cant really tell what :: is doing, like its doing something, but cant put my finger on it yet

[boreal bough]

reducing both weights

[autumn forum]

ohhh yeah i see whats its doing to your first prompt

[boreal bough]

(part1 of prompt:0.5), (part2 of prompt:0.5)

[upbeat summit]

really nice. I'm just realizing that you can very often recreate a similar composition with a different seed. You can now start adding style words or post production effects. SDXL is way less random - at least with some prompts.

[dusk mica]

does adding "4k,8k,XDR,HDR" do anything or is it just snake oil?

[boreal bough]

yes - but dont add them at the same time. add them one by one, since they actually do things (influence bokeh, influence composition, etc...)

[dusk mica]

ok thanks

[upbeat summit]

we are going to make a lot of vector graphics

[upbeat summit]

damn. it's working 🙂

[upbeat summit]

I'm not so happy with most upscaling workflows because of over-sharpening and other artifacts. independent of the really great workflow used here, these images look excellent. the grain is especially nice and makes them really cinematic.

[vast narwhal]

Did you saw my first work? The dog? It's only 3472x3472

[autumn forum]

just tested it, doesnt really do much but change the image up a bit. maybe helps idk. i added ", 8k, high resolution, HDR, Best Quality" in the base text prompt.

[civic sigil]

I would say thats a big difference for better or worse

[west breach]

https://tenor.com/view/offended-whyy-shocked-what-shookt-gif-17241626

[autumn forum]

here it is with the text in the refiner not the base

[upbeat summit]

overall sharpness and colors are great here! great work. I hope I can build a workflow like this!

[dusk mica]

right one looks much better

[upbeat summit]

I have said most - not all!

[civic sigil]

The worst qualities of the other two lol

[autumn forum]

looks like it! using it in the base prompt seems to have a better effect and could potentially make the picture better

[vast narwhal]

Your other prompt. I can share my workflow when it's ready. No problem with that

[west breach]

my workflow still fails at human faces most of the time

[upbeat summit]

yes, please do. it's always exciting to push fidelity! I guess some of the fine textures are coming from juggernaut?

[west breach]

but can do so many things really well, then human face.. blerg!

[uneven dove]

4096px

oops

that one i realised is a lil gruesome lmao

[vast narwhal]

Since this is the original 1024x1024 i can say it from your prompt

[autumn forum]

as long as the cat is okay! he kinda looks like he killed someone tho..

[boreal bough]

john trapped in a web of lies, neon, highly detailed, juxtaposition, euclidiean, explosive, arachnid, abstract, detritus, ((psychotropic)), radiosity, volumetric, sanguine colors, bold cursive lines, stencilled, acid etched, anthropomorphic, interlaced, double exposure, ((halftone)), zoom blur

sry john

[upbeat summit]

try this prompt. this is my main SD 2.1 portrait prompt build. works for landscapes and portrait images. it's a cinematic analog look - so not too modern but very moody. Just fill in the placeholders:
cinematic movie extreme close-up still of an epic scene of a [ETHNICITY] [OCCUPATION] in the [SEASON] at [DAYTIME], centered, looking into the camera, fog atmosphere, volumetrics, photorealistic, from a western movie, analog, very grainy, film still, kodak ektar, fujifilm fuji, kodak gold, cinestill 800t, kodak portra, photo taken by thomas hoepker

[boreal bough]

also john

[upbeat summit]

yeah. one of my favorite prompt builds since SD 2.1 for sure. the high res version really is impressive. I so want that 😄

but it's 1004x1004

[uneven dove]

looks like i saved that one wrong lol

[autumn forum]

risograph playing cards

[west breach]

I wonder if having an ethnicity helps to draw the face better

[ionic dragon]

is there any model, which can be trained by a lot of prompts?

so that when we use it, we can get a good stable diffusion prompt

[uneven dove]

RoBERTa

[uneven dove]

easy to fine tune

[ionic dragon]

rather than telling each and every detail to chat-gpt?
chatgpt also doesn't give the perfect prompts too

[ionic dragon]

the prompt model?

