✨｜sdxl

[wild ice]

What interrogator did you use? Are you using Comfy? Automatic1111? What checkpoint? I'm trying to understand what influences the "interrogator"...other than the image itself which is being interrogated.

[mossy canopy]

https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[lusty wolf]

Watching...

[mossy canopy]

watching you watching

[stone fossil]

Lets jam.

[uncut fiber]

Mr. Hightower

[stone fossil]

https://www.youtube.com/watch?v=JbHI1yI1Ndk

[floral island]

he looks angry. somebody probably stole his prompt 😄

is that the nightvision model?

[stone fossil]

Lets cook some vram. 🙂

[stone fossil]

Noice one.

[fleet veldt]

we have controlnet\inpainting working with sdxl yet? was away on vaycay

[mossy canopy]

we have both

[fleet veldt]

nice. time to pull on auto1111 and go hunting the model files.

[mossy canopy]

yes

[floral island]

dude putting out major models

i feel they work really well

[mossy canopy]

overfitted..

[floral island]

probably a bit, can't quite tell yet where it's overfitted the wrong way

but overfitted models (like any anime model) can be used

[mossy canopy]

it can't do artistic very well.. just photoreal

[floral island]

it can't do artistic?

oh the nightvision?

protovision can do realistic and artistic

[mossy canopy]

not very well compared to others..

[fleet veldt]

and in my brief testing it does photoreal NSFW the best out of what i've downloaded and then smack with custom trained lora's via adetailer

[floral island]

use protovision 😄 (you vs the girl he tells you not to worry about)

protovision is a lot more creative

although i like the realistic output from nightvision you've shown

[mossy canopy]

look at the details here..

[floral island]

add 3d render (or 3d,render) in your negatives

[floral island]

🔥 ❤️

[stone fossil]

Yes more cooking vram. 🙂

Vram chefs died? 😦

[floral island]

model remembers what e30 looks like. I am pleased!

[stone fossil]

xD

[stone fossil]

Thats nothing place that back pff.

[floral island]

ketchup is involved

[stone fossil]

Now thats boombastic. 🙂

[floral island]

gudshit

this fairy is done with humans and their wars

[stone fossil]

Very nice broski. 🙂

[wild ice]

Help...I'm using automatic1111 and I've followed instructions for activating CLIP_stop_at_last_laers. I restarted stable-diffusion/automatic1111. Nevertheless, the "control" at the top of the UI has not appeared as expected. Ideas?

[uncut fiber]

Recheck settings if everything properly there. If you did it somehow and pasted here, there is missing Y.
Easiest way is in setting display all and there activate it @wild ice

[stone fossil]

Lets make snakers.

Drop some.

Fila*.

Nike.

Merge.

[uncut fiber]

i know therefore ✅

[stone fossil]

Les go.

[floral island]

just going through old prompts

[floral island]

very nice

[wild ice]

😦 I don't understand what you are saying. "There is missing 'Y' "? Maybe you are asking me to paste a screen shot?

[uncut fiber]

yes that is proper way how to do it, but they mess with UI, so it probably can affect this

[wild ice]

@uncut fiber Ugh! Well, never-mind. I have now refreshed the UI for the 3rd time...and THIS TIME the setting finally appeared.

[floral island]

❤️

[uncut fiber]

CLIP_stop_at_last_laers <- here is missing y 🙂

those flowers, impresionist expresionist and abstract, not technique mentioned 🙂

[floral island]

???

[wild ice]

Now my ultra bad question. I still don't understand 'y'. Is that letter short for something? I'm new to all things AI-art.

[uncut fiber]

no, just laers instead of layers 🙂 @wild ice

with foot included 😄

[floral island]

still working on old prompts... at 20k out of 80k or so...

[uncut fiber]

nice emerald eyes

[floral island]

sdxl based models slap hard

[uncut fiber]

80k prompts?

[floral island]

something around there

the amount of images i generated before i got to sdxl or so?

[uncut fiber]

thats a large number if each prompt is unique

[floral island]

well, i switched to my randomizers at some point

so having a lot of text prepared makes everything unique, and not at the same time

[uncut fiber]

yes me too using it, sometimes with for me very nice results

[floral island]

but trying old prompts really gives me a feel what sdxl models like and not

[floral island]

my randomizers being my own wildcard scizo prompt fueled with verser.ps1 song lyrics lol

[uncut fiber]

i tried add price tag, but it was completely ignored.

[floral island]

meh, really got to go to bed

alarm still goes in 5 hours lol

[stone fossil]

Use the force. 😄

[strong field]

Best goddam sneaker ever seen in my life fr

[wild ice]

Well, this is an artistic community! I still don't know what you mean. I think you are saying 'y' is short for "layer".

[uncut fiber]

@floral island GN!

[floral island]

#justonemore!

[uncut fiber]

simply, what command you posted, there were missing y in word layers. Thats all. My english is bad, so i cant better describe it 🙂 @wild ice

[stone fossil]

Ik zal morgen aan je denken als ik een jj op de bank rook slaap lekker gap. 😛

[floral island]

with this, it's time to go to bed 😄

sleep well my dudes ❤️

[stone fossil]

Ok so where where we. 😄

[upbeat summit]

@stone fossil you've been posting some great stuff tonight 🙂

[stone fossil]

Thanks heh, I am being supported by you and @sturdy abyss , @floral island , @dapper dragon .

Lets cook some more VRAM.

A love letter to SD.

🙂

[native knot]

This one. This one right here.

[wild ice]

In this youtube video. The fella is making a primitive hand "sketch", then de-noising. However, in my automatic1111...the "sketch" doesn't let me hand-sketch as he demonstrates...I can only mask. What am I doing wrong?

[stone fossil]

Hehe yeah some are very emotional heavy.

🙂

[stone fossil]

Lets click gen some more heh.

[native knot]

Been screwing around with PPF noise and using combined latents to get some interesting results. Essentially, if I generate the same info for with and without combining the latents I can get similar poses but simpler backgrounds with the PPF mixed in. For instance, here's a knight. Top image is with PPF mixed in and bottom is without.

