✨｜sdxl

[high skiff]

ew, but look at that minimum :p

what CPU?

[nimble heart]

I have no clue what that is. I never saw the counter go below like 190

[high skiff]

oh! I wasn't running minimum

I had maxxed out crowd density

dies

[nimble heart]

shit I might have too

[high skiff]

I closed cyberpunk already

[nimble heart]

a pretty good one

[high skiff]

what CPU do you have?

ahhh yeah, fair haha

the single core is another step up from my new one

[nimble heart]

I was coming from a 6700k so its smooth

[high skiff]

I came from the 6700k's bastard brother, the 7700k lol

[shy kelp]

I play cyberpunk on a mac w three translation layers, windows to macOS, x86 to arm, and directx12 to metal, and still get higher min fps if running at lowest 😭

[high skiff]

oh wow, the single core isn't nearly as much better than I thought

the multi is almost double tho haha

[lilac wren]

1 year ago, I had my 4670k OC at 4.5 Ghz, all the time 100% on Ableton, a nightmare. since I have a 3700x

[high skiff]

my 7700k always would get trashed in FL studio, and it has an even more optimized ASIO than abelton, so I bet

[upbeat summit]

hey just have to freeze every track 🥴

[delicate grotto]

When are we going to change x86

[high skiff]

my audio work was so intense that it would take upwards of several minutes to freeze some tracks

[lilac wren]

Ah, misunderstood, yeah, freeze 40 tracks every time...

Problem were big VSTs

on a shit old CPU

[upbeat summit]

but you wanted to change the attack in that one synth...

[high skiff]

oh my god

I just realized

I did lose my crystaline rack

oh my god

[shy kelp]

Dam, prompt, settings?

[lilac wren]

That's why i clone my drives with Macrium Reflect, from old ones to Samsung 970 Evo/870 Evo.
Fresh meat

[nimble heart]

they must measure it weirdly cause this is max settings + full RT. My min is way lower and my max is wayyy higher than I observed on 2 separate fps counters. Maybe they start counting before the scene finishes loading or something

[upbeat summit]

so good in full res... those palette knife paint strokes

[high skiff]

this main bass I made here took 100% of my CPU for like 3 minutes just to render cause of all of the effects and automations

hell, pitchmap on its own can nearly pin an m1 mac lol

[high skiff]

very weird

[upbeat summit]

are we going to post our music before the mods get here?

[high skiff]

its dead hours, our convos are doing them a favor towards activity lol

[west breach]

positive: "apache helicopter, oil painting, thick, textured brush strokes, impasto, glazing, visible canvas texture, blending, rich colors" negative: "watermark, Smooth, flat surfaces, quick drying, transparent colors"

[shy kelp]

Nice, thanks

[nimble heart]

my homemade music player is small enough I can upload the whole thing to discord

[high skiff]

I love seeing the 30 minute brainstorm version of something compared to the several hour rendition of the same thing lol

music to tickle your GPU's core voltage to

[nimble heart]

we should just migrate progressively further off topic until Fruit or someone comes back in the morning

[high skiff]

lol

[high skiff]

right lol

without off topic, there is no activity :p

this is what the death of a CPU's single core sounds like

Could easily flatline a 13900ks with the amount of processing I dropped on this bad boy lol

4 synth layers, 3 harmonic re-pitchers, 41 or so miscellaneous effects, 3 different notes, and 16 real time automations running at 320kbps resolution lol

[nimble heart]

train an AI to use send midi input into fruity loops for a live DJ session

[high skiff]

that would be amazing

[west breach]

my new cpu ❤️

[high skiff]

SDJ

[high skiff]

okaayyy flexxxxxx lmao

[nimble heart]

tryna show me up or something with your lower boost speed having ass

[high skiff]

||Inb4 I return my 12600k for a 13900ks just to flex||

[high skiff]

lmfaoooo

[nimble heart]

nah screw the 13900ks. My 7900X hits 5.9 GHz on single core loads

[high skiff]

the shade

[nimble heart]

simple undervolt and you're good to go

[high skiff]

13900ks hits 6.0ghz on single core :p

[nimble heart]

I don't think that's stock

[high skiff]

for the KS it is

[delicate grotto]

Is intel going to take the lead back?

[high skiff]

not likely, AMD and Intel are both doing really good ATM

[nimble heart]

ohhh the KS

[high skiff]

Intel is dominating the budget space tho

[delicate grotto]

Yep

[nimble heart]

I was thinking the K. Didn't realize the KS was out already

[high skiff]

i got my 12600k for $170

[delicate grotto]

Its the first year that intel is actually a decent offer

[high skiff]

faster than the 7600x and its like $70 less, I think?

[delicate grotto]

So far i saw the ryzen 3 and up as superior

[nimble heart]

how do you see receipts on amazon

nvm

yea I paid $440 for mine

[high skiff]

I skimped on my CPU cause I don't need it as bad as a 3090 haha

[west breach]

I think SDXL needs to go back to kindergarden and learn how to use playdoh. That's supposed to be an apache helicopter

[high skiff]

also, a 12600k is still pretty damn fast

[nimble heart]

Rust code takes ages to compile so I spent the extra money

[high skiff]

oh for sure, makes 100% sense

[nimble heart]

males 100%

[high skiff]

all I need my CPU for is single core, and not a pathetic multicore haha

[high skiff]

sorry

had to

love that emoji lmao

[nimble heart]

sus

[high skiff]

its uhhh

[nimble heart]

do you have a sergal one

[high skiff]

from a very NSFW furry server lmao

[high skiff]

I doooo!

hell, my old brand name used to be Scyth Sergal lmao

traded it in for Sytan

[nimble heart]

Quite the rebrand

[high skiff]

yeah haha

I went from a love for sergals, to a love for Satan lol

[high skiff]

its a slippery slope

[nimble heart]

my old screen names were too cringe to share

[nimble heart]

Imma be real with you chief, that's more of a cliff as far as general perception goes lol

[high skiff]

lmao, fair enough haha

[indigo carbon]

So, what's the meta for SDXL? Vlad's UI? Or ComfyUI? I heard a1111 is not supporting use of refiner and base simultaneously

[nimble heart]

Comfy with a good node setup

[high skiff]

Comdy UI allows you to use better workflows

[soft zealot]

@visual glade FYI, going back to my weird zoom issue from last week (the one where zooming in comfy in W11 was fine but the same PC in W10 it was wildly all over the place) .

Bought a new mouse at the weekend and that works fine in both W10 & W11.

Oddly the original mouse now alos works fine in both W10 & W11.

I can only imagine there was some sort of driver refresh when I added the new mouse

[nimble heart]

No one here seems to use vlad's so I cant say how that does

[high skiff]

Vlad doesn't use Mixed diffusion (which BTW, mcMonkey told me my way of using it is actually the best way to use SDXL and they "planned to use it like that from day one")

yet they had no info about it lol

[nimble heart]

Mixed diffusion
is that the official name for the return noise method?

[high skiff]

and I had staff in my DM's really interested in what I was doing

[high skiff]

thats what @uneven dove coined for us when we passed it off to diffusers, so is uppose so

or wait

I think I may be using the wrong name

just a sec

[nimble heart]

partial diffusion?

I heard that earlier

[indigo carbon]

Ah, nice! So, you made an efficient SDXL workflow?

[high skiff]

found it

Partial Diffusion

my bad

Thats the name its coined under

[soft zealot]

A1111 also doesnt split the positive promt into what @high skiff has tagged "Linguistic Positive" & "Supporting Terms"

[indigo carbon]

I see, how many it/s is this in comparison to 1.5 models?