[uneven dove]

yes..

[ionic dragon]

how can we?

[uneven dove]

lol google

[upbeat summit]

yeah I think so. for the placeholders I use wildcard files. the prompt creates mostly high coherent portraits imo

[ionic dragon]

https://huggingface.co/Gustavosta/MagicPrompt-Stable-Diffusion

there's already one

[vast narwhal]

I think i need to make changes for this kind of close up, the last phase burned a bit. But the reflection in the eyes is something. It's a brazilian farm in autumn

[ionic dragon]

@languid scarab are you the dev of this?

[upbeat summit]

yeah, it looks a bit overcooked and some haloing / glowing is going on. but I can still see the fine details. so if the settings would be fine tuned, it could be a great highres image.

cool - thanks for doing those!

[upbeat summit]

hey - it works in ultra-wide 🙂

[west breach]

the skin is so wet in a lot of these 😄

[upbeat summit]

yeah just a bit more and it might get uncomfortable 😄

[vast narwhal]

This is the stage before. It's very good,

[upbeat summit]

overall much cleaner

[west breach]

yikes!

[upbeat summit]

maybe you need to tune your negative? skins aren't that moist in my portraits heh

I'm using this classic negative prompt for the portraits:
Photoshop, video game, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, 3d render, photograph

I build on that negative almost all my images for SD 2.1. also works with some SDXL prompts pretty well to enhance realism and fidelity

[west breach]

this is my current negative: illustration, cartoon, 2d art, trypophobia, low resolution, crushed cans, crumpled paper, blurry, mangled hands

[high skiff]

[upbeat summit]

for SDXL I need to refactor my prompts for sure. some stuff isn't working at all as before. like always - it's like learning a new language and exploring the latent space is a lot fun

[high skiff]

My new PC will be here tomorrow

[autumn forum]

ayyy

[high skiff]

[upbeat summit]

how are you @high skiff?

[high skiff]

wrong emote, but I will keep it lol

[high skiff]

good, just endlessly tired is all

[upbeat summit]

I know the feeling

[high skiff]

I doubt I will slepe much with my anticipation for tomorrow lol

nooooooo

My GPU is getting here laatttteeee

I'm gonna miss SDXL 1.0's Launch

[autumn forum]

hey sytan guess what..

[high skiff]

What am I guessing?

[autumn forum]

https://tenor.com/view/matt-amodio-its-working-gif-23487713

[high skiff]

WHAT IS!?

[autumn forum]

gifs lol

[high skiff]

oh-

I thought you meant SDXL LoRA's lmfao

[autumn forum]

LMAO. sorry if i got you excited

[high skiff]

puts my part hat away

gifs are good too tho- 🥲

[west breach]

adding wet skin to the negative

[upbeat summit]

smoothed it out

[high skiff]

moistn't

[autumn forum]

kitty

[boreal bough]

kay. it's 8am. I'm off to sleep XD I'll see ya tomorrow later

[west breach]

scuse me, you have something in your hair

need to figure out how to level up to 4k, it uses up like 39gb of memory for 1 image

[timid sonnet]

@west breach There are programs made for gaming (not known ones, I can't recall the name but look into setting affinity and ram usage to improve FPS), that would help you be more efficient

[west breach]

I'm talking about generating a 4k image with SDXL

[sharp robin]

Damn SDXL rocks, cant wait for 1.0

[timid sonnet]

Is it possible for comfyui to be created into an actual Python Application with internal web view, html, css?

[autumn forum]

i really want some donuts right now..

[molten gull]

is it possible yet to traing embeddings in sdxl ?

[ionic dragon]

if sdxl is available in a1111, then how can we use base and refiner?

[mellow dome]

they actively working on it like crazy: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11757

[sharp robin]

refiner refines

photorealistic looks better

[ionic dragon]

i know what a refiner does, but how can we use 2 models in a1111, we can use multiple models in comfy

but i am not sure how we can use it in a1111

[sharp robin]

then u dont know, use it as an img2img model, create ur gen, then img2img it w refiner

[ionic dragon]

oh ok

does a1111 support 0.9?

