✨｜sdxl

[trim orbit]

1.5 i think is similar to a community fine tune of 1.4. it's a finetune that runwayml did themselves

[deft notch]

When is sdxl 1.0 releasing ?

[raw prairie]

July 18th

[sage basin]

Tuesday™

[urban fjord]

Horror is fun, but there is also something fun about weird movie scenes like this.

[raw prairie]

Do you think it'll work on 8GB VRAM?

1.0 I mean

[autumn forum]

its an extra step after the base model that adds fine detail to the image.

[urban fjord]

Yeah, the base should work on 8GB.

[raw prairie]

It's because I need a laptop for video editing and probably will install sd locally

Weird thing, the rtx 3080 laptop model has 16GB while the 4080 has 12GB only

[trim orbit]

there were 3080 laptops with 10 and 12gb if i remember

[raw prairie]

And different power consumption

[uneven dove]

wes craven style lol

[raw prairie]

Bart's dopplegänger

The last one

[uneven dove]

Bort

[raw prairie]

There was an episode where bart meets a rich kid who looks identical to him and they switch lives

That was the reference

[urban fjord]

Oh no...

SDXL is struggling with the simpsons with me, a lot of broken outputs.

I asked for Bart in real life...

[eternal fog]

lol

[raw prairie]

Very midjourney-esque

[upbeat summit]

so clean

[west breach]

@upbeat summit wildcard now supported by my prompt with style node

[upbeat summit]

this may be my setup but I just realized - I'm sending the pos + neg prompt also to the refiner. when I use wildcards from what I just tested, the refiner textencode evaluates it again and it uses a different wildcard :/

[urban fjord]

Krusty the Clown works better as it's less overbaked in the base model

Not sure how you untrain the overbaking in the base model.

[lusty raptor]

just use an autohotkey script, ezpz

[uneven dove]

furby version of ewok village lul

[eternal fog]

Huh this can very nearly do a few cosplayers

it at least gets the "Professional" cosplay style from it

[winter raptor]

How big of a leap do you guys think 1.0 will be?

[eternal fog]

It won't be a leap

It might be a little better at some things

[winter raptor]

Workflow wise, is it gonna change the way we use the refiner?

[eternal fog]

No idea they've not said

[sharp robin]

problably not, refiner will still be used

[urban fjord]

Hopefully we won't be plagued by too many almost the same merges.

[urban fjord]

Given that Emad announced a new refiner it means that they were not successful in getting rid of the refiner which means the advances is probably quite small. At worst even more concepts are overbaked.

[high skiff]

I hope we get to a point where we can make images like this lol

[trail bay]

<48 hours left until 1.0 release right?

nice

[winter raptor]

I’m having a ton of difficulty with eyes in SDXL, anyone have any ideas?

[urban fjord]

Try over up the eyes so you cannot see them?

Btw, what are your eye issues as I can't see any.

[winter raptor]

I’ll share when I’m back on my computer, but the eyes just like melt into the face, have weird irises and pupils, etc

With what? I’ve tried just using an upscale model, but that’s actually been making it worse for me for some reason

Ahhhh

What model have you been using with that?

[indigo carbon]

Yo, is SDXL1.0 going to release tomorrow after all?

[uneven dove]

the balkan beat box

[winter raptor]

So to recap, use the refiner as the model for sd upscale? And which upscaler model are you using with it?

[uneven dove]

reverse side

[urban fjord]

No, you use the refiner for the last steps of the sampling.

[indigo carbon]

Emad tweeted SDXL1.0 is releasing on July 18th, is this still going to happen?

[winter raptor]

What model do you plug into sd upscale?

I’ve seen some people use juggernaut, etc

[uneven dove]

you can read about this, in my book

[glad fulcrum]

can we post something with blood?

[eternal fog]

She looks a bit long. I've found increasing the text encoders height and width, both normal and target by 2x seems to alleviate this somewhat.

[uneven dove]

an impoverished pokemon town that suffered the fall of pokemon harvesting

[urban fjord]

I recommend just getting this book

[uneven dove]

mine will teach you the dark arts

[quasi remnant]

is that just for upscaled/stuff in different aspect ratio? i haven't really tinkered with the wXh/target wXh, works fine for me on 1024x1024 gens

[glad fulcrum]

can someone give me an img to img workflow for comfyUI, as I've only used it once

[eternal fog]

Yeah if you change aspect ratios you need the clip to match in my experience

[glad fulcrum]

If there is an im2img working workflow

[eternal fog]

If you don't move from 1024*1024 you don't really need to mess with it

You can get some interesting results if you do though.

[quasi remnant]

might have to play around with it once 1.0 is out... so close now

still 2 days tentatively?

[eternal fog]

Just use a normal workflow, but pass an image into the first ksampler and turn down the denoise.

[glad fulcrum]

normal workflow?

[eternal fog]

Well 1 and a bit

[glad fulcrum]

I don't know what a normal workflow would be

[quasi remnant]

lol, of course xD

[glad fulcrum]

can you share an image so I drag and drop

[eternal fog]

I'm not at my pc because I'm going to bed

[quasi remnant]

or do you mean a planned midnight release or something?

cuz that'd be even neater

[eternal fog]

I assume they'll do it at some point during the working day on Tuesday

[quasi remnant]

same

[eternal fog]

Maybe slightly after

[quasi remnant]

super excited to test the full capabilities with the finished model, so many things i still wanna test or re-test that were just almost there on 0.9

but came out well on bot

[quasi remnant]

do you have a link to that tweet? all I've found are people who had early access/insider information from working with devs who said stuff about july 18th

tbh i actually haven't seen an official post about july 18, just been assuming + mid july

the official stuff has said mid july iirc

[high skiff]

I have awoken

[quasi remnant]

https://twitter.com/EMostaque/status/1672701298846887936 https://twitter.com/EMostaque/status/1676222592762486787 emad never directly said july 18th either

Emad (@EMostaque)

@tsi_org Monday for 0.9 mid July for 1.0

Twitter · Jun 24, 2023, 8:20 PM

Emad (@EMostaque)

@alfredplpl We are working on trainers and guides for mid July release with Kohya and others

There will be a fine tune service in due course

Twitter · Jul 4, 2023, 1:33 PM

just mid july

[high skiff]

I am sicc

[quasi remnant]

get sum zicam

[indigo carbon]

https://twitter.com/EMostaque/status/1676923097159712768

Emad (@EMostaque)

Paper up on Arxiv for #SDXL 0.9!

Target open (CreativeML) #SDXL release date (touch 🪵) 18th July 🎯

Team looking at different, more advanced, variants to pick release candidate

Research access for 0.9 will widen shortly please apply here if interested https://t.co/gIUX9s46CZ

Likes

189

Twitter · Jul 6, 2023, 11:56 AM

[quasi remnant]

Ohh

i didn't see that

[high skiff]

So many staff members and official publications have stated July 18th for a hot moment now

[quasi remnant]

i've only seen mid july stuff officially til now, thx

[indigo carbon]

They are talking about SDXL 1.0, right?

[quasi remnant]

yeah

[high skiff]

I used to use it, then I looked into its FDA claims, and apprently they got sued for literally ruining peoples sense of smell permanently in like 2012, and also they have been called out by the FDA for having legit 0 supported research that they do anything for people who use them

[quasi remnant]

i totally don't pay attention to news

all i kno is it shortens my cold duration by like 3 days xD

but i guess that could be a deterrent

[high skiff]

All I have to say is the FDA has said numerous times in their studies that it does a whopping nothing lol

[indigo carbon]

I'm just never sick =]

But when I was little I was sick about 7 or 8 times a year

[quasi remnant]

i'm always sick, it's why i do remote work from home and never leave the house xD

[wicked frigate]

July 18th has been the target. Bear in mind target doesn't mean guarantee

[indigo carbon]

For 1.0, right?