[high skiff]

Yeah, instead of making an image from the base then img2img fixing it, my workflow does a portion of the image with base, sends the incomplete image to the refiner, and goes from there

[nimble heart]

Aight maybe that'll intimidate comfyui newbies a bit less then telling them to "use the noise return method"

[high skiff]

That as well, which can give huge quality improvements if used properly

[high skiff]

Thats why I have .jsons on my github with it all set up with documentation and an easy to follow naming scheme

[nimble heart]

way slower in 512, possibly somehow faster @ high >1024 res according to people that arent me

[indigo carbon]

Odd

[high skiff]

yeah, i have found that SDXL is pretty comparable to SD 1.5 at 1024x, and shockingly faster at 2048x

[nimble heart]

its slower but it scales better

[high skiff]

something to do with the way its optimized

also, it uses way less VRAM additional to increase res

from 1024x to 2048x only uses like 1GB more VRAM to gen

[nimble heart]

the first unet doesnt blow up into infinity when you crank the res or something like that

its more comparable to batch sizes in 1.5

VAE still 100% blows up though

I've got like minute long decodes @ 2k^2

[high skiff]

when I get my 3090 tomorrow, I will 100% be testing LoRA's for SDXL immediately

[indigo carbon]

Huh, is it working with all optimizations?(CUDA11.8, CUDnn6.8, Xformers, sdp attention, etc..)

[nimble heart]

yes

[high skiff]

not sure if I shared my early test for my new generation Na'vi LoRA I made with 1.5 or not, but the SDXL results should be even better

results like this from just a frction of my data set for the SDXL LoRA at only 1/4th the res

[indigo carbon]

Is this going to be implemented in something like A1111? Not everyone are like us and know how to mess with node systems

[high skiff]

its a hueg step up over my old one, which looked better, but was hard locked to guys that looked identical, with no way to add clothing to them, and also nos upport for females

[soft zealot]

@high skiff used your workflow as the starting point for the monstoristy I have now lol (a lot of the additinal steps are slected using "jumpers" so they can be added in as required)

[high skiff]

your node lines are gorgeous

we need straight line pinned node connectors in base comfffyyyy

[soft zealot]

its a .js file I cound somewhere, one moment

[high skiff]

I unfortunately can't use it ._.

not for the offical release at least

Comfy told me my release on the wiki can only use stock features, which I totally understand

[winter raptor]

Dear God what is this lol

[high skiff]

heaven is what it is

[ionic dragon]

How comparable is SDXL with MJ?
which ver of MJ does SDXL correspond with?

[soft zealot]

save the attached to \ComfyUI\web\extensions and rename to .js
0 is straight lines
1 is slightly curvy
2 is spaghetti

justresfresh in browser

[shy kelp]

Whatever this is, it's sorta cool tbh

[ionic dragon]

how the fuck are you able to make the workflow clean?

[ionic dragon]

@high skiff

[shy kelp]

It looks unlike anything stablediffusion normally generates

[soft zealot]

#✨｜sdxl message

[high skiff]

too hard to compare honestly

[ionic dragon]

someone was saying it compares with v5

is it v5?

[high skiff]

no

[winter raptor]

@soft zealot so what are the “jumpers”? What do they do?

[high skiff]

its better at some things, worse at others

like I said, too soon to compare

[ionic dragon]

a rough estimate

[ionic dragon]

i am applying for some ai intern, they said they were looking for MJ artists, so i wanted to convince that SDXL is better

[high skiff]

its better than MJv4 for sure, but its not as good as MJV5 at some things

[high skiff]

oh god, please don't its likely a scam

what is the company's name?

[soft zealot]

enable me to turn on/off functionalitty. Basically I split an output using 2 reroute nodes and then "jump" them together to enable or remove the link to disable. Bit like old school dial up modems

Might be easier to see the workflow directly

[upbeat summit]

thanks for sharing it 🙂

[elder dagger]

hello there, do you people could inform me about the specs required to run SDXL models ? like especially for ram and is there people that make it works on AMD GPU ?

[ionic dragon]

he is an YT'er

[high skiff]

system RAM is likelt 32GB minimum for a smooth experience (tho some have said as little as 16 works), and a 6GB NVIDIA GPU is as low as I would recommend

[high skiff]

which one?

[ionic dragon]

you wouldn't know him

lemme share

[winter raptor]

Thanks, I’ll take a look in the morning when I’m back on my computer. I appreciate the help, I’m pretty new to all of this so I’m trying to pick up as much knowledge as I can

[high skiff]

I'm telling you, its likely a scam

I got scammed to hell and back with an AI art startup like that

[elder dagger]

yeah I use 16gb ram with 16gb vram AMD card and SDXL models make my ubuntu crash

[upbeat summit]

cool solution 🙂

[soft zealot]

works for me 🙂

[upbeat summit]

I need to do some fine-tuning to the snapping

[ionic dragon]

exactly at this timestamp
https://youtu.be/wKRAjI5Kf_c?t=2015

[soft zealot]

yes I know you can "mute" nodes but tI just find this easier

[high skiff]

yeah, i would not trust him lmao

[ionic dragon]

he is just looking to create yt thumbnails

[soft zealot]

just the thinbnail feels like a scam

[high skiff]

I am advising you don't do it, it looks like a scam

[ionic dragon]

dude, he has around 2million subs on his main channel

[high skiff]

right

[high skiff]

yeah, and LTT had 18mil when they pumped a crypto scam

[ionic dragon]

he is one of the most famous guys in my country

[ionic dragon]

his channel was hacked

[high skiff]

do what you feel like man, but I do not trust him at all

[ionic dragon]

he is trust worthy

[high skiff]

98% of all AI artist jobs are a scam

[ionic dragon]

anyways, i aint sending money, i will send send my ai art

[high skiff]

best of luck to you, you have multiple people here suggesting against this

[ionic dragon]

he is not offering a job, he is just looking for someone to generate art like an intern

[high skiff]

and I worked for a company that has over 400 sales and over 100 5 star reviews on etsy, and they ended up fucking me over, so don't trust anybody man

as somebody who got got by one of these pump and dump schemes, I recommend against it

[nimble heart]

alright some actual hard numbers this time using a 7900 XTX on pytorch 2.1.0 nightly rocm 5.5:

1.5 @ 512: 17.49 it/s
1.5 @ 768: 5.72 it/s
1.5 @ 1024: 2.34 it/s
1.5 @ 1536: ≈0.55 it/s
1.5 @ 2048: ≈0.19 it/s

XL @ 512: 12.36 it/s
XL @ 768: 6.33 it/s
XL @ 1024: 3.13 it/s
XL @ 1536: 1.07 it/s
XL @ 2048: ≈0.48 it/s

Also note SD 1.5 @ 2048 my system VRAM usage was @ 21.4 GB, while SDXL was only 14.2 GB.

This is only SDXL base. The refiner runs slightly slower.

[hard sun]

You should take a look at some Coffeezilla coverage of influencer scams - having a lot of subscribers doesn't make you trustworthy at all, maybe even to the contrary

[ionic dragon]

i totally understand but he's not gonna do scams, like why would he run a scam and kill his reputation, he is a very famous guy, he even has podcasts with celebs and billionaires

[high skiff]

I guess famous people never run scame then?

[ionic dragon]

what scam did they commit?
like did they charge money?

