✨｜sdxl

[west breach]

1.0

[uneven dove]

@upbeat summit

[high skiff]

ooOoOoOooo

[trail bay]

it is like a frog monkey superhero

[uneven dove]

ooh

[eternal fog]

lmao

[upbeat summit]

a stunning portrait of the cookie monster

[eternal fog]

Theres so many people where it's really close, but just off

[trail bay]

like there was a 3way between the hulk, lesser monkey, and a frog

hey does anyone know the hour/s of the day that most of the significant 'announcements'/releases occurred?

by Stability

[upbeat summit]

SD 2.0 came out about now in November

[trail bay]

I want to see what they'll say for 'recommended' workflow

like how to use the models and how to do the styles, etc

[sharp robin]

• 😱

[high skiff]

The workflow I started using for SDXL is not recognized as the "correct" way to do it by staff, so I will be hoping they suggest mine haha

[trail bay]

well I think it is great that we can do whatever the hell we [all] want, but I also would rather start at some place that is pretty good based off of the bazillions of votes and data they gather from here haha

[dense chasm]

have u tried your workflow with load image and vae encode nodes,i mean img2img,i just tested.it can't work well

[trail bay]

and then move from there especially when I see more rigorous comparisons

[livid basin]

I'm also collecting SDXL prompts, currently collected like ~15000 over the last week

[trail bay]

woah

[livid basin]

There is also DiffusionDB, with 14 million prompts

[trail bay]

I'm guessing that has/will have a SDXL 1.0 section/tag/filter soon?

[livid basin]

I hope so

[trail bay]

community is about to go crazier than it already is very soon

lol

[livid basin]

I'm getting like 2000 new prompts per day, since I only collect Showdown prompts

The regular bot-X prompts are super nsfw

Showdown is like 7% NSFW by my calculations

[trail bay]

soon to be like 50% 😛 hahaha

[livid basin]

I hope not

[trail bay]

well j/k not sure they would do that here anyways. but NSFW makes the world go around AFAIK. but I can appreciate NSFW and SFW.

[high skiff]

I have not at the moment, however I do have some ideas as to how I could potentially get it working p

[dense chasm]

can't wait for you sharing

[molten breach]

How long after SDXL drops can we start asking about SD 3.0?

[trail bay]

When SD 3.0?

[sharp robin]

So u know what works for sdxl nsfw

[livid basin]

Yeah I do

[molten breach]

Soon(tm)

[trail bay]

SDXL 1.0 Soon®

[livid basin]

I can tell you that just adding 1girl increases the image's NSFW score quite a bit

In the prompt I mean

[livid basin]

And also women outnumber men 4:1 in the showdown images

[eternal fog]

It's not difficult at all to get nsfw on SDXL

People still crying about it because their "gape" tag doesn't work

[sharp robin]

So danbooru tags?

[livid basin]

yeah

that tag alone can strongly skew the image

So please, don't make images with that

Cuz I want clean images in the showdown for my dataset

[trim orbit]

people are gwoss.

[livid basin]

gape might not, but leg spread sure does, I've got hand crafted filters now

And people asking for "anatomy"

As in literally, anatomy

[frail mist]

Ok, it's christma-- I mean it's SDXL day!

Daaad are we there yet?

[upbeat summit]

gn8 all and have fun! ttyl - today's gonna be the day!

[eternal fog]

When I tested. It would do upper but not lower.

[livid basin]

yea it introduced changes to the top

[west breach]

i found a way to fix the eyes!

[livid basin]

wow so realistic, please teach us

[high skiff]

Alright, I am here specifically talking for now

mom and i are out to dinner and I can't use comfy without a flat surface haha

[autumn forum]

pumpkin grimlen

[deft quiver]

Any advice for killing the text that likes to pop up in various places? It's present in almost all generations seemingly regardless of sampler.
I've tried heavily biasing the negatives with things like
"symbols, watermark, text, writing, signature, artist signature, "

but it seems to be an almost ever present part of certain prompts, seemingly regardless of seed

[uneven dove]

sounds like something old people do lol

[west breach]

Hi Helen, how ru?

[high skiff]

I will old people your face >:C

lol

https://tenor.com/view/cat-eating-fast-eating-eat-gif-25232258

[high skiff]

damnnnn, this restaurant only has one waitress right now

[uneven dove]

#🌶｜off-topic

[high skiff]

I wonder what happened to the others, they are super overwokred right now

[uneven dove]

he has no SDXL available to him right now

he is bereaking the rules

[high skiff]

come on man, don't be as bad as the mods lmao

I almost said something a little problematic, so I changed my message before sending

also ew, I do not recognize you

Your PFP looks like a botched Bea Arthur lmao

[uneven dove]

its betty white

stunning portrait of an overworked waitress at a restaurant, cinematic, epic

[west breach]

helen mirren

[high skiff]

Really? It looks so much more like Bea Arthur lmfao

[uneven dove]

that'd Bea a really bad fuckin guess

Arthur any others you'd have?

[high skiff]

well it sure as hell doesn't look like betty white lmao

[west breach]

i'm pretty bad with celeb names, but that's what came to my mind

[uneven dove]

or Bea

[high skiff]

Thats not what Bea Arthur looks like

[uneven dove]

doesnt look like her at-fucking-all

correct

neither are

[high skiff]

I think it looks like it could pass for a gen of Bea lmao

[uneven dove]

wood carving of betty white. see? it's my PFP face

i think

[uneven dove]

i'd totally eat cookies out of that

[uneven dove]

@high skiff

[high skiff]

chapples

[uneven dove]

hmm, bear cat

[civic sigil]

Talking Shid Taking Shit

[uneven dove]

@high skiff bear cat 😁

[hot dagger]

What time is SDXL dropping?

[winter raptor]

What is the deal with the pruned fp16 model I keep seeing? What are its advantages, and is it publicly available?

[uneven dove]

it's already out! we've already been using it

[hot dagger]

Link?

[uneven dove]

bear cats

[winter raptor]

He's doing a little bit of trolling

[hot dagger]

🥲

DONT play with my feelings woman

[wicked frigate]

pruning and converting to fp16 are ways to make a model easier to load, and you can do em yourself with just a lil bit of technical experience and/or any updated UI that supports it (eg auto webui's checkpoint merger can do it).
If other people are reuploading models like that, uh, I don't think that's currently permitted for the research release due to the temporary strict licensing so probably avoid those for now

[civic sigil]

Good to hear confirmation that that's temporary lol

[winter raptor]

Ah ok, thanks! For doing that myself, will the old version of a1111 work for that, or is there an updated version I would need to use?

[winter raptor]

This is fantastic!

[west breach]

mcmonkey is typing

[uneven dove]

hindustani ghostbusters

[winter raptor]

Wait to be clear, would I be permitted to do that for myself and not upload it?

[civic sigil]

If it's strictly for research purposes I'm sure

[old mulch]

Where is the 1.0 release

Wasnt it supposed to be july 18

[civic sigil]

Yeah

[old mulch]

Its already 5h late in london time

[trim orbit]

WHERE THE HELL

(╯°□°)╯︵ ┻━┻

🔥

riiiiot

[civic sigil]

They didnt say 12AM

[trim orbit]

THEY SAID THE 18TH and LIED

[high skiff]

its still the 17th here lol

[trim orbit]

so meanwhile, i'm using this pheonix optimizer on 0.9 loras and wowwweeeee it works good with 10 images

[high skiff]

I also think people should note that its not guaranteed to launch today lol

[high skiff]

oh? Whats this you are speaking of?

[uneven dove]

set your clock forward to get access to 1.0

[high skiff]

I am gonna go home and mess with some SDXL LoRA's

[trim orbit]

newest kohya has a new optimizer thats a derivative of dadapt

[high skiff]

grated, 1.0 should be releasing very soon lol

[west breach]

It's July 18th 6:15pm in Kiritimati Island

[wicked frigate]

as long as you're not uploading it you can do anything you want lol

[trim orbit]

0.9 loras should be able to merge with 1.0 loras so not bad to mess with

[hearty ginkgo]

does anyone have a good sdxl workflow?

[uneven dove]

wink wink

[high skiff]

yes

I am away from my PC at the moment tho

[hearty ginkgo]

ok nw

[high skiff]

I am gonna go home and write up expanded documentation on how to prompt

[shy kelp]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 🙀 🙀 🙀 🙀

[trim orbit]

we're on the cusp of a release event. what happened if they had dropped it at midnight. not very prepared

[winter raptor]

Do you plan on adding a text l field to your custom node? I'm loving the styles you created btw

[civic sigil]

Wtf 40 gb?