[timid sonnet]

Maybe add something about riding stance

[mellow dome]

project vayne

[timid sonnet]

A1111 doesn't work for me for SDXL, the preview loads and shows steps being added, then it makes a black image

[mellow dome]

haver u try on this branch https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11757

git switch sdxl

[upbeat summit]

did you try running it with --no-half-vae?

[static prawn]

i cant get into sdxl, probably will stay with 1.5... tried so many times now but any of my results are extremly bad.

[winter raptor]

Can SDXL effectively render images with lower resolutions than 1024x1024, or is there no point in really trying?

[static prawn]

it can but i guess everything is trained on 1024x1024 so it makes more sense using this res.

[timid sonnet]

@mellow dome trying now @upbeat summit ill try that next

[winter raptor]

I really gotta get it running on a1111 instead of comfyui, comfyui hurts my brain lmao

[mellow dome]

i was waiting for a1111 too

[static prawn]

comfy with a nice workload is pretty much the same

[winter raptor]

I've just gotta figure out how to outpaint, increase batchcount, etc with comfy. I do really like that comfy automatically uses tiling if you run out of resources

[static prawn]

so many tries, always terrible results like here...

[upbeat summit]

what's your cfg?

[static prawn]

7

but tried allready lower cfg and higher cfg

everything is completely full of grain

[upbeat summit]

with some prompts I go down to 3 - 4

[static prawn]

or extremly blurry

1.5 looks so clean compared

[upbeat summit]

the soft depth of field is very present but you can counteract it a bit in the negative

[winter raptor]

You routinely get some of the most impressive results I've seen with SDXL, awesome stuff

[upbeat summit]

Thank you, Max! Much appreciated

[ionic dragon]

@upbeat summit can you please share your workflow if you are comfortable

[static prawn]

idk its probably bec. i generate on sd 1.5 with 512x512 and then highresfix to 1024x1024 , that removes most of the grain and blurriness

a out of the box 1024x1024 seems to be worse

[dusk mica]

whats the time on a1111 for sdxl?

[upbeat summit]

right now it's a total mess with lots of experiments, broken routes and weird cabling :D. I will make a cleaner version soon that I can share, but this workflow by @stray mantle uses some of the stuff how I've build mine: #🌠｜show-and-tell message

[winter raptor]

I'll give that a try, and I'm waiting with bated breath for your release. Where should I keep an eye on for when you share it? And thanks for all the help!

[upbeat summit]

sure, anytime

[ionic dragon]

nvm, comfyui is always messy it meant to be messy
tq

[timid sonnet]

A1111 with SDXL still only making black images

I git from the SDXL branch, new folder

Using --no half vae thing

[upbeat summit]

I haven't set up SDXL on a1111 yet, but on github some people fixed it with --no-half-vae, but not everybody had success using it :/

[timid sonnet]

comfy works fine for me though

[winter raptor]

I have a feeling once I get used to comfy I'll love it

[autumn forum]

After having used comfyui for some time now, I have to say I like it better than auto1111. I have fun using it lol.

[winter raptor]

The ability to share workflows is awesome

[sharp robin]

Lot’s of customizability.

And it’s teaching me a little bit of Python

[lilac wren]

A1111 is working with SDXL now??

[rancid junco]

i assume the sdxl model is still not openly released?

[upbeat summit]

you should be able to easily sign up here: https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9 and get the pre-release version SDXL 0.9. the current official release date for SDXL 1.0 is July 18th.

[rancid junco]

i shall wait for the normal release, and hope we will be able to train or merge

[lilac wren]

This version need lot of vram i suppose

[winter raptor]

I've been fine with 6 gigs, but I'm used to sluggish speeds

[upbeat summit]

it works on many gpus using comfyui. I think the a1111 dev branch with SDXL support needs currently more resources.

