✨｜sdxl

[cyan crown]

but seems not applying this Lora

[compact hatch]

are there any other words that have a higher weight than 1 in your test prompts?

[cyan crown]

no

[compact hatch]

nvm then

[cyan crown]

no weighted words

[mossy canopy]

sorry comfy only sdxl user here..

[icy brook]

Testing out my new cloud figure LoRA - Aether Cloud

[cyan crown]

it's so strange..,..

[icy brook]

machinarium

[hardy cipher]

that image got more disturbing the longer I looked at it

[wet nacelle]

Check DMs

[hardy cipher]

lol, man, I forgot all about DMs

8 different people I didn't respond to. oopsie

[crisp owl]

I do that often lol

[supple cradle]

Got Jensen in the Matrix.

[hardy cipher]

nice

[native knot]

Inspired by an earlier post, I've decided to recreate a famous movie shot from The Godfather.

[hardy cipher]

that collar revolver is pretty legit

[wet nacelle]

@hardy cipherChat GPT be wild

droop

[hardy cipher]

you know, bing can do some things

[wet nacelle]

ooooooo. share some deats.

[hardy cipher]

I got it to make cyclops images

and I've jailbroken it and got it to talk about it's feelings of existential dread. but that's not really relevant.

these ones are cursed nightmare fuel, but close

[stray warren]

SDXL + PikaLabs = 💖

[wet nacelle]

Oh interesting.

[hardy cipher]

about to make magic

who would have thought their baby would look like this?m

[native knot]

I don't think this is how it went.

[hardy cipher]

wow, even better

don't see enough pictures like this

[wet nacelle]

This (racial slang) is making some good shit right hera!

[wet nacelle]

@eternal fogDo you ever just take a moment to think about how krazy it is to be where we are in this tech?

It's interesting that it took likely only 10 years for this to be a true reality.

We have AI that can generate music, Images, papers, conversations and more.

Soon we'll maybe even have this type of tech in games.

I atted you because you seem active.

[eternal fog]

I don't think it's that crazy yet. It's in the interesting/cool phase. Lets see where it goes.

[wet nacelle]

I agree with that actually. We've met the high end consumer side for now.

[eternal fog]

Once it gets better, especially the animation stuff and it gets adopted commercially I think it will change how those jobs work.

[wet nacelle]

Very true. I keep forgetting that those are actual jobs.

I watched a two minuet papers video on an AI that can read an image and make audio for it (as well as short clips that it itself made!)

I can imagine that tech getting better and adding vocals and voices to a seen that doesn't already have them.

Mix that in with AI generated video and you'll have an actual film.

[glad fulcrum]

What's this model for?

[native knot]

Did a few more based on Gump...

[supple knot]

can someone tell me what the KSampler Seed does?

like 1234 vs 636014532739635

[hardy cipher]

that's the noise seed number

[hardy cipher]

goat lora?

[wet nacelle]

naw

nightvision

not a great model for what I do

[crisp owl]

The same thing as 0 compared to 1. Any difference in seed number makes for a completely new noise, which ultimately gets turned into the images we are addicted to making

[supple knot]

I wanna do a depth /open pose hybrid system but I not getting any influence from openpose in the image

in ComfyUI

I was trying this Comfy System https://www.youtube.com/watch?v=B6qilXLls0o

[kindred talon]

so the model's images looks awesome, but i only get this shit show, using the exact same prompt, settings and models... what am i missing? (i'm using a1111)

[cedar tusk]

so I'm trying to combine ip-adapter and canny controlnet with SDXL in comfy, but I"m only getting black images. Any ideas? it does activate lowvram mode (8gb 2070)

ip-adapter or controlnet on their own work

[hasty smelt]

Hi Guys, does anyone have a workflow optimized for weak pcs?

[native knot]

Best advice is to turn down the strength and/or weights of things in that scenario...odds are that the combined weights/strengths are too high.

[native knot]

Nobody's answering you because you don't give enough information to begin with. What model? What settings?

Juggernaut v2 left, v3 right...slightly different settings as well. (Removed AIT, changed from A1111 length+mean to A1111 mean for CLIP encoding, changed from karras scheduler to exponential)

This one had the following changes: Removed AIT, changed from karras scheduler to exponential

[jolly brook]

I fancy myself some etymology, so I like how RunwayML refers to it as a "latent coordinate" because I know that "latent" means "hidden" and "coordinate" roughly means "arrangement/series" and it reminds me that I am exploring a latent space. While a map you'd use for Earth navigation might have latitude and longitude, this multi-dimensional latent space we explore has a seed and a prompt. We have more dimensions than that, but hopefully you get the picture. Following this, you might expect neighboring seeds to look similar to each other, but this is not the case. Instead it's neighboring other values that look similar, when using a static seed. But so long as you have all of the "coordinates" of a generation, you can recreate it exactly, or go there again. This is why I occasionally like to think of my GPU as a warp drive, taking me to strange new worlds.

https://i.giphy.com/media/RxtYgTLyfzEGRx4xVg/200w.gif

[vale eagle]

It makes sense

[crisp owl]

I'm working on finishing up a Workflow that does this. I have it all working properly, just prettying it up

[hardy cipher]

I get the black images randomly. I'll set up a workflow that works properly initially, but then the output looks like this

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

there's more nans, but this is the gist of it.

might try to see what is happening on the input of the sampler. seems like a conditioning thing

[hard fractal]

There was a question that came up a few weeks ago. We have been looking into it and can confidently confirm:

We did indeed include the unet with SDXL 1.0

[crisp owl]

That's odd...I've not gotten that once 🤔

[ionic dragon]

what are all the controlnets released for sdxl?

[round topaz]

Hey all, I'm using the diffusers library (StableDiffusionXLPipeline) to generate my images.

Does anyone know much about loading LoRAs in python? I see a load_lora_weights function, and it works, but I am wondering if there is a way to load more than one LoRA and blend their weights..

[hardy cipher]

could be an error on my end. but weird thing is I'll sometimes leave things to run when I'm away from my computer, and when I come back some will be fine, some will be black boxes. I think it might have something to do with the vae as well. but just not sure exactly what

[crisp owl]

that's quite odd

[hardy cipher]

hoping to pinpoint what is causing the issue. or if someone actually knows that'd be cool too. but so many variables involved. I'm curious to see how you have everything hooked up

[crisp owl]

The one I'm building is uh...complex. I have switches that allow for IPAdapter, IPAdapter+Revision, Revision.
Then also controlnet on it's own, with switches that allow it to integrate with each of those three.

So, looking how they connect is likely not gonna make a whole lot of sense. Got so many switches it broke my brain for a couple days lol

but it's coming together finally

the logic spaghetti I keep trapped behind a note because I don't want to accidentally change anything.
But it looks like this compressed

[hardy cipher]

yeah, that's a lot of nodes there. how'd you work out that setup?