[grizzled warren]

Imagine that MegaModel guy on Civitai right now 🤣

[high skiff]

I will nuke you lol

I am at an akward middle ground where I have no idea if the final release will have a refiner

[high skiff]

if not, then my workflow is useless now and diffusers changed their code for nothing

[thin nova]

half expecting them to merge sdxl into sd 1.5

[high skiff]

Cause I would put out my workflow later tonight, but I have 0 idea how compatible it will be with 1.0

[trim orbit]

it'll blow people's computers up. they're all gonna laugh at you

[high skiff]

Thats not even possible 😅

[winter raptor]

Teasing us

[urban fjord]

Emad mentioned a new refiner on Twitter, so trust that as you want.

[west breach]

The way it works now is it concatenates the prompt from the styles.json file with the user prompt and feeds that into the clip_l

[high skiff]

He said its done, which means it could be done as in done training, or done as in this is the end of the road for it

[thin nova]

just average the weights, bro. just add them all up and divide by two. new model is torch.tensor([3.1253])

[old mulch]

They are based in london. So they are late but not that late lol

[trim orbit]

they'll release the beta refiner renamed as 1.0 lol joking, but it still works if that's the case

[high skiff]

Yes, but the latent scaling would be very different

[winter raptor]

ahhhh i see

[high skiff]

SD 1.5 is a square 512x model, SDXL is a 1024^2 equivalwnt multi crop model

[grizzled warren]

Chances are the 0.9's refiner should work 1.0's latent images, so even if 1.0 refiner won't be a thing, we still can use it to add a little bit of details.

But that might not be mandatory

[high skiff]

That is true

yeah, I am in an awkward middle ground

cause on one end, I want the refiner otherwise my contributions to comfy and diffusers are kinda pointless, but at the same time, training LoRA's and such would be so much easier with no second model to worry about

[trim orbit]

will people train loras for the refiner you figure?

[grizzled warren]

I'll be in the same awkward middle ground until tomorrow lol

[high skiff]

Thats the biggest concern

[trim orbit]

i had embeddings working in comfy but can't get them working in auto

ones i trained

[high skiff]

The idea that eveyr LoRA would need to be trained for the base and the refiner

[civic sigil]

The license tho

[high skiff]

embeddings as in TEs?

[trim orbit]

yeah

[high skiff]

*TIs?

[urban fjord]

I can't read this as anything but done training.

[high skiff]

oh, thats very nice to hear, how have they been working in comfy?

[trim orbit]

kohya has support to train sdxl TIs

pretty well! i could tell it had effect but my training params i know sucked

[high skiff]

alright, in that case I would likely agree, tho Emad is about as reliable as a house of cards in amn earthquake

[trim orbit]

"you're finished! DONE! you hear me? You'll never refine in this town again!" -emad prbably

[high skiff]

I would not be surprised

[civic sigil]

We made a new type of refiner not compatible with Sytans workflow, we just finished it

[high skiff]

also, flowwolf, what happened to you? Why are you new here now?

[urban fjord]

The refiner has been sacrificed to this thing and is now finished.

[high skiff]

lose your account/leave?

[grizzled warren]

Well, I am not using it commercially now anyway, and it's unlikely that the model police breaks into my house to remove my fp16 pruned safetensors copy, in case SAI won't release the updated version xD

[trim orbit]

left. i was wasting a lot of time here and got fed up. but now i'm excited again

[high skiff]

Ok, you don't have to come for my throat with such hot guy gens, SMH

[high skiff]

fair enough haha

alright, so when I get home I will be writing better documentation on what I have learned for SDXL in general related to my workflow

including info on why I made the choices I did, and some ways to better suite other styles than just realism

[grizzled warren]

That's good.
I've heard someone claiming 1.5's TIs work, but my testing shows they are extremely erratic, since they only affect Clip_L. Some of them are erratic, others are just... Goofy?
So even though they do work, they aren't useful with SDXL 0.9

[high skiff]

I would be interested in trying out TI's if they end up working as good as some people are theorizing

[trim orbit]

well, like, one of the text encoders is the same clip used in 1.5 right?

legitimately asking, i dont know if thats right

[grizzled warren]

Nah. Even in theory, they'd only work for style embeddings.
And even those don't work properly for some reason

[high skiff]

Oh also, I would like to mention that I will be releasing a more complex workflow, and a much more simple workflow with only 1 text encoder for those who seem to think my workflow is terrible cause it can't work well with single word prompts 😅

[high skiff]

yeah

I believe its Clip L

IIRC, clip L is the same from 1.5, and clip G is the new natural language TE

[grizzled warren]

I have quite a collection of TIs to test. There might be some random lucky one that does work, but that's pure luck. Most of them either don't work at all, or they reduce the output quality.

[trim orbit]

textual inversions on clip L i could see being some kind of effective. maybe there is a translator that couldd be made to affect the dual encoder with old embeddings

[high skiff]

which is how they are split and addressed in my workflow, and I will be adding more documentation towards that as well

[uneven dove]

errr

if you have TIs to test, are those from SDXL?

or are they 1.5 TIs? because those will not work, you have.. ya know, nvm. i'm too lazy

[trim orbit]

i've heard the murmurings that people have old embeddings working too, but thats one of those i'll believe it when i see it things

[grizzled warren]

No, I am talking about 1.5 TIs claimed to be working.

Technically that claim is correct, to a point. But it's not really useful.

[west breach]

prompt: snoop dogg, style: padam

[civic sigil]

It would make sense i they work for CLIP-L I guess

[west breach]

crayon?

or color pencil

[winter raptor]

color pencil, it's a fantastic style. I wanna thank you again for making your workflow available

How do I update it btw? Can I just git clone it again? @west breach

[west breach]

open a terminal in the folder and run git pull

[winter raptor]

ignore the horrifying faces in the background

[west breach]

nice!

I need to work on the modeling-compound style

[winter raptor]

I will say, if you manage to get a photorealistic style that preserves a deep depth of field, you will have solved a ton of peoples' problems. I've seen a ton of people trying to get it working @west breach

[west breach]

do you mean without the background blurred?

[winter raptor]

yeah

[dense chasm]

i doubt sdxl has a violation on celeb generation

[winter raptor]

From what I've noticed, the prompt "shot on iPhone" OCCASIONALLY works

or at least helps

[west breach]

ok interesting. I'll give it a go! 🙂

[winter raptor]

Thank you!

[high skiff]

already at 709 words in my documentation for SDXL, and I have a ton more to share

[autumn forum]

thought id share some of the images i made tonight

[trim orbit]

icky!

[fair crow]

Would be interested in this as well!

[ionic dragon]

i am half day through and still didnt get 1.0 😭

[high skiff]

@hearty ginkgo@fair crowhttps://github.com/SytanSD/Sytan-SDXL-ComfyUI

Here you gusy go. Please stay tuned as I work on new information, and a couple new workflows to upload as well, including one that plays with my fractional offset diffusion for better tonal contrast, and also a simplified workflow that utilizes just one positive text encoder for those who do not want to mess with dual text encoders

[fair crow]

Thanks man, truly appreciated!

[high skiff]

just reached 1000 words on my new documentation I am writing up

[high skiff]

No problem! Hope it serves you well!

[fair crow]

They are in London right? So probably still a few hours until they release it

It’s like 6:30 am over there

[ionic dragon]

hopefully

[autumn forum]

is there suposed to be something new on your github?

[high skiff]

not at this moment

just sharing for other people who don't have it yet

I am working on new stuff tho

[autumn forum]

oh im dumb yeah i see that now lmao

[high skiff]

my SDXL documentation is now 1130 words

growing rapidly

[autumn forum]

nice, thanks for your efforts! they will not go to waste.

[high skiff]

Thats the hope haha

1400

[hearty ginkgo]

Thanks very much I have been using this one since last night added the ultimate upscaler node also I sent you a dm so u have it if you don't mind when you finish the other workshops and don't mind sharing please give me a dm or @ me

[high skiff]

WIll do. I will have another workflow with ultimate upscale baked directly in at some point, though I need to figure out some strange issues with it before opening that to the public

[hearty ginkgo]

thats what mine looks like lol

[plain flax]

Does SDXL understand natural language? I was using 1.5 & 2 mostly with comma separated keywords but no setence.