[quasi remnant]

wat kinda suggestion is that. part of it is that i'm allergic to everything outside, maximum grass allergy, every type of animal, etc.

vitamin d comes in tablets kekw

i literally can't go outside. i have a compromised immune system and allergies to everything in the natural world xD

nope, underweight

[indigo carbon]

Bro, leave him alone. No one is judging you

[eternal fog]

Both

[indigo carbon]

@quasi remnant, I get that. When I was little I wasn't even able to go outside so I just studied computers all day, and here I am today.

[eternal fog]

This sounds like a precursor to a delay lol

[quasi remnant]

lol, i do remember back in sd 1.5 release there was a sliiiight delay xD

but it seems like there's been nothing contradicting the release date to this point

[glad fulcrum]

why you answer? is that a workflow?

mimi

[quasi remnant]

reaction image to the image you posted lol

the reaction of a man who sees a cow sitting at the dinner table xD

[glad fulcrum]

oh

[quasi remnant]

it does have my workflow in it though, yea

lol

[glad fulcrum]

I need an img2img workflow

[urban fjord]

That doesn't meant there can't be delays.

[urban fjord]

Cows get hungry too

[quasi remnant]

reminds me that i've seen like vids of tame bears eating at a dinner table way back

lol

[eternal fog]

No target needs changing too

I mean I can't tell where they are going. But that's what I've been doing and it's been working ok

Same as you

Have a play with it. I'm going to bed. It's 4:30am lol

[quasi remnant]

i'm guessing it still doesn't support SDXL specifically, but for img2img i prefer the ease of use of webui, especially for inpainting & stuff otherwise you gotta upload manual masks n such

[eternal fog]

Lol

But do you have work in the morning

[uneven dove]

@high skiff

the handy bear of bear hands

[wicked frigate]

auto webui has an in-progress SDXL-support update going

[quasi remnant]

ah

that's good to know

[eternal fog]

For img2img in comfyui you just have to pass in an image instead of an empty latent. It's easy.

[quasi remnant]

i'll stick to comfy for txt2img but would be nice to use webui for inpainting

[wicked frigate]

it will be 1.5.0 but you can try it now via the dev branch

[quasi remnant]

ty for the heads up

[glad fulcrum]

I know its possible but I am not smart at this

[high skiff]

death

AYOOOO

[quasi remnant]

pretty sure it'd be something like load latent with the images you want in input folder(?) haven't tried it yet.

[high skiff]

MY GPU IS GETTING HERE 2 days early!?

[quasi remnant]

then use that instead of empty latent image

[high skiff]

yeah haha, I have a whole new PC I built yesterday

from a 4 core 7700k, 32GB DDR4, A 3080, and a 650 watt white PSU

To a 10 core 12600k, 64GB DDR5, A 3090, and a 850 watt gold PSU

[eternal fog]

Load image node, into a resize node so you can make sure you don't put an 8k image through. Then a VAE Encode node and then that into the sampler.

Some prompts do seem to cause that sometimes.

[glad fulcrum]

Ok I will close auto and try in a sec

[uneven dove]

@urban fjord was saying polar bears are hard to make look real

[urban fjord]

Show your failures too pseudo:

[uneven dove]

i hardly even notice them anymore

[west breach]

wes anderson style

[urban fjord]

SDXL is getting so good at hands that to flex it adds more hands than needed in the image.

[high skiff]

thats a weird issue, I have only had that a couple times across 1000's of images

[quasi remnant]

something like that i'm guessing

[eternal fog]

I think that looks normal she's just missing a belly button

[high skiff]

yeah, all the time

2:3 up to 1:2

[eternal fog]

All my portrait stuff is generally fine though

Whoops

Clicked on a none portrait

Press buttons until it works

That's what I do

[uneven dove]

bloody hell that's terrifying what SDXL thinks Ronald Reagan is

can make him the cryptkeeper tho

[urban fjord]

I really want to know what's actually in the training data. What are those things... Are those scrimps or...

[eternal fog]

Try Boris Johnson. He looks slightly worse than real life

[quasi remnant]

what's are some celebrities that don't get baked by the cfg?

i know Chuck Norris just works

lol

[quasi remnant]

that looks good too

so chuck and boris

o yea i think trump works

[uneven dove]

john cusack

[quasi remnant]

trump's face is always messed up tho

[uneven dove]

that's baked

[quasi remnant]

that's baked in a different way

lmao

[eternal fog]

A lot of celebrities I've tried work. But I also only know the very famous ones.

[trim orbit]

it does his mess of a hair very well

[eternal fog]

Strangely I've seen a few celebs where it looks better in the preview whilst it's stepping through and then by the time it's finished it doesn't look like them anymore.

It's almost like it has them at an intermediate level and then steps over it

[uneven dove]

just request 'handsome' trump

[eternal fog]

Over cooked maybe?

[quasi remnant]

donald trump but it's actually david hasselhoff(how do u spell that last name)

[uneven dove]

dont hassel the hoff

[trim orbit]

more noise and gaps in the details lead the brain to think "hey thats pretty good" since the brain is super good at recognizing faces

the ol less is more adage

[quasi remnant]

for once this extra hand kinda fits lmao

[eternal fog]

One thing I struggle with is getting older pictures of celebrities.

I was trying to make some Stargate stuff with Amanda tapping, but it kept generating how she looks now and not what she looked like then.

[uneven dove]

the golden parachute

[high skiff]

I have a dude who keeps trying to slam my workflow on my github cause "It doesn't do good with prompts less than 10 words"

[uneven dove]

block him lol

[eternal fog]

Lmao. Nothing to do with the workflow.

[quasi remnant]

okay donald trump but it's actually david hasselhoff gets burned tho

[uneven dove]

he's blaming you for how crappy SDXL's prompt following is

[high skiff]

hes trying to suggest people use his workflow, which is just a .25 denoise pass with no A scores or dual text encoder lol

[eternal fog]

He must be using ass prompts

[quasi remnant]

but burned in the best possible way? LOL

[eternal fog]

I can get good results with minimal prompts

[high skiff]

no his thing is, clipdrop and his workflow work way better for extremely low effort prompting

here, I have an example

[quasi remnant]

same

lol

[high skiff]

hes right, but like

prompt is

Diagrams Of How To Draw Hands

clipdrop

[glad fulcrum]

thanks, the img2img worked

[high skiff]

his

what he got from mine

[eternal fog]

But that's just seed and steps

[high skiff]

but then I changed the prompt to:

"A pencil drawn diagram of how to draw hand"

and got this

[quasi remnant]

i'll run this exact prompt thru mine for 4 images and see what pops up

[high skiff]

I have told him multiple times now, he just doesn't follow how to prompt for my workflow lol

[eternal fog]

He's cherry picking his good ones and your bad ones to try get people to use his instead

Just block him

[high skiff]

no, he was right

my workflow is bad for that

[quasi remnant]

first 2 images

[high skiff]

but thats cause I am 99% sure hes using one text encoder

I have a whole long note on how to prompt my workflow lmao

Prompting: For the linguistic prompt, you should try to explain the image you want in a single sentence with proper grammar. For example:

A photograph of a (subject) in a (location) at (time)

then you use the second text field to strengthen that prompt with a few carefully selected tags that will help, such as:

cinematic, bokeh, photograph, (features about subject)

Full prompt example:

Linguistic: A cinematic photograph of a pretty woman with blonde hair and blue eyes in a park at sunset

Supporting: clouds, nature, bokeh, f1.8, cinematic lighting, entered composition

Photographic tips

1. Try and use words or entities that tie into realistic imagery, such as National geographic, Vogue, ethnographic, portrait, and similar words

2. Do not use words like "realism" or "realistic" when prompting for photographs, as they are associated with art styles that depict "realistic" things, and such will weight towards traditional or digital art

3. Feel free to use compositional words, such as "centered subject" or "(object) to the side", this can help a lot for achieving different and interesting results

4. I recommend not going over 6 on the positive aesthetic score for the refiner when going for realism, I find that the higher you go, the more your images are likely to go into the direction of aesthetic paintings and

Its legit IN THE WORKFLOW lmao

[quasi remnant]

2nd 2

[uneven dove]

using your method

[quasi remnant]

i only used clip_g

[eternal fog]

I mean I don't use the 2 prompts on the way you instruct and it works fine for me

[high skiff]

Also, I am sure he used negatives, but I changed mine to have no negatives

[quasi remnant]

if i add diagrams/tutorial to clip_l i have...