[high skiff]

I have said my peice, I wish you best, but I will offer a strong I told you so if this goes bad

[ionic dragon]

no, i am saying that he is a guy he doesn't run a scam

[ionic dragon]

even if its a scam, nothing would happen, as I am just sharing some of my best generations

definitely not gonna pay

[high skiff]

they tried to lie between me and a client (which is fraud), they pose as a non AI store (which is fraud), they witheld my payment (which is illegal), and they overall pay absoluetly disgustingly bad wages

you don't have to pay to get scammed dude

wait a sec

when you say intern, you mean for free?

cause if so, then mannnn, fuck that lmao

[ionic dragon]

not really, he might pay a bit

[hard sun]

"Might"

Why even bother

[high skiff]

at that point thats not even a scam, thats just bad judgement on your end

[ionic dragon]

dude, i am saying he is a very trustworthy guy

[high skiff]

you don't know himmm

[ionic dragon]

he would pay

[high skiff]

did he say he was going to?

[ionic dragon]

i have been watching his videos from the past 2 years

[ionic dragon]

yeah

[high skiff]

ok, and I have been watching youtubers for over a decade, so that means I know everything about them and how they are not gonna scam me

come to think of it, I am sure people go out and say "I am not trust worthy, I will scam you" before they do

They TOTALLY never pose as trust worthy

look man, we are trying to look out for you, cause this seems disgustingly sus

some random youtuber with millions of subs reaches out to some rando to do free work for him?

[hard sun]

Look at the story of Logan Paul's CryptoZoo project. He has 23M subscribers yet screwed over both his fans and the developers that worked for him

[high skiff]

hell, I did a background check on the company that hired me,a nd I STILL got scammed lmao

[high skiff]

exactly

famous people do scams alllll the damn time lmao

[hard sun]

They don't know you and they don't give a damn about you

[supple knot]

Coffeezilla did a story about that scam

[high skiff]

not surprising, that man is over EVERYTHING ALL OF THE TIME lmao

[wooden pond]

SDXL officially realeases tomorrow yes?

[high skiff]

day after

or wait

the 18th

might be later for you depending on time zone

oh shit, nevermind

its 2AM lol

[wooden pond]

OK thanks 🙂

[high skiff]

so yes, tomorrow

as in the 18th

[supple knot]

will it be different or just official?

[high skiff]

(assuming they didn't lie and overhype it, which honestly, yet to be seen)

[high skiff]

different

we don't know just how different yet

[nimble heart]

soon™️

[high skiff]

I mean, its Emad we're talking about

I trust that youtube asking for free MJ images more than Emad lol

[nimble heart]

I never mentioned CPU in that bench?

[delicate grotto]

ah yeah, the xtx is a gpu

[nimble heart]

it keeps leaning towards SDXXL and I'm starting to think it's a feature not a bug

[delicate grotto]

my mistake

[high skiff]

lol

[nimble heart]

All good

[high skiff]

ahhh, I am getting butterflies in my stomach thinking about my 3090 :>>>>

[delicate grotto]

when is it arriving?

[high skiff]

later today

[upbeat summit]

my very first SDXL test image 😄

[high skiff]

it was gonna get here on launch day, then it was gonna be the day after launch, but I got an update saying its a day before launch!

[delicate grotto]

for a goddamn reason i also started with a clown

[nimble heart]

well I cranked the CFG and it fixed the sign but now I feel very uncomfortable

[high skiff]

lmao

[nimble heart]

moooom its doing it again

[delicate grotto]

how much steps do you do on the refiner?

[high skiff]

this was the original prompt for my SDXL beta workflow json lol

[delicate grotto]

and do you do you send to the refiner with noise or without?

[supple knot]

so 3090 has 24gb vram what are you running now? @high skiff

[high skiff]

3080

10GB is just not enough

and a month ago I had a 3060ti lol

its actually on my desk right now lol

[nimble heart]

ComfyUI?

[supple knot]

wow 3090 should cut thru this like hot knife thru butter

[high skiff]

yessir

I don't care about speed

my 3080 is already really fast at SDXL

I need more VRAM

[nimble heart]

yea selling a $600 gpu with 10 gigs of vram is criminal

[high skiff]

I wanna do SDXL LoRA's and finetunes, and i need the beef for that

[dusk mica]

1.0 drops tomorrow right? Which time?

[high skiff]

I got it for $320 basically brand new to be fair, but yeah

[high skiff]

its supposed to drop on the 18th, no time confirmed yet

It might not even drop, but we will see

[nimble heart]

gtx 1080 had 11 gigabytes so I'd still say it's sufficiently bad

[boreal bough]

just realized... but with enough psu, you can 34gb vram XD

[high skiff]

true lol

[high skiff]

yessir lol

I was thinking about it

still kinda am

cause I have my old 650 watt as well

[supple knot]

have you looked into water cool mod for gpu? it does cover vram only the main chip.not sure if I should do this

[boreal bough]

I'm jelly. You can both train, and test the resulting lora properly...

since comfy can just run on the second gpu

[high skiff]

could string another connector off the 850 topower one for the 3080 and then use the 650 for the rest of the 3080 power

[nimble heart]

having the 2nd gpu on a 2nd psu 100% works. I did that a lot at work. You get to flick two switches to turn on the PC like an airline pilot

[high skiff]

if I can find a way to pad out/make some money, iw ould keep both

as it is now, I bougth the 3090 under the pretense that I would be selling my 3060ti and 3080 to pay it off

[nimble heart]

24 is more than 90% of people running SD will have so I think it'll be fine

[high skiff]

yeah

[nimble heart]

people try to cram 1.5 into the steam deck of all things

[high skiff]

hell, people run 1.5 on phones now lol

[supple knot]

these gpu price change so much .hope by dec there even lower

[high skiff]

TBF, some android phones have 24GB shared RAM now

so thats better than most consumer GPU's lmao

you know

[nimble heart]

my gpu now is $100 less than it was back in february

[high skiff]

if I can get a good bite on my 3060ti (cause its a special one), maybe I could squeeze out keeping the 3080

[nimble heart]

the 20 gig version is down to $790 damn

[high skiff]

or I can just ask my mom to float my funds for a couple months, which she usually does cause I don't aks for it often

[nimble heart]

would pooling the two gpus also increase speed or no

[high skiff]

no

I wouldn't pool

I would use 3090 for training, and 3080 for generating

[indigo carbon]

I doubt you can run SDXL on A PHONE

[high skiff]

people have

[dusk mica]

Fr?

[high skiff]

oh, SDXL, IDK

but yeah, SD 1.0 and 2.1 just fine

[indigo carbon]

What, with 4bit degradation? Heck no

[nimble heart]

I wouldn't bother then tbh. 24's more than enough for everything except maybe absolutely crazy finetunes

[high skiff]

like I said, it would be to run one training all the time, and the other one testing the training as they are coming in

[nimble heart]

eh

[high skiff]

could get 15 minutes into a 10 hour training session and notice "hey, my settings are fucked"

[nimble heart]

I'd just pause training for a hot sec when I wanna use the card for other stuff but that's just me

[high skiff]

you can't pause training

not that I know of

and I suppose I could use CPU to test LoRA's as well... hmmm

[indigo carbon]

Does Vlad's UI also support SDXL with all of the optimizations?

[high skiff]

not all the optimizations

[nimble heart]

If his uses Diffusers wouldn't it have partial diffusion now too?

[high skiff]

won't find my partial diffusion or split text encoder/A scores

[high skiff]

oh your know what, maybe in that case

I didn't think of that

[nimble heart]

or does he use a pinned version that doesn't have it yet

[high skiff]

I am not too sure how they are implementing it

thats @uneven dove 's info to share, as he is the one that made the new code to implement it

we make a good team, I get the ideas, he gets the code haha

[indigo carbon]

Well, A1111 doesn't seem to want to support running SDXL+refiner, so Vlad is probably the only option for people that don't like ComfyUI

[ionic dragon]

can I post images of Jinping as Winnie the pooh here?