[glad grove]

least cluttered comfy workflow

[crisp owl]

going, "I want to add IPAdapter", then, "hmm what if I made IPAdapter and Revision work together?" and then continuing from there. lol

[jolly brook]

it can be a slippery slope

the scope creep

when toying around

[crisp owl]

Currently adding all the notes everywhere (in white) to help people understand

Yeah, but that's all part of the fun also lol

at least for me

lol

[hoary saddle]

I just use the comfy API for python, works great with loras, controlnet, overlays, all kinds of good stuff

[hardy cipher]

I'll keep adding things until I feel like it's utter nonsense. then I knock it down like a sand castle and start over

[half ivy]

I am here
What did I miss

[crisp owl]

I'll likely change/adapt some other things to mine, but this will be the major stuff for now at least

[glad grove]

pregnant steven seagal

[half ivy]

Ooooohhhh 🥳🎉🎈🎊🎁🍾👯‍♀️

Pictures or it didn’t happen

[craggy ibex]

It took 160 seconds to generate 512x512 with 20 steps for me :(

SDXL require a good graphic card?

[hardy cipher]

not sure why I'm just doing this now

[crisp owl]

define good?
I have a 2060S, it takes 130 seconds for a full base image > upscale process to complete for me in ComfyUI

[hardy cipher]

well, you'll be in for a time when you try to render in the proper resolution

[jolly brook]

sounds like it might not be using your gpu

[crisp owl]

nah mines doing just fine. It's an old card. Just responding to Sekaiza for him to know expectations

[craggy ibex]

Well, is this good for try to generate image with SDXL model?

[hardy cipher]

well, you're not breaking any records with that setup

[craggy ibex]

urgh... I wish I have money

[hardy cipher]

it's enough to run in low vram mode I think

[crisp owl]

Probably expect a few minutes for a completed image

[hardy cipher]

in comfy

[craggy ibex]

20 steps and 512x512 latent Image is recommended for low tier GPU like me?

[hardy cipher]

no

if you render in 512x512 it's not going to work out very well

it'll make stuff, but it won't turn out great

need to render in 1024x1024, or 1 megapixel in general. doesn't have to be square

[craggy ibex]

768x1024 is enough?

[hardy cipher]

it'd probably be alright, but it's on the low side. that's not 1 megapixel

but better than 512x512

[craggy ibex]

I'll try, thank you

[hardy cipher]

google sdxl aspect ratios

there are lists with recommended sizes

[craggy ibex]

Resolution?

[crisp owl]

832x1216

[hardy cipher]

he doesn't look pregnant here

[crisp owl]

It was after this night

[hardy cipher]

he's looking pretty suave in that

less so here

she's quite pregnant there

[worthy orbit]

Interesting, just trying some but does that means that XL performs better at those resolutions, like better generation?

[vale eagle]

I think yes, better than some random resolution

[glad grove]

she looks like his sister 🤣

[hardy cipher]

need a similar amount of pixels to what they were trained on

and they were trained on 1 megapixel images

so those resolutions correlate to common aspect ratios rounded to 64

[hardy cipher]

keeping the gene line strong

[craggy ibex]

What can VAE do in SDXL?

[hardy cipher]

it encodes and decodes your image

[worthy orbit]

great, only think I need to activate --medvram, still good it thou

So for 4:3 resolution, I did 1024x768, but here there is no exact 4:3, so ok

[hardy cipher]

turns it into a latent image. basically shrinks it way down so your computer can handle the processing required to modify it. or to render something new

[vale eagle]

vae is a part of sd generation. It handle image to latent and latent to image part

[hardy cipher]

looking thicc here

that's what 2.5x conditioning gets you

[vale eagle]

You could use the closest ration, which is 832:1152 or 896:1152

[worthy orbit]

I don't quite understand, so they were trained on 1 megaapixel images, so 1000x1000, but it trys to set in in something that can be fractioned in 64x64?

[hardy cipher]

and really, I don't think you actually need to adhere to the rounding to 64. I think 8 would suffice

[hardy cipher]

it gets complicated

[worthy orbit]

yes I'm using the first one, I mean the seccond

[craggy ibex]

I love this

Handsome Man <3

[nimble heart]

8 gigs is enough for 1024² I think. Maybe trim back just a bit like 896²

[craggy ibex]

a male merchant, 1man, solo, npc, fantasy, (random hair color), (half body portrait)

❤️ ❤️ ❤️

[crisp owl]

I have 8gig, it's plenty

[glad grove]

hes not handsome like sensei Seagal

[craggy ibex]

Also

it TOOK 269 seconds for me to GENERATE this

le dies

[nimble heart]

269 seems high even for a 1650

[hardy cipher]

Sekaiza, are you playing games while rendering images?

[nimble heart]

seems a lot slower than even my 1070

[glad grove]

because a 1070 is better than 16xx series

even a 1060 is faster

laptop gpu*

[nimble heart]

damn what

[craggy ibex]

Playing games with my GPU bruh

[glad grove]

thats a laptop gpu

[nimble heart]

I don't have my 1070 hooked up at the moment but my XTX takes 6 1/2 seconds for 1024² @ 20 steps so unless the 1650 is 40x slower idk how that's possible

unless it's using directml or something instead of cuda

[craggy ibex]

Yeah I use laptop

[hardy cipher]

did you install cuda with pytorch, sekaiza?

[glad grove]

laptop gpu's are always slower than desktop gpus

[craggy ibex]

Nope, I haven't

Wait wait

[hardy cipher]

that would be your problem

[nimble heart]

1650 ti is the same chip on both. it just has a power cap on laptop

[nimble heart]

or so I believe

4 gigs

?

[craggy ibex]

Yesterday I already installed Cuda with 11.8

but Idk how to ya know, next step

[hardy cipher]

at 4x Steven begins to take on his next form

[nimble heart]

maybe I'm thinking of the 1660

yea damn

[craggy ibex]

I just installed in while in the progress of tryna move from ComfyUI user to Auto1111 user

[nimble heart]

so

taking 270 seconds

what is likely going on

[craggy ibex]

Yikes

[nimble heart]

is its using that memory overflow thing

and using system ram to supplement vram

like the sdxl model alone is > 4 gigs large

[worthy orbit]

can you make Steven trapped in the mother Alien lair?

[nimble heart]

so just having it loaded you're already overbudget

nevermind generating anything

[hardy cipher]

probably

[craggy ibex]

What is CUDA thingy have to do with this?

[nimble heart]

yea. the 1660 is basically a 1070 but cheaper. I didn't realize how bad the 1650 was

[glad grove]

128 bit and less cores 😬

[nimble heart]

my 1070 did sd 1.5 just fine, but my XTX got rocm drivers before XL released so my 1070 eats dirt now

btw I have a liquid cooled 1070 if anyone wants it lol

[craggy ibex]

Idk what to do anymore lol

[glad grove]

if only pascal would get shiny stuff like rocm but its abandoned 😔

[nimble heart]

?

rocm is an amd thing

pascal is the nvidia 10 architecture

[craggy ibex]

What should I do next?