[uneven dove]

i think he just purposely set it up to do the worst test to get attention

[high skiff]

yeah, my point is I think he purposely ignored TE number 2, which will nuke all the results

[uneven dove]

just block him

[west breach]

vaporwave

[quasi remnant]

if he added two words

to style text

he would've gotten this

lol

[eternal fog]

Yeah like I said he's doing it on purpose to get people to use his workflow

[uneven dove]

^

[quasi remnant]

or to clip_l i mean

[uneven dove]

ego shit like that guy who CONSTANTLY spams his shitty youtube tutorials everywhere you look

[quasi remnant]

idk this is what i'd expect

[high skiff]

there, I blocked him

[uneven dove]

empowering

[high skiff]

I am not able to take down peoples issues, am I?

[sharp robin]

So is 1.0 delayed?

[quasi remnant]

all i did was add 2 words to clip_l and it turned the low-mid prompt to mid-high

[uneven dove]

you can delete them lol

might have to dig for the button

[high skiff]

his other one was "Its not good when I prompt "Elvis Presly"

like

Its just

Elvis Presly

[uneven dove]

with the typo?

nice

[quasi remnant]

bruh

i just misspelled it too

[high skiff]

I likely spelled it wrong lol

[quasi remnant]

i just misspelled it a diff way

LMAO

Presley"

xD

[high skiff]

I don't remember what he prompted lol

[uneven dove]

it's elvis prezli guys

god damn it

[high skiff]

Elphis Prestleen

[eternal fog]

Elvis Pretzel

[quasi remnant]

misspelling it is meta guys

[uneven dove]

@high skiff we'll have you rejecting issues like a pro in no time

[high skiff]

He looks like billie eilish lmao

[quasi remnant]

lmao

[trim orbit]

So many people want to turn the tools into a contest. Branded releases. Mine is bigger than yours. I'm more popular.

[winter raptor]

gah the eyes, its always the eyes

[eternal fog]

I mean he's that famous just Elvis would work.

[urban fjord]

Just take the auto1111 approach and have over 2000 open issues on GitHub.

[trim orbit]

Lets just make cool workflows that suit each of our needs and can blend and sing together in harmony

[quasi remnant]

this looks cool

[uneven dove]

2k open issues means "reach" and "popularity"

[quasi remnant]

messed up eyelid tho

[sharp robin]

Idk what they are Auto has never not worked for me.

[trim orbit]

i use sytans cause its got the dual encoder. but i wonder.. why people be so aggressive about that

[uneven dove]

here we go, an AUTOMATIC1111 fanboy to defend it like clockwork

[quasi remnant]

wait

[uneven dove]

just go look at the list lmao

[quasi remnant]

sytan homie

[sharp robin]

Jokes on you ive been on Comfy now but have to give credit. Im no ones fanboy.

[uneven dove]

a1111 still doesn't have complete SDXL support

[quasi remnant]

WAS HE NOT SATISFIED WITH THIS? your wofklow literally gave the best result

[eternal fog]

My assumption is most people don't understand how to use comfyui and just want a workflow and because it's not "easy" they cry

[quasi remnant]

versus all the other ones

conceptually

the single best result was from your workflow

lmfao

[quasi remnant]

that is actually following the prompt

more than any of the other ones on the page

[urban fjord]

This is a really bad take, you shouldn't have to build your own user interface. The default package should just work.

[quasi remnant]

dude tried to make a point and looks like a clown xD

[civic sigil]

Were the most active A1111 devs given SDXL as early as comfy was? It seems like an unfair point to bring up

[sharp robin]

Ive even got up scalingand img2img workingnow on comfy. But still Auto has always worked for me. I didn’t like nmkd but like his upscalers.

[uneven dove]

haha he didn't get it because of how he behaved with earlier leaks

same reason he's not on this server anymore

[winter raptor]

My 3 fave workflows rn are Searge's, Sytan's and Mike's. And I love each and every one of you for sharing your workflows ❤️

[quasi remnant]

?

he just isn't active

he's still here though

his last post was 6 mths ago

[sharp robin]

Nope.

[trim orbit]

i think those people are fakes. his real handle is voldy

[quasi remnant]

still, his behavior with the earlier leaks progressed SD's community efforts so much

[high skiff]

@uneven dovewould you be ever so kind as to help me close this guys issue?

[trim orbit]

https://rentry.co/voldy

[sharp robin]

Why can’t we all get along and agree both tools have helped

[eternal fog]

Different tools for different people. Pretty sure comfyui was built as a backend first. The only reason people are trying to use it that don't understand it is because it was the first to support SDXL. And they aren't willing to learn it's intricacies.

[sharp robin]

And they are tools like the refiner is

[uneven dove]

https://github.com/bghira/SimpleTuner/issues make one here and i'll see how it's done

[quasi remnant]

i liked putting the time into research and figuring out the best node setup to get good results

[civic sigil]

I like comfyui but it shouldnt need this industry plant campaign for it to get popular

[quasi remnant]

comfy is neat cause it's way more tailored to your own design

[high skiff]

Submitted lmao

[quasi remnant]

and has way more accessible features

[high skiff]

it also has support for my mixed diffusion, and the special text encoders

tho diffusers also implemented mixed diff on behalf of Pseudo and I

[quasi remnant]

sytan don't even listen to that dude who opened the issue, lol. the result from your workflow in that one was the peak, even if it was just one of the images

[high skiff]

dope

[quasi remnant]

it was one of the only ones that fully followed the prompt

[high skiff]

it gone

thank you :>

[uneven dove]

np

i never have to do that, the projects i am / was community liason for, simply closed things and did not ever delete them 😛

[high skiff]

did you at least appreciate what I had to add to the convo? lmao

[trim orbit]

node based editors are inherently more technical and less "default push to go" they're like, professional camera equipment vs the point and shoot cellphone apps

[urban fjord]

I have hopes for SD.Next fork, but not tried it yet and will probably look more at it when 1.0 has been out for some days.

[uneven dove]

well Trotskyu isn't one of the unburnt

[uneven dove]

in the ticket? i did not read

[high skiff]

it was ticket issue OwO? with the description "that is all"

[uneven dove]

so used to just closing issues without reading them

[trim orbit]

its still gradio. meh.

[uneven dove]

my code is perfect

[trim orbit]

i want to move on from gradio. its a hog

[uneven dove]

any opened issues were usually a mistake

[high skiff]

all code is perfect when it isn't imperfect

[urban fjord]

Why does everyone hate gradio so much?

[winter raptor]

@high skiff do you have any recommendations on upscaling using your workflow?

[high skiff]

from what I know, they are pretty scummy people

[uneven dove]

"If you never did anything wrong, or right, and something somewhere in between - people might not 100% be very certain of the fact of if you whether, performed anything of an act at all" - Wayne Gretzy

[high skiff]

currently working on a special upscale focused beta workflow

no promises when it will be out, as I am sick

[wicked frigate]

those aren't fake, the first (automatic1) is actually him

(the other is a troll or something idk)

[trim orbit]

gradio trying to do a full studio like auto1111 and forks of it, is like using a wrench for a swiss army tool. hammer, saw, all of it.