[hearty ginkgo]

dont ask why the steps and batch is so high lol

[civic sigil]

Great, Comfy bugged out and refuses to work for my workflow now

[naive goblet]

It won't load any workflow for me

[jaunty valve]

I love this art style you've achieved

[civic sigil]

Probably going to have to throw away 30 minutes of tedious work because of the buggy primitives

And I dont even know if I can add all the features I want now

It's almost impossible to add inputs to existing nodes

[lusty raptor]

i know this feel 😦

[winter raptor]

Thanks!

[civic sigil]

I wish it would give an error instead of the queue prompt button just not clicking

[autumn forum]

yes

[civic sigil]

I can get it to run now thankfully even though its not an elegant solution

[high skiff]

yes, one of the text encoders is made for natural language

[civic sigil]

Maybe someday I will clean up this massive animation workflow and put it online

[high skiff]

I need prompt ideas people :p

open to all styles now

not just realism

[civic sigil]

Go through your outputs from 1.5 lol

[high skiff]

preferably nothing too extreme that will really stress SDXL lol

[vital python]

When will it be publicly released ?

[keen charm]

so cool

[high skiff]

they are all hot guys and realism gens lmao

[high skiff]

July 18th, likely the date where you are

[vital python]

Oh so today

I'm new to this and not sure how it works but, once released, can I use the checkpoint or safetensors in Automatic 1111 ?

[high skiff]

@west breach @boreal bough You two fella's got any ideas for test prompts?

[hearty ginkgo]

Linguistic: A thrilling photograph capturing the fierce battle as a pirate ship sails through stormy seas, cannons ablaze and crew members engaging in a dramatic fight against rival vessels.

Supporting: maritime adventure, historical reenactment, naval warfare, dramatic lighting, crashing waves, billowing sails, intense action, immersive storytelling, high energy.

Fundamental Negative: Blurry motion, lack of dynamic composition, unconvincing battle scene, dull colors, poor detail, lackluster atmosphere, uninteresting framing.

[high skiff]

oh wow, you got it all formatted as I need and everything haha

[west breach]

a man doing a cartwheel on a bicycle 😄

[civic sigil]

Idk about those negatives bro

[high skiff]

I will have to trim this down as it is pretty intense, but I can work with it!

[hearty ginkgo]

try them

[civic sigil]

Maybe I'll give them a shot but they seem sus

[hearty ginkgo]

i was lazy lol i just inputing systan prompt guiding and the orginal postivitve supporting and netigtive terms into chatgpt

[deft quiver]

anyone else having issues with the refiner adding text to seemingly everything?

[lavish sinew]

@slender coral You excited for SDXL?

[clever verge]

@deft quiver Yes, when CFG is high I've noticed it.

[hearty ginkgo]

i just did those it was ass ngl

[deft quiver]

Even at cfg around 3-4 it seems to crop up alot, you can see the patterning crop up in many places as well. Really unsure how to fix

try using an unprompted unguided refiner pass but that actually made things worse

[high skiff]

I am redoing the prompt susbstantially, but you have me the inspiration haha

[hearty ginkgo]

lol

[high skiff]

trying to turn it into a wood carving mural lol

I wanna give prompts that have very different styles for people to poke at

I am pretty happy with this haha

[hearty ginkgo]

this is the best i got with midjournery that your competition @high skiff

[deft quiver]

yeah no idea, it's almost like the prompt is inherently too noisy

[winter raptor]

I love this

[high skiff]

I don't try to compete with midjourney. These are supposed to be prompts just to show people how to use SDXL

[sharp robin]

wow that a lot of text everywhre

[high skiff]

oh wow, no joke

that is very very weird

[hearty ginkgo]

midjournery is a pain in the ass to compete with

[deft quiver]

yeah I have no idea how to fix it

[clever verge]

What CFG is this?

[deft quiver]

8, 8, and 6

for both refiner and base pass

[deft quiver]

would someone sanity check that this isin't just my generation setup
"Cosmic Shark Octopus Warrior, volcanic explosions in the background, cinematic masterpiece, massive scene, best quality 4k, 8k, cinematic art by Tom Bagshaw"

[high skiff]

its likely the 4k and 8k tags

[deft quiver]

I've added and removed them

to no effect

[sharp robin]

running but my pc is slow so give me like 3 mins

[toxic vortex]

call me crazy but try "background volcanic explosions" rather than "volcanic explosions in the background"

[high skiff]

I would recommend "With a volcano exploding in the distance"

[spring fulcrum]

over processed but still cute... any suggestions on how to get it to be less smoothed out?

[clever verge]

Maybe it's Tom 🙂

[high skiff]

like my corgi :>

[lusty raptor]

oh, maybe get rid of background entirely? -could be hitting wallpaper instead

[civic sigil]

Wow now I get the MJ hype 😅

[high skiff]

oop

my corgi :>

[hearty ginkgo]

well the only reason i actaully use mj is because my dad pays for it for his businness instead of paying like 500 a month to istock for images

[deft quiver]

yeah even like permutations of prompts and artists

does not seem to do much

lol

chronic text

[deft quiver]

hmm I wonder if it's my generation setup

[clever verge]

Nice carving!

[high skiff]

trying to diversify my 5 sample prompts

[hearty ginkgo]

https://imgur.com/a/ZdVUZtJ

was to big to share on discord after upscaling every image is like 20 plus mb

[winter raptor]

This is a fantastic style

[hearty ginkgo]

Sytan what upscale model are you using? Is the 4x ultrasharp good?

[nimble heart]

if it's a markdown/rich text format 100% create a summary with refs to all the categories

[high skiff]

my PC crashed lmao

[nimble heart]

how

[high skiff]

I got too intense on my OC lol

[nimble heart]

3090 or 3080

[high skiff]

+200mhz core and +900mhz on VRAM lmao

3090!

I have it now :>

[nimble heart]

how is it

[high skiff]

amazing

now that I can run both models cached in VRAM, and also benefit from the new nightly torch VAE that is basically instant, I can run images more than 2x as fast as on the 3080 while using the same power

[nimble heart]

nightly torch vae?

[high skiff]

from 11.6 seconds to 5.2 seconds

[sullen bison]

2 more weeks, i guess

[nimble heart]

nice argument why dont you back it up with a source

[high skiff]

there is a nightly torch build that allows for a BF16 VAE, and it can do even like BS 32 near instant

for me, BS 8 is less than 1 second to decode

[west breach]

@winter raptor how are these? any prompts to try?

[nimble heart]

bf16 vaes have worked for me for a while. < 1 second on 1024 but > 1 minute on 2048

[sullen bison]

i have no source because it's the most obvious route

[winter raptor]

Those are really nice! Try generating a man standing in a cyberpunk city street, that has given me a ton of trouble

[high skiff]

why? lmao

[sullen bison]

they pushed it back, no way it's coming anytime soon

[high skiff]

and where did you see that?

[sullen bison]

i don't think im allowed to share it here, will send you a DM

[high skiff]

if you can't share information, then don't message me lmao

[dense chasm]

finally

[nimble heart]

oh shoot rocm 5.6 was marked as ready for merge in torch an hour ago. New nightly tomorrow hopefully 🤞

[dense chasm]

D-ID ban for celeb uploading image,damn

[sullen bison]

i can't but i must not, but i will

[high skiff]

bro

lmao

[sharp robin]

LMAO cap send me the info

[high skiff]

bro is just flat out lying lol

[sullen bison]

i believe it

[gentle mirage]

what he lyin about?

[sharp robin]

let's see it

[peak dove]

Has 1.0 dropped yet?!

[sharp robin]

ROFL

dm me

haha

[sullen bison]

it's getting delayed apparently

[sharp robin]

not the 18th yet,

at least not in the USA

[gentle mirage]

huh

[sharp robin]

u have whole 24 hrs

for delivery

[peak dove]

In the UK it is 18th already! 🙂

[sullen bison]

you asked for it and i'll blame you if this gets me in trouble
https://www.reddit.com/r/StableDiffusion/comments/152oey2/comment/jsf2oih/?context=3

[sharp robin]

no timezones matter but hte USA

[hearty ginkgo]

It's the 18th in flroida USA but it dosnt even matter the model won't drop till atleast work hours start if not later during the day ingeneral

[sharp robin]

sad truth

[high skiff]

even then, it was just a goal, not a promise lol

[sullen bison]

expect dissapointing official news in the morning

[gentle mirage]

Thats unfortunate

[sharp robin]

florida, texas, EST, CST is already 18th but not PST

[civic sigil]

Yeah they are killing refiner lmao

Unrefined chads ftw

[high skiff]

and where is this from?