[high skiff]

I would love to see lol

just make sure you don't get shot lol

[nimble heart]

@indigo carbon there's some good comfyui workflow examples on comfy and sytan's githubs to get started fast. I have a hi res fix version too that I use personally if you want

[dusk mica]

Whats the best comfyui workflow rn?

[high skiff]

would love to see your high res fix implementation

would be very interested in combining it into the base workflow

[high skiff]

mine is one of the best for simple use

[indigo carbon]

Nah, I will probably switch to ComfyUI, I'm asking for people that don't like ComfyUI

[ionic dragon]

are you planning to release your OG workflow?

[high skiff]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[dusk mica]

Thx

[high skiff]

OG? no, its changed and gotten a lot better now

[nimble heart]

have at it. Nothing revolutionary. I just save steps by running the low res on dpm++ karras since it converges fastest so I can stop it earlier, then it's just regular sdxl stuff with the correct values for width/height on the encoders.

[supple knot]

ComfyUI was made for SD? or does comfy have other uses?

[nimble heart]

it's set for 1.5x cause 2x takes like 90 seconds per image on my card

[high skiff]

peeping

[nimble heart]

I use it mostly @ 1.5k^2 and 1920x1080

[ionic dragon]

this is the new OG or demo?

[high skiff]

thats a beta demo

I will hopefully have a full basic release done by 1.0 release

[ionic dragon]

oh, the one you sent a few days ago?

[high skiff]

yeah, i have not changed the one on gitbug yet

[west breach]

vector art

[high skiff]

I may mangle this into something easily compatible with my workflow, if thats alright with you?

I am testing it right now

wait

how

how did you

[nimble heart]

Go for it. If it winds up in git just @ me in the commit or something.

[high skiff]

wha

[nimble heart]

I have access to powers many might consider unnatural

[high skiff]

AHA

YOU CAN MAKE THEM HORIZONTAL

I HAD NO IDEA

[nimble heart]

yea

[high skiff]

OoOoOoO

[nimble heart]

you can't make them go backwards though

[high skiff]

that should help the cleanliness of my workflow

[nimble heart]

so that's why I have them doubled up in places

[high skiff]

I am testing to see how well this works, then I can see if its worth adapting into mixed diffusion or not

also, holy shit this new CPU flies through the text encoders

[nimble heart]

"mixed diffusion"?

I already use partial diffusion in the high res pass

[high skiff]

damn it

partial

ok, yeah, i do see that now

you are just using one TE instead of 2, which is an easy edit

[nimble heart]

I tried my damndest to return noise on low res but latent upscale always mangles it beyond what the refiner knows how to fix

[high skiff]

ok hmm

I see what you are doing, but I have a proposition for a potential fix

[nimble heart]

thought that was before I was on ddim maybe it's different for that

[high skiff]

or improvement

[supple knot]

in ComfyUI on KSampler nodes How many step do you set it for?

[high skiff]

ok yeah, i think I might have some tricks to get this working better @nimble heart

[supple knot]

Im dong 30 then 50 but cant tell what its doing

[nimble heart]

does it involve return noise on the first base?

[high skiff]

testing first to see

[nimble heart]

it'd be badass if it could work

im trying it again now cause you've inspired me

[upbeat summit]

of course this would need some manual work, but here is a quick vector trace in SVG format using adobe illustrator (reduced to 16 colors)

[nimble heart]

also interestingly the area upscale has no artifacts its just blurry. Refiner doesnt like it though so

or was it bilinear

[west breach]

wow awesome, not a whole lot of difference between sdxl output and what you did there

[nimble heart]

yeaaa refiner no likey

[upbeat summit]

not much 🙂

[supple knot]

That roll back driver worked for me. But my max outputs are 1024 x 576.without crashing from enough vram

Kurt Cobain using SD 1.0 from a cell phone 🙂

[high skiff]

@nimble heartAha!

you were using partial diffusion wrong

you were time splitting from base to base, it needs to go base to refiner, I have a new idea to potentially get better results

[nimble heart]

it does though?

base -> base is img2img not partial diffusion

first base doesnt return noise

it just ends early to save time

2nd base -> refiner returns noise properly partial diffusion

[high skiff]

I do get what you are saying, but I am getting considerably better results at the moment by doing base into refiner into refiner

[nimble heart]

You can't latent upscale with returned noise or else it breaks partial diffusion (afaik), so you have to run a 2nd base on top as i2i to re-add the noise before going to partial diffusion

So you're just img2img-ing directly into refiner?

[high skiff]

kinda sorta

testing more things

my hands and brain move faster than I can formulate what I am doing lol

what we need is a hacked version of ultimate upscale

if I could learn how to hack it, I could get the best results possible

trying some things, some things I am trying

[nimble heart]

feel like I'm the only person never used ultimate upscale

[high skiff]

very few people use it with SDXL, but it does get the best results from what I have seen

but, I know how to make it better

I just don't know how to code well enough

[nimble heart]

even in 1.5 on my 1070 I just cranked the img2img res and ate popcorn while it rendered for 10 minutes

[nimble heart]

it's python you can just kinda fake it till you make it I believe in you

[shy kelp]

What's ultimate upscale?

[high skiff]

maybe I can talk pseudo into doing this for me, cause I guarantee that if I can get this to work, it will be the best upscale option at the moment

[nimble heart]

i img2img'd using only the refiner @ 50% and it sure did dondely refined it hard

[high skiff]

grrr

[upbeat summit]

refiner fixes some anomalies for me, but very fine details from the 1st base pass disappear as well. I'm still not sure what the best settings are. right now I'm doing base -> refiner -> refiner - I guess I don't have the right settings yet

[high skiff]

I wish I knew enough python

[nimble heart]

just write some things for the sake of writing them it'll come naturally

[high skiff]

I need to reimplement partial diffusion based into the code

[nimble heart]

copy past

[high skiff]

no no, my implementation isn't in the code

its through nodes

I would need to figure out how to code it in the first place

[nimble heart]

I mean from diffusers or something

[supple knot]

refiner refiner means you have a second refiner node?

[high skiff]

diffusers works very differently to comfy ._.

[nimble heart]

shh

copy past

[high skiff]

I am big brain storming right now

[nimble heart]

its python it'll work somehow

[high skiff]

thinkin up a STORM

[nimble heart]

it at least wont crash

[supple knot]

can Chat

gpt write this code you need

[high skiff]

not likely, it would need a lot of context as to how comfy handles all of his diffusion based code

[nimble heart]

also check your weird ass double refiner upscale on text. My first few upscaler workflows garbled text hard. Text has to be gotten right in the low res base then not ruined in upscale

[high skiff]

I need it to early terminate

[trim snow]

Hey how did you use image2image? Is it possible to use it with one of the bots and how? Sorry I’m new

[nimble heart]

Not on a bot. I'm using it directly on my computer with a $1050 graphics card.

[trim snow]

So it isn‘t possible here in discord? That’s sad

But thanks

[shy kelp]

Why would you img2img, aren't you supposed to feed latent data into the refiner not the actual image?

[nimble heart]

high res fix

latent upscale ruins the partial diffusion workflow

[high skiff]

@nimble heartok, so this is one of the images I just tried with my limited upscale workflow

base

after my workflow

it is just a crop

but yeah, its not the best

tho I know how I could make it better, I just don't know how to implement lmao

[nimble heart]

by base do you mean mine unaltered or the literal base

[high skiff]

unaltered

this is the image

[west breach]

https://tenor.com/view/spongebob-squarepants-one-hour-later-two-hours-later-three-hours-later-gif-17166012

waiting for A1111 to start up

[high skiff]

lol

[nimble heart]

are both ours the same res or is yours 2k?