[high skiff]

but I have one coming out for fractional offset as well

[wicked frigate]

voldy is just what 4channers call him from way back then

[uneven dove]

mmm he's gone from the HF server, not here my bad

[sharp robin]

Is there a discord he is active in? He mentioned something ab discord in one of the SdXL comments

[trim orbit]

cool thanks for confirming. i did account that there were huge name change movements recently too

[winter raptor]

oooooo awesome, super excited. And I hope you feel better

[wicked frigate]

There's a restricted devs-only discord he owns

[high skiff]

thank you :>

[winter raptor]

What's that mean? Sorry, pretty new to all of this lol

[urban fjord]

I really don't get you here, gradio has nothing to do with Auto1111 as far as I know.

[sharp robin]

Ahh damn thanks guess the answer is no then.

[trim orbit]

a1 is built on gradio

scroll to the bottom you'll see

[urban fjord]

Yes, and how is that a fault of gradio?

[trim orbit]

😵‍💫

[uneven dove]

some people like to be very trendy and hate on something simply because it's widely perceived as being bad. they might not have a true understanding of the tech, but they just go with the flow.

[high skiff]

its an experiemental change for my workflow that adds timesteps that are misaligned by a fractional different between 13/21 and 13/20

It allows for a much higher level of detail and fine textures, but cna introduce some unwanted issues at times, so I am still trying to optimize it

[uneven dove]

i really like how gradio lets you build a UI application that automatically has API access to it, when you use it correctly

[trim orbit]

metaphors are hard. i called gradio a wrench. wrenches are good for some things. auto with all the extneions people like to add to it is trying to achieve a full studio experience.

There are better platforms for a full studio experience. Other ui's are trying to build it out already. i don't see any threats to worry about

auto is popular. dw

[sharp robin]

too much salt being thrown around can we go back to being friends and posting corgis or something?

[trim orbit]

corgis are way too default

[high skiff]

without fractional offset

[winter raptor]

can we switch to dachshunds

[high skiff]

with it

much more fine granular texture

[trim orbit]

sparkls

[urban fjord]

If ComfyUI turns out to be the most performant I'll probably just slap a gradio interface instead of the node interface and use it that way.

[high skiff]

please share if you do! I may not like gradio, but I would take anything at this point to be able to use it on my phone lmao

[wicked frigate]

if only there were some easier interface over the comfyui backend that'd be available soon

IF ONLY

[winter raptor]

👀 exciting

[eternal fog]

Have a look at https://github.com/space-nuko/ComfyBox

[trim orbit]

a presentation layer would be nice butttt, gradio ? that'll take a lot of jiggering to repurpose that tool to that job

[winter raptor]

And where might I keep my eye out for said hypothetical interface in the future?

[nimble heart]

since comfy caches the base latent you can just leave it off by default and drop the refiner step by 1 when you need it without re-rendering the first 2/3 of the image

[dusk mica]

Yeah i noticed, comfy is so much faster than a1111

[urban fjord]

I'll gladly share but be warned that the code may not be for human consumption.
There is a good way of doing it and there is a fast way of doing it.

[uneven dove]

sounds like you need to write your own sampler that can adaptively scale its 'sigmas'

[high skiff]

Yeah, I do that at the moment

[civic sigil]

Dont worry the comfy brigade will let you know

[high skiff]

all good lmao

[urban fjord]

Actually no, it's barely an inconvenience.

[nimble heart]

For the public git I might add a note somewhere near the refiner node explaining it.
I 1000% wouldn't make it the default unless it's a photoreal only workflow

[high skiff]

@uneven dove I found a new song that would probably give you a heartatach, with your weak ass boomer ears

[wicked frigate]

sure is a shame no such thing might exist around when xl launches or anything :/

[urban fjord]

But yeah, I hope someone else releases it so I don't have to make it myself.

[high skiff]

oh no, it will be its own experimental branch

[urban fjord]

As long as it support any workflow that is possible in ComfyUI

[dusk mica]

I mean, creating good interface isnt easy

[high skiff]

it is not fully baked for sure

I am curious how this song embeds

Warning, may be very loud lol

https://open.spotify.com/track/01dAV8cnaC8TkJKbX46Zrn?si=529d20f4cb9e43cc

[uneven dove]

rip ears

[wicked frigate]

just utterly tragic that i don't have exactly that working in front of me right now :'(

[winter raptor]

Is the node public?

[high skiff]

ah, right at the buildup haha

[uneven dove]

type /nick hypemonkey for me plz lol

[high skiff]

jesus, spotify embeds are like 4 bit lmao

[wicked frigate]

unfortunately there's nothing for me to hype

truly a shame

[uneven dove]

https://giphy.com/gifs/bruh-washington-wizards-john-wall-Uss9jg5C9REgo

[wicked frigate]

can't release if joe murders me first

[uneven dove]

have you ever lost a USB key, mcmonkey?

like, over the internet?

[high skiff]

The alternative universe of my corgi prompt lol

[west breach]

trying to create instagram filter styles. tis tricky

[high skiff]

@uneven dove

https://open.spotify.com/track/3Tv9I9tuUvxYGjbKpEGspV?si=f9ae0f5bf6a74aa4

have another lol

OK THAT ONE EMBEDS LOUD LOL

[wicked frigate]

funny story about the time i slipped and fell and uploaded a couple files way back in late july 2023

[high skiff]

thats the last one I have lol

[winter raptor]

yay a dachshund

[nimble heart]

I'm not sure it's ever going to be fully baked. From my admittedly limited playing with, its seems the 1 step trick basically refines new details from nothing which always results in artifacts/distortions. For really busy photoreal images it kinda works but otherwise it makes the images less coherent, and this is only 29/30 instead of 19/20

[uneven dove]

that's in the future, ya damn geriatric

[high skiff]

oh come on, the song embed legit ends right when it drops lmao

[winter raptor]

[high skiff]

I am not using full step offsets

[nimble heart]

Ah

[high skiff]

I too found that a full step adds too much, so I started using fractional misalignments

[nimble heart]

so something like 1/2 step, equivalent to 39/40 ish

or what have you

[high skiff]

so sampler 1 does 13 steps of 21, and the refiner does 7 steps at 20

So there is a misalignment between the fraction end of 13/21 and the fraction start of 13/20

so the difference is whatever the difference is between 13/21 and 13/20

[urban fjord]

Sytan: What's your github so I can try to slap a gradio interface tomorrow for testing?

[civic sigil]

Sounds like a very small difference

[high skiff]

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

[high skiff]

it is, but it can do some heavenly things

just a sec

[urban fjord]

Thanks.

[trim orbit]

https://github.com/jac3km4/comfyweb i've seen this fork of comfy and okay.. 4 months since a commit. might not be up to speed. the work that was done is stellar looking though

[uneven dove]

uh oh @wicked frigate someone is going to beat you to the punch and release a subpar thing that becomes the gold standard

[nimble heart]

ah no that's what I'm doing yes. Refiner's total/end @ 20 while base total is 21. I always found the distortions displeasing on anything except busy photoreal images.

[urban fjord]

Don't worry I will probably procrastinate it and end up slower.

[high skiff]

I should also note it is EXTREMELY Sampler and schedueler dependant

[wicked frigate]

that's just an outdated fork of comfy with some css

[nimble heart]

it only works really at all on Euler or DDIM. dpm just produces smudgy jank

[high skiff]

yeah, and it works much better with certain scheduelers as well

just a sec, I have an example

[civic sigil]

Would this work with two passed of the base model as well?

[high skiff]

well, I am making one lol

[wicked frigate]

i wish we named models that well

[high skiff]

it should

[wicked frigate]

with... versions and stuff that make sense

wait what

[uneven dove]

yeah i can't wait for stable diffusion 1.0

[trim orbit]

one thing i wish i knew how to set up in gradio is switches. like, if i mute a lora, i have to rewire it to make it work. i just want to turn it off and have things pass through.

just wish there was a way to set a bunch of check boxes for various circuits

[wicked frigate]

oh, no, lol, wrong file, you grabbed the Diffusers model

the non-diffusers model is sdxl_vae.safetensors

[high skiff]

lmao

[wicked frigate]

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

[trim orbit]

looks amazing. thank you

[high skiff]

@nimble heartThis is what it looks like now with my better tuned settings

Without vs with

it adds some detail, but less overbearing than before

yes

right is with

[grizzled warren]

You reduced the refiner steps percentage in relation to the base, right?