[sullen bison]

4chan, which got it from reddit.

[boreal bough]

cries

[high skiff]

ah yes, the most credible source known to man

we will see when it happens

[shell oasis]

i mean joe commented in the reddit thread

[deft quiver]

4chan and reddit? I couldn't imagine two more credibles sources of information

[sullen bison]

don't expect much

[shell oasis]

[gentle mirage]

why are u acting tone-deaf joe is literally in that screenshot.

[winter raptor]

Wow!!! This is awesome

[gentle mirage]

the man is literally there

[high skiff]

tone deaf is not the right term, and my bad, I forgot that people haven't faked screenshots about SD before

[trim orbit]

screenshots of text. usually fake

[high skiff]

where is this post?

[shell oasis]

https://www.reddit.com/r/StableDiffusion/comments/152oey2/hi_guys_is_sdxl_10_released_today/jsf7p64/

[gentle mirage]

its litreally above u scroll up 2 times

[gentle mirage]

blud

[sullen bison]

the reddit thread is real, someone from the staff commented on it.

[peak dove]

Vlad AUTOMATIC1111 SDXL 0.9

[sullen bison]

its

over

[gentle mirage]

idk what would someone even gain from faking smth like that lool

[shell oasis]

giggles

[high skiff]

alright, that's a little more concrete in that case

[slender coral]

Hey guys is 1.0 out yet.

[boreal bough]

oh well

[high skiff]

bro, people in this community have faked shit before lol

[boreal bough]

no drop today, just more info tomorrow T.T

https://www.reddit.com/r/StableDiffusion/comments/152oey2/comment/jsf2oih/?utm_source=reddit&utm_medium=web2x&context=3

[deft quiver]

if more time = better model, sweet

[gentle mirage]

unfortunate though

[slender coral]

I was just messing around. 😄 my bad.

[gentle mirage]

i've some datasets ready to be trained

[civic sigil]

Well the last version of the refiner to model to ever release is out so like 50%

[high skiff]

finme by me if they are finding a model thats better than the current base 1.0, cause I am still not too happy with base 1.0

[trim orbit]

screenshots of text are always faked. woven bullshit usually. wonder what joe talking about but screenshots of text are just dumb

[high skiff]

tho it does seem to be that they confirmed that the second pass is staying, which is good to see

[winter raptor]

I think you really nailed it

[sharp robin]

wow @sullen bison was right

holy shit

by Joe

[nimble heart]

guess I just play more skyrim for a while

[deep dagger]

I am going to sleep lol

[slender coral]

What exactly is 1.0 releasing that 0.9 doesn't have? Just more training data?

[sullen bison]

an indefinite delay is an indefinite delay, a better model might come but it also may not

the chance is there

[sharp robin]

havnt decided to delete refiner or keep it yet

[high skiff]

yeah, tho it seems that now we would be able to see what its like

[civic sigil]

I expect they will release it today anyways but I still need a little bit more time to work on my workflow so I wouldn't mind

[high skiff]

welp, oh well, gives me more time to write my docs

[sullen bison]

we can't ignore the possibility of cacellation

[high skiff]

its not gonna get cancelled lmfao

[civic sigil]

Yeah we can lol

[west breach]

if you git pull, there is a new style called deep-field. it's still hit or miss. had to put landscape and panoramic keywords to stop the blur, so it does affect the composition away from portrait photos a lot

[shell oasis]

bruh there is absolutely 0 shot it gets cancelled

[sullen bison]

then you are just being a fool

[gentle mirage]

I doubt it. we can pretty much finetune 0.9 to improve quality. they're just making a better base with good parent data so finetuning is easier

[high skiff]

like, sure, delayed, I'll give you that

Cancelled? Thats just stupid lmao

[sharp robin]

i have doubts if they really think they came upon a breakthrugh then they will wait, they are scared of 2.x reaction lol, but also pls train more nsfw

[nimble heart]

I mean never say never but dumping hella money into a model that's been performing well just to cancel it is kinda wild

[deep dagger]

Its clearly delayed but it will not be canceled lol

[sullen bison]

if that's what you want to believe, so be it.

[deft quiver]

oh snap, the text problem was being caused by xformers memory efficient attention

[sharp robin]

nah i belive it

[civic sigil]

What would the motive be for them to release research weights and then cancel, trying to get people to use their paid API while the weights are wrapped up in a restrictive license?

[gentle mirage]

sure. look at base 1.5 its dogshit. finetuning can literally do wonders

[nimble heart]

text is more memory efficient than pixels obvs

[sullen bison]

you said it, i didn't

[gentle mirage]

how does that not cross your mind lool

[deft quiver]

pog, it's not ruined anymore

I can sleep tonight

lets ignore the text on her shirt

[boreal bough]

oh well RIP joe. nothing like a leaked info to make your workday easier, for a decision you dont make yourself ^^'

[nimble heart]

person who makes images of lovecraftian sea horrors

[slender coral]

Anyone know what teh diff between 0.9 and 1.0 may be?

[nimble heart]

at least 0.1

[sullen bison]

and be stuck in a model finetuning hell until the cycle repeats? sure beats being stuck on 1.5

[civic sigil]

No refiner for 1.0

[slender coral]

😄

[slender coral]

It's built into the prompt?

[sullen bison]

0.9 exists

[civic sigil]

Only base model

[slender coral]

I like the refiner though.

[soft zealot]

Surely additional training for NSFW should come from the Community rather than SAI ?

[winter raptor]

I'll give it a try, it's not a huge deal if it prevents close ups because typically the background will matter less there. If I come up with anything I'll let you know, I really appreciate all of the help and work you have done!

[gentle mirage]

nothing wrong with finetuning a model to improve. finetuning in repeats are good. (it just crossed my mind that you're pretty but clueless abt finetuning. are u new or just never has finetuned anything before)

[high skiff]

https://tenor.com/view/straight-to-jail-straight-to-jail-meme-linus-tech-tips-linus-ltt-gif-25530127

[gentle mirage]

but you do you g

u can stay on 1.5 if u wan

[trim orbit]

and so begins the delves into #conspiracy for the next week

[sharp robin]

it would be better if it came from SAI built into model, it would make it easierfor community/finetuners to train it,

[nimble heart]

thought we couldn't do gifs

[high skiff]

they added it like 2 days ago

[sharp robin]

easy to train a concept it already knows vs training froms scratch

[high skiff]

I am about to test that theory

[boreal bough]

still a SFW official work server - wrong server for this type of conversation

[civic sigil]

I wouldn't mind a push back, like Ive been mentioning I would like to see SAI finetunes like Anime, Photography, and Traditional Art

[sullen bison]

i remember when that wans't an option

[high skiff]

now that we have some more time before 1.0 launches, I'm gonna make a LoRA

[soft zealot]

but remeber what may be klegal in one poart of the world may not be legal in anaother. Why should SAI expose themselves (no pun intended) to excessive risk of complaint?

[civic sigil]

And well trained Controlnet weights

On day 1

[magic dawn]

Is it delayed?

[sullen bison]

have fun, god knows it's the only thing we can have nowadays

[gentle mirage]

finetuning has always been an option. Optimizations come as soon as they could. shelling 100 bucks on a few A100s to tune doesn't make u poor.

[sullen bison]

yes

[gentle mirage]

or skill issue ig

[high skiff]

seems to be

[magic dawn]

Source?

[trim orbit]

i dont think that's the problem. they're in the UK and tey're safe there for ML

[gentle mirage]

SAI is not gonna hand hold you. they're made as base model to be tuned

[boreal bough]

https://www.reddit.com/r/StableDiffusion/comments/152oey2/comment/jsf2oih/?utm_source=reddit&utm_medium=web2x&context=3

[deft quiver]

this argument makes no sense

[magic dawn]

Omg not surprised at all...