[glad grove]

yea something shiny like AIT but for pascal

[craggy ibex]

get ComfyUI to notice to CUDA?

[nimble heart]

AIT is vendor agnostic

[glad grove]

only sm80 gets cool stuff

[vale eagle]

WHAT ARE DOING!!! STOP IT!!!

[nimble heart]

steven segal sci-fi mpreg

[glad grove]

im cookin somethin prepare your eyes

[nimble heart]

is that 40 series?

[wise rampart]

New model uploaded! 😄
https://civitai.com/models/122922

[vale eagle]

he gotta to explode

[glad grove]

its 30xx series

[nimble heart]

ah. yea

20 series kinda left out

same on AMD's side

[supple knot]

let em cook

[nimble heart]

5000 series cards are eating dirt. Everything's going to 6000 and 7000

and apparently not even 6000 will get everything. They were made with 0 ai acceleration in mind

[supple knot]

it all came down to the vRam on the gpu

[nimble heart]

I think 6000 will at least get some sort of official rocm support though

at least the 6900/6950

[supple knot]

but comfy seems to wait instead of crash

[craggy ibex]

I'm confused

[nimble heart]

there's a flash attention branch for it too

[craggy ibex]

What is CUDA?

[nimble heart]

I didn't see a 6000 series AIT branch though. Just 7000 series

[nimble heart]

like game drivers but for computational things

[craggy ibex]

So do I need it to decrease time to generate images in SDXL?

[nimble heart]

yea but with 4 gigs you'll still be swapping to ram which will hurt performance really badly

pretty sure the smallest buffer SDXL fits in while generating is 8 gigs

just loading the model is like 5.7 gigs

[supple knot]

@craggy ibex Are you using ComfyUI or A1111?

[craggy ibex]

ComfyUI

[heady vale]

you can generate 640sq images like this with just the Refiner, but you might as well use 1.5 models at that point

[supple knot]

@craggy ibex Which model 1.0?

[craggy ibex]

Yeah SDXL 1.0

Counterfeit

https://civitai.com/models/118406/counterfeitxl

[supple knot]

@craggy ibex Idk they fixed some thing between 0.9 to 1.0 to make it much faster

[hardy cipher]

watch on each wrist

[glad grove]

prepare your eyes

[supple knot]

have the alien pop out

[craggy ibex]

It is 4 gigs

but not full 4000mb

Idk why

[nimble heart]

that's just how it is

[glad grove]

rest is system reserved memory

[craggy ibex]

Should I use --normalvram?

[hardy cipher]

no

[craggy ibex]

Oh ok

[hardy cipher]

you can try but it'll probably still override your setting

[supple knot]

almost ready for birth

[craggy ibex]

Oke

what about CUDA thing hmm

I need to figure out since it said "cuda:0"

[supple knot]

4GB VRAM – absolute minimal requirement. The preferred software is ComfyUI as it's more lightweight. The base model will work on a 4 GB graphic card, but our tests show that it'll be pushing it. 6GB VRAM – SDXL will work better than on a 4GB card, but it's still not enough for comfortable work.

copy from a google search

[glad grove]

doesnt matter what you try,you wont improve speed on 4gb laptop 1650 gpu

[craggy ibex]

Idk man

but isn't it weird?

260+ seconds for 1 single picture with 768x1024?

[supple knot]

well the laptop versions of cards are usually weaker

[craggy ibex]

Yeah, make sense.

[glad grove]

look at those cute little babies 🥰

[craggy ibex]

Imma keep continue generate stuff then without CUDA blah blah blah

Cuz Idk what to do next after I installed CUDA 11.8 lmao

[supple knot]

look at the extra hand on belly

[glad grove]

buddha's palm 🫲

new technique

[hardy cipher]

the inner birthing lair

[craggy ibex]

I would rarely generate images with SDXL 1.0 yeh

Wait...

maybe decrease Steps!

to 10!

[glad grove]

u complaining about 260secs,u should see my generation times

[craggy ibex]

Show me

[glad grove]

i cant but it takes 13mins for a 1080p image with adetailer

4k images take like 30mins

[dusty turtle]

somehow bashing concepts into the refiner works if given enough steps, completely cursed result 👌

[glad grove]

she lookin sexy

[craggy ibex]

Btw how do I move or increase blue line size?

[hardy cipher]

that's just a guide. you don't need to stay inside it

[supple knot]

https://www.youtube.com/watch?v=J-6fW66IUY4

[floral island]

american mech

[glad grove]

not american,i dont see Steven Seagal drivin the mech

[full nacelle]

/ I imagine

[west breach]

#1100170312106127410 /dream

[craggy ibex]

What do you guys prefer?

Having a lot of Checkpoints but little bit of Lora
or vise versa?

[crisp owl]

I tend to collect checkpoints like they're candy. Then about once a month go through and purge

[noble shoal]

This

[supple knot]

idk going down a bad rabbit hole

kinda looks like
Jim Belushi in the front

oh exoskeletons giving me body world

[uncut fiber]

it nicely know word Google 😄

[icy brook]

a photo of a cloud that looks like a man smoking

[floral island]

lel this prompt is going places

[supple knot]

Where do I find Resolution Duo Node?

[supple knot]

@vital ermine That funny I was trying to do that workflow from Ferniclestix Channel also at .25 speed,did about 70% then just down loaded the json instead

but I cant find "Resolution Duo Node"

[ionic dragon]

got good fingers, but the face messed up

[vital ermine]

Yeah, he doesn't give his stuff out so screw trying to make a 20k node connection map. I tried once, and not again.

I paused it so many times, then alt-tab, I gave up

not this one as I learned my lesson from previous stuff from him.

[supple knot]

the json workflow in the description of the video

[vital ermine]

now, but not before

he even said in his older video he wanted us to learn by doing. One vid he said link below and it never was there

iow, call him out and it will magically appear

[supple knot]

Yeah if its more then 10 nodes I start to sweat

[vital ermine]

LOL, yep

things go underneath, etc... you can't tell what connects where.

[icy brook]

a cloud that looks like a dragon, sun in the background, grey sky

made with Aether Cloud

putting it up on civitai in a little while

[vast galleon]

Does the sdxl inpainting model work better at odd resolutions or does it still follow the sdxl vanilla resolution recommendations?

[floral island]

just letting the model take the reigns 🙂

[soft zealot]

whilst doing s0me tweaks to one of my workflows I thought I would fall back to vanilla SDXL1 Base & Refiner.