[stone fossil]

Yeh. 🙂

This is my love letter to SD show me urs, output wise. 🙂

[native knot]

@upbeat summit - What do you think about this?

[upbeat summit]

it's looks very interesting. I'm not sure yet if perlin is the way to go for detailed compositions but it might be great for textures or complex patterns - how it's used in the industry. for a couple of days now I wanted to do some experiments with perlin noise - but haven't come around to it yet. I just saw that WAS updated his perlin noise node so I have to try it out very soon.

[native knot]

There's a great Blend Latents node here:
I'm trying a third box with different operations just to see what kinds of outcomes are there. The top result from earlier was done using a Latent Composite.

[stone fossil]

Dance some pictures. 🙂

Dance with me.

[native knot]

New workflow with all 3 images, plus the 3rd upscaled knight image

[upbeat summit]

heh yeah. need to make some images - but still have to do some work but I just finished including a new prompt building workflow in my setup. so looking forward to try that out

[native knot]

That was using "difference" as the operation, which is interesting.

[upbeat summit]

there's some interesting stuff going on for sure

[native knot]

I think this could be a pretty nice way to make some very small tweaks as well. The PPF noise can be refined very very finely.

[upbeat summit]

now I just need to try it out with revision, IPAapter and ControlNet

[native knot]

Yeah...a lot to do there.

Oh, and just as an FYI, the blend and composite are combining fine while I'm using AIT, so no issues there.

[upbeat summit]

yeah. I have experimented with different noise latent images and it should be definitely explored

are you still on the old commit to use AIT or is it fixed now?

[native knot]

I'm currently running on up to date builds and I think the only thing that's not currently working for me are TTN's nodes.

YMMV, of course

[upbeat summit]

cool. thank you. making a backup now and then git pull 😉

[native knot]

(((cracks knuckles)))...time to stage these into passes.

[zinc cargo]

is there a snap to grid option for comfy?

it drives me crazy that i'm not placing the nodes aligned

[heady vale]

there is

[zinc cargo]

what am i missing?

[heady vale]

cant remember how I did it now

probably an extension

not sure which one

[heady vale]

found it

[zinc cargo]

thanks!

[shy kelp]

https://i.postimg.cc/NfSjFxHH/6-Q3g-C9651u-KQAPo-P3l-VK-1-2wgxn.jpg
https://i.postimg.cc/zXVfqQVN/8u-ZHy4f-Lxc332-Y3u15za-1-pgcnu.jpg
https://i.postimg.cc/x1mTzbvF/Xp-Cyq2-ILcv4-Mvt9-YQz0c-1-cxokh.jpg

The same prompt in Adobe FireFly creates more or less the same thing, except that there is much less reflection on the ground and Adobe doesn't place houses at the borders. I don't want the houses, but SDXL placed them thinking that I wanted a street crossing a village. I said a village at the distance, far away.

SDXL also understood "chinese light poles" better than Adobe. Adobe placed lanterns hanging from the trees or lens flare floating as if they were fireflies

[hardy cipher]

My favorite thing about firefly is that sweet watermark

[native knot]

lol

I'll do you one better in a sec.

What's this, what's this?

[native knot]

Impressive...it knew to knock down the number of fingers.

[crisp owl]

reminiscent of Kingdom Hearts

[native knot]

Aye

[hardy cipher]

Nope, don't like that

[hardy cipher]

Need to enhance those pictures hanging up

[wild ice]

How, presumably with a prompt, do I control how close or far an object is? Or its position? I want "forest surrounding wise man" and I want the man to be somewhat centered and consuming perhaps 30% the height of the image.

[hardy cipher]

Just experiment. There's no one right approach. You can literally type in anything. Make a prompt out of what you just typed here. You can use terms like medium image, centered, etc. Be as descriptive as possible. The better you describe the scene the better the ai will know what you want

[hardy cipher]

Sometimes I make up words. People need to do more of that

Stable diffusion always knows what to do

[native knot]

I've done that a few times...usually I'm merging a few words together that make up what I mean...it's odd how well it understands that.

[hardy cipher]

Well I've thought about creating a node or something that will pit prefixes and suffixes and things together in ways that are syntactically correct but aren't used in the real world

But in latent space anything goes

I wonder how well it grasps latin

[native knot]

When do we all get to realize that we live in latent space and can fly?

[hardy cipher]

Hopefully soon

[native knot]

plzI'mdyin'ovahere

[crisp owl]

I just checked my hands, 6 fingers on one hand and 7 on the other.
I think that confirms I'm in SD Latent space

[native knot]

pix or it didn't happ....wait...

[crisp owl]

Oh I responded to the wrong post lol
But you still got it haha

[crisp owl]

Perhaps Peele was right in his "You Can Do Anything" video. He knew back in 2012 that we can do anything, we can literally fly

[uneven gazelle]

Hey, i'm having this issue a lot of time meanwhile i do nothing and it appear for no reason. Anybody know how to solve that ?

[native knot]

Sounds like a bad extension in A1111.

Take a look here:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/12155

Maybe it helps.

[hardy cipher]

"pip install tqdm" perhaps?

https://github.com/tqdm/tqdm

or install tqdm_auto

a1111 really taught me a lot about troubleshooting errors

making some conditioning modification nodes. not sure if they'll be useful, but you know, they'll be something

[soft zealot]

Problem narrowed down. It looks like a conflict with https://github.com/failfa-st/failfast-comfyui-extensions

[hardy cipher]

what's the problem?

[soft zealot]

follow the message link
#✨｜sdxl message

[hardy cipher]

ahh, didn't even notice that you quoted yourself and a youtube video. nice. I've not had that issue myself. but I guess that would make sense since I don't have those nodes either.

had them, but when I clean installed I didn't download as many things

[woeful patio]

quick question, Is there any news on controlnet for a1111?

[supple knot]

There was one that looked cool called "DW Pose "

[hardy cipher]

has a1111 kept up on anything in the last few months?