[high skiff]

So much longer that the old narator got tired and they had to hire a new one

[high skiff]

mine is 2k

I could try 1.5k

[nimble heart]

cause besides the pixel count I really can't tell much difference tbh

least from the window

[high skiff]

areas like the hair are much clearer

[nimble heart]

are those both 1.5k?

[high skiff]

no, 1024 and 2048

[nimble heart]

well like

[hard monolith]

1.0 release tomorrow?

[high skiff]

this is just my workflow

[nimble heart]

2k is gonna be clearer than 1k

[high skiff]

thats... the whole point?

I am upscaling a base image lol

its the base vs the upscale

[nimble heart]

thought it was comparing to mine unaltered

[shy kelp]

I think so

[high skiff]

oh no

[nimble heart]

you can run my workflow in 2k I'm pretty sure. Change all the 1536's to 2048's

get an apples:apples

[high skiff]

I am not comparing it to yours

[nimble heart]

aint that why you peeped my private nodules

[high skiff]

yeah, I just wanted to see

now I am messing with my own

[nimble heart]

Ah.

Any developments on the actual sampling?

[high skiff]

nothing outside fractional offset

well actually, not true

I have some small things I need to test more

[nimble heart]

I've found its good 20% of the time, meh 30%, and bad 50%

[high skiff]

it does good for my style of realism, but I also know thats not how most people use it of course

[upbeat summit]

nice!

[nimble heart]

daily reminder to make sure your workflow can still handle text

[high skiff]

mine car for sure

I will not be able to get much better results until I can sucker somebody into editing the code for ultimate upscale for me

[fresh path]

miku hiring

[west breach]

kaleidoscope style

[high skiff]

I'll bother pseudo with it tomorrow, if it works how I predict, it should produce significantly better results

[nimble heart]

Will that even be necessary on SDXL? 1440p is smaller than 2048^2

If the vae speed blowup is fixed maybe even native 4k wouldn't be impossible

[high skiff]

I always love high res stuff, its been my passion in SD since I started

[high skiff]

its not an issue with ultimate upscale, and its awesome

[nimble heart]

ah like those 100 mb images you can zoom in on?

[high skiff]

yeah

I have another idea, just a moment

[nimble heart]

All the sd upscalers I tried were basically just img2img with some mild hacks to try (and failed) to keep it consistent so I gave up

maybe the refiner could help since you could use a stronger denoise without butchering it

[high skiff]

ultimate upscale with controlnet tile was goated

[nimble heart]

never really used controlnet

[nimble heart]

What are we comparing?

[high skiff]

I was making high quality 4096x images with controlnet tile and ultimate upscale in like a minute flat

[nimble heart]

Looks like right on the kitten might be the refiner timescale shrink thing

ah nvm its hires

everyone copy-pastes their images so I can't peep them

[nimble heart]

damn I should probably sleep insead of just microwaving the whole house

im gonna go die @high skiff ping me if you finalize your hires so I can peep it

[high skiff]

will do

[upbeat summit]

well you gotta do what you gotta do

[high skiff]

likely not to be looking much better until I can get somebody to mess with it for me

[nimble heart]

my room is like 85*F right now im probably not gonna sleep

[high skiff]

same here lol

my room is 84

[nimble heart]

my PC is fighting our 2021 model AC unit and winning

[high skiff]

I don't even have an AC unit installed yet

its 77 outside

gonna open my window and try to cool it down some

[nimble heart]

83 here

outside

@ 330 am

no wonder the AC is struggling to hit 77

or hold, rather

[high skiff]

holy shit

I didn't realize Comfy could get gen data off of A1111 images lol

[sharp robin]

I need to make a github and learn to make repo for dl nodes wte.

[spring fulcrum]

Speaking of holy shit... I just watched the nun a few hours ago... here are a few images I generated for that creepy experiment... Enjoy

[peak dove]

Vladmandic version of AUTOMATIC1111 using SDXL 0.9 - Inuyasha gives a great anime look

[high skiff]

that doesn't look anything like anime

But that does look like a dope style based off of more traditional japanese art!

[peak dove]

Thank you Sytan 🙂

[nimble heart]

wait what. can you just drag them in or something?

[high skiff]

yeah lmao

I just did it

[peak dove]

Some more ... SDXL is fantastic!!!

[nimble heart]

...I already wrote a script to read it manually and now I'm just wondering why.

[peak dove]

I see all y'all are using Cumfy - anyone using Vlad AUTOMATIC1111 at all?

[nimble heart]

this guy is

[peak dove]

Way to go! I am using the base diffuser - the refiner version just refuses to work!

[nimble heart]

the refiner not working is why we use Comfy

among other but that's a big part

[peak dove]

Is its output substantially different to the base model?

[nimble heart]

When done correctly yes

[peak dove]

Thank you for the insight!

[nimble heart]

Try the partial diffusion workflow from Sytan's git if you want a decent example including proper refiner use

[peak dove]

I have both safetensor and diffuser - not sure why? Can I use them both?

[nimble heart]

just safetensor

[sharp robin]

Depends on ur settings it could be minor or it could be a footballs field differet

[high skiff]

ok

I

don't know what to say

but comfy is following my prompts lightyears better than A1111 ever did

[peak dove]

I will leave my setup as it is - base diffuser works OK - refiner crashes> Vladmandic is working to get the whole thing right ...

[high skiff]

for this specific model, its just like

its not even close

[peak dove]

I find the results are attractive. I must get used to prompting "for an exact result"

[sharp robin]

Run same everything in comfy and auto. U likely have different settings and prompt crafting has likely gotten better including negatives.

Because that shouldn’t be how this works

[nimble heart]

left is improper refiner use, right is "partial diffusion" aka how it's supposed to be used. You can drag the right image directly into ComfyUI to see how I made it.
#✨｜sdxl message

[peak dove]

I am never using negative prompts

[nimble heart]

I don't either tbh

[sharp robin]

Ok. Cool story. But sometimes they can make or break your gen.

[nimble heart]

unless something keeps showing up and I need it gone

[sharp robin]

Like for most part i do no neg but if it’s being stubborn one word in negative significantly helps it.

Sdxl is great about no negs tho.

[peak dove]

ART DECO, BELLE EPOQUE, DARIA PETRILLI, INUYASHA, BILLIE ZENGEWA, MARCI MCDONALD, 20 iterations, cfg 6, DEIS also some LATENT MIRRORING extension

[fresh path]

release soon™️

[upbeat summit]

true. normally I'm hacking and confusing the latent space with negative prompts (SD1, SD2). I'm not saying I'm not doing this with SDXL as well , but I've got many interesting results completely without a negative prompt. really cool.

[fresh path]

I never think negatively

[ionic dragon]

oh, so you have created a formatter?

[soft zealot]

I didn't create it, found it 🙂

[ionic dragon]

it looks great

especially the straight lines

[peak dove]

SDXL images are 4 x larger than SD - so its taking about 1 min/picture

[ionic dragon]

@high skiff do you know any node formatter?
like a beautifier

[peak dove]

So I need a bigger rig - aiming at 24Gb VRAM ... and 128 Gb DDR6

[ionic dragon]

it depends on the workflow

[ionic dragon]

just get a used 3090

[peak dove]

Mebbe

[spring fulcrum]

wait ddr6 isnt a thing yet

i just got ddr5

[peak dove]

An A1000 would be nice ... 🙂

[spring fulcrum]

an a6000 would be even nicer... 48gb vram

its a server grade card

its also like 5000 bucks used

[peak dove]

Let me win the lottery!!!