[nimble heart]

@high skiff left is one of my detailed but not realistic bench images right is with the step less trick. Like it definitely adds a lot more details, it's just the details aren't really good if that makes sense

[winter raptor]

Now to figure out what those fields actually do lol

[uneven dove]

i agree tbh

[high skiff]

interesting

[civic sigil]

I feel like the refiner should have been more of an upscaling model than anything and might as well have been the same size as the base

[trim orbit]

i feel like i need to buy two 3090s

[uneven dove]

yep, here's 2, 3, and 1 steps on at the end

vs the beginning image

[wicked frigate]

more gpus is better gpus

[high skiff]

the way you do it is different to me

[quasi remnant]

is this vae already pre-baked into sdxl 0.9 or are we suppoed to use it alongside it/refiner?

[uneven dove]

it's not different 😛

[wicked frigate]

it's in the model, you don't need it separately

[quasi remnant]

ty

[high skiff]

ok, which do you guys prefer here?

[quasi remnant]

both have their pros/cons

[lament vapor]

A girl with white wings, long hair and blue eyes.

[civic sigil]

I wonder if anyone will pipeline the 2.x upscale model like we are doing with refiner

[quasi remnant]

i don't have a definitive winner of those 2

[grizzled warren]

you hear a thunderous screech coming from thousands of GT 730s 🤣

[high skiff]

NAUURRRR

[quasi remnant]

2 is more detailed for sure

but is also changes facial structure so much

removes ear

[sharp robin]

1

[quasi remnant]

and some of the leaves don't look as good in 2

[uneven dove]

in a thumbnail view, the right one just looks like it has a bunch of "noise" left? am i the only one. but i've seen all of mine that i've tested that technique on, look the same way

[civic sigil]

It's nice to have variations, this is the sort of thing you take as two layers in photoshop and mask out the parts you dont like

[quasi remnant]

o yea 2 also messes up shoulder

i think 1 wins

[high skiff]

alright, so the more detailed one is clearly the fractional lol

which is why I am not shipping it in main

it really is just a built in add detail LoRA

sometimes you want it, sometimes you don't

[uneven dove]

it's just that when it goes poorly i'm like, holy fuck i wish i never tried this

[grizzled warren]

Perhaps we should try a workflow which uses the refiner locally, similar to adetailer in a1111...

The baroque method

[nimble heart]

yea it's almost like the shortening of the refiner timescale makes it turn the 'leftover' latent noise into random details. Like it's expecting 7 steps of noise but really there's 8 so that extra noise is interpreted as part of the image and refined upon which has the effect of creating more random small 'noisy' details.

[uneven dove]

yeah they're not necessarily correct details, i found that prompting does the same thing more reliably

[nimble heart]

slap intricate details into the positive and profit

[uneven dove]

family photo time

why they so far apart

[nimble heart]

looks like the power rangers embedded one of those giant villains into the ground and left him there

[uneven dove]

she looks reaaaal standoffish

[high skiff]

thats the father in the middle, his husband on the left, and his ex wife on the right lol

[uneven dove]

ohhh

it's a modern thing

[high skiff]

lmfaoooo

[wicked frigate]

i was thinking the two on the outer edges were divorced and have to be a minimum distance apart

[uneven dove]

restraining orders. good thinking mcmonkey

[civic sigil]

The divorce lawyer in the middle

[high skiff]

alright, I have another test

[nimble heart]

restraining orders are usually > 10 feet though

[high skiff]

OH OR THAT LMAO

[uneven dove]

slowly working his way into the relationship with the dad

[high skiff]

lmfao

[wicked frigate]

ngl i'd want a divorce too

[uneven dove]

no one noticed the monster tho

just sayin'

[winter raptor]

Woah this is fantastic

[high skiff]

we noticed it lol

[high skiff]

which do you prefer?

[nimble heart]

:(

[high skiff]

there are 3 different things happening there

[uneven dove]

that dudes' gotta do huge poops

[winter raptor]

Either the first or second. The face is a more realistic size in the second, but I like the overall look of the first

[nimble heart]

Is all the red spots a Diffusers thing? I havne't noticed it on my own images

[high skiff]

yeah

[sharp robin]

i wonder how big dinosaur poops were

[uneven dove]

comfy breaks the license and doesn't put the watermarks on

[west breach]

need a negative to stop this ugly texture that kinda looks like sugar granules or something, kinda grosses me out haha 😄

[high skiff]

I noticed it happens on diffusers and the SDXL API

[uneven dove]

you're technically supposed to watermark images you share

[high skiff]

gross :p

[nimble heart]

I thought the watermarking algorithm was supposed to be non-perceptible

[uneven dove]

it is supposed to be, but like i keep telling you all, computer scientists stare at a screen all day

there's a new 'tree ring watermark' being looked at to replace it with

if you care about the result, go test it lmao

[west breach]

🤢

[uneven dove]

i'm fucking blind, so

i do not care about rEd DoTs

[nimble heart]

are you saying like we as computer scientists notice cause we stare at it all day or the computer scientists who made it don't notice cause they fried their eyeballs with blue light from looking at screens all day

OK thanks for clearing that up

[uneven dove]

@west breach how do you even get that dirt?

[nimble heart]

typically that happens when the dog pees in the same spot a lot

[west breach]

in pictures without grass 😄

[uneven dove]

yours has grass

[sharp robin]

happens on only base or trying to do base + refiner and trying to upscale

[uneven dove]

no refiner

[west breach]

if there is no dirt, it's fine, but sometimes the dog is sitting on a path or bit of dirt in the garden

[uneven dove]

seems to prefer to blur his dirt lol

[west breach]

it's similar issue I saw in 1.5. Any stone, dirt, concrete texture had that weird black dot fuzz grossness

[nimble heart]

"black dot fuzz"?

[uneven dove]

you might need to reduce CFG to 3.0

[civic sigil]

SDXL loves extremely narrow depths of field

[uneven dove]

that's just the base

no refiner

[sharp robin]

dog looks fine but htat grass is like fake shag rag

[nimble heart]

the ground is one large chia pet

[uneven dove]

compare it to the grass in the image above, frame three lmao

[sharp robin]

top looks like that fake grass carpet, but looks better imo

[uneven dove]

well, everyone has their vices i guess

[nimble heart]

I actually have artificial grass and it looks a lot different tbh

[uneven dove]

i hate how poorly the composition was done at 5.4, looks pasted on like MS Paint

CFG 3.0 goes way harder

those are just the first images that came out. i could keep going and make better ones

[nimble heart]

in 1.5 I had my default CFG @ 4 and just ran ridiculous batch sizes until the images came out good for that very reason lol

[indigo carbon]

Also, does a1111 support SDXL with all the optimizations already? Or is it still just ComfyUI

[uneven dove]

vlad's does, but A1111 is still struggle-bussing

[indigo carbon]

With max efficiency?

[uneven dove]

vlad's? yes

[indigo carbon]

Oh, great

[uneven dove]

you can get vlad's down to about 3.5GB of VRAM use and take a long time to gen lol

[lilac wren]

Limewire? (the worst software that never exist)

[nimble heart]

A1111 main doesn't even get 1.5 to the same speed comfyui does so I'm not sure how fast max efficiency will be over there

[uneven dove]

better than KaZaA

[indigo carbon]

With refiner? Or do you need img2img with Vlad's

[uneven dove]

well AUTOMATIC1111 is saying that his platform will not load refiner/base at once

he didn't think it's a good idea

Vlad's does it though

[nimble heart]

but high vram is the default anyways???

[lilac wren]

Ian Mckellen in GTA V

[indigo carbon]

Great! Is it a similar it/s like with 1.5 models?