[woeful patio]

nah, but the latest update fixed a lot of things for sdxl tho

[hardy cipher]

that's cool. kind of a bummer to watch them drop the ball with everything. it's always had lots of errors and all that, at least for me. but I always reminded myself that it was free open source software, new technology, lots of things having to work together, etc. but when I realized they were falling behind in everything I had no incentive to keep using it

[supple knot]

yeah I change to Comfy UI

[woeful patio]

yeah all true, I still can't deal with comfy. It's just not my style of program. All those custom nodes and none of it makes any sense. Just let me use my a1111 in peace lol

[hardy cipher]

I'm not one of those weirdos that cares what other people use. but personally I wish I'd tried comfy sooner. but stupid redditors naysaying it, and that had been my main day to day source for stable diffusion information in the past

[supple knot]

But once you learn it , It show your more under the hood. and the images png have the meta data of the node in it. so can load a Comfy UI thur a png image made

[hardy cipher]

it just satisfies my brain

yeah, thing is, you literally never have to create your own workflow if you really don't want to

but as soon as I acclimated myself to it I immediately started creating workflows that probably didn't ever need to be made

[supple knot]

Any idea how to put the render image in progress under KSampler ? I saw someone have it in a video

[hardy cipher]

there you go

[crisp owl]

Not necessarily that one I'm hovering on, but those turn it on

[supple knot]

Thanks, Oh But its Slower,Ill test thanks

[crisp owl]

Yeah just see how you like it. I ended up sticking with just preview nodes for the time being.

[hardy cipher]

yeah, I don't run the live preview. I like to delay my disappointment

[supple knot]

What is badge and channel for?

[hardy cipher]

we get badges?

[crisp owl]

badge are those at top right

[hardy cipher]

ahh

I need to look at my settings. I forgot all about that

[woeful patio]

do you use the refiner when using a lora?

[hardy cipher]

no one knows the right answer, marksouls

[crisp owl]

My refiner is on for every run. But I don't use Lora's hardly. Just the offset.

I would probably default to keeping it on for using others though, then adjust if needed

[meager sparrow]

why there is no reference for best parameters we should use for training a model on SDXL dreambooth ? 😦

[hardy cipher]

because no one knows

isn't there default settings though?

[meager sparrow]

But there is lots of content for LORA SDXL.

[supple knot]

badge :looks like it counts the nodes , is nickname if someone add a nickname to a Node?

[hardy cipher]

well it's to help you figure out where you got it

[meager sparrow]

I'm not sure if I can use the same parameters that we use on lora for dreambooth as well.

[hardy cipher]

does dreambooth not train loras?

[supple knot]

Sick Helmet

[meager sparrow]

Good question, don't know 🙂

[supple knot]

eye yi yi

hows he eat?

[hardy cipher]

remember that southpark episode?

where people stopped eating through their mouths?

[supple knot]

not sure, the human centipede one?

[hardy cipher]

maybe. I think it had martha stewart in it

and she ate an entire thanksgiving turkey in an unconventional manner

[supple knot]

Theres some AI south park project out there

[hardy cipher]

like entire shows? or images?

[supple knot]

There a episode written by AI, and a Automatic Run one which is like ChatGPt linked to text to speech and the cartoon. there a similar one on twitch of Rick and Morty

[upbeat summit]

impressive

[analog spade]

Anyone tried making youtube videos with sdxl? It's pretty good even with creative prompts https://www.youtube.com/shorts/dpLSKQroYCk

[supple knot]

Yeah ,run them thru Pika

https://youtube.com/shorts/Vk3mG8u3sGY

[analog spade]

Any good image to video apis out there yet?

[supple knot]

this one called Pika its on discord

I dont know how to send a link for it

[zinc cargo]

gen 2 is always some kind of slot machine

[hardy cipher]

can you make a video with this?

[supple knot]

yes for sure, mama little baby like short bread

[hardy cipher]

that's it

what other dumb things could I do to conditioning output?

[supple knot]

@hardy cipher

https://www.youtube.com/watch?v=4Y28w2ErHj4

[hardy cipher]

grease bread. I forgot that I knew that song, lol

[supple knot]

@hardy cipher So in Pika I put in prompt :looking, blinking, smiling gen2 will do something crazy with this ill try it

[hardy cipher]

not good, my buds

decoded (tensor([[[[nan, nan, nan],
          [nan, nan, nan],
          [nan, nan, nan],
          ...,
          [nan, nan, nan],
          [nan, nan, nan],
          [nan, nan, nan]],

         [[nan, nan, nan],
          [nan, nan, nan],
          [nan, nan, nan],
          ...,

more inpainting

[supple knot]

@hardy cipher So Gen2 found image NSFW and didn't process

[hardy cipher]

lolol. nsfw?

how?

[supple knot]

idk its all automatic

[hardy cipher]

well I appreciate the effort

yeah, just wondering how it triggered the algorithm or whatever

[supple knot]

I get better results with pika and its free

[hardy cipher]

I wonder if either of these would make it through the filter

[supple knot]

Are those Sea Monkeys?

[hardy cipher]

rainbow goblins, silly

[supple knot]

yeah I already see the NSFW, they need pants

[hardy cipher]

fair enough, well rainbow goblins are free range and they choose not to wear pants

[supple knot]

Pants Lora

[hardy cipher]

the bottom right one in both images is extra special

[supple knot]

they have hats on or are there head shaped like that?

can you put them in New York?

[hardy cipher]

hmm, maybe. I have a few more. I just cursed those ones

[supple knot]

ill try it in pika

[hardy cipher]

they're wearing clothes here

[supple knot]

@hardy cipher

[hardy cipher]

beauty. I'll turn that into a gif and send it around. excellent work

[supple knot]

ill process the other 2

@hardy cipher

[hardy cipher]

https://tenor.com/view/dragoncon-ono-michio-otis-farthawk-gif-24453258

[supple knot]

peace out

[hardy cipher]

https://tenor.com/view/steve-carell-the-office-yikes-this-is-the-worst-cringe-gif-21219465

[hardy cipher]

thank you, gentlemanscholar

[low hill]

I keep an eye on you!

[willow bane]

i want to export to 2k

what should i do

?

@uncut fiber

[uncut fiber]

i dont know. Would send it to extra tab and try esrgan or swinIR probably. It will keep picture unchanged. Hires fix will change it a bit.

[echo temple]

Thanks a lot!