[spring fulcrum]

lol i wish

[soft zealot]

depends how you measure it
SD1.5 is (genrally) based on 512x512 images
SD2.0 is (generally) based on 768x768 images
SDXL is (generally) based on 1024x1024 images

Which is 2x bigger on each dimension which is 4x bigger on area 🙂

[ionic dragon]

if you can afford

[peak dove]

Yes SDXL is 4 x 512 i.e. 4 squares each 512x512

[spring fulcrum]

I had to switch off of the demonic shit... the eyes kept getting closer and creepier.. lol onto the demon warriors

Any suggestions for improvements here?

[upbeat summit]

nice compositions!

[spring fulcrum]

thanks 🙂

[upbeat summit]

for me personally they are a bit over-sharpened but that's always a trade off with upscalers

[peak dove]

Snoop Dogg's Crib - his bathroom! 😄

[spring fulcrum]

He actually hires a guy that just rolls blunts for him all day. I didn't know that until recently. lmao... that must have been an interesting job posting on indeed.

[west breach]

more styles. glam, nonagon and haunting

[peak dove]

I really like Snoop Dogg's crib 😄

[spring fulcrum]

Evil Snorlax sleeping in the park

[upbeat summit]

nice glow on the lightning

[smoky patrol]

love the glasses and green

[grizzled warren]

SDXXL, oh no, rip my VRAM 🤣

[peak dove]

Peacock Love

[spring fulcrum]

Tron Dogg... lmao

[paper phoenix]

wonder if tomorrow will turn this channel into inane tech support or just a wailing wall of malcontents.

[west breach]

can i run sdxl 1.0 on my windows phone?

[upbeat summit]

"why does my prompt work differently now?"

[spring fulcrum]

so are you saying the real deal 1.0 is dropping tomorrow?

[upbeat summit]

this is still the official statement

[sudden cliff]

It's already occasionally a wailing wall of malcontents.

[paper phoenix]

"It doesnt shoot lasers at my enemies and give me money for using it. this sucks and you personally, devs suck."

[paper phoenix]

i typed "SDXL" into a word document and shouted at that word document. where do all the pictures live? what kind of tech support is this?

[upbeat summit]

having some wildcard fun surfing the latent space? 🙂

[west breach]

just the styles 🙂

[west breach]

now i'm exploring with wildcards 😄

[molten gull]

say, is it possible yet to train embeddings on sdxl ?

[molten gull]

a1111 now can do sdxl, too, i read ? so can you train embeddings like you could with 2.1 ?

[rustic garnet]

I found embeddings surprisingly less powerful than in 2.1
It's hard get a good embedding without loosing coherence in the images

[peak dove]

The original d/load of Vladmandic AUTOMATIC1111 does not have a folder for embeddings ... but I've yet to try? Normally, I can produce a usefull embedding after 3.5 hours - if SDXL is 4 x Larger - then embedding times will be 12 hours + ???

[plain basin]

sdNext has embeddings in Models\embeddings rather than in root folder

[upbeat summit]

sdxl embeddings when

[shy kelp]

24h to 1.0 leggo

[west breach]

https://tenor.com/view/24-clock-gif-3535520

[rustic garnet]

it has. It's in the model directory.

[uneven dove]

a1111 cant train sdxl

[high skiff]

I need to figure out how to use inpainting in comfy UI

hhmmmmm

ohhhhhhh

no way it actually works like that

you have to edit the image in another program, delete the background, and then export as a png?

[sage basin]

You can right click on an image, open in mask editor, do your business, save it to clipspace, make a load image node, and then right click and paste clipscape (or something like that)

[high skiff]

I generated this super cute guy, but I need to inpaint his face cause it is ROYALLY messed up lmao

[high skiff]

oh god, you life savor lmaop

[sage basin]

Took me a couple of months to figure it out 😅

[molten gull]

i m a bit confused now, so can i do embeddings in sdxl yet somehow? and if, what do i have to do/download ?

[delicate grotto]

yes, you can use embeddings with SDXL, but i didn't see how to train them yet

[high skiff]

I am not sure what is happening, but inpainting in comfy is slow as hell

mmm yes, 6s/it

[delicate grotto]

also, the img2img is weird

[high skiff]

oh

I realize why now

oops

I upscaled my upscaled image again

bruh

[delicate grotto]

we need to go further

[fading kayak]

Happy SDXL release week? 👀

[ionic dragon]

@high skiff @uneven dove how can we make extra-circulars out of SD?
so that they would be useful for me while applying for collleges?

[uneven dove]

that seems like a question for Emad

ping him and ask

[ionic dragon]

as of now, i just generate images, so i dont think i can show off my work other than just creating a showcase website

[delicate grotto]

ummm you want to apply to a college by using SD?

[ionic dragon]

i dont think he would reply, his chat might be flooded with pings and DMs

[delicate grotto]

you best bet is to get good grades

[peak dove]

Vladmandic AUTOMATIC1111 SDXL 0.9

[ionic dragon]

no, how can we make a good EC using SD

[high skiff]

you evil ass lmao

[uneven dove]

lmao yeah ai art isnt going to make you money

[delicate grotto]

EC?

[ionic dragon]

i can get good grades, ECs matter a lot so thinking of them

[ionic dragon]

extra ciriculars

[ionic dragon]

i dont want to make money

[delicate grotto]

2 sec i need to translate what that means

[uneven dove]

knowing how ai training works might make you money but once they train an ai to train ai you are done for

[uneven dove]

then do nothing, it is easiest

[delicate grotto]

ahhhh, your country accepts EC's then? that's odd
here it is solely grade based + special exam

[lilac wren]

? x)

[delicate grotto]

why not make a simple ai from scratch using python?
some simple tasks like that can help you with EC right?

[ionic dragon]

i want to make something which can be helpful for me to get into a good clg

[ionic dragon]

i am applying for US unis

[uneven dove]

good colleges are doing ai art?

[peak dove]

There is no embedding folder under Models in sdNext - but I can place an embeddings folder there?

[ionic dragon]

i dont know how to make ai tools

[delicate grotto]

mmm your best bet and im not joking then is to lie that you are a native american part hindu or whatever

[ionic dragon]

no, good colleges need some kinda ECs

[uneven dove]

#🌶｜off-topic

maybe ask gpt. i dont know

[ionic dragon]

why would i lie?
anyways my report cards would definitely be from my native country, they'll kick me off if I do that

[high skiff]

alright, inpainting is seriously annoying in comfy lmao

[delicate grotto]

everyone lies there... the sad reality is that if you are white/chinese - you are being extremely discriminated and there is nothing that you can do

[high skiff]

I just took mental damage from reading that

maybe don't go into radical racism theories in the SDXL chat

[delicate grotto]

its not a theory sadly

[high skiff]

don't start with BS like that man, come on

[ionic dragon]

does sdxl 0.9 support inpainting as of now

[high skiff]

man, inpainting in comfy is a pain in the ass

IDK

I am using 1.5

[uneven dove]

its damn monday

[delicate grotto]

i can send you some stuff in private about it...

[high skiff]

I genuinely might have to download automatic just to do this stupid inpainting

[uneven dove]

no

[delicate grotto]

and the reality is quite dumb

[uneven dove]

just stop

[high skiff]

I am not interested in your radicalized rediculous theories bro. Keep me out of your stupidity

[ionic dragon]

wait, you have never used a1111?

[high skiff]

I have used it

[vagrant plover]

thats where #1101520089980882994 comes in handy lol

[uneven dove]

roflmao

[high skiff]

right lmao

guess you saw that too lmao

[uneven dove]

this server in general needs more moderation

[ionic dragon]

whats deep floyd?
how is it diff from sd?