[lilac wren]

How are we supposed to use this? (huggingface link)? With A1111?

[winter raptor]

Speaking of comfyui though, am I shutting it down wrong or something? Every time I close it, I can't reopen it without reinstalling it

[upbeat summit]

ComfyUI https://github.com/comfyanonymous/ComfyUI
sd.next https://github.com/vladmandic/automatic
dev branch of a1111 https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/dev
or online tools like the official discord bot, dreamstudio, clipdrop

[lilac wren]

Yes, I already use Comfy but with the leaked models that are ~6GB each (base and refiner), I don't understand. There is only the base link on the huggingface link (with leaked models, yeah i assume)

[static prawn]

which workflow are u using atm ?

[lilac wren]

mine

[static prawn]

i mean those json files for comfy 😄

[lilac wren]

talking to me?

[upbeat summit]

the huggingface URL is the official Stability AI account

[lilac wren]

I know, i just asking if i just need to load this model present in files, in Comfy?

[upbeat summit]

in comfy it just works out of the box

[lilac wren]

Like before with leaked ones

and, where is refiner ? i just found base model 13 GB

[upbeat summit]

I guess so. I would download the models from the original source. than you can be sure that it works

[upbeat summit]

https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-0.9

[lilac wren]

Yeah i'm talking about the files present here. There is a Base model (13 GB) but no Refiner one

[upbeat summit]

official SDXL 0.9 refiner download: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-0.9/resolve/main/sd_xl_refiner_0.9.safetensors

[lilac wren]

And, the Base model works with only 6GB of Vram, as @winter raptor said?? Or he was talking about leaked one (6GB file), because, a model of 13 GB should at least load 13 GB of Vram? No?

[upbeat summit]

I know that the official model works with many GPUs. 8gb and even lower afaik.

[winter raptor]

I'm using the official model

[lilac wren]

PErfect, did you saw a difference between leaked ones and officials??

[winter raptor]

I haven't personally used the leaked ones, so I couldn't tell you

[lilac wren]

1st gen looks ok 🙂

And i can confirm, with same seed, image is exactly the same

[timid sonnet]

Best output so far

[lilac wren]

I don't understand why if Huggingface released an official version of SDXL, it's still not working on A1111. Have you had any news about this?

[lilac wren]

REfiner from huggingface give weird results

[timid sonnet]

@upbeat summit using A1111 I merged Base + Refiner into a checkpoint file

.5 weight, no interpolation

[nimble heart]

dear god why

[timid sonnet]

and I'm running it as a secondary base into a refiner

[upbeat summit]

nice experiment :]

[nimble heart]

If it's somehow not completely broken I'd test it with a higher base:refiner ratio since the refiner is pretty specialist and doesnt seem to handle a lot of things well on its own

[nimble heart]

so refiner only really works well when using euler or ddim in the noise return method from my experience

[lilac wren]

noise return method?

[timid sonnet]

it works just fine

[upbeat summit]

great image!

[timid sonnet]

@nimble heart here's another one of my results

[nimble heart]

something like this

[lilac wren]

Actually i do :
x1 base > refiner
x1base > x1.5 Base
x1.5 base > Refine

I could use x1 refined > x1.5 base

[nimble heart]

Not sure what you're referring to here. I'm not talking about upscaling

[timid sonnet]

wow my output

its so weird

[lilac wren]

Yeah i see, you use steps to steps generations

[nimble heart]

yes with return leftover noise. the noise return method

that on ddim or euler are where the refiner works properly

[lilac wren]

I have used it in the past, I have not found any real change since I stopped, on the contrary, I feel like I have better results.

[timid sonnet]

@nimble heart if I had "add noise" to the second modified CKPT base, it creates literal noise

so I only select add leftover noise

[nimble heart]

left is improper noise return workflow using I believe dpmpp karras, right is ddim. Works very well when set up right. You can even see the seeds in the flowers.

also the right image has comfyui metadata so you can try it out if you want @lilac wren

[timid sonnet]

Ive used dpmpp karras, it has good results

[nimble heart]

I use dpmpp karras if I'm only using base

[upbeat summit]

it likes to do collages sometimes. you can also ask for it specifically.