[uncut fiber]

two different rain-snows 🙂 many thanks to Thaevil1

[craggy ibex]

Does Automatic1111 run super slow?

[willow bane]

n

why

[craggy ibex]

I am trying to switch from ComfyUI to Automatic1111

but it's really slow

Idk how to make it faster hm

[willow bane]

but results are awesome

like deforum

https://imgur.com/hsRu4pq

see

[craggy ibex]

ok

[stone fossil]

Deforum is fun, till you do it 8+ hours in a row and your head explodes. 🙂

[ionic dragon]

how realistic is this?

[strong field]

Looks really good, slight mismatch on the eyes/pupils

The SD curse

[uncut fiber]

just sun and shadows on shirt? Dont know

trying dithered color, not sure it is working....

[vapid roost]

https://followfoxai.substack.com/p/part-4-advanced-sdxl-workflows-in
my weekly post - this time about building workflows in Comfy

[uncut fiber]

Could be real inspiration, bell tower in the middle of the square 🙂

[dense chasm]

anybody have a better idea how to implement the stage design?loRA or controlnet?i'm afraid the stage lighting is an obstacle and disturbing of final render

[upbeat summit]

new lora?

[vital ermine]

yeah

[upbeat summit]

nice!

[vital ermine]

I love the style of it

[uncut fiber]

i posted different 😄

[vital ermine]

it has some issues still so needs more training (steps)

[upbeat summit]

me too. looks a bit like TV Funhouse 🙂

[vital ermine]

tv funhouse?

[upbeat summit]

animated shorts as part of saturday night live

[vital ermine]

oh.

it has that 1950s vibe to it

[upbeat summit]

yeah

[vital ermine]

this is what it did to my robot 😦

[upbeat summit]

retrofied

[vital ermine]

yeah, but not a cartoon and made it have a skin face

[upbeat summit]

very nice 🙂

[vital ermine]

Very much like a childrens cartoon

[uncut fiber]

Lucky Luke!

[stone fossil]

Yep lora by @dapper dragon

[uncut fiber]

Not using Lora, only one tempting me a bit is Audrey Hepburn ❤️

[vital ermine]

how is it that no matter how much memory I have I am still getting a huge page file?

I crashed due to it yesterday with a WHEA uncorrectable error. Ran a mem test etc... all good

Does anyone still use Lexica to gen?

[uncut fiber]

any originality here?

[upbeat summit]

you made it - so it's an original 😉

[vital ermine]

Well, lookee there

bottom right

[upbeat summit]

and is artgerm in your prompt?

[vital ermine]

nope

[uncut fiber]

sdxl has nice steps and banisters.

[vital ermine]

No matter what I neg that sig remains

[uncut fiber]

i have with signatures issues as well, once i had impresion that blur corners helps, but not sure now.

[vital ermine]

I lost those issues when I left 1.5 for 2.0 then 2.1 now they are back

[vital ermine]

got it

[uncut fiber]

@vital ermine How?

[vital ermine]

(signature:3) killed it

[uncut fiber]

o.k. very strength value

[vital ermine]

yes

trying 2

nope, too weak

2.5 it is fading

[uncut fiber]

and no impact on picture?

[vital ermine]

2.7 is the magic value

[uncut fiber]

2,718 🙂

[vital ermine]

2.7

[uncut fiber]

2.718 is e 🙂

[uncut gull]

Working on a book cover, and I decided to ask SDXL to render the title, just to see what would happen. I didn't think it would work, because there's no way that these title words were anywhere in the dataset. So, here's my question: How does this work, exactly?
Does SDXL understand the concepts of spelling, letter-joining in words it has never seen, line breaks, and even typographical balance?
(Title "The Heavenly Ruins")

[vital ermine]

Diffusion has issues with text, period.

[uncut gull]

To be clear: SDXL can accurately spell words it has never encountered before, and make the text look beautiful in any style. Not with 100% accuracy, but I only rendered 20 to get a perfect render.
How is this possible?

It understands how to write.

It's not copy-pasting words its seen.

[vital ermine]

no, never

it is an issue with diffusion that is being worked on for all types of various diffusion based schemes

[uncut gull]

It is using an internal world model governed by rules. That's the only way this could be possible. 😕 (But I need to get back to work.)

[vast galleon]

messing around with the comfyui mask editor. got a zoom function working kinda/sorta. still some functionality breaking bugs to work out.

[uncut fiber]

Can you post same picture with and without (signature:3) ? I think it washing detail a bit.

[vital ermine]

I did it was the one before with just signature

[uncut fiber]

o.k.

[vital ermine]

you will never ever get the exact same picture if signature, or whatever, is in the neg as it is a filter

[uncut fiber]

yes

[vital ermine]

would be cool if it did, though

[sand pier]

🕯️

ERROR: LOW VRAM MODE NEEDS accelerate.
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4090 : cudaMallocAsync
VAE dtype: torch.bfloat16

I guess i need a better graphic card for sdxl

[vital ermine]

we all need 128GB cards now

even better

Oh, wow this is the 1960s TV Godzilla

[uncut fiber]

try Gapa 🙂

[vital ermine]

gapa?

[uncut fiber]

or gappa

[uncut fiber]

https://www.imdb.com/title/tt0061550/
Gappa

[vital ermine]

that was what xl says is gappa :/

[uncut fiber]

nice text 😄

[vital ermine]

LOL, yea

Ahem

that is Mr. Fabulous

2.7 not enough

5

must be listed as something not signature

[wild ice]

I'd like help with a prompt (or some other tool/extension?) that would cause an image to have an overall increase in firelight-orangeness added...as if an orange-like cellophane was overlaid on top. Ideas?

[floral island]

it's obvious what i'm trying. getting it to do it... is harder

[shy kelp]

only sdxl pro get this type of photorrealism.

I had to make another account to try again.

[thorny frost]

SDxl pro? What is that @shy kelp ?

[shy kelp]

1.0 "pro users only"

[lapis gale]

hm... "a car with (donut tires)"

[floral island]

SUCCESS!

[shy kelp]

this is way beyond what I can achieve with Adobe's FireFly. It's so life like.