[high skiff]

oh well, if anything, I should feel much better by the time it does release, as I am sick right now

[nimble heart]

I feel like a lot of people are misremembering how base 1.5 handled nsfw because 90% of finetunes enhance it in some way

[deft quiver]

the whole point of the base model is exactly to hand hold you

[soft zealot]

Its still a rosk regardless of where based,

[sullen bison]

expect dissapointment and lockdown tomorrow

[sharp robin]

they use or used to use loian database, they dont get in trouble the db would

[trim orbit]

thats not a source really

[sullen bison]

we'll see in a few hours

the chance is there and it's the most likely scenario

[magic dawn]

So what's the point of the bot voting if they just say nevermind we have another candidate?

[soft zealot]

anyways elsewhere, because I;m a pervert I switchd base to point at he refiner model and refiner to point at the base model

[trim orbit]

it's just such a weird thing people do posting screenshots of text in an html inspector world

[spring fulcrum]

whose a gooboy

[nimble heart]

they should just release all 3 candidates as SDXL 0.97, SDXL 0.98, SDXL 0.99

[sullen bison]

ah yes, using html inspector on discord mobile

[sharp robin]

that is because of the bot, u dont read, that is what post said, the bot u are voting on had very well liked photos, and those photos correspond to other models they have, so in essence the bots or "you" voted for this

[sullen bison]

how didn't i thought about that

[sullen bison]

they'll never do that

[nimble heart]

I know

[boreal bough]

yeah, just wishful thinking 🥲

[sullen bison]

like i bet we almost didn't got 0.9 out there

it was a miracle someone got it out there

[gentle mirage]

It wasn't. The guys knew the leak was coming. LMAO

[boreal bough]

the alternative is to just say screw it, and release everything I have for 0.9 and start posting the training guides XD

[soft zealot]

I think its fair to say that SAI only released an official SAI in order to make it easier to stamp on the leaked versions

[trim orbit]

joe literally annoucned the leak

[gentle mirage]

mcmonkey was right the the moment it got leaked in the SAIL discord

[boreal bough]

finetune levels of training can be achieved with lora now - so not like it would be a hindrence...

[gentle mirage]

asked him and he wasn't suprised

[magic dawn]

Since I don't read but you do, where in the screenshot was the bot mentioned?

[trim orbit]

"they're cancelling it" still has me laughing

[high skiff]

same lmao

[naive goblet]

https://tenor.com/view/side-eye-dog-suspicious-look-suspicious-doubt-dog-doubt-gif-23680990

[trim orbit]

not because its whimsical and cute, but because someone said it with conviction

[sullen bison]

yeah...

[gentle mirage]

https://tenor.com/view/get-real-cat-skate-funny-meme-gif-18666878

[high skiff]

Bombastic side eye

[magic dawn]

Where's that screenshot from? It's not in this server

[nimble heart]

I think that's the future tbh. A strong enough base then just a bunch of experts in the form of loras.

[cosmic compass]

is gigapixel still the best option to upscale those creations ?

[sharp robin]

hope u understand arrows

[trim orbit]

no. its an ancient gan by today's standards

[gentle mirage]

gigapixel as always been shit. find yourself a good upscaler from the upscale wiki

[sullen bison]

it must be an internal message or maybe a group, the message says that he will @ everyone tomorrow

[trim orbit]

topaz has been out of date fora long while

[high skiff]

[boreal bough]

its a bit of effort to get the datasets properly set up - but the training is easy. just takes a while (20~40 hours on rtx4090)

[magic dawn]

Thanks you saved my day

[boreal bough]

already have 2 finished as a proof of concept :/

[sharp robin]

[high skiff]

@boreal boughHey caithhhh

I have my 3090 nowwww, and I wanna dick around with some LoRA'ssssss

[boreal bough]

XD

[high skiff]

also, good day to you

[deft quiver]

okay this time I'm definitely actually sure, the source of text was in fact the refiner

[high skiff]

I am gonna try kohya

[nimble heart]

Oh no I don't train I just download a hundred of them from civit and mix them together after performing a ritual to the old gods

[boreal bough]

I'll get the guides ready tonight - and give you access to my google docs.

[civic sigil]

Please try a full finetune

[deft quiver]

trained a few sdxl loras with it, works great

[nimble heart]

If you play with the lora strengths you can mix like 8 together in one inference without too much destructive interference

[civic sigil]

Mcmonkey said it should be easily doable

[deft quiver]

you want the training configs sy?

[high skiff]

That is for later when Pseudo and I break in his A100

[boreal bough]

for now just use kohya (either directly, or via derrian)
kohya ui works... but... did odd things when I ran it - so I'm back to using kohya directly for now (due to the hourly updates)

[sharp robin]

that may have been a little passive agressive, it was all in good spirits, hope it is worth it if they do delay...

[high skiff]

Same, I can open my stuff to you as well

[heady halo]

Sdxl releases today right?

[sullen bison]

no

it was delayed

[heady halo]

To when

[trim orbit]

no rumors. we dont know yet

[sullen bison]

indefinite

[civic sigil]

I will get a 3090 if full finetune actually works

[trim orbit]

wait for an anouncement

[cosmic compass]

^^ a good upscaler to use on my own gui ? Curretnyl i havent figured the ui for my own pc

[boreal bough]

needs (multiple) A100 - also, not needed if you just want 1.5 levels of 'finetune'

[magic dawn]

No problem, I really didn't read carefully...

[sullen bison]

https://www.reddit.com/r/StableDiffusion/comments/152oey2/hi_guys_is_sdxl_10_released_today/

[trim orbit]

this guy is trolling

[deft quiver]

it does but it's extremely slow

[gentle mirage]

every 5 mins another blud asks for sdxl 😭 check the announcement post

[civic sigil]

Have you done it

[nimble heart]

esrgan and a tune of it called real-esrgan are the two most popular

[high skiff]

its not announced in this server

[gentle mirage]

for gui use ChaiNNer and for upscalers get it from the upscale wiki

[deft quiver]

I got 2000 steps in and gave up, it's too slow on batch size 1 for me personally

but it does work

[gentle mirage]

they'll announce it if they release later duh

[high skiff]

on what GPU?

[deft quiver]

3090ti

[high skiff]

hmm

[trim orbit]

this is stability's server. why woudln't it be?

[high skiff]

for full finetuning?

[deft quiver]

ye

for loras it's great

[high skiff]

ah, alright

thought you meant LoRA's

[deft quiver]

but full fine tune it's reallllly slow

[cosmic compass]

what is a "tune of it" ? like some extra model ?

[high skiff]

excited to see how SDXL works with hyperdreambooths

[sullen bison]

watch them announcing a mid august release, then a mid september, then mid october, then mid novemeber, then mid december, then mid 2024

[trim orbit]

hyperdreambooth if google ever releases code

[nimble heart]

yea just an alternate version

[civic sigil]

Thanks, Ill buy the 3090 then

[cosmic compass]

❤️ apreciate it will try

[boreal bough]

it does - but not how you think it does
basically, the bigger the batch size - the better the fine tune
with 24gb vram, you can do batch 1.

[winter raptor]

@west breach I got this out of it which is pretty cool

[civic sigil]

That's fine, as long as it works. Not trying to make the next NAI lol

[boreal bough]

I mean you'll get worse results than with lora training

[civic sigil]

That's fine

[boreal bough]

in that case go for it. 24gb vram is enough

[rustic garnet]

I have a 3090TI. I think its quite good for AI due to the large vram and still much less expensive than 4090

however, you don't need all the vram for training SDXL. Just train Lora. Its better than finetuning anyways

[civic sigil]

I need another PC build anyways

My animation workflow is gonna involve a lot of rendering and SD and waiting around if I try to do it on one computer locally

[soft zealot]

I use a 1080ti that cost <£200 with an ek waterblock fitted, sure uits slower than a 3090 but its also a lot cheaper.

Tortoise & Hare 🙂

[rustic garnet]

but its cool anyways to be not so limited on vram. Like when you want to use BLIP2 or llms you will need large vram again

[gentle mirage]

if all you do is prompt 1080ti is plenty

[soft zealot]

when did they release the 1090ti lol

[high skiff]

Thanks for reminding me! I can run Vicuna 30b now!

[gentle mirage]

its typo duh

[rustic garnet]

I don't think thats a big thing

[soft zealot]

I know hence the lol

[rustic garnet]

they released a similar paper months ago and nobody was talking about it afterwards

[high skiff]

I imagine it's one of those shitty re-releases

Like the 12GB 2060

They slapped 22GB of 64bit bus VRAM on a 1080ti, called it a 1090ti and raised the price by 2.5x lol

[peak dove]

Where is the LOG file located in Vlad AUTOMATIC1111?