[nimble heart]

if I'm doing part denoise in base and part in refiner then ddim + uniform is the best imo

[lilac wren]

sorry, what image?

[nimble heart]

dpmpp using noise return makes the refiner kinda janky as shown in my comparison

[nimble heart]

that old post of mine I replied to

[lilac wren]

i will search, ty

[nimble heart]

#✨｜sdxl message

notice when using the refiner properly with return noise on ddim the flowers look super clean for ai generated stuff.

should be able to drag it straight into comfyui from your browser

[timid sonnet]

@nimble heart here is dpmpp karras with noise added and leftover noise

[sharp robin]

The workflows have gotten complex.

[lilac wren]

Generating,
Personnally i don't like have lot of process and no preview, it take long time for a potential miss x)

[timid sonnet]

@upbeat summit I can turn my Ascore positive number WAY UP to 9 and it won't ruin the photoreal output

with my 3 pass method

[nimble heart]

that workflow shows previews just fine

if you're using mine

[lilac wren]

I just see a unique output

[nimble heart]

run comfyui with --preview-method latent2rgb and each sampler node will show a realtime preview

[lilac wren]

Ahhh dude, i was looking for this since long time x)

[nimble heart]

so you can watch it all the way from the first sample

[amber fulcrum]

does anyone has the workflow of img2img including both base and refiner?

[nimble heart]

so the karras scheduler has such a strong denoising it kinda minimizes the effect of the return noise method I've found, and all the dpm samplers make it artifacty as well

[upbeat summit]

yeah you can build very elaborate workflows especially with ComfyUI. in a couple of days I'll bet there will also be more simpler setups that work great too when everyone is experimenting.

[nimble heart]

try ddim or euler on all 3 return noise stages

[timid sonnet]

without using the add noise and return noise, it only gave a blurry unreadable output

I only use 2 refiners with return noise

[sharp robin]

Mone has gotten chaotic. Im 1/4 of the way to Beric.

I think

[nimble heart]

on something like that? I can't run it since I don't have your merged ckpt

[timid sonnet]

My FAVORITE generation using my 3 Pass modified base method

[nimble heart]

using dpmpp karras sorta stacks more work on the first pass in my testing

[upbeat summit]

do a clean up session, reorder, rewire, snap everything with shift and a day later you do the same since there is so much new stuff to try out 😄 but of course that takes all time away from making images

[sharp robin]

I think im having more fun getting it to work than actually genning

[upbeat summit]

yeah me too. I try to take turns.. one day making images, one day building the machine

[sharp robin]

Lmao. I gen 2 images and adjust one setting then everything messes up nd then im hacking away.

That’s my mess

[timid sonnet]

@snow trellis

merge base and refiner in A1111 then add in comfy

[sharp robin]

Did u mean to ping me?

[nimble heart]

my version with the same settings

[sharp robin]

Ive got a better merger thingy that does math. Ill so it tomorrow

[lilac wren]

@nimble heart with your workflow

With mine, 16/9 and upscaled

[timid sonnet]

@sharp robin Oh okay I'm getting good results with Base + Refiner merged at 0.5

I'm using the modified base in the second pass before third final

[nimble heart]

if you're using my upscaled one it should do 16:9 fine as well

just set all the sizes correctly

[sharp robin]

Interesting. The recommended % is 20 or 30 of refiner ill do 50 20 and 30%

[timid sonnet]

Oh I didn teven know people were doing that. I just thought to try it when I was trying to get A1111 to work

but yeah I used 0.5 I thhink with no interpolatioin

sorry lagging

[sharp robin]

Auto works good but slow

[nimble heart]

yea you can also have the refiner use 1 step less on max to reduce it's [timescale?] which add lots of small details

[sharp robin]

Just download the branch in separate folder

[timid sonnet]

@sharp robin

[sharp robin]

That’s really nice!