[vagrant plover]

yea it does

[high skiff]

last time I brought that up, I was kinda told I was being hostile, so I now have no input on that matter lol

[vagrant plover]

its way better but its very hard to have the recouces for

[ionic dragon]

resources of?

[vagrant plover]

which part?

[vagrant plover]

vram

[high skiff]

the server needing more mods

[ionic dragon]

whats the min vram req?

[high skiff]

liek 12GB I think

[uneven dove]

the best kind of hostility. the accurate kind

[high skiff]

right lol

[vagrant plover]

ahhh yes i have on a nother server and they hate it lol

[high skiff]

this inpainting is driving me insane lmao

[vagrant plover]

the truth doesnt care abt your feelings

[vagrant plover]

no 16

24 is reccomneded

[high skiff]

ah

I will have 24 later today :>

[vagrant plover]

oooh 4090 today?

[high skiff]

3090

[vagrant plover]

i have a 4070 ti

[uneven dove]

3090

[high skiff]

got it for a damn good deal lol

[vagrant plover]

ohh how much

[uneven dove]

he spend 6500 on it

[high skiff]

Zotac Trinity OC 3090 with repadded VRAM and repasted core, in the original box for $600 shipped with tax

[ionic dragon]

then why are 3090/4090 owners running sd?

[vagrant plover]

dangggggg

[high skiff]

yeah lmao

[vagrant plover]

idk

[sage basin]

Literally this 💯 If you have WASasquatch's node suite, there's a CLIPSeg node that allows you to generate masks from a prompt. It's not perfect, but I find it helps a lot

[vagrant plover]

i got my 4070 ti for 900 lol

[high skiff]

the mask isn't the problem, I am just cofused how to use it

I use the mask from the mask image, and the pixels from the normal image

and then the result is blank unless I use like .7 denoise

[covert herald]

Relative new to comfy but what's bothering me is that we can't fix faces or fingers like the A1111 i can use ADetailer and boom problem solved. especially with SDXL were it produces bad hands and faces if the subject is not zoomed in, so i take the image to Photoshop to finish it. Any ideas???

[high skiff]

which at that point destroys the image

[uneven dove]

not using the refiner?

[vagrant plover]

@high skiff whats your full system spec

[uneven dove]

like 600tb of ram

xenon cpu

[sage basin]

Ohhh, right, yeah, the inpaint encoder will replace everything in the mask with empty latent noise. If you encode for inpainting and then decode it again, you can see what it's doing (everything in the mask should be grey)

[uneven dove]

nvidia 3090 x 8

[high skiff]

with the new GPU today, it is

12600k, 64GB 6000mhz DDR5, 4.5TB NVMe, and a Zotac 3090 all running on an 850 watt gold PSU

[vagrant plover]

hm nice

[high skiff]

just built it day before yesterday lmao

[sage basin]

If you want to denoise the original image, you have to composite the latents

[high skiff]

ohhhh, is that how it works?

testing!

[uneven dove]

he forgot to take the plastic protector off the cpu prolly

[high skiff]

I did not, thank you very much :p

my CPU temps are great actually :p

[vagrant plover]

can i see a pic of it?

[uneven dove]

its okay if you did, we will not make fun

[vagrant plover]

what cooler

[uneven dove]

=p

[high skiff]

Its not in a place to be photographed

[vagrant plover]

lol weve all done that

[high skiff]

cooler is Perrless Assassin by ThermalRight

[uneven dove]

lmao take the cpu cooler off to photograph it

[high skiff]

its the new go to for cooling higher power CPU's

[uneven dove]

show us the pins

[high skiff]

265watt TDP, 6 heatpipes, 2x120mm fans for $30

[vagrant plover]

ill bend them for him

[high skiff]

the pins in my CPU?

[uneven dove]

no the pins in your grandmothers broche thing. of course the cpu

[covert herald]

sure i do

[uneven dove]

never had bad eyes tho here

[high skiff]

do you happen to know how to go about compositing Latents by chance?

[uneven dove]

hands, limbs in general, a horror show

[high skiff]

OW MY KNEE

[covert herald]

the fingers are the worst

[high skiff]

[uneven dove]

just negative prompt for hands

[high skiff]

lol

make them amuputee's

[uneven dove]

i dont think inpainting them will work without controlnet

we dont have that yet

[clever verge]

Out of at least 1000 images a handfull was not a horror show.

[uneven dove]

sounds like you have bad settings then

[clever verge]

Negative prompting does not make any difference

[uneven dove]

or you are running 1.5

[high skiff]

what the hell is up with all of the political things today lmao

[clever verge]

Maybe I'm too picky. If fingers are good, something else is bad, like nails.

Or then eyes are bad.

SDXL

[uneven dove]

,horror show'

its literally making images out of gaussian noise, pal

[clever verge]

I got better detail in the image when I removed the negatives, not sure why.

Yes, I know.

[uneven dove]

do you? lol worrying abut fingernails?

[plain basin]

yes it should work

[clever verge]

Why not?

You have to find the current limitation, I find it intresting to know.

[uneven dove]

as you have seen, you only got a few you consider acceptable out of 1000

[sage basin]

You want the bottom one

[peak dove]

Vladmandic AUTOMATIC1111 SDXL 0.9 has inpainting ...

[high skiff]

absolute legend, thank you immensely

[uneven dove]

thats just how it is, you are too picky. sdxl is state of the art. if it aint good enough, what youre looking for is called a camera

[clever verge]

With some combined knowledge of human anatomy it will be very powerful.

Don't get me wrong, I have no issues with it.

[uneven dove]

lmo here we go again with the 'anatomy' discussion

[peak dove]

Vladmandic AUTOMATIC1111 SDXL 0.9

[uneven dove]

if you have no issues why did you call it a horror show

[void skiff]

It say to vote on what best match the prompt, but "Insanely detailed; is that super duper over the top with graphical details that almost cover the image like American x-mas decoration on a tree, or is that insane as in dead rabbits and so on the fingers?

[clever verge]

Used your words, dude.

[uneven dove]

i am talking about the issues i have observed. you used my words to describe something else, which i would not describe that way. as it is i think the answer for you is 'you cannot'

[peak dove]

I don't think SD understands "insanely detailed!!!" 😄

"Bonkers nutcase dementedly detailed!!!" 😄

[uneven dove]

'intricate details' is like royalty

[high skiff]

@sage basinalright, I do have it working, so a hueg thank you for that, though its not very good still unfortunately

I think I will have to pop my images back over to auto for proper inpainting unfortunately

[uneven dove]

auto uses controlnet

[clever verge]

It does hair well, it's impressive. Much about it is impressive.

[high skiff]

it doesn't have an option to do a lower res image over where the face is, so its doing like a 1536x image and only using a tiny little bit of that latent space to inpaint the face

[void skiff]

I have to test now.

[high skiff]

absurdly hyper fideli-detailed

[clever verge]

Is there a way to know if the model knows the words that goes through clip and in to the model?

[heady halo]

Is sdxl 1.0 releasing tomorrow?

[void skiff]

Prompt. "rainbow unicorn in 8k glory and bonkers nutcase dementedly detailed!!!"

[peak dove]

Tee hee

[ionic dragon]

Yea, but the workflow isn't as flexible as comfyui and the outputs aren't as good as compared to comfy

[high skiff]

I am not mentally prepared for the sheer amount of people asking basic questions in here once SDXL drops

[peak dove]

I am at home with the 'flu - SDXL 0.9 is a good way of passing the time 🙂

[high skiff]

I am sick too!

[thorn lion]

it's gonna be so hard for me to find the good information. I hope we use threads tmrw

[high skiff]

same here

[peak dove]

I know a "lot" about SDXL, mainly cuz I've had it crash and crash and crash - learned a lot putting it straight!