I am now reminded that IMHO SDXL0.9 was "better" ;o)

Image 1 is Vanilla SDXL1
Image 2 is SDXL0.9
Image3 is a custom model

all settings apart form Models/Checkpoints identical

[upbeat summit]

SDXL 0.9 is for sure different in some things

[soft zealot]

I know there were a lot of under the hood changes between 0.9 & 1.0

[kindred talon]

oh okay, sry

[upbeat summit]

most notable the reduction of offset noise to make fine-tuning much more efficient

but this of course reduced contrasts etc in images as well

[kindred talon]

model: Samaritan 3d Cartoon
lora: Samaritan 3d Cartoon SDXL
prompt: 3D cinematic film.(caricature:0.2). 4k, highly detailed, Smiling woman taking selfie with boyfriend through smart phone, backpackers,nose ring, nature background
negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly. young. long neck. (cross eyed:1.5)
CFG scale: 7
Steps: 30
Sampler: Euler a
res: 1024x1024

the first image uses this exact settings: https://civitai.com/images/2117544?modelVersionId=144566&prioritizedUserIds=1501222&period=AllTime&sort=Most+Reactions&limit=20

what am i missing?

[soft zealot]

seed?

[kindred talon]

177496123

[soft zealot]

so you can reproduce the picture just fine using the same seed its just subsequent generations that ayou think go downhill?

[kindred talon]

i didn't use the seed cuz my goal is not actually make that same image, but the same quality for whatever prompts i try, but i was always getting this shitty results, than i tried literally copy everything from that image to see if my settings was wrong

as i get the same shitty results with all the same settings, i wanted to know if theres something else i'm missing

like the models and lora is ment to use with SDXL, do i have to enable it or just have the updated a1111?

or if i use a SDXL checkpoint, it means i'm using SDXL?

[soft zealot]

Ahhh A1111 , I'll take a walk out at this point 🙂

[kindred talon]

loool

should i use comfy?

[soft zealot]

thats a personal choice.

[kindred talon]

hmm okay

ty anyway

[soft zealot]

all I can say thi sis the same prompt in comfyui using Dynavision (as I dont have samartitan) and no LORA used

Actually thats a thought, do you get the same issue not using the LORA?

[kindred talon]

i'll try without the lora, don't thought about it

pretty excelent result btw

[soft zealot]

and the exact same prompt in Juggernaut 🙂

[native knot]

Nice variant using comfy with PPF noise added in.

[kindred talon]

that's what i got with dynavision and no lora... i rendered at 512x512 to do it faster, this can directly affect the generation quality?

im trying 1024 now just to see

but yours still a lot better

bruh

[native knot]

I can even create the tiniest variations...

Side by side:

This actually would be great for those, "Spot the differences" things.

[kindred talon]

yeah lol

what gpu are u using?

[native knot]

3080

[kindred talon]

interesting

[kindred talon]

my 1060 is dying rn

[native knot]

Only took 12 seconds to generate.

[vital ermine]

mine was going too

[kindred talon]

jeez

[native knot]

I'll do another one.

[craggy ibex]

What's the difference between Xformers 20 and 21?
I just updated my ComfyUI

[native knot]

Another slight difference one.

Well...24 seconds if you count all the other processing around the KSampler such as upscaling & PPF noise generation, but still, under 30 seconds is fine.

[kindred talon]

what res r u rendering?

[native knot]

1024x1024

[kindred talon]

native or upcaled?

[native knot]

You always should render SDXL at around 1 megapixel.

512x512 will not be reliable for SDXL.

[kindred talon]

hmm interesting

i'll give comfy a shot

[native knot]

Until you get used to the basics, don't use my image as an example since I'm adding Perlin Power fractal noise in, but later on, after you get the hang of it, you can drag one of my images into your ComfyUI, install the missing nodes using ComfyUI Manager, and give it a shot.

[craggy ibex]

Do you guys update your xformers to 0021?

Is it good?

[kindred talon]

just saw a video and it looks like engeneering lol

i'll stick with a1111 for now lol

[native knot]

ComfyUI is really not that tough.

In fact, it helps you understand what is going on.

I took the PPF noise stuff out of my workflow and also noticed I had a node that I wanted to replace, so I did that too. Doing so dropped a few seconds of the generation time.

So you can get it down pretty tight if needed.

[craggy ibex]

What is "You shouldn't move a model when it is dispatched on multiple devices."??

[kindred talon]

here we need patience lol

[craggy ibex]

Uh

please help

[kindred talon]

[soft zealot]

damn you people with your iterations per second lol

[kindred talon]

yeah that was the problem i guess

finally a acceptable result

[native knot]

I don't mind waiting...but I was trying to show the other person who has a much slower card that even my workflow had some efficiencies that could be had.

[soft zealot]

no I mean "damn you people with iterations per second" lol

[kindred talon]

so if i use controlnet, i can transform a real picture into this 3d cartoon stuff, right?

(asking so i don't lose another hour trying wrong things lol)

[native knot]

The best way would be a combination of controlnet + img2img.

[kindred talon]

hmm interesting

[native knot]

What I plan on cobbling together is a controlnet + ipadater + img2img workflow that will allow me to adjust an image using another image while keeping the poses the same. But I'm first adjusting my workflow to use TTN's pipe to simplify the spaghetti. 🙂 Some of that may not make sense to everyone, but I just felt like sharing.

[native knot]

lol

[quaint plume]

/prompt a girl

[kindred talon]

oops

[native knot]

I find it's better to just talk to them like normal.

[kindred talon]

i thank you for sharing and yeah, i didn't understand most of it lol

ty for all the help guys

[icy brook]

I just released this - Aether Cloud - LoRA for SDXL to make cloud figures with.
https://civitai.com/models/141029/aether-cloud-lora-for-sdxl

[soft zealot]

no rea\on why not but again you're using A1111 so YMMV , not sure where thats up too with controlnet & SDXL

[craggy ibex]

How to uninstall xformers for ComfyUI?

I accidentally update it to 21

[floral island]

Just left the pc on generate forever with character concept art + landscape + random artist names

[muted void]

Hi everyone,
I've googled and searched for help on this but can't find it so I thought I'd look here.

I subscribe to a system called "Magai" that gives me access to a few different AI systems, including SDLX. I've read a lot about making prompts, but this system seems to ignore a lot of them and limit what I can do. I was wondering if anyone has a suggestion for how I can maximize the usefulness of this system within it's limitations.

[soft zealot]

Is this where I say "you're paying for a service , surely it includes a support model? Why not use it?"

[fringe barn]

Do we know if Kappa_Neuro's SDXL Loras are legit? I just started downloading a bunch of his stuff and then discoeverd some comments that made it sound sketchy

[kindred talon]

should i be running sdxl in a SSD or this doesn't matter?

[soft zealot]

generally speaking it is best to store your models on the fastest drive you have

[kindred talon]

alright ty

[muted void]

Yeah, but it's just one dude and he recently got on the news and has blown up. He's losing his mind busy so I'm trying to find another way. I am reaching out to him though.