[gentle mirage]

dreambooth is just shit. tag your images and finetune. relying on 1 tag is pretty stupid

[boreal bough]

in theory, the same could be achieved with a bit longer training, an additional regularization set of faces + a 2/1 lora. right?

[civic sigil]

You can use tags with dreambooth tho? You always could

[rustic garnet]

HyperDreambooth has nothing to do with Dreambooth

[deft quiver]

the big thing with hyperdreambooth is the tiny size you get in the end

but I don't know how many people really care about small size

[gentle mirage]

yes you can set 1 token to recall training data

[rustic garnet]

no, it isnt

the size is just like a lora

[sharp robin]

It is but isn’t. Gens take forever on higher scales

[deft quiver]

it's 10,000x smaller than the original dreambooth

allegedly

[rustic garnet]

the difference is that they train faster than lora, because the initialize the lora using some hypernetwork

[civic sigil]

Wdym you can use full prompts with your dataset

[gentle mirage]

I was being nice 😭
ofc 1080ti is dogshit in current standards

[rustic garnet]

its cool if you have some cloud service or online game or something like that where your users can create images based on a photo. Because here it would be a huge difference if training times take 10 seconds or 10 minutes

[gentle mirage]

no dum dum. when you're training u dont actually get to describe each images to train the text encoder you use a singular token to recall the training data

ofc you can prompt

finetune exists dreambooth is just if you're really lazy

[rustic garnet]

for us its nonesense. We can just train 10 minutes. We can even train 4 hours and then get results that are 100 times better than Hyperdreambooth

[civic sigil]

No? Lol

[west breach]

Awesome 👍

[civic sigil]

Not sure what you're talking about tbh

[rustic garnet]

I think he is talking about Dreambooth as it is described in the original paper

we nowadways call every finetuning "Dreambooth"

[sharp robin]

It isn’t dogshit either. That’s what’s on my pc and i can do like 3k res on it it’s just slow as a pig. It’s a powerhouse but is old. It could do 1.5 models just tine no complaints but SDXL is a whole nother beast.

[gentle mirage]

then what you're talking about isn't dreambooth. dreambooth uses a token to recall data. u can still prompt. finetune is much better since u can tag all images

[rustic garnet]

but actually, it was a technique how to train on your face by using a special token ("sks") to describe your face

[void loom]

according to who?

[civic sigil]

Is that the proper meaning of Dreambooth? I never heard it used so restrictively

[gentle mirage]

Im just not willing to wait the time. i had a 3070 before i had my 3090 and it was dead slow on gens

[gentle mirage]

Bruh

so the guy doesn't know what dreambooth is

[boreal bough]

its complicated - but in sdxl, you can train over 100 concepts - fully intact, with a 43mb lora, using 24gb vram, with around 20 hours

[gentle mirage]

[boreal bough]

"fully tagged"

[gentle mirage]

you're talking abot LoRA

[sullen bison]

@void loom Joe a.k.a Mystery Guitar man, part of the stability Ai staff

https://www.reddit.com/r/StableDiffusion/comments/152oey2/hi_guys_is_sdxl_10_released_today/

[civic sigil]

Dreambooth as it existed when it was first ported over to A111 at the very least could take full prompt tags, that is the colloquial meaning of the term

[rustic garnet]

Dreambooth is just a fancy name of a technique from google. Its referring to this strict workflow of tagging your images with "photo of sks person" an then finetune it. We later used this term repeatedly for describing full finetuning in contrast to, for example, Loras

[austere bay]

If SAI is delaying 1.0 because there is not enough info from the bots to choose a candidate, it is at least suspicious, the bot works pretty simple, it makes you 2 images, you choose 1, that is the better image, and not only that, there is a lawsuit currently and the plaintiff's lawyers said Emad that if the plaintiff does not get everything he wants he will "burn down" SAI, just saying.

and I quote Emad: "Furthermore, if Cyrus is unable to profit from this situation, he seems intent on harming the company. His counsel has communicated that he would "burn it down" if he ultimately recovers nothing."

https://twitter.com/EMostaque/status/1680774535342358528?t=qHQed3zpXby_FXEiG-glxw&s=33

Emad (@EMostaque)

In the interest of transparency and to set the record straight, I must address the unfounded lawsuit filed by my former colleague, Cyrus Hodes, against Stability AI and I on Thursday.

He alleges that he was deceived into selling his shares to me when he voluntarily left the…

Likes

665

Twitter · Jul 17, 2023, 3:00 AM

[gentle mirage]

in training it cant. you use an instence prompt to describe your training data, you can still prompt like normal.

thats why dreambooth sucks

[civic sigil]

Wrong

[gentle mirage]

you do you g

[sullen bison]

I see, thanks

[civic sigil]

Yeah that sks crap originated from textual embeddings first yeah?

[rustic garnet]

no

[sullen bison]

I guess it could mean that not enough people are voting for their images

[rustic garnet]

I mean thats the point

[civic sigil]

I remember that exact tag mentioned for embeddings

[rustic garnet]

textual embedding + finetuning would make much more sense

[civic sigil]

Yeah but they tested it very early and said it was unnecessary

It was the working theory

[rustic garnet]

oh, that I totally don't believe ;D

[civic sigil]

Well somebody tested it, i dont remember who but I remember it was early on

[rustic garnet]

the limitation of dreambooth is that they only train the unet and so, they have to use some special token like "sks" that is not associated in the unet with anything to train the unet on your face

if you train the unet + the textencoder, it wouldn't matter. You could just use your real name as prompt and the text encoder would learn to embed it properly

[gentle mirage]

Dreambooth is that you have all your training data and you would use 1 tag called an "instance prompt" to describe that since captioning .txt images were not a thing in Dreambooth. you can still prompt as normal but u would have to still use that instance prompt to recall that concept it learned.
then the dreambooth guy implemented finetune later on which allowed you to caption describe your images individually which is better overall.

idk what @civic sigil is confused abt

[boreal bough]

@civic sigil but are you just arguing for the sake of arguing about the definition of a word and its lack of past ability?
Cause if you want progress, we have it
If you want to complain about new technologies having bad names - a legit argument, but not a long one.

[gentle mirage]

I feel like he's clueless

im not sure if hes trolling or genuinely idiotic

[civic sigil]

I am trying to learn the terminology you all are using, stop acting so childish

[gentle mirage]

🤷‍♂️

[civic sigil]

If you dreamboothed a bunch of red cups labelled as "cup" and then prompted for a cup with the resulting model would that cup not be red?

[rustic garnet]

@gentle mirage I don't think that "you cannot caption the image" is dreambooth. i mean, this are just implementation questions

[austere bay]

Why would they wait until the last moment to announce it them?

[gentle mirage]

sigh.

[stoic nexus]

Hi! This is the day we get to run SDXL at home, right?

[rustic garnet]

yes, but then all your cups would be red. The idea is that you can teach it a special kind of cups by calling them "sks cup"

[civic sigil]

Yeah, and that's not how you have to use it. Calling it bad overall because of how the people who wrote the paper forsaw it would be used is plain dumb in this field

[sullen bison]

Idk, i guess they were expecting more votes by the last moment but i guess they weren't enough? Idk it is quite suspicious

[sharp robin]

It’s the opposite some of their images are getting really good and received. Nd those images are from later models after their 1.0 contenders. U also no read.

We need a reading comprehension class.

[rustic garnet]

I agree partly if its about this "you are not allowed to make a caption for your training data". They never claimed that in the paper

[civic sigil]

If you want to dreambooth an art style for instance I dont see why you cant have frozen TE? It works fine for me

[rustic garnet]

but the point is that dreambooth refers to the idea of using a special token like "sks" to describe your concept

if you don't do this, its not dreambooth but regular finetuning

[civic sigil]

Alright fine, I concede the argument

Nobody uses the term that way anymore though lol

Bro is just trying to show off his history knowledge

[boreal bough]

• 'Dreambooth' is a word reused multiple times - often it refers to completely different (newer) techniques - some website even call all training 'dreambooth style training' - when all they're doing is LoRA in the background
• LoRA - a newer method of training, that works by applying a lot of theories all at once - and many of them work well. (however many features or lora remain unused, since they make as good as no difference)
• Lycoris - derivative of LoRA -> basically many versions of LoRA on drugs. Some good drugs/some bad drugs. Your mileage may vary - but useful if you know what you're doing
• Finetune - a full training of everything at once, hence why batch size is so important, since you dont want one picture to affect the outcome of thousands
• Merging/Extraction - its own topic, too long for one sentence

[rustic garnet]

I checked the paper. They evaluated dreambooth against textual inversion. So they only checked if dreambooth works better than textual inversion alone. This is stupid of course. You cannot compare "training a token" versus "training the unet". They didn't checked if textual inversion + training the unet works better (and I'm VERY sure it works much better than dreambooth)

[rustic garnet]

that's true, though

[civic sigil]

My understanding at the time was that they tested training a textual embedding as well as finetuning the model so that they wouldn't have to overfit the model as much. Maybe my interpretation was wrong since I was still new to SD back then

Probably cause Textual Embedding and Text Encoder have the same acronym lol

[vast peak]

Could you please let me know if the WebUI already supports SDXL?