[floral island]

lol

[wet nacelle]

room designed by john van damme

[thorny frost]

kappa_neuro is attacking again with his countless loras of artists already included in the base model. This time they work though.

i am downloading some, they seem to somewhat improve final results.

[uncut gull]

That would require training a noise offset lora. It's the same impossibility as trying to get a "bright" or "dark" image by prompting. https://www.crosslabs.org/blog/diffusion-with-offset-noise
The poorman's solution: I tried sepia orange vintage film, * subject * and that gave a reasonably good base. After that initial render, you would need to use photoshop / gimp / whatever to adjust the colors further.

(Prompt subjects courtesy of a Llama2 AI running on my PC. 🙂 )

[noble shoal]

Is that Owen Wilson?

[dapper dragon]

nope, sorry bro 🙂

[dapper dragon]

lonesome cowboy

[hasty smelt]

Does anyone know if a workflow for weak systems has been shared recently? I haven't been following the chat in the last few days. @high skiff had said that they would make a workflow available.

[floral island]

sorry, my lightsaber whiffed

[wet nacelle]

I LOVE SPAM! I LOVE BEING ON AN EMAIL LIST!!!!! YAAAAAAAAAA!!!!!mnbj we rtfljkhlvbhferjl;jhkv ewf

https://tenor.com/view/sam-hyde-samhyde-gif-24206394

This is not a stable thing. BUT FUCK ME MAN!!! WHAT DO I FUCKING DO!!! I CAN'T DO SHIT HERE!!!

YEARS OF THIS!!

Alright I'm done with the caps

[noble shoal]

🤔

[vast galleon]

a little more progress made on the zoom functionality. drawing is still buggy, but it now has a pretty background canvas.

[mossy canopy]

tf is going on with sytans workflow..

[floral island]

she's a scarred fire survivor?

[mossy canopy]

give me ur workflow.. mine is broken for some reason..

[floral island]

mine is A1111 + protovision (probably overtuned) model

[mossy canopy]

damn..

[floral island]

sorry my dude/dudette ❤️

[floral island]

fuuu that's really bad

i mean, refiner can't even fix that

[mossy canopy]

not sure if its the model

[floral island]

she quite literally stole his heart

[floral island]

which model u using?

[mossy canopy]

https://civitai.com/models/139565?modelVersionId=154593 this one

[uncut fiber]

How to get rid of donuts sticking to car, any idea on negative prompt

[floral island]

it might be that ur prompting style just fucks the model over....

[floral island]

what is the pos prompt?

[mossy canopy]

simple one sentence prompt.. no nonsense

[uncut fiber]

mmnt

Real car, (tires made out of donuts with toppings:1.6), seen from side, solid background.

turn down 1.6 do not help

[floral island]

"real car, donuts as tires"

i'd use that as my first prompt

or smth like that

[uncut fiber]

o.k. will try thank you!

[floral island]

just to make sure the model understands what i'm trying

some models have better comprehension, some have worse

[upbeat summit]

"homer simpson" as a positive token

[floral island]

my personal 1.5 model had great comprehension (but is a toddler compared to sdxl)

[floral island]

the worst part of SD in summer? the heat.

[floral island]

taylor swift vibes

[uncut fiber]

p.d. official desktop 🙂

[floral island]

i feel this image has 5kg of botox injected lol (perhaps not far from the truth nowadays, but what od i know)

[mossy canopy]

found the mistake. i have to use super low cfg with this model.. or else..

[floral island]

in my experience, that means the model is very knowledgeable

[mossy canopy]

or overcooked..

[floral island]

that might be tho

[floral island]

glorious donutmobile

[ionic gulch]

now you made her hide her face

[mossy canopy]

getting better

[shy kelp]

clipdrop generates the same style that I get with nightcafe. Except that it doesn't put houses by the road's side.

However. I keep saying "negative prompt, aslphalt" and it insists on asphalt damn

[soft zealot]

ok the perspective is off lol

[mossy canopy]

clipdrop pushes default prompts style..

[uncut fiber]

try asphalt 😄 @shy kelp or in positive cobble stone

[shy kelp]

I tried reimagine in clipdrop. Everytime I seem to get worse images.

[mossy canopy]

not sure if they are using revision or just an unclip model

[shy kelp]

I'm installing sdxl. Installed python, then invoke ai

the photorrealism in clipdrop seems to be a bit too hyper real compared to nightcafe

[supple knot]

With Comfy, why use Google Colab in your workflow?

[mossy canopy]

good luck installing invoke

[shy kelp]

there are better options than invoke?

[mossy canopy]

automatic1111, comfyui, sd.next, fooocus

[mossy canopy]

countless

[shy kelp]

I went for invoke because it was mentioned first in the installation guide

[cyan crown]

comfy for sdxl at the moment is the fastest and still better

[supple knot]

ComfyUI works great

[shy kelp]

when I install it. Is my PC generating the image as opposed to a remote service with much more processing power?

[supple knot]

remotes better I mean they have server size room computers

But Im getting great results with comfyui local, You have a lot of control with your image

[uncut fiber]

and what about official Stableswarm-UI?

[supple knot]

https://youtube.com/shorts/i1iePQ8PnGg?si=FJxSAsEg5IeAYwJX

[shy kelp]

uh now I don't know how to remove invokeai

[orchid pendant]

this trait is very beautiful

[supple knot]

there's so many AI systems usually its all in one folder

[shy kelp]

well, clipdrop doesn't have weight control

[orchid pendant]

what is the model?

[cyan crown]

standard + personal Lora

SDXL standard + a personal trained lora

[shy kelp]

oh damn... I installed py 3.11.

have to rollback

[cyan crown]

And this is another Lora

[orchid pendant]

I like to do this old comic strip

[ionic dragon]

@mossy canopy do you know any ai showcase websites?

with prompts

[mossy canopy]

https://pixai.art/

[ionic dragon]

are most of them waifu?

[mossy canopy]

https://lexica.art/

[mossy canopy]

https://prompter.fofr.ai/

[ionic dragon]

why are lexica images too bad, like they are too polished

[mossy canopy]

just take the prompt.. https://majinai.art/

[mossy canopy]

https://search.krea.ai/

[mossy canopy]

lexica was always bad..