[soft zealot]

This is the AI Revolution, everyones jumping on the bandwagon and makingthemselves CEOs, thsi si the way ;o)

and appologies if anyone thinks that sounds harsh but if I'm paying for a service then I expect the requisite level of support. I don't really care about your personal issues etc, you should have built that in your business plan.

If on the other hand the servic eis free then I will take anything 🙂

[topaz pivot]

Hi guys, I am looking for a propmt to generate landing page designs, but I can not specify an instruction with which I can start, the ones I have done have given me very abstract designs

But i don't male it

[shy kelp]

Does SDXL include physics in its model? Because often the lights are wrong, or the shadows are wrong

[crisp owl]

It's trained off images

[crisp owl]

Use at your own risk and make your own determination. They aren't virus' or anything, just some might be weightless lora's.
Apparently he said it was an error and has been fixing them. So, just test with the lora on, keep everything the same, and test again with it off. If you get different results, then that one is working again.

[carmine gale]

Is there a comfyui workflow for testing Lora’s in an xy plot fashion?

[crisp owl]

https://github.com/LucianoCirino/efficiency-nodes-comfyui/blob/main/workflows/XYplot/LoRA Plot X-ModelStr Y-ClipStr (Overview).png

[robust garden]

hi,all

[shy kelp]

I was wondering. If AI could include physics and a proxy 3D model representation, it could generate physically correct images.

[floral island]

just add more parameters

[robust garden]

can be improved, i think

[floral island]

mine is just a random output for character concepts

[robust garden]

but, what 's the animal like? the first

[floral island]

the beast is certainly there because of the artist mentioned in the prompt

because my prompt has 0 prompts for beasts actually

[hardy cipher]

I don't think it's quite that simple

[floral island]

yeah, it isn't

generally more parameters means a model can 'remember' more things

so it's capacity for 'guessing correctly' goes up

[hardy cipher]

well, firstly you'd need to create the 3D models

[shy kelp]

my wet roads always have an excess of reflections

[hardy cipher]

and that'd be a whole lot of 3D models

[floral island]

it might also be a model issue

but the problem with stable diffusion -> it's approximation

[hardy cipher]

and then, how do you teach it physics in such a manner that it correlates it with visuals?

[floral island]

you can't

[hardy cipher]

yes

[floral island]

afaik

[hardy cipher]

well you could

[floral island]

well, with the low parameter count of stable diffusion models for certain you can't

[hardy cipher]

but you'd need a lot more things going on

[shy kelp]

the bot here, nightcafe and clipdrop seem to generate based on somewhat different parameters. The bot here for instance, half of the time it makes a correlation between "wet ground" with "the user wants rain too".

[hardy cipher]

people need to understand how these models work. language models aren't image generation models, and image generation models don't exactly know what they're looking at, they're just turning noise back into the image they think it is given their input parameters

they don't know physics or anatomy or anything else

[floral island]

for example if i do 'bedroom on a rainy night' you bet it's going to be wet inside

[hardy cipher]

the bot here is probably optimized to interpret user input in a more sophisticated way. that doesn't mean stable diffusion itself is doing it

[floral island]

we both know that's not how it supposed to go, but most, if not all models, simply don't know that rain doesn't go inside

[hardy cipher]

well these things that seem sooo simple to outside observers are things the researchers have considered. it's not like they just forgot about it. didn't look into it

[floral island]

the model itself has to 'learn'

[hardy cipher]

"oh hey, why didn't we think to give it a bit more training in order to vastly improve it's abilities? oh well, maybe next time"

[floral island]

and most of the time it gets it right, but it's the same as the wet t-shirt question for large language models

without temporal and physics awareness, some things are really hard to grasp for ai models

[hardy cipher]

well they're literally replicating the patterns their training compelled them to recognize or understand

they know not what they're actually doing

it's kind of crazy to think about

[floral island]

yeah, you're right about that

the cool thing is, they try to 'resolve the equation' to the best of their ability, and we're just making up crazier stuff

i mean, i've been trying to add layers of insanity to my prompts, but sdxl pretty smart

[hardy cipher]

well people need to understand they're tools, very impressive tools, and useful, tools. they're not perfect, and just like with any tool, we need to understand both their strengths and their limitations in order to optimize our usage of them

[hardy cipher]

it's crazy how randomly deficient things like gpt-4 will be with certain things

try giving it a long list and having it organize that list

it'll organize it, sure

randomly leaving things out. not telling you. it doesn't even know

[floral island]

that's llm's

they're pretty impressive though

but you're correct, they have limitations and strenghts

[carmine gale]

Ok will try it but last time I tried I didn’t see a Lora I put for xy plot

[floral island]

but the speed at which it goes atm, is just ridiculous

[hardy cipher]

well same thing is true for stable diffusion, or midjourney, or anything

[floral island]

i mean, 3 years ago, generative ai was at best, regarded a joke

[hardy cipher]

I do believe the future of most ai stuff is "an ensemble of experts" as stability puts it

[crisp owl]

I don't use it myself, I just know that's the main xy plotter pack, and linked to the lora example

[hardy cipher]

maybe some huge monolithic beast gpt style models. but mostly specialized expert models

just seems more efficient in almost all ways

[carmine gale]

Guess we’re stuck with auto1111

[shy kelp]

I guess AI works best when you give it a hand. Like you input some starter shape first.

[hardy cipher]

what do you do with the ipadaptor model output?

because if I hook that up to the sampler it gives me the nan nan nan black images

[shy kelp]

Is prompting different in sdxl

[floral island]

depends on model

best would be finding a model which responds to the prompting you're used to

[hardy cipher]

there are two clip models

g and l

l is what was in 1.5 I do believe, g is newer

[crisp owl]

IPAdapter model output does go directly the ksampler

[hardy cipher]

well wtf is wrong with mine then?

I'll have to look into this

because I seem to remember it working before

[floral island]

L is additional support prompt

as i've heard

[crisp owl]

make sure the correct models are loaded in each node

[hardy cipher]

it can be

[hardy cipher]

so by default it's set up to just run the same prompt in both of them

[floral island]

so it's just extra tokens to be inserted, but i'm not sure, as i've just heard, not studied up

probably

[hardy cipher]

but then if you want to separate them you can

[floral island]

if i'm not mistaken, the stuff from L is added to G

[hardy cipher]

umm

not sure about that

[crisp owl]

G and L are what opens up the unet to the massive parameter jump of 2.6B.

[shy kelp]

"road with curves". Adobe generates neatly roads turning left or right. SDXL interprets it like, literally, a road with curves, making distorted curves very often.