[rustic garnet]

I remember they later tested training text encoder as well and found that it works much better

[gentle mirage]

git switch sdxl

[sullen bison]

Yes but you need a lot of vram

If you have less than 8Gb of vram, you can still use comfyUI

[vast peak]

But I have 4090.

[sullen bison]

How much vram?

[boreal bough]

4090 master race ❤️ (24gb vram)

[rustic garnet]

but yeah. Thinks are complicated as everybody is using the terms differently. Just take caiths " LoRA - a newer method of training, that works by applying a lot of theories all at onc". Acutally, the term LoRA is very well defined, but there are so many new techniques and variations and sometimes they call themselve LoRA, sometimes they come up with a new fancy name

[sharp robin]

24

[vast peak]

24gb

[sullen bison]

Nice i wish I had one of those, i only have 6Gb, love my little PC

[hearty ginkgo]

I want the a100 ngl lol

[boreal bough]

emperor race be like

[high skiff]

Training an SDXL LoRA!

Also, looks like I can use much bigger BS than I assumed

BS 4 is using 17.8GB VRAM

[vast peak]

Does Kohyass support it?

[high skiff]

Yeah

[rustic garnet]

yes, its crazy. BS 12 is using not that much more vram than batchsize 1 oO

[high skiff]

Kohy-ass lmfao

[boreal bough]

8 should be the max if you wanna keep using the pc without noticing the training
10~12 is the true max limit

[high skiff]

Dope, thanks

I am just doing a low effort 20 step 1 epoch training to test the waters right now

See if it's grabbing onto anything, as I am messing with settings out of the box

[vast peak]

So, do they have to release SDXL after work?

[boreal bough]

stay clear of faces/anatomy/words that describe the body like "legs" - for your first lora. everything else works easy ^^

[rustic garnet]

I found that lora training on SDXL is as crazy as in previous versions. Sometimes it works, sometimes not. Sometimes you change a few training images and immediately everything gets much better. But overall, results with SDXL are just much better than with 2.1 or 1.5

[high skiff]

I have done 50+ 1.5 LoRA's, so I hope my knowledge carries over haha

[rustic garnet]

one thing I found strange is that when you train textual inversion then the anatomy really quickly can really go wrong

like if it doesn't like to get its tokens pushed out of the comfort zone

[boreal bough]

same. its why I'm still doing unet only

[stoic nexus]

is 8GB of VRAM enough to run SDXL?

[high skiff]

Yeah

[vast peak]

I need to train some Lora to redraw scenes in product images, specifically areas other than the product. Is there anyone who can teach me some methods?

[rustic garnet]

I then trained textual inversion on a initializer token (instead of random initialization) and with very low learning rate and epoch and then switched to text encoder training and unet training. That solved the anatomy issues for me. But maybe it also just good or bad luck. I would have to train more to understand that better

[boreal bough]

I've pretty much got anatomy training down now - but was hoping to test 1.0 if it changes anything T.T

[rustic garnet]

yeah... the reason I stopped training was because I thought we would get 1.0 soon -_-

[boreal bough]

🫂

[hearty ginkgo]

Would a k80 24gb be any good?

[rustic garnet]

but given that Emad is usually writing hyping twitter messages every day before a release and we hear nothing is telling me we won't get SDXL so soon... ;_;

[vast peak]

I'm still struggling to collect training datasets.

[rustic garnet]

yeah, thats always the most difficult and most important task

[civic sigil]

It grows...

[vast peak]

I tried using Midjourney, but when I made some changes to the subject, the image changed too much. It's completely unusable in the same training dataset.

[slender coral]

Use mj to bot your prompts, then use those to train.

[fair island]

today's the day? maybe?

[boreal bough]

am I the anomaly here with 300k images saved? xD

[vast peak]

The same prompt can also generate images with different decoration styles.

[slender coral]

I have over 5 million so far.

[lusty hornet]

i heard that sdxl1.0 will open source today,is that true?

[boreal bough]

noice!

[peak dove]

(Cannot locate LOG file in vladmandic/automatic1111?) 🙂

[high skiff]

Alright.

I'm a little concerned about my new graphics card

[civic sigil]

Why

Who was your seller so I can avoid them lol

[boreal bough]

see ya'll later today

[high skiff]

It just blue screened for the second time

With the same error

I'm going to hope it's just a bad driver, but I'm not sure

Video TDR Failure nvlddmkm.sys

[peak dove]

You got CUDA problems?

[high skiff]

No, I have video engine problems

[lusty hornet]

brand new card?

[high skiff]

Brand new to me, used from somebody else

[stable kiln]

https://www.reddit.com/r/StableDiffusion/comments/152qs8j/sdxl_delayed_more_information_to_be_provided/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=13&utm_content=2

[vast peak]

I've also encountered an issue where after generating 100-200 images, it starts outputting solid-colored images. I was wondering what might be causing this.

[boreal bough]

check for any overclocks before anything else - disable all and see if it goes away/ then turn them back on over the days, one by one
(stock OC is still OC, and can be set to normal in bios)

[civic sigil]

"Whether of not 0.9 Loras will work" is so funny

[high skiff]

We know

[civic sigil]

I hope its fake just for that

[peak dove]

This TDR fault is when GPU switches from one power state to another ...

[high skiff]

Is that so?

Is there any recommendation to help remedy it?

[lusty hornet]

oh no

[civic sigil]

What's your psu wattage? I heard 3090 takes a lot

[high skiff]

850, it should be just fine for it

[civic sigil]

Idk then but I will be spending $700 on my card just in case lol

[dim steeple]

what do you guys want made from the community most when 1.0 drops

[high skiff]

I was running my 3080 on it for the last 2 days, and it uses more power

[boreal bough]

bios update if there is one

[high skiff]

Good idea, it's a brand new motherboard and I forgot to check

Every time the PC turns on, I immediately get prompted to download some form of gigabyte bloatware for bios updates

Might as well just bite the bullet

[peak dove]

TDR Power Management Issue ...

[high skiff]

I'll inspect the cables on the GPU to ensure that they are plugged in properly

Although I will say, neither blue screen occurred while the GPU was doing anything above idle

The first time it blue screened while I was just scrolling on Amazon, and the second time was while I was going through my file explorer with just music in the background

[hearty ginkgo]

What powersupply u have?

[high skiff]

Cooler Master 850 watt gold plus MWE I believe

[hearty ginkgo]

Are you blue screening with the overcloacks only?

Overclocks*

[high skiff]

The first time I had a very severe overclock and I figured that was the issue, but now it was running completely stock, not even achieving its full boost clocks

But on top of that, it's not crashing when it's under load, it's only crashing at idle

[peak dove]

Power Management - two competing clocks

[high skiff]

It's a very weird phenomena, it crashed about a minute after I was done with my LoRA training

[high skiff]

Do you have any recommendations on how to fix this sort of thing?

[old mulch]

Ill answer on behalfof everyone:
Realism model
Anime model
More detail lora
Nsfw lora

This should cover everyone. Ill add a selfish one, a web design lora

[high skiff]

I'm currently heating up my food right now, but I will be taking a look at All of my power connectors to ensure that there are no problems there

[peak dove]

That TDR picture I posted - the original poster says it was fixed after gBIOS file update from his GPU card manufacturer

[high skiff]

All right, that makes sense

Although the bios on the GPU should be up-to-date, as I got it from somebody else who is using it just a couple days ago

But, some people in a different community pointed out that it might be a motherboard BIOS issue, as I did not check to update the BIOS on my motherboard

[peak dove]

I guess if the gBIOS update is not the answer - then suspect "secondary stock"

Bad hardware

[old mulch]

Who fking knows. Its like 8h late at this point, im starting to lose hope

[hearty ginkgo]

Check if you 3080 has that bios switch

[high skiff]

Good idea!