[ionic dragon]

yeah

[ionic dragon]

are the prompts good?

[mossy canopy]

i dont use them.. i write my own or just clip interrogate it if its unknown to me ..

[lapis gale]

only problem is that the old prompts lose their relevance with new technologies really quickly.
remember when it was basically impossible to get decent weapons or tools from SD? now you just prompt for them normally and SDXL does the rest

just do your own prompts instead, describe what you want, and be a bit creative
e.g. this is "dark batman standing on top of a tall building at night looking over gotham city, made of bats, explosion shards frozen in time, cinematic"

[mossy canopy]

i always used controlnet to get them properly..

[lapis gale]

too much work for me 🙂
i just prompt for new stuff and don't try to replicate existing work or finetune my creations

oops, cape on his front side, this will be inconvenient

[lapis gale]

what i like about SDXL is that you can use the style selector (basically adding stuff to the prompt and negs) and get a fresh look for the same prompt

whoa, psychedelic strikes

dystopian indeed - the batman sign shaped bats are a nice touch

and so on 🙂

[mossy canopy]

where are u using sdxl?

[lapis gale]

local auto1111 on my 4090

5s per image 🙂

[lapis gale]

"((flat 2d cel shaded studio ghibli style)) dark batman standing on top of a tall building at night looking over gotham city, made of bats, explosion shards frozen in time, cinematic"

ok last one 🙂

[cyan crown]

ahahah this was an error!

btw this was done with 1.5 model trained with my face

[mossy canopy]

model?

[cyan crown]

sdxl + lora

[crisp owl]

My wife says she approves 😎😉

[cyan crown]

these are done with sdxl + Lora of my face

[crisp owl]

"mmmhmmmmmmm, yup"

[mossy canopy]

show some abs

[cyan crown]

abs?

[mossy canopy]

for his wife

[cyan crown]

😄 yes but I don't know abs what stands for

[shy kelp]

larger than 512 x 512 gives me out of memory error

[mossy canopy]

gpu?

[shy kelp]

8GB

3060 Ti

[cyan crown]

512x512 is not a good res for sdxl

[mossy canopy]

invoke?

[shy kelp]

automatic1111

[cyan crown]

512x512 +sdxl = bad results

[mossy canopy]

oom

[cyan crown]

yes but sdxl can't work with 512x512

[uncut gull]

A1111 can do SDXL 1024x1024 on an 8GB card though... I'm 90% sure.

Launch with --medvram or --lowvram

[mossy canopy]

add some optimization arguments other peoples are using like he said..

[crisp owl]

For 8gb, use --medvram
I haven't used A1111 since SDXL came out though, so can't speak to how my 8gb 2060s works with it.

[shy kelp]

much lower quality than clipdrop

[uncut gull]

You're using SD 1.5. Clipdrop is on SDXL. (Also, your prompt needs some work.)

[shy kelp]

so there is no offline sdxl 1.0

[zinc cargo]

make a mexican wrestler

[mossy canopy]

coming up next

[zinc cargo]

pure prompt?

it's beauitufl

[mossy canopy]

just prompt

[crisp owl]

Installed locally, yes.
Automatic1111, Comfyui, Swarmai, etc etc

[zinc cargo]

what i was meaning to say is....
make her a mexican wrestler

[mossy canopy]

whom?

[shy kelp]

dreamstudio is more or less the same result as nightcafe

[mossy canopy]

https://tensor.art/ try this

[shy kelp]

it's funny that I say "village in the distance" but it places houses at the road's borders.

[cyan crown]

Las tone before going to sleep. DRacula and his castle

[shy kelp]

so I added "houses at the road's borders" to the negative prompt and it worked

[shy kelp]

the only thing left is to fix the excessive lighting. It's placing just too much light.

[ionic gulch]

some results of ip adapter experiments with my workflow

[steel mist]

It’s a pretty nice image as is. I’d honestly drop it into an image editor and adjust the lighting and color temp with curves if prompts aren’t getting you where you wanna go

[orchid pendant]

this was perfect damn

[hardy cipher]

what is going on here?

[shy kelp]

I got that in img2go. I asked for Elon Musk practicing moonwalk and it made two elon musk, one with three legs

what are those merged body errors?

[hardy cipher]

these are all cursed images, my friend. that is the explanation

[glad grove]

theres a red pepper with a human leg in the background 🤣

[hardy cipher]

that collage randomly has a big bird/howard the duck demon on the left

[orchid pendant]

anyone who knows an unlimited generator for free? really good

[hardy cipher]

you have to ask yourself what the incentive would be to create such things

[orchid pendant]

give life to my ideas, and frustrations

[hardy cipher]

use the bots on this server

[orchid pendant]

it's giving a very annoying problem, when I try to use a prompt that I've already used it appears to download in BIN

then I have to rename it to PNG, but then I keep downloading a bunch of poorly made images and deleting them

[shy kelp]

VPN and multiple accounts

[hardy cipher]

change the prompt slightly if this is an issue with you. add spaces or punctuation. problem solve. I really have no answers for you other than that

firefly is so lame

stable diffusion dgaf. stable diffusion will make images from the string of characters left when my cat naps on my keyboard

[orchid pendant]

I am going to try

[hardy cipher]

good luck with all that on discord

[nimble heart]

graphics card

[hardy cipher]

my old account got banned because they didn't like my vpn and I was labeled as a spammer, lol

[empty osprey]

hallo, i was wondering if what the difference between controlnet rank128 and and 258?

[hardy cipher]

never once spammed, but you know, that's not the important part

and their customer service people helped a lot by doing nothing at all to help me

[nimble heart]

did you ever start on that gmic node?