[nimble heart]

Yes and no

No on low res but it gets better with higher res

[uneven dove]

it ... depends on the GPU

[nimble heart]

yea I saw some 1660 numbers that were super rough

[uneven dove]

it's honestly a lot slower across the board

[indigo carbon]

Hm.. I guess I'll find out when I get back home

[uneven dove]

it's not terrible, on higher end kit

8gb vram, fine

12gb, 16gb, better, 24gb is perfect

[nimble heart]

if you can run bf16 well then that helps with the VAE as well

[lilac wren]

Yup, 8GB is largely enought (i have a 2070 Super)

[uneven dove]

bf16 runs 3% slower than fp16 and uses more vmem. just run the fixed VAE

[nimble heart]

fixed VAE?

[uneven dove]

you can reduce it by 3GB

[lilac wren]

BF16 needs a RTX 3xxx at least

[uneven dove]

madebyollin has one

on huggingface hub

go nuts, im off to bed

[nimble heart]

aight

[winter raptor]

outpaint when

Wait does that work well?

honestly im just super excited for fully featured controlnet with SDXL, ik it'll be awhile, but boy will it be useful

[shy kelp]

maybe theyll release this bad boy early and drop it tomorrow

[lilac wren]

ControlNet Useless? Nah, just for OpenPose it not will be
i'm dumb

[winter raptor]

Eh, it was super useful for getting great results out of ultimateupscale, for outpainting, for a lot of things really

[lilac wren]

Yeah, i'm big CNET user too, can't wait

I just fear the worst, that with the size of the big SDXL models, loading more CNet models will not be enough with only 8GB of Vram

[winter raptor]

oh i see what happened lol

[sharp robin]

rate?

[winter raptor]

true. I'm on 6gb, I feel your pain brother

[pure portal]

Hello, does anyone knows that SDXL in available for Dreambooth's fine-tuning? I'm using v1.5 model and wondering I can use SDXL for better result or not

[lilac wren]

Do you know MultiDiffusion extension (A1111)?

[winter raptor]

I don't think so

[lilac wren]

There is a "Tiled Diffusion" fonction on. You can save a great amount of Vram with this

for the same gen

[winter raptor]

oh really?

[lilac wren]

y

with my poor 2070 Super, i can do a 2000x2000 easyly

[wicked frigate]

custom training SDXL models is very much a thing. LoRAs are probably gonna be more of a thing than full tuning on dreambooth though. Almost equal quality, way less filespace and way lower training requirements

[winter raptor]

I will say, comfy is infinitely more kind than A1111 to my 6gb of VRAM

[wicked frigate]

i haven't played with em much on current sdxl arch but yeah that's a thing too yes

i think kohya's trainer supports em already?

yeah https://github.com/kohya-ss/sd-scripts/blob/sdxl/sdxl_train_textual_inversion.py

[lilac wren]

Nop

There is a special branch, but not works on 8 GB

[wicked frigate]

well idk the vram reqs, but, uh, it works lol

even for SDv1 wasn't TI a 12GiB+ thing?

[winter raptor]

I've given up on ever training loras, I assume there is no universe where my 6gb vram will be enough

[lilac wren]

Kohya on SD v1 works well, i can train a face on 7 minutes

[wicked frigate]

lot easier to use TIs than to train them. at usage time they're basically free

[pure portal]

Oh thx. I followed the tutorial on below. But, I couldn't have good result with LoRA
https://www.shruggingface.com/blog/self-portraits-with-stable-diffusion-and-lora

Do you know any good resources?

[wicked frigate]

you're so close :(

only needed 2 more

[nimble heart]

Well the patched VAE reduced my decode by 15 seconds @ 2k^2 but it still takes longer to decode than it does all 3 stages of sampling so that's fun.

[wicked frigate]

maybe future optimizations will get there

but yeah not rn

[winter raptor]

I'm determined to push my gaming laptop from college as far as it will go lol

ROG Zephyrus G14, still going strong

[wicked frigate]

i'm not really the one to ask tbh, my attempts are much worse quality than nitrosocke or any of our other experienced trainers

[wicked frigate]

oof. Yeah the VAE scales really hard with resolution increases

one day we'll replace that blasted thing

[nimble heart]

It's like that on NVIDIA too? I was starting to wonder if it was a ROCm issue.

[wicked frigate]

oh it's probably better on nvidia than amd lol

[nimble heart]

1k^2 it decodes in <1 second

[wicked frigate]

but, yeah, it hurts regardless, just a question of how much

[nimble heart]

2k^2 it takes like > 1 minute

[lilac wren]

After depending on the faces you want, people have already trained them and posted them on Civit.ai, you have to look in the LoRA category

[winter raptor]

Alright well I'm gonna get some sleep. @wicked frigate you have no idea how excited that hypothetical Comfy frontend has made me lol

[trim orbit]

frostwire was a good fork once it went sour

[pure portal]

Thanks, Im trying to create my own images like Lensa, so Civit.ai might be different..

[winter raptor]

Wait what? So there is hope for my 6gb of vram? The 2060 max-q lives to fight another day

[lilac wren]

YOu have 6 GB?

[nimble heart]

we need to go back to the days where you could add extra ram sticks to your GPU

Not on consumer cards afaik

Everything's soldered as close to the die as possible to reduce latency

[pure portal]

I'm using paperspace.com to generate so I can choose wihch GPU to use

[lilac wren]

Christopher Lee as big banker of Los Santos

[nimble heart]

or you can buy a lower binned one and use a bios hack to unlock it lol.

the waifu is indeed XL

[civic sigil]

I wonder if WD is gonna restart for 1.0

[winter raptor]

Turns out I didn't sleep, and the eyes have once again ruined my image

[west breach]

black spotty grossness in the dirt gone! looks a little weird, but much prefer it over the way it normally renders that texture

[steady grove]

SDXL tanning on a beach with gigantic muscles, under a sun shade

[winter raptor]

Finally got an image with acceptable eyes

[west breach]

acceptable everything 😄

[spring fulcrum]

Is there a way to get sdxl in comfy ui to use a deeper depth of field. It keeps using a shallow dof like f/2.8 I would like to get everything in focus like a f/11 shot would be. Any help would be appreciated.

[winter raptor]

I've desperately been trying to achieve this. Hold on, I used a workflow yesterday that was able to achieve it most of the time. I can't take credit, I didn't create the technique or the workflow

@spring fulcrum

drag an drop that into comfyui

[high skiff]

am bacc

am also sick, and not feeling good

so I think I am gonna play some games lol

I haven't done that in a hot moment

[trim orbit]

factorio getting an expansion pack soon

[spring fulcrum]

how do i save my current workflow?

[lilac wren]

every image have workflow sttings inside

[spring fulcrum]

ohh

[lilac wren]

but you can do SAVE to .json

[spring fulcrum]

yes how can I do that?

[spring fulcrum]

Awesome... thank you. Now I will play around with some new workflows.

[lilac wren]

And you can copy/paste Nodes from a workflow to another, with 2 different tabs

[winter raptor]

Wait you can???

[lilac wren]

Yup

did it 100 times

[spring fulcrum]

how do you do that?

[lilac wren]

Ctrl+C/V

[high skiff]

@visual glade Sorry for the random ping, but I wanted to let you know I HAVE installed my other drives into my mobo slots lol

Just thought you would feel relieved at that lmao

[winter raptor]

You just changed my life

[autumn forum]

try shot on iphone

[lilac wren]

Another tips, if you do a drag&drop from a connexion to empty, Comfy will recommand only compatibles nodes.

[visual glade]

so it's working?

[lilac wren]

And you can do Ctrl+clic to select several nodes, Shift+Clic to move them all

[high skiff]

yeah, but I found out I lost all of my data from my root drive from my previous PC, and I am trying not to get overwhelmed at the thought of how much stuff I lost

[autumn forum]

if you dont remember it doesnt matter right? 😛

[high skiff]

Oh I remember, thats for sure, just trying not to

[autumn forum]

ah

[high skiff]

lost years worth of work I will never get back

[spring fulcrum]

Have you tried to put your old drive into a linux system to see if you can recover... I use linux all the time for recovery when even acronis won't recover on a windows system

[high skiff]

100's of 3D models, dozens of programs, 100's of sounds and samples

[visual glade]

how did you lose it?