[hearty ginkgo]

ik i have it on my 3070 not sure about the 3080 ik some do but it could also vary depning on brand

[high skiff]

This is a 3090

I don't see a switch

[ionic dragon]

SDXL 1.0 is not releasing today?

[placid edge]

just saying that gigabytes garbage software is highly insecure

same goes for asus

[peak dove]

SDXL 0.9 Prompt - lowpoly 8k matte painting grand guignol venice carnival jazz dance streamline moderne tamara lempicka mads berg peregrine heathcoate catrin welz-stein georgy kurasov alenka sottler arief putra itzchak tarkay rené magritte victo ngai henri rousseau vladimir kush remedios varo - 20 steps, Euler A, 1024x1024, cfg 6

[high skiff]

the GPU is Zotac

[placid edge]

i know i just thought u meant u were gonna install it

lol

[hearty ginkgo]

https://www.zotac.com/hk/news/dual-bios

It's a button

For the zotac gpu's

[ionic dragon]

https://www.reddit.com/r/StableDiffusion/comments/152qs8j

is this true?

[high skiff]

seems to be

[placid edge]

disappointed but not surprised if true

[ionic dragon]

oh so, their finetuned models are better than 1.0?

and figuring out which one to release?

[high skiff]

thats for 4000 series

let me check mine

[hearty ginkgo]

Have you googled the blue screen error b4?

[high skiff]

yeah, says maybe "outdated drivers"

but they are brand new

[hearty ginkgo]

GeForce experience?

[high skiff]

I stay away from geforce experience

[peak dove]

Best drivers are DIRECT from manufacturer's site

[olive flare]

Of course it’s delayed

[high skiff]

NVIDIA drivers? I don't think thats how that works-

[peak dove]

Although GeForce has always been good to me - I use the Studio version of their drivers

[hearty ginkgo]

I get my driver's either from GeForce experience or from the nvidiea site itself https://www.nvidia.com/download/index.aspx

[high skiff]

I used to use studio drivers, until i found out they cut out some very key features

[high skiff]

same here

[hearty ginkgo]

Only use game drivers lol

[peak dove]

Game Drivers kept crashing my Photoshop - so Studio Drivers was a godsend!

[high skiff]

I used to use studio drives cause I did a ton of rendering before, so it kept it nice and stable

1000's of hours of rendering, and it never crashed once

[hearty ginkgo]

Wait your driver's are the newest ones?

[high skiff]

ok, maybe not PLURAL thousands, but at least 1000

[peak dove]

Studio Drivers also allow a quieter fan noise 🙂

[high skiff]

yes

just installed the GPU like... 7 hours ago lmao

[hearty ginkgo]

Maybe it could be blue screening because the newest drivers is Cuda 12.2 and torch is only up to 12.1 I think

Because I got atleast 3 bluescreens and thousands of errors until I found a working version of torch that worked with my Cuda version bc I have the newest drivers

[high skiff]

I have cuda 11.8

[hearty ginkgo]

Duh fuck

thats mine

[high skiff]

weird, all I know is SD likes 11.8

[hearty ginkgo]

idk mine seems to work fine atm the version of torch im using is 2.0.1+cu118 also

only one that worked with cuda 12.2 for some reason took me like 10 trys unistaling and reinstalling torch

[dusk mica]

I thought sd likes the newest cuda

[high skiff]

@visual glade any statement on this?

[soft zealot]

I have no issues with SD and using CUDA 12.1

[hearty ginkgo]

Idk I got a bunch of errors that pytorch didn't support my Cuda or sum shit and then just found a torch version that worked

[high skiff]

hmm

[leaden knoll]

Does anyone know what time sdxl 1.0 will be released? I remember it was meant to release sometime today.

[high skiff]

its been delayed

no info currently on when it will be out

[leaden knoll]

Sad…

[civic sigil]

I dont think they would wait till last minute to delay it

[fair island]

did they say anything about that? where?

[hearty ginkgo]

Officially?

[soft zealot]

meanwhile before I forget :
@visual glade well impressed with the update I grabbed yesterday. A basic 1024x1024 using @high skiff workflow has gone from 80-90 seconds to around 50 seconds.

Plus I can actually run batchs of 8 without any issue 🙂

1080ti

[high skiff]

https://www.reddit.com/r/StableDiffusion/comments/152qs8j

thats one of the posts

https://www.reddit.com/r/StableDiffusion/comments/152oey2/hi_guys_is_sdxl_10_released_today/

this one too

[hearty ginkgo]

So rumers?

[high skiff]

no, they are from Joe Penna, one of the heads of SAI

[naive goblet]

https://tenor.com/view/blindspot-jaimie-alexander-jane-doe-nail-bite-anxious-gif-4555950

[high skiff]

but, they are saying its being delayed cause they have a few new tuned versions of it that are "blowing them away"

and they want them to cook fully

[leaden knoll]

Why would they delay it on the day? Seems like pretty bad timing.

You answered that actually, lol.

[civic sigil]

He isnt even saying it's delayed in this post lol

I am fine with a delay but why wait to announce it

[naive goblet]

Milk or cream how about some sugar

[high skiff]

now that I am not sure about

I agree

I would rather wait longer for a better product

don't want a cyberpunk 2077 of the AI world lol

ohhh

[dusk mica]

Wait

[high skiff]

windows event manager is seriously telling me off right now

[wicked frigate]

okay since that's leaked can i just say

the new models mentioned there?

hhhhhhhhhhhhhhhhhhhhhhhh

AAAAAAAAAAA

That is all

[high skiff]

better live up to hype :p

[wicked frigate]

that's not me trying to hype it that's just genuinely my own personal reaction

idk what others will think but those codenames are hiding some things i've been waiting to see happen

@sour obsidian just waltzed into my dms the other day like guess what i did and AA im happy

[dusk mica]

Kinda sad, really wanted to see 1.0 today. but i hope its worth the wait

[high skiff]

Same here

[dusk mica]

Hope its just a rumor tho

[high skiff]

the 1.0 tests in the bot's were... not looking so hot lol

[high skiff]

McMonkey is a dev for SDXL

@wicked frigatewith that said, the model is currently delayed?

[shy kelp]

I hope the delay won't be too long because the people who thought "I won't finetune on 0.9 because 1.0 is only weeks away" might reconsider, most people have been holding off on releasing any 0.9 finetunes, but if we're going have to wait more than a few weeks for 1.0, that might change.

[true wing]

Stability really needs to get their shit together PR wise. They do a great job hyping things up but their messaging around delays, leaks, errors, and controversies (like when they took that user's discord and subreddit) leave a LOT to be desired. If they're really about to announce the day of that there's a delay it's the worst possible move. They should have announced at least a weak before about a possible delay of there was even a chance. And all of those problems are actually exacerbated by how good they are at hype. Do they have a PR team? If so I'm sorry but you guys need to either step up your game, reorganize, or hire some more people on that front.

[wicked frigate]

as joe said, more info coming tomorrow (or, well, today technically now that it's 2am)

[true wing]

https://tenor.com/view/spaceballs-confused-huh-what-camera-gif-17264159

when will then be now?

[wicked frigate]

(sorry that was re Sytan's question)

[ionic dragon]

😭

[wicked frigate]

I'm just a dev I'm not in charge of PR stuff, schedules, etc. so I defer to the higher ups on that

[warm hazel]

You'll get your model guys.

[ionic dragon]

but when?

[wicked frigate]

Soon™️

lol

[ionic dragon]

im sed

[wicked frigate]

details tomorrowtoday

Tuesday

the current day but like give it somewhere between 1 and 24 hours idk how many

[shy kelp]

Yeah, Stability has time and again made a PR mess from things that really shouldn't even be a mild speed bump - a delay is fine and reasonable but announcing it (or rather, not even announcing it but just telling a select few people in discord) the day of the release is just terrible.

[warm hazel]

Gotta break out ol' reliable for this one...

[ionic dragon]

why is it delayed

[soft bone]

I guess SAI had to uphold their reputation

[true wing]

Apparently there are three models and they might be too good or some shit? Super vague messaging.