[hardy cipher]

well, I'm figuring out the basic workings of conditioning. but I would say I'm not very far yet. I'm learning the arithmetic so I can move up to the algebra

[orchid pendant]

I'll keep renaming the downloads, multiple accounts is even more annoying

that's it

[hardy cipher]

I have a conditioning normalizer node. not sure what benefit that serves, but I made it

[nimble heart]

auto does something like that

see if your results are comparable when both using cpu seeds

[shy kelp]

#1100170312106127410 message the bot is also great. Except that I missed some parameters

[hardy cipher]

my problem lately is I keep getting this sort of thing

samples {'samples': tensor([[[[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]],

[nimble heart]

in general or just with your custom nodes

[hardy cipher]

man, I'm having a hard time telling. tbf I haven't had a chance to spend a bunch of time with them. but can't tell if it's a vae thing or the custom nodes. realized I'd been modifying the conditioning tensors themselves, and then trying to figure out if that carries over to new image generations or what. so just trying to get my bearings with it all before I step up to anything more complicated

it's like opening a can of worms for me. it's all stuff that will be simple in hindsight, but isn't yet

[nimble heart]

assuming you're using that new 0.9 vae, make sure you're not autocasting to fp16 when starting comfy

[hardy cipher]

ahh, that could be an issue. because the 0.9 vae keeps giving me black boxes

[nimble heart]

beyond that idk. I've only NaNd a whole tensor in a custom node by mutating it directly without a clone which caused each subsequent run of the same seed to compound the modifications cause of the comfyui caching

so the numbers just kept growing until it broke lol

[hardy cipher]

yes, this is what I did. in place operations, lol

but then I fixed it, or thought I did

[nimble heart]

if you're directly mutating a tensor using pytorch/numpy ops just clone it first thing in the function

else you mutate the cached version it seems

cause comfy doesn't clone its own version for the cache

which makes sense ig since python doesn't discriminate between read-only and mutable data

probalby why Comfy talked about rewriting the whole ui in C lol

[hardy cipher]

yeah, I just have a problem with learning all the ins and outs of things in detail before I jump in

so end up breaking stuff sometimes

[nimble heart]

nah that's just how it is in the hacking worl

d

break things before you fix them

just like

back up your shit before you do something that could potentially brick your data

and you're good

[hardy cipher]

I'll start it up and see what sort of issues it gives me now. keep tweaking things here and there

other thing that discombobulates me is all the different data forms and understanding what requires what. tensors, lists, tensors that contain lists, tensors that contain tensors, tensors that contain tensors and lists, tensors that contain tensors that contain lists

alright, adding "--force-fp32 "

hopefully that helps

[hardy cipher]

I set it up to clone and create a new tensor when I realized what was going on. but what exactly was it even changing before?

Forcing FP32, if this improves things please report it.
...
VAE dtype: torch.bfloat16

wtf

so is it overriding what I'm tring to do?

[nimble heart]

so comfy doesn't build a new tensor every run. It only makes a new one as far back as the prompt changes.
so if the prompt doesn't change leading up to your node it feeds your node the same tensor every time, and if you mutate that tensor directly that's how you get the compounding effect

[hardy cipher]

ahh

so that's not that bad I guess. not what I want though

[nimble heart]

on nvidia 3000+ cards it adds --bf16-vae by default

idk how to remove it

if there's an --fp32-vae maybe? BF16 shouldn't make NaNs though...

[hardy cipher]

I don't know. I'll see what happens now that I changed the code a little bit

sweet sweet errors. alright, I need to figure this one out now

TypeError: can only concatenate list (not "tuple") to list

boo hiss

[shy kelp]

hey. For some reason my prompt keeps generating straight roads

I even tried negative prompt "straight road"

hmm.... Swapped "sakura trees" with "willows" that made the whole image different. It seems that the AI thinks that if you want sakura trees, it means japanese architecture.

[supple knot]

looks like sakura a anima character also ,try SakuraArchitecture

So SakuraArchitecture, trigger the Anime character for me

[shy kelp]

gta 7

[glad grove]

sensei

[glad grove]

theres a guy there with face made of fur 🤣

[hardy cipher]

lots of perimeter fur in that one

[dim steeple]

hello, I have been away from SD for a few months. Can someone help me install and setup SDXL in auto?

[steel mist]

https://www.youtube.com/watch?v=1Ldc5kZKKFw&t=150s

[dim steeple]

Also, is SDXL pretty much better than 1.5 in everyway? or no?

[glad grove]

its better in some ways but not everyway

[hardy cipher]

well more training parameters or whatever, but it was trained on 1 megapixel images, so 1024x1024, vs 1.5 being 512x512. this means that in order to get good results that mostly lack things like extra limbs you'll need to render images that are around 1 megapixel vs .25 megapixel images with a1111. so 4x the number of pixels. basically takes longer to make images. but they're normally more detailed, and a higher percentage of wins. 1.5 is more of a numbers game, make a bunch, hope a few work out

[native knot]

Really, the only serious detractor right now is the time to process vs 1.5. All the patchwork and add-ons that were used to make 1.5 good are in the process of being brought over to XL, which will only continue to make XL better. Speed will come over time as there are new processes and new hardware. It's inevitable that 1.5 will become outdated over time. But hell, eventually the same thing will happen to XL.

[glad grove]

yes but what about the porn

[native knot]

We have to go... deeper ...

[dim steeple]

so 1024x1024 for sdxl?

Also basic question but where is the checkpoint to download?

Also is the raw checkpoint best rn?

or are there good fine tuned models

[native knot]

Or any other resolution where the sides are divisible by 64 and the total pixels are close to 1 megapixel. I posted a table here: #✨｜sdxl message

There are some really great models for XL out right now above and beyond base. Everyone's got their preferences...I like these (among others):
Copax Cute
Copax Melodies
Copax Timeless
Copax Vivid
CrystalClear
DeepBlue
Dreamshaper
Dynavision
Juggernaut
Nightvision
Protovision
NijiSpecial

huggingface has the base model and refiner

[hardy cipher]

some people aren't into it, but you might want to download the refiner model as well

as soul just said

[native knot]

https://giphy.com/gifs/HBOMax-hbomax-gordita-chronicles-on-uB9rh22TTr82ilZ4Hm

[hardy cipher]

no one really knows what it does, but it refines, so there's that

[glad grove]

maybe its like high res fix but laggier

[native knot]

I'm checking Civit like multiple times per day with my main view filtered to XL checkpoints...lol

[hardy cipher]

man, whoever came up with the structure of things on there has problems. I set my preferences to only xl. and then I go there and it shows me all sorts of cool looking loras which I download not thinking. and they're 1.5 loras

what would be the benefit of doing such things?