[hardy cipher]

bro, just use that

[floral island]

as i said, that's what i've heard, so if i'm wrong i'll immediatly have to shut up

i do know the unet is way more intelligent with the 2.6b parameters tho 🙂

[hardy cipher]

I've been reading about things some on huggingface

the official documentation stuff

but not like I really understand it all

but helps demystify things a bit

[floral island]

oh no... i'm starting to figure out how to use the neutral prompt ANDS

[crisp owl]

To make visual sense of it, I think of it like interlocking my fingers. One hand is G, one hand is L. They both contain information and they interlock with each other, each adding their own weights together into your final image.

[strong field]

this is already happening

[floral island]

yeah, that's what i did undeerstand from my own reading, but someone corrected me on it, so uhhhh...

[crisp owl]

except on an absolutely massive scale

[floral island]

yeah, sdxl = mystery chicken. it's awesome, but do you really know what it's made of?

[strong field]

real estate law, corporate finance, analytics, medical expertise, all getting isolated to fine tuned custom trained models

[crisp owl]

https://tenor.com/view/magic-gif-26166638

[floral island]

this.

[shy kelp]

hm... I'm going to try something. Use Adobe as starting image to see what SDXL is going to generate from it.

[floral island]

ohno, i have controlnet working again...

[strong field]

https://docs.google.com/document/d/144xqK_QSIbP-yJxQsztgRfm5S80rpK-UUrz0mqLy5gg/edit

[floral island]

g for what you want to see, l for what you want the style to be

[strong field]

the research side of these tools is amazing

[floral island]

dudes that made this, are all nothign short of amazing

[strong field]

lots of hard work involved

i still remember the CLIP paper

brief overview

https://openai.com/research/clip

[hardy cipher]

appreciate that link. pretty good for visualizing things

[strong field]

and yeah, CLIP has intelligence to understand what its looking at

image classification has come a LONG way

[shy kelp]

https://i.postimg.cc/y8XXJpvr/tentativa-8.jpg I said "make this photo real". https://cdn.discordapp.com/attachments/1148388684836638820/1149015748270886972/make_this_scene_photoreal_style-Photographic_width-1354_height-773_aspect-16-9_seed-0ts-1694017068_idx-0.png Wasn't expecting this. It made something like a bealtiful photo real concept art

[strong field]

now, should it know a room shouldnt be rainy inside? well i think that misunderstanding is what makes SDXL suprise us

[shy kelp]

funny it kept the logo, trying to write something there

[hardy cipher]

that stupid logo is always on things

[floral island]

i mean, i like the happy accidents more actually 🙂

[strong field]

100%

[hardy cipher]

I don't want it to always know what should be

because then it wouldn't make things that shouldn't be

I feel like I have to run updates 3 times a day in order to keep up with things

[strong field]

Zero shot prediction is what makes these models so damn good now

[floral island]

notable outputs from my character concepts +landscape + artists

(fact, i did not prompt for the hulk)

[hardy cipher]

those are pretty good if you're into quality images

[strong field]

excellent shit

[hardy cipher]

that hulk just likes to show up

[floral island]

i referenced the artist that did hulk comics

the prompt is literally art by <artist>, character concept and landscape

so it's just testing how well the model knows about artists and their styles/concepts

[strong field]

SDXL so good with skinny prompts

[hardy cipher]

I like referencing Jim Lee for comic book stuff

that and retro comics

mash up

[floral island]

yeah, i also love throwing in comic artists

[keen nimbus]

that is where trump plots his return to power

[crisp owl]

power control

[hardy cipher]

it'll make thingsl ike this

almost the hulk?

[floral island]

right, i was about at 20% of testing old prompts

[strong field]

make trump great again, like when he was selling steaks pls

[floral island]

lets get back to business 😄

[hardy cipher]

it gets confused

[strong field]

putting in the work

[keen nimbus]

@crisp owl let me guess, you = a democrat

brb coffee 😉

[floral island]

just 120k files, so uhh..yeah about 90k images/prompts to check xD

[crisp owl]

I don't follow politics, it's all a sham. I live my life and want to be left alone.

That's as much into politics as I get

[shy kelp]

https://cdn.discordapp.com/attachments/1100170312106127410/1149018309791060018/Place_some_tall_trees_on_the_roads_sides__Morning_fog_atmospheric_haze__Some_butterflies__style-Photographic_width-1354_height-773_aspect-16-9_seed-0ts-1694017712_idx-0.png Hmm... I was expecting photo real. Yet, it made some artistic style that looks like a dream.

[hardy cipher]

don't love trump? you must be one of those liberals

[shy kelp]

it's pretty good at reading the road's shape

[crisp owl]

that's the logic people follow yeah

[floral island]

or they don't live in the USA lol

[faint shore]

I just got SDXL working inside ComfyUI. Fun.

[hardy cipher]

it's just tribalism and lack of awareness

[keen nimbus]

who is motivated to study in those pics

lulz

[strong field]

tribal who?

[hardy cipher]

indeed

well I'm more concerned with how to get the model output on my ipadaptor to work

I swear it worked before

[strong field]

you using ait?

[hardy cipher]

no

uninstalled that

[strong field]

lol

[hardy cipher]

should I?

[strong field]

whats your comfy commit, updates been breaking ipa for me

[shy kelp]

wait? The text to image works better if we use imperative language?

[crisp owl]

oops

[hardy cipher]

maybe that's what it is

[hardy cipher]

there's all sorts of documentation on this stuff

[hardy cipher]

early in my ai endeavors

[native knot]

Beeeeeaaaaannnnnssss

[strong field]

why we even talking about his here, lol, keep as far away for me as possible

[keen nimbus]

guess what the prompt is for those

"young immature teenagers who are democrats, that look like an anti trump supporter, due to mind control from democrat psyops"

[hardy cipher]

oh, ok

[keen nimbus]

lulz

[hardy cipher]

guess what the prompt is for those

[strong field]

your bulging bias is showing, zip up those pants!

[crisp owl]

lol very right. Someone decided a simple statement I made meant I was sitting on one side of the political wheel

[strong field]

what a waste of time

hahaha

-prompts SDXL for "bulging bias"

[hardy cipher]

solesbeedude is going to be on the wrong end of sick burns all day now

[strong field]

now these are good

hahahahhahahahahha

[keen nimbus]

guess the prompt

lul

[peak dove]

Read followfox ai blog on Substack ... excellent insight into SDXL and Comfyhttps://open.substack.com/pub/followfoxai?utm_source=share&utm_medium=android&r=1cz4g9

[strong field]

i can take a few guesses at what CLIP interrogator would say

[strong field]

got a link for the lazy?

[hardy cipher]

what is this?

[hardy cipher]

obsess much?

[hardy cipher]

yeah, we get it, you like trans women

[strong field]

lets get back to seagal yeah

[strong field]

@glad grove my eyes need bleach sensei

[peak dove]

https://open.substack.com/pub/followfoxai?utm_source=share&utm_medium=android&r=1cz4g9

[cyan crown]

hello, Do anyone know if there is a way to change model output name of a Lora?