[autumn forum]

ok yeah that would be hard to lose.

[visual glade]

dead drive?

[high skiff]

no, it won't get anything, I tried about 10 different softwares, software that worked for me before

[visual glade]

or overwrite?

[winter raptor]

Is it possible to use clip interrogator in ComfyUI?

[high skiff]

My PC had the wrong ID assigned to each M.2 slot, so the slot that was coming up as my D drive was different on the physical board

So I ended up formatting the wrong drive for windows

It ate everything, and I didn't realize, so like 300GB has been rewritten over. Tried tons of recovery softwares, and they all get just a few files, even with extreme deepscans that I let run for hours

[spring fulcrum]

I'm sorry to hear that... I have kind of a crazy setup at my house... I have a server with 108TB that I use to back up all of my systems at home among other things that it runs

[high skiff]

so on my mobo they were drive bay 0 and 1, and in windows, they were assigned as 1 and 0

so my 0 drive, which is what said was my D drive in windows, was in my 1 slot on my mobo

@lilac wren If you really think its funny that I lost years worth of work, then screw you man. Have some compassion FFS

[autumn forum]

need a nas my man. something to back up to

[shy kelp]

Woo I just got access to SDXL on huggingface

[lilac wren]

Sorry, but it's my job so it makes me laugh. Reminds me of all those times it happened to me. I know that feel

[boreal bough]

ufff rip ❤️

[lilac wren]

or a simple Raid 1

Everyone takes NAS as if they have ultra sensitive data from NASA. The truth is that people are paying a fortune for NAS for nothing.

[boreal bough]

NAS does have huge advantage of being queryable even if the pc is off

[steady grove]

NAS is just network attached storage.

[boreal bough]

good for hosting movies for TV

[autumn forum]

you can get a nas for like 200 bucks lol

[high skiff]

luckily I was responsible enough to back up my super important files, otherwise I would be ruined. But I didn't have enough space to backup everything. I was able to store the most important stuff on that drive

[steady grove]

can be any configuration you want. my dream nas would be raid 69

[lilac wren]

And HDDs for double

[high skiff]

oh wait

maybe I can make my old PC a NAS

[autumn forum]

maybe if your buying a 20tb drive

[lilac wren]

For a NAS, it is rather recommended to have a low consumption config, and a linux base installed on it

[boreal bough]

you can just get any NAS enclosure from QNAP that meets your demands - then put in any HDD/SSD you get from amazon.
It's not like you're in a corporate environment where you're accessing it 24/7

[high skiff]

I am not gonna be doing linux any time soon

my old PC I have would have to work, its either that, or nothing

[west breach]

i bought an old office computer which i use as my 'nas'

[high skiff]

my only concern is backup times lol

but this new PC has enough cores to not be completely bogged down when data transferring

data transfer on my old PC would screech everything to a hault

[spring fulcrum]

I run an MSP for B2B IT and I had a younger guy working for me. He was trying to get this guy on Microsoft 365 from a local email setup. He didn't backup the PST file for a CEO of a multi-million dollar corporation. It was corrupted on the transfer to the cloud and this guy lost about 10 years of really important company email. Caused me a lot of headaches and awkward meetings. I know the feeling.

[high skiff]

its not the SUPER important stuf, thank god

[west breach]

when upscaling goes horribly wrong 😄 (mouth)

[high skiff]

I just lost about a month of music stuff

which compared to the 7 or so years of stuff I have still is not that bad

but I did lose my cutting edge new bass processing rack

thank fucking god I did not lose my crystaline rack, that would have been devastating

[spring fulcrum]

some of these are coming out pretty good... check this out.

[high skiff]

The ones that hurt the most at the moment is all of my 3D printer config files, all of my fusion 360 models/project files (including the one I am like top 50 in the world for), and my cyberpunk 2077 save file, which is seriously hurting right now. I was playing that with one of my best friends. Like 60 hours of meticulous grinding and gameplay gone

oh thats not good-

it crashed lmao

OH

I have my Ai overclock on it lmao

man, I am gonna be able to play a youtube video and not drop big frames in games anymore

[spring fulcrum]

all right folks have a good night i'm out for today

[high skiff]

holy hell

my 3080 is WAY FASTER in this new PC

I must have had some severely fucked GPU drivers or something

Before in CP2077, Ultra with no RTX was pinning my GPU at 100% and stuttering

Now I run ultra WITH RTX at just 75% GPU usage

[shy kelp]

I used to play maxed out cp2077 with rtx on medium at 1440p dlss on just a 3060 laptop

Got like 40fps but thanks to freesync it's not bad at all tbh

[boreal bough]

'maxed out'

is a big statement XD

I get 14~16 fps on genuine maxed out with a 4090 and i9 13gen

pathtracing go brrrrrrr

[void loom]

Any idea what time the official SDXL models will be released tomorrow for download? will it be like 12 midnight UTC or some time in the day? ... or a pre-leak today perhaps? XD

[ionic dragon]

you got a 3080?

you said that you were getting 3090

[upbeat summit]

greetings

[high skiff]

I have a 3080 right now

the 3090 comes in tomorrow

[boreal bough]

we didn't get any official statement
odds are high they'll just do it during the workday once everyone has given the final ok XD

[nimble heart]

are the DLSS settings the same both times you tested?

[high skiff]

yeah, both on auto

hell, let me turn it off

[nimble heart]

maybe this time it auto'd in favor of lower res

[high skiff]

it was like 30 FPS on the other GPU without DLSS

[nimble heart]

3060?

[high skiff]

3080

I always thought the perf seemed sketchy

yeah, right now, ultra maxxed with no RTX and no DLSS, I am still only using 70% GPU to run 75fps locked at 2560x1080 lol

before even with DLSS it was struggling hard

I can only assume I had some royally fucked drivers, cause my GPU also lost the ability to idle a few days before I upgraded

[nimble heart]

wonder if the power limit was messing it up some. If it gets too low sometimes its not just a performance cut, things straight stop working properly

[high skiff]

I did have to power limit the card, as it would turn my PC off, but even then

this is using 220 watts to run, I had the card limited to 280 watts before

its very weird

and I even remember being super disappointed when I got it, cause I used to get better perf on my 3060ti with RTX on on the same settings, but it all makes sense now lmao

I figured they just updated the game a ton and it was much slower now

[nimble heart]

3090 should about be able to do that with but raytracing on iirc

[high skiff]

yeah, I will test tomorrow when it gets here for sure

hell, I should be running the CP2077 benchmark

[nimble heart]

think my XTX does 4k @ "balanced" FSR which is 1080p upscaled I believe, and XTX's usually bench similar to 3090's with raytracing depending on the game.

it's a 4090 with 3090 raytracing basically

[high skiff]

oops, started the bench with DLSS maximum perf on lol

[nimble heart]

maximum perf is basically soup

[boreal bough]

voltage? You did mention underclocking once

[high skiff]

gonna try it with the most maxxed out settings in the game with no DLSS and see how it does

[high skiff]

yeah, i did have to do that, but even then, it used more power for like half the perf

it was very weird

Ok, doing max bench with no DLSS

[nimble heart]

yea max everything first before RT. The RT performance hit is so strong you dont really lose anything else by having everything else on ultra

[high skiff]

I just wanted to see

[boreal bough]

oh yeah. in that case definitely a power issue.
Dont forget that your PSU can handle significantly more than advertised - it just REALLY shouldnt for extended periods

means you were running way above max

[high skiff]

Cause with psycho maxed out RT and no DLSS, it's still getting 45FPS

Which makes my unstable 75fps in the last PC with no RTX and full DLSS really weird

[nimble heart]

80+ plat's can usually go 100 or 150 watts above rated

[high skiff]

no i wasn't

this GPU just has massive transients that were peeking it

[upbeat summit]

@boreal bough So how did your test training go for your DND Character Maker XL? 🙂

[high skiff]

GPU would use like 300 watts on a 650 watt GPU, but peak at like 800 watts for a single transient

the transient triggers surge protection

[nimble heart]

?