[ionic gulch]

is there a way to prevent a node from delivering / using cached results? i wanted to use a node from "wlsh nodes" to get the time for a filename, but it somehow delivers cached values, so even if i generate minutes apart, it always delivers the time from the first start. this is the code: https://i.imgur.com/9LjfcSx.png

[hardy cipher]

hmm

[native knot]

Are you filtered like this for LoRAs?

[hardy cipher]

no, I just had it set to sdxl 1.0. but it had loras and checkpoints. I'm assuming 1.5 checkpoints as well

[native knot]

It shouldn't show you 1.5 stuff unless the uploader was stupid.

Sometimes there are releases where they've got both a 1.5 AND an XL version, so when you go to the page, you need to look at the top bar to see which one you're looking at and download the right one.

[hardy cipher]

I think it caches everything until it's triggered to change. example being conditioning. if you don't change any of the parameters impacting the conditioning it runs off the cached version

I figured this out when I was modifying the original conditioning data rather than creating new. that gets out of control quickly

[native knot]

For instance, this one has both an XL and 1.5 version:

[jovial arch]

Does anyone know how many images SDXL was trained on? 2 billion?

[hardy cipher]

but, perhaps you could just change the cached time?

[ionic gulch]

but the seed somehow knows it has to run again. guess that kind of behavior must be transfered to that get time node

[hardy cipher]

you can modify the cached data. but it won't overwrite unless it deems it necessary

[hardy cipher]

well it was trickery. they weren't showing me what I had my setting restricted to. it was just "things you might like" or something along those lines. makes about as much sense as the search setup on there. I turned all that crap off though. or as much as I could

well I also double check what version it's made for.

it's not a big deal, but wasted my time

[dim steeple]

this is the model download right? https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

[native knot]

yup

[hardy cipher]

that's the official base release

good place to start really

[dim steeple]

So I should download that, the refiner

anything else i should know?

[native knot]

Using base for a bit will get you used to the way you should prompt.

[hardy cipher]

might want a vae

[dim steeple]

can you train models with dreambooth with xl?

[native knot]

Just try things. That's honestly the best advise to start out with. Once you have it up and running, just try shit.

[hardy cipher]

but they're also baked in sometimes

soul is correct. just get it working and then once you're acclimated and comfortable with the base expand to other things

[heady vale]

Ive noticed the same thing. Look at the top left of the image if it has Lora XL before clicking

[dim steeple]

which one?

[hardy cipher]

2

[native knot]

middle

[hardy cipher]

3 is a lora

[dim steeple]

cool, thanks for the help

[hardy cipher]

you probably want it

but it's not required

[dim steeple]

what is the lora version gonna do?

[hardy cipher]

well that's not a version of the model. it's seperate.

basically it adjust the noise offset. just download it and play with it. it'll change the dynamic of the resulting images

but like I said, it's not necessary, just helps with finetuning your results

[dim steeple]

How to get the refiner?

[hardy cipher]

should be on huggingface

[dim steeple]

Also wow it takes way longer to render than 1.5

on a 3080

[hardy cipher]

4x longer. 4x the number of pixels

maybe a little longer than that actually. but haven't done the math

[native knot]

I'm running on a 3080.

[dim steeple]

its taking maybe 10-20x for me

maybe just bc its the first job idk

[hardy cipher]

well is interdimensionalplanetarytv on a1111?

[upbeat summit]

I guess my lora weight is too damn high

[native knot]

lol

[hardy cipher]

goodness 😂

[glad grove]

by the time sdxl is done rendering i would have done at least 4 pics on 1.5 with high res fix on ( 1080p)

[dim steeple]

oh i think it was bc i had high res fix on

[glad grove]

😬

try it with a base 1080p then high res fix to 4k

[hardy cipher]

nice hands in that one

[glad grove]

bro needs a new haircut

[hardy cipher]

I think it looks alright

[wet nacelle]

what are you using to get these?

[native knot]

That one is nice, Mack.

[wet nacelle]

So are most of yours.

[dim steeple]

what "switch at" setting is good for the refiner?

[wet nacelle]

Okay. So I use a base step of 20 and then 5 for the refiner.

I don't know much and base my info from what I use of others.

[native knot]

When using base+refiner, it's a bit of a "play with it and see", but like Mack, I was running around 60-80% on base, then refiner for the rest. I did toy with a workflow that was base, then refiner, then base again on top...that gave an interesting and polished look to certain images.

[wet nacelle]

I did the same here actually. I forget whose it was but it was a really interesting way of doing it.

[dim steeple]

how long does an average image take for you guys?

I am getting like 2+ min

[native knot]

Depending upon what you're doing, that may vary. You're using A1111, I assume?

[dim steeple]

yes

[wet nacelle]

I get one every 20 seconds. I use COmfy UI

[native knot]

Ah...a lot of us are on ComfyUI.

[dim steeple]

i should probably switch

[hardy cipher]

comfy is the way

[native knot]

With the right workflow, I can put out a really nice txt2img in under 30 seconds on a 3080.

[wet nacelle]

@hardy cipherYO!

[native knot]

Run that same thing but against the 640x1536 resolution.

I bet that will look insane with the taller res.

[wet nacelle]

bro thinks I know how

[native knot]

lol

[native knot]

Seriously, though. Whatever that prompt actually is would look great on that tall aspect ratio.

[wet nacelle]

cyberpunk exterior house art by barclay shaw

The issue is that I'm using a model that is forcing bullshit so the art style won't commit

[native knot]

Ah.

[native knot]

Nightvision, from what I can see, eh?

[wet nacelle]

Yeah. It's a great model but is so not what I thought it was going to be. He made something great here for sure.

[hardy cipher]

File "C:\AI\ComfyUI_windows_portable\ComfyUI\comfy\sample.py", line 42, in broadcast_cond
    copy += [[t] + p[1:]]
TypeError: can only concatenate list (not "tuple") to list

stupid tensor

[native knot]

Yeah, I stopped using Nightvision for people and started using it for other things...a similar thought came to mind.

[wet nacelle]

HEY!!!! HOW DID YOU MAKE THESE!!!

[native knot]

He's totally ignoring ya.

lol