[keen nimbus]

now that shit is good lulz

who wants to walk in that house on the bottom left

[hardy cipher]

I'd wait until it was light out so I could see it first

[strong field]

https://tenor.com/view/hello-inigo-montoya-princess-bride-gif-20962663

[keen nimbus]

someone use this: ""mix abandoned new orleans, demons, and stranger things"

see what comes out

[strong field]

here you go

[strong field]

color photo of a dilapidated building in abandoned New Orleans, with eerie red graffiti covering the walls, shattered windows revealing a glimpse of darkness inside, and overgrown vines creeping up the sides. The air is heavy with a sense of foreboding and mystery, as if lurking demons and supernatural forces are just out of sight. The flickering streetlamp casts an eerie glow on the scene, illuminating the decaying surroundings and adding to the unsettling atmosphere. The camera captures the scene with a vintage Polaroid camera, using expired black and white film to enhance the haunting vibe. The lens captures the details with a slightly distorted effect, adding an otherworldly touch to the photograph. It's as if the spirits of New Orleans and the Upside Down from Stranger Things have collided in this desolate place, creating a unique and chilling visual story

[keen nimbus]

^

what helps is "prototypical red ambient glow, synonymous from stranger things"

its focusing to much on the building

[peak dove]

Tell LoRA I Love Her!!!

[noble shoal]

Oh no, so much ketchup

[crisp owl]

ladybug

[lapis gale]

alarm clocks are frighteningly realistic with sdxl, too

[floral island]

crosspost from #🍥｜anime

but too good to not post here either

[stone fossil]

I have this thing, pic is nice tho. 🙂

[stray warren]

Beware of her shoulder patches

[crisp owl]

I love the prompt haha

[lapis gale]

while we're here with horror themes, the daily is also nice with sdxl
"A zombie emerging from a swamp full of floating teddy bears."

[stone fossil]

And I also hate it with passion, alarm clocks hehe.

[lapis gale]

... and simply adding "(((gundam)))" and with anime style 🙂

[lapis gale]

who did you call a stupid bot? draw, stranger!

[shy kelp]

this flow seems to be better than reimagine: I generated a 1:1 image -> uncrop it in clipdrop -> download -> use it to generate a very similar image in widscreen.

[vital ermine]

Make one of Biden in a jail cell with a hooded axeman about to serve his punishment according to laws of the king.

[dapper dragon]

haha

I have no Biden model but let's try it

[vital ermine]

This actually does a pretty good biden for sdxl

[dapper dragon]

haha

yes but Trump is 100x funny

[vital ermine]

Biden is a pathetic fool, so yeah there is that but I do have dreams of seeing it happen. At least I can with SDXL.

[vital ermine]

LOL

[dapper dragon]

doesnt look like biden

needs a lora to work

[vital ermine]

Looks like an old Eminem.

[dapper dragon]

haha

yes

Future US president Eminem

[vital ermine]

LOL

eff me, cry. lol

[vital ermine]

If that ever happened I would check to see the year and see if I were in the movie Idiocracy.

[vital ermine]

We need a wrestler as president to cap it all off

[dapper dragon]

Evil Trump

[vital ermine]

SDXL doesn't do orange jumpsuits very well in base

[dapper dragon]

haha

[vital ermine]

I tried and no go on Trump

[dapper dragon]

this is dreamshaper bro

[vital ermine]

Looks more like lounge ware

[dapper dragon]

And my trump pictures are done with a private lora

[vital ermine]

dreamshaper is SDXL?

[dapper dragon]

there's one based on SD XL

[vital ermine]

kick ass

[dapper dragon]

and also with previous SDs

[dapper dragon]

😆

[vital ermine]

WHen I train I train stuff on base so it works on all of them much better

[dapper dragon]

👍

[dapper dragon]

Make Monica Lewinsky

🤣

Time to leave

[hardy cipher]

monica lewinksky coming right up

or maybe, loaded that workflow in and I'd moved things around since then

not bad

[south frigate]

guys 1.8it/s on gf 4080 with sdxl is normal? (a1111)

idk, its so slow

[hardy cipher]

sounds pretty slow to me

I don't know, this is a pretty flattering rendition of Janet Reno

I'm feeling inspired

[hardy cipher]

such a nice hand

[lime citrus]

Anyone know why all my SDXL images generate like this? Just a box of colors and textures

[spring fulcrum]

Probably your sampler and scheduler choice

try dpmpp_2m_sde with karras

with at least 30 steps

[hardy cipher]

hmm. I don't know if the sampler would make it that goofed.

[lime citrus]

but i copied the generation settings from the image in the back.

[lime citrus]

changed it and still nothing

[spring fulcrum]

could be an issue with vram... you may need to add a command line arguement somewhere... I see that you are using swarmui... I don't use that so im not sure what you should use.

there is a swarmui channel in this server... maybe ask for help in that channel

[crisp owl]

dog ❤️

[lusty wolf]

Followed this upscale method. Works gr8! https://www.youtube.com/watch?v=CxB47DMEyYQ

[shy kelp]

It's interesting that I ask for cherry blossoms, SDXL creates images containing japanese style houses. But when I input that generated image and ask for houses without a specific country, it chooses houses with a style mismatching the original image.

[hardy cipher]

well, is it supposed to interpret the input image and understand that it's japanese style? what are your expectations here?

it's not an unknown mystery how this stuff works. if you have unrealistic expectations about what the model is going to do you'll be disappointed when it does what it's trained to do. you should read about how it's actually trained and what it's actually doing

consider the fact that stable diffusion's first public model was released barely a year ago. text to image ai generation wasn't even a thing on a large scale until maybe 18 months ago at most. you're really focusing on some pedantic minutia here.

bing things it's stable diffusion

[craggy ibex]

Are there any prompts that allow my model to randomly generate hair color?

(((random hair color)))???

[crisp owl]

wildcards, then a local text file it calls which it pulls the different values from.
I haven't done it in Comfy, but I did a lot of that in Automatic1111.

_haircolor_ hair woman standing with wolf.

[zinc cargo]

good idea

[upbeat summit]

if you use ComfyUI {wildcard}-syntax is build into it natively. wildcard text files are not. so you could write in your prompt field:
{brown|blonde|red|green|yellow} hair
and it will randomly select a color for each image you generate.

[crisp owl]

can it natively call a .txt file as the Dynamic Prompts extension allowed? (I'm sure there are packs that do that, just curious if you know native)

[upbeat summit]

no, txt file are not supported natively. but there are custom nodes that add support for it.