[timid sonnet]

Thanks! thats using the Merged ckpt on 2nd pass (2/3)

[timid sonnet]

what do you mean 1 step less like 30 total steps, with 29 on the max input?

[nimble heart]

using your earlier prompt,
left is base total 30, start 0 end 20, refiner total 30 start 20 end 30.
Right is the same except refiner's total and end is 29.
The shrinking of the time the refiner operates on makes it like cram more details in faster which can be nice for photoreal. Both images have embedded workflows so you can drag them into comfy and check it out

[sharp robin]

Base does 30 next node does 29.

[sharp robin]

Or ex 50 then 49

[nimble heart]

yea also it's not really desirable for flatter art I've found. I have the steps matched by default and I just manually turn down the refiner by 1 if I think it needs it.

[west breach]

a new lora?

[lilac wren]

deep Throat lora, dirty

[hollow halo]

Locon this time, but same idea 🙂

[timid sonnet]

so do I got 1/3 : 0 - 30, 2/3 : 13 - 29, 3/3 : 20 - 28?

[sharp robin]

Yeah makes it try and look human/realistic

[nimble heart]

No idea if that's correct since you use that 3 pass thing. Open the right image in comyui it'll show it pretty clearly

[nimble heart]

sometimes even then it can overcook things. The magically added details get distorted fast.

[timid sonnet]

@nimble heart like this

[nimble heart]

probably not.

[nimble heart]

I think you're trying to do something like this

[timid sonnet]

it looks midjourney'd out @worn vale

[nimble heart]

give this a try

[hollow halo]

lol think you tagged the wrong person

[upbeat summit]

very cool contraptions you've made there! I worked on stargates / portals earlier too

[timid sonnet]

@hollow halo i did lol

[glass acorn]

please someone enlighten me. I want to learn everything about how SD works. longform youtube content recommendations? While i'm at it.. i have no clue what the conditioner is but why does it have 2 text inputs L and G instead of just one?

[nimble heart]

should look more like this.

[timid sonnet]

@glass acorn maybe start with sd 1.5 and A1111

[sour obsidian]

depending on how deep you want to go I recommend this if you want to leap straight into the deep end: https://course.fast.ai/Lessons/lesson1.html
python experience recomended. Pt 2 basically breaks down the entire old sd model

[glass acorn]

is all this text stuff optional? does the series of videos pretty much cover it?

i absorb info like a sponge

[rustic garnet]

because they could not decide for one

[lilac wren]

YOu should absorb this : #1072220168534642768

[glass acorn]

i know how to use them. im ready to start deconstructing them if you know what i mean

[timid sonnet]

@upbeat summit

[glass acorn]

thanks. maybe i wasnt clear. im looking specificaly for youtube or audiobook type content

[rustic garnet]

the text encoders are freezed. They were trained independently from SD and are not changed during SD training.
So it matters which text encoder you use as input and it seemed that the Clip-g was better in prompt understanding while clip-l was better in style understanding. Instead of choosing one they just took both

[timid sonnet]

@upbeat summit the image is less important, look at the +prompt I used

[upbeat summit]

uhh - interesting

[timid sonnet]

It actually generated based on what I was announcing to it

We should consider using words like radius

[lilac wren]

https://letmegooglethat.com/?q=begin+with+stable+diffusion

[glass acorn]

dude. youre obnoxious

[upbeat summit]

yeah. SDXL impressed me a couple of times already how well it can follow a description, effects and storytelling

[glass acorn]

just dont help if youre not going to help

[timid sonnet]

along with starting out with an initial shape, then describing what you want around or in the shape, etc.

[glass acorn]

its just toxic

[timid sonnet]

@glass acorn just block him

[lilac wren]

Sorry, I had to put it

[glass acorn]

you didnt

[rustic garnet]

I would say look into the papers 😅 but there are also several blog posts about how diffusion works. For sdxl you can look into the source code

[glass acorn]

actually the one linked before has a whole series on stable diffusion so i think im golden. thanks everyone

[rustic garnet]

maybe ai find the motivation to write a blog post about how SDXL works..