[boreal bough]

that was my first one XD it had the most fails since I tried using old methods
but it works great currently 🙂

[nimble heart]

does the 3080 use the 12 volt atx cable? how would it ever draw 800 without melting

[high skiff]

thats why modern NVIDIA and AMD gpu's have such high PSU requirements

[high skiff]

oh yeah, they do all the time

when they boot up, they hit a massive power transient for like 1ms before lowering considerably

[upbeat summit]

just don't turn your computer off 😉

[high skiff]

you can see one here

a small one compared to some

3080's peak over 650 commonly, thats why their minimum recommended PSU is 750 to not trip surge protection

[boreal bough]

that spike checks out. 800W is not normal

[high skiff]

and I have a super OC'd 3080 at stock

this is from LLT for a 3080 they were reviewing

close to 850 watts right there

[nimble heart]

I had watch cat /sys/.../power1_average running for a while when I first got my XTX and never saw it above like 420 ish watts

[high skiff]

you wouldn't be able to see it

they are sub 1ms peaks

you would have to be running above 1000hz sampling rate for it to show up on a graph

here is another

3080 hitting 600

(The non OC'd version of my card actually)

[boreal bough]

OC thing maybe? cause over double should not happen - as that would turn off your pc, as you encountered

[high skiff]

my GPU stock is rediculously OC'd

[nimble heart]

is that monitored @ the pci cable or is it for the whole system?

[high skiff]

it can hit like 2100mhz if you give it enough power, completely stock

they also note that they are measuring at 20mz, so their spikes are severely cut down compared to what the PSU sees

[nimble heart]

think mine hit 3200 at one point 👀

[high skiff]

so the 1ms spikes are measured and spread out over 20ms

[boreal bough]

can vs should
I also run my ram at slightly above average, rather than the advertised max OC thats on the package

It's not about if it runs - its about diagnosing the rare issues that can pop up, which you'd never attribute back to OC

[high skiff]

the OC on this card is stock, so I had nothing to do with it

[nimble heart]

i tried to run my ram a slightly above and it bricked my bios

[high skiff]

this 3080 has all the perf ringed out of it

hell, most 3080's are 300-320 watts, I have seen this bad boy hit over 400 watts average draw

it uses more power than my new 3090

[nimble heart]

my averages are weird. Cyberpunk can average like 400 or so if I leave it uncapped but blender is like 270

[high skiff]

yeah, different pipelines use drastically more or less power

like furmark, furmark is a power virus

its designed to max out your total power to make sure your PSU can keep up

[nimble heart]

pytorch is a sorta roller coaster between like 320 and 400

[upbeat summit]

you should at least run furmark and prime95 for a bit to check system stability

[high skiff]

ok, gonna test identicla settings to on my old PC with this same GPU just to see

[nimble heart]

I did. Ran everything I could, heated up my room by like 15 degrees.

[high skiff]

Alright, running the same settings as on old PC with no DLSS

[nimble heart]

I think PSUs are oversold. Everyone at my work was calling my a psychopath for running an overclocked 7900 XTX with an overclocked 7900X on an 850w but its absolutely fine.

[upbeat summit]

as it should be :]

[high skiff]

Yeah, 100FPS locked with no DLSS, before on my old PC, I was getting severe frame drops and stuttering at same settings with DLSS set to auto

I can only assume it was drivers before

cause like I said, my GPU lost the ability to idle power drop a few days before I moved to this new PC

[nimble heart]

I actually got rid of my space heater. I built my PC in february and it produces so much heat I just do some stable diffusion batches or something to heat up my room. Better than just wasting electricity lol

[high skiff]

no matter what I did, it was always at max core voltage and max clock, even at 100% idle

[nimble heart]

"damn it's cold in here. Let's use A1111 to make a 16x16 grid of every sampler in the entire UI across a bunch of prompts to see which is best"

[upbeat summit]

haha - yeah. when SD 2.0 came out in november I didn't have to turn on heating at all 😉

[high skiff]

For me it's a curse

[nimble heart]

speaking of I haven't checked but I swear ddim in auto produces garbage results compared to comfyui. In comfy it's basically the same as euler on normal scheduler

[high skiff]

I have to power limit my PC as much as possible to not kill myself with a heat stroke

[nimble heart]

it got to 110*F today so im not exactly in a good spot either

[high skiff]

I live in a stupid hot place, and installing my in window AC unit is hell in this new house

[high skiff]

Feel you lol

[nimble heart]

damn you in SoCal or something lol

[high skiff]

And my room is the last on the AC route, and I have sun from the front at sunrise, above at mid day, and to the right at late day lmao

[nimble heart]

my phone charging cable in my car this afternoon was all soft and pliable. was kinda unnerving

[high skiff]

High desert

Devil's ass crack lol

It was still 98 a few days ago at midnight lmao

[nimble heart]

DAMN okay

[high skiff]

It's 81 today at 1AM

[upbeat summit]

had 100° F / 37° C 2 days ago = no thanks

[nimble heart]

82 @ 1 am here. 98 @ midnight is wild though

[high skiff]

eheheheeyehehehe

[high skiff]

that was an abnormally hot night, to be fair

so I need to get my AC unit installed in my window as soon as possible to keep up with SD gens

without having a heat stroke lmao

[nimble heart]

we're pretty close actually

[boreal bough]

meanwhile I experienced Ragnarok yesterday.
Sirens went off at midnight as if a war had started. shit got blown up by thunder everywhere. random hail. power outages cities wide XD

[high skiff]

slaayyyy

yeah, so with DLSS balanced, my GPU is doing 2x the perf it was doing in the old PC with DLSS auto (which should be able to make up for all frame drops)

[nimble heart]

I'll take the hail if you don't want it

[high skiff]

same

its so dry here

💀

tho the dry is fantastic for my 3D printing

[nimble heart]

yea we cant even have grass cause its too dry

and my 3d printer is still in a box so Im not even benefiting from it

[high skiff]

feel bad for people who 3D print in humid environments

Having to spend $100's on humidity reduction and air dryers and stuff

had my filament on display open air for nearly 3 years, never had any moisture issues haha

[nimble heart]

downsides: 3d printing bad
upsides: pet gecko

[high skiff]

I wonder what my PC would get at max DLSS with the worst settings lol

[nimble heart]

a bottleneck

[sharp robin]

Why is humidity a problem for print?

[high skiff]

filament

[boreal bough]

unhappy plastic

[nimble heart]

sticky goo

[high skiff]

3D printing filament loves to suck in water from the air and boil when printing

[sharp robin]

I thought plastic likes watter. More bendy

[boreal bough]

dryer + 30 min time solves it though - but as sytan said, costs between 30~100

[high skiff]

Ok, def got a bottle eck lmao @nimble heart

[sharp robin]

Ooooohhhhhh boil. Steam

[high skiff]

Tho I have to say

[sharp robin]

Ahhh got it.

[high skiff]

The game doesn't look bad at worst settings lmao!

[boreal bough]

oven is also a solution, if you're brave enough 🤣
assuming it can stay at 60, and not randomly spike up

[nimble heart]

most games dont nowadays. all low with shadows and textures on medium and you're most of the way there

[high skiff]

yeah, bottleneck lol

god

so many boxes

2x the threads, and 2.5x the cores of my old CPU, tis heaven

[upbeat summit]

nice - enjoy 🙂

[high skiff]

I am

tears of happiness

WOOOOOOO

[shy kelp]

Majestic suss jet 😭

[high skiff]

indeed, very majestic sus jet

[nimble heart]

I am the storm that is approaching