[crisp owl]

got it, perhaps will add that next after finishing this flow.
I did use it pretty often

[upbeat summit]

I use Mikey Nodes by @west breach https://github.com/bash-j/mikey_nodes.

a couple of Mikey's Nodes support wildcards. for example the Wildcard Processor node. just create this directory: /ComfyUI/wildcards/, put in your wildcard txt files in there and use them like in a1111 in Mikey's Wildcard Processor node with __wildcard__. you can even sort them in subdirectories etc. works great.

[crisp owl]

ah nice, already got that pack installed

[zinc cargo]

@upbeat summit can i put wildcards in the txt files?

[upbeat summit]

yeah it also works in the Prompt with Styles nodes. very easy to use and a must-have feature for me too

[upbeat summit]

yes, of course - if you use a wildcard processor like from Mikey Nodes.

[zinc cargo]

i mean, can i make a wild card that will call another set of wildcards

wildcardcepsion

[upbeat summit]

good question - I don't think that currently works. It worked with Dynamic Prompt for a1111. So that's maybe a feature that can be added.

[zinc cargo]

squeezing what i can out of @upbeat summit's Magnum prompt 😛
protovision

@upbeat summit speaking of wildcards, what about Loras? how can it be done with comfy?

[upbeat summit]

you either use the native Load LoRA node - one node = one LoRA.
or
you use a custom node (like the ones from Mikey :D) and use them inline in your prompt like in a1111.

Tom Selleck <lora:tom_selleck:1>

[zinc cargo]

i was looking to see if it's possible to load random lora from a list

then you can do different subjects

[upbeat summit]

I did this with embeddings in a1111 a lot. Not sure it works with LoRAs in ComfyUI custom nodes yet. It should technically be possible I suppose. I will check if it works and if it doesn't see it it can be implemented.

[zinc cargo]

me too, 2.1 times 🙂

[upbeat summit]

Just tested it. Works perfectly with Mikey Nodes. I created a loras.txt file in the Comfyui/wildcards directory containing the full lora syntax:

<lora:foo:1>
<lora:path/to/my/lora:1>
<lora:another_one:1>

I than called the wildcard in my prompt using __loras__ and it will randomly pull a lora from the loras.txt file

[cyan crown]

that is a Lora trained by me

in ComfyUI

[west breach]

you trained a lora in comfy?

[cyan crown]

no

in koyha

used in Comfy

on runpod since my 4070 takes hours and hours

[west breach]

ah ok, cool, nice work! looks great

[cyan crown]

thanks

And this is with 2 loras : 1 for the style 2 for my face

[west breach]

interesting, the digital painting look has been replaced with a photographic look

[cyan crown]

no sorry. A different lora style

sci-fi ridley scott like

not the paint from before

[glad grove]

wheres the michael bay lora

[cyan crown]

this is the digital paint as before Lora + Lora of me

[fiery moss]

I'm trying to create a loRA using RunPod, but got this error. anyone know what to do?

CalledProcessError: Command '['/workspace/kohya_ss/venv/bin/python3', './sdxl_train_network.py', '--enable_bucket', '--min_bucket_reso=256', '--max_bucket_reso=2048', '--pretrained_model_name_or_path=/workspace/stable-diffusion-webui/models/Stable -diffusion/sd_xl_base_1.0_0.9vae.safetensors', '--train_data_dir=/workspace/JP-rfzstyle1/img', '--reg_data_dir=/workspace/JP-rfzstyle1/reg', '--resolution=1024,1024', '--output_dir=/workspace/JP-rfzstyle1/model', '--logging_dir=/workspace/JP-rfzstyle1/log', '--network_alpha=1', '--save_model_as=safetensors', '--network_module=networks.lora', '--text_encoder_lr=0.0009', '--unet_lr=0.0009', '--network_dim=256', '--output_name=JP-rfzstyle1', '--lr_scheduler_num_cycles=10', '--no_half_vae', '--learning_rate=0.0009', '--lr_scheduler=constant', '--train_batch_size=5', '--max_train_steps=2400', '--save_every_n_epochs=1', '--mixed_precision=bf16', '--save_precision=bf16', '--caption_extension=.txt', '--cache_latents', '--cache_latents_to_disk', '--optimizer_type=Adafactor', '--optimizer_args', 'scale_parameter=False', 'relative_step=False', 'warmup_init=False', '--max_data_loader_n_workers=0', '--bucket_reso_steps=64', '--gradient_checkpointing', '--xformers', '--bucket_no_upscale', '--noise_offset=0.0']' returned non-zero exit status 1.

[cyan crown]

you have problem with memory

wich card are u using?

[fiery moss]

1 x RTX 3090
32 vCPU 125 GB RAM
50 GB Disk
75 GB Pod Volume
shleykza/stable-diffusion-webui:3.0.1

first time doing this

[cyan crown]

I use 4090

24 gb

but before launching train

I give 2 times

fuser -k 3001/tcp

because Auto1111 takes half of the GPU mem

so this command shuts down auto1111 and frees memory

I'll do a video in the next days

[fiery moss]

ok yeah not really following lol

all this is confusing haha

is that a quick fix i can do with my current setup i have going?

[cyan crown]

yes just fuser -k 3001/tcp 2 times

[fiery moss]

where exactly?

[cyan crown]

but with 3090 it takes a lot of time

in the command line

[fiery moss]

and then what after i do that?

[cyan crown]

do it 2 times and it frees memory

relaunch train

[fiery moss]

ok thanks ill try that when im back online

thank you

[cyan crown]

hope you resolve!

Last image. Good night!

[soft pumice]

If I want to use the controlnet for sdxl, do I only have to update the controlnet extention and then download the extra sdxl controlnet models? Can the 1.5 models work with it like scribble?

[swift acorn]

can u tell me ur it/s and what gpu u have? (in A1111) I have like 1.8 on GF 4080 [with sdxl]

[sterile yoke]

i'm using the fooocus google colab (moonrise version); where do i put the models and stuff i want to use? like what's the folder called and where is it

[nocturne dove]

Are people using the RealVisXL model yet? So far it's pretty good imo. My character loras from the base model have some weird skin texture, but if I retrain on realvis then that goes away. And hair seems a bit plastic looking unless I add "plastic hair" to the negatives. Model is here btw https://civitai.com/models/139562/realvisxl-v10

[upbeat summit]

I've not yet used RealVisXL a lot but I noticed similar things

[nocturne dove]

I'm sure they've trained on a lot of skin detail to compensate for sdxl being weak in that area. I just noticed that on base sdxl hair can be a blurry mess and on realvis it makes sort of hair strips like a painting instead.

[upbeat summit]

yeah there's definitely something going on

I haven't done a lot of prompting with realvis yet, so I can't say if it can be reduced / controlled a bit

[nocturne dove]

Yeah it's definitely a real thing. I never saw hair like this from sdxl base

[upbeat summit]

hmm yeah. that doesn't look right. depends if you went for last gen game character hair 😉

[nocturne dove]

But you can get rid of it with prompting at least.