[thorn lion]

facepalm I just realized an option I have though

[peak dove]

Github says ask SDXL questions at Discord - but Discord often says "SDXL is unsupported - so any software you use may not work!"

[median prism]

When's SDXL supposed to release again

[high skiff]

I am already tired of this, and its not even out yet lmfao

[peak dove]

https://www.reddit.com/r/StableDiffusion/comments/14sthgn/sdxl_release_date_july_18th/

SDXL by all reports drops tomorrow July 18

[indigo carbon]

Yeah, 1.0

I asked a Dev, they said that is the targeted release date for the full release of SDXL1.0

[peak dove]

So it'll be just a case of d/loading the new diffusers/safetensors?

[indigo carbon]

Likely

Also, is anyone using SDXL on Vlad's UI?

[ashen vale]

Are the style options for sdxl bots just style prompts or are they different individual models?

[void skiff]

We can now test a new and better AI text to image generation to create amazing art. first thing natural stupidity do, "Waifu 69, hyper realistic octane render anime".

[indigo carbon]

?

It should theoretically work with all optimizations

And Vlad's UI uses diffusers instead of safetensors, meaning it's the same optimization as ComfyUI (for the most part).

[peak dove]

I am using Vlad AUTO1111 SDXL

With base diffusers

[indigo carbon]

Is it as fast as 1.5?

Also, is that with the refiner?

[peak dove]

SD at 512x512 generates at about 20 seconds per photo on my rig. SDXL 1024x1024 is 4 x 512x512 - so each generation is taking about 60 seconds...

[visual glade]

yeah diffusers should be as fast as comfyui but it's a different implementation

[peak dove]

The refiner crashes Vlad AUTO1111 - but the base diffuser works just fine

[indigo carbon]

As far as I've seen, results are similar if not identical

[indigo carbon]

Really? Damn

[visual glade]

yes results should be the same between both

[peak dove]

Base Diffuser results in SDXL 0.9 are spectacular anyways ...

[indigo carbon]

If I recall correctly there is an SD_refiner checkpoint next to SD checkpoint

[peak dove]

I have the safetensor (base and refiner) and both diffusers

But I am using only base difuser

[indigo carbon]

I heard Vlad's UI supports SDXL with refiner just fine =\ guess it was a hoax

[indigo carbon]

Did you try with the SD_refiner

[peak dove]

Mine won't work! Help?! 😄

[indigo carbon]

I have no idea! That's why I asked. I just wanted to know what's the optimal inference for SDXL

[peak dove]

I get ERROR messages with the refiner diffuser

[indigo carbon]

Open an issue on Vlad's GitHub

[peak dove]

Where do I ook to check inference?

[indigo carbon]

Screenshot?

[peak dove]

OK, thanks

I can make a screenshot - but under which tab?

[indigo carbon]

Txt2img, full screen. I want to take a look how Vlad's UI looks like

[peak dove]

Prompt for SDXL 0.9:- a realistic beautiful autumn queen, fibonacci, headshot, close up, night time, autumnal mood, venice carnival, fauvism, zentangle, grand guignol, wavy hairstyle, white hair, character concept art, created by victo ngai henri rousseau vladimir kush coles philips elizabeth catlett arief putra john currin alenka sottler itzchak tarkay

[indigo carbon]

Easy fix. Go to quick settings, add SD_refiner

It should use both models simultaneously

It will add another drop-down menu next to SD_checkpoint

One that you add the refiner.

[peak dove]

Do you mean SETTINGS? Is there a Quick Settings?

The refiner diffuser sits in the models folder, but when I change to that diffuser ... ERROR 🙂

[indigo carbon]

Yeah, there is a tab in the settings

[trail owl]

hi

[peak dove]

Don't see one I'm afraid!

[trail owl]

can someone help me i got a error saying (modulenotfounderror no module named 'diffusers' )

[high skiff]

hey peopleeee

[indigo carbon]

I don't know how to navigate through Vlad's UI, if you want me to help screen shot the settings tab.

[high skiff]

does anybody here know how to get past dubious ownership errors?

[autumn forum]

Become the owner

[high skiff]

I am

Its from my previous PC lmao

[visual glade]

on windows right click folder and set permissions to everyone?

[peak dove]

Select the Green Text which specifies the directory/file, copy that file name, then open a cmd prompt and use git config --global --add safe.directory (add the green text here)

You must do this for each and every dubious ownership

[high skiff]

yeah, I started doing that, but its gonna be like 50 files lol

I am trying to mess with permissions

[peak dove]

I did it about 30 tiimes

[indigo carbon]

@peak dove , did you get the refiner working?

[peak dove]

Not yet, but I will keep trying 🙂

[high skiff]

ok yeah, permissions did nothing

time to gety to copy pasting

[molten gull]

THE MAGIC CUT OFF HER FINGERS !!! poor girl !!!

[autumn forum]

Can it be tomorrow already ugh.

[indigo carbon]

Show all pages

[visual glade]

if you want to play around with SDXL ComfyUI with this workflow https://github.com/SytanSD/Sytan-SDXL-ComfyUI is probably the easiest way to do that

[indigo carbon]

It's somewhere there

[molten gull]

but HEY ! the magic gave her fingers back ... gave her even EXTRA fingers !

[high skiff]

oh sick, only had to do 7 dubious owners

[molten gull]

[ back in 2020 ]: ... one day kids, and it will be in just like three years, we will be able to make AWESOME pictures with ai ! art ! portraits ! photography ! the possibilities are endless !! ... the only thing that won't be possible is hands

[peak dove]

Thanks for the help - my SDXL is busy now. I will try and get the screenshots later 🙂

[thorn lion]

first thing I will be do tmrw is testing racial diversity and social bias. I can't pretend realism = anatomical structural homogeneity. The previews seem good though so I'm not too worried

[peak dove]

An Odalisque SDXL 0.9

[uneven dove]

nope. i fine-tuned SD 2.0 on 10,000 high quality images of hands. and wouldn't you know it? they started appearing randomly in stuff

[thorn lion]

ask ChatGPT to write a script to run a certain command recursively along all contents of a folder

[uneven dove]

but that script will have permission issues

[high skiff]

lmao

[thorn lion]

oh I don't have access to code interpreter or whatever so I'm literally copy pasting code blocks into my SDE...sde...no....like VS Code or Fleet or whatever....that thingy

[high skiff]

was able to do it in like a minute, only needed 7 dubious owner corrections

[thorn lion]

if the file is created on the device, would it still have permission issues? hasn't been my experience unless i'm doing something absurd

[indigo carbon]

@peak dove I figured it out! Look for a setting called "Quicksettings list"

It's in one of the tabs in the settings section

[paper phoenix]

i think even with a million photos of hands in every possible configuration act and angle there will still be hand issues. i mean humans can see missing thumbs/got your nose/sleight of hand tricks and have to correct themselves.

[uneven dove]

the thing is, people just get upset about hands for no reason. look at random images from photographers like Charles Cushman. most of the images look like "incorrect hands", but they're from a camera

[paper phoenix]

i reckon we should just get influencers and celebs to wear those 3 finger one thumb mickey mouse gloves, put it down as a trend and then the AI will be fine.

[thorn lion]

down

[uneven dove]

you'd need a DALL-E2 size model to do hands correctly

theirs uses 30B parameters iirc

[paper phoenix]

even then.

[distant zodiac]

https://tenor.com/view/leonardo-dicaprio-clapping-clap-applause-amazing-gif-16384995

[indigo carbon]

@peak dove it's in the user interface tab, scroll down until you find "Quicksettings list"