[glass acorn]

https://www.youtube.com/playlist?list=PLfYUBJiXbdtRUvTUYpLdfHHp9a58nWVXP

[rustic garnet]

lol, ai find Motivation is also great. But I meant "I"

[dense chasm]

i set the prompt as below:cinematic movie extreme close-up still of an epic scene of a blackman basketball player Michael Jordan in the summer at morning, centered, looking into the camera, fog atmosphere, volumetrics, photorealistic, from a western movie, analog, very grainy, film still, kodak ektar, fujifilm fuji, kodak gold, cinestill 800t, kodak portra, photo taken by thomas hoepker

[lilac wren]

@glass acorn check your pm

[dense chasm]

but why the image generation of michael jordan looks different with MJ's photo.not the same person

[dense chasm]

how to generate the real celebrity with proper prompts?

[sour obsidian]

the video stuff is pretty great

[glass acorn]

yeah i think this is exactly what i was looking for. thanks so much

[timid sonnet]

+Prompt
A photograph of Michael Jordan five-feet off the ground mid-dunk in his famous pose

Supporting +prompt
chicago bulls jersey, nba stadium, dunk of the year

[west breach]

trippy

[dense chasm]

i just wanna to make a dialogue video between Mj and Kobe Brant,just wanna to make them close up looking staring at the camera individually,how can i do with the proper prompts?

[timid sonnet]

@west breach How do you make such a wide image like that

[sour obsidian]

[west breach]

1600x640

[timid sonnet]

use standard SD, find a model that has celebrities like Realisticv20 and use control net

[west breach]

this is sdxl saying, 'wtf are you on about'

[sour obsidian]

https://tenor.com/view/head-on-gif-4936632

[dense chasm]

u mean sdxl 0.9 is not the appropriate models generating celebrity images?

[timid sonnet]

@dense chasm for what you want just use a 1.5 or 2.1 model that was trained on celebrities

with control net

[glass acorn]

This might be a bit simple for someone who watched that entire course but this was a nice high level video i enjoyed https://www.youtube.com/watch?v=1CIpzeNxIhU

[sour obsidian]

ah yea love that vid 😄 such a good overview

[timid sonnet]

d

[sour obsidian]

love the computerphile group

[glass acorn]

they rock

alright its bedtime. see you when im an expert or something broken

[timid sonnet]

gn

[upbeat summit]

such a great image and fidelity

[west breach]

are you doing another base + refiner on the 2x, or just one of the checkpoints?

[upbeat summit]

I worked on a couple of films featuring Beemers / BMWs. Totally dig the imagery.

[lilac wren]

could you share us your workflow?

[sour obsidian]

1.0+refiner there,

[west breach]

1.0 -> upscale -> refiner ??

[sour obsidian]

pretty much lol

[timid sonnet]

You're too good

[upbeat summit]

uhhh you made it

[timid sonnet]

This one came out alright

[hollow halo]

Nah it's my Locon doing the heavy lifting lol, my prompting skills need work

[upbeat summit]

great coherence!

[languid swift]

how do yo use de sdxl 1.0?

[sour obsidian]

I made it haha (along with the rest of the incredible team )

[timid sonnet]

WoWWOWOWOOWOWOWOW

@hollow halo

[languid swift]

omg, awesome, which is the prompt and how did you use it?

[timid sonnet]

Your image inspired my new prompt

[hollow halo]

Nice!

[sour obsidian]

oh gosh the bike, something about a bike on a road with mountains in the back, fujifilm? haha wasnt many words

[languid swift]

for default is set de sdxl 1.0 or you have to write something special?

the*

[sour obsidian]

just good sampling, not doing any wacky tricks there haha

[languid swift]

you're getting images much betters than mines, hahah I will review my prompts thx

[upbeat summit]

you've increased your upscale quality and resolution 🙂

[upbeat summit]

nice - huge upscale as well