✨｜sdxl

[floral island]

probably the people who are used to using 1girl... (i hate booru prompts)

[royal python]

extremely wrinkled man, with HUGE pores, many face scars and a LOT of stubble, sweaty skin, high cheekbones and deep laugh lines,

[cunning geyser]

where do i put upscalers in comfyui

[floral island]

reject waifu, embrace skelly!

[deft coral]

this isn't a bad shot

[floral island]

load upscaler model, insert image into upscale, rescale to desired resolution

[royal python]

then in negative put smooth skin, makeup

[deft coral]

thanks for the help

[deft coral]

this was my result

[royal python]

nice

[azure oxide]

whoahh we like makeup tho 🙂

[wet nacelle]

@indigo vinemedic

[azure oxide]

i keep mascara in positive prompts whenever i prompt for a lady, its such a nice detail for eyes

[edgy sierra]

what do you use as negative prompt?

[deft coral]

so 1,0 cant do full body

?

[echo pine]

it can, just do full-body in positive and closeup in negative

[azure oxide]

it should do it better than 1.5 due to how they trained it

rule of thirds photography as a positive prompt worked nicely for me in 1.5, i wonder how sdxl reacts to it

[trim orbit]

i think too, doing just a little token towards the eyes, draws attention to the eye and the latent space provides more details in those focused attentions

[wet nacelle]

wicked

[deft coral]

has anyone been able to train on 1.0 yet

[whole cloud]

/settings

[trim orbit]

"Can't" is a bad default assumption. You should rather lean towards "I'm prompting wrong" when you're unable to achieve a result

[flat ibex]

Does anyone know when dreamboothing will be supported for SDXL? I tried both the A1111 extension and the Joe Penna dreambooth repo to no success. I know it's brand new and devs are still working on things, just curious if anyone has had any luck 🙂

[deft coral]

@trim orbit

as anyone been able to train on 1.0 yet

[candid pond]

Lykon seems to fine-tuned Dreamshaper XL

[trim orbit]

all the loras i've made for 0.9 work on 1.0 and i trained one last night. worked fine. people doing full models too.

[azure oxide]

😉

relevant post

[candid pond]

How much Vram will I need for a full model?

[peak dove]

@floral island

[deft coral]

i can't train for some reason

[deft coral]

i keep get ting missing model/unet

[candid pond]

Will renting a 3090 with 24GB's on Runpod be enough?

[deft coral]

Missing model directory, removing model: G:\stable-diffusion-webui\models\dreambooth\SDXL Model test 1\working\unet

[trim orbit]

nice. i do love relevancy

[floral island]

you used something from my prompts to make that? 😮

[trim orbit]

use kohya-ss to train instead. i don't like how poorly the extension updates

[deft coral]

that was on kohya-ss

wait

[trim orbit]

back when i used the extension, everytime i updated it'd need to be a full reinstall with me making sure the requirements.txt files didn't conflict manually

[deft coral]

my mistake

[floral island]

this, is actually amaznig.. it's hard to get pareidolia effects

[peak dove]

Err no, sorry, I added you by mistake 🙂

[deft coral]

ill download it

[trim orbit]

https://github.com/bmaltais/kohya_ss this one is up to date and is wrapped wiht a nice gui

[peak dove]

It was a random effort, rather than something like Arcimboldo

[floral island]

look into my beady eyes 😮

[deft coral]

im gonna sell my soul

[tropic turret]

ComfyUI crashed my PC with too high of image generation settings and now I get this issue trying to load it up. Any solutions?

[trim orbit]

not in this economy. supply and demand. everyone wants to sell out now

[deft coral]

[amber pecan]

ok, I have some ideas for a future version of stable diffusion

1. stable diffusion needs to become a 3d editing program, possibly based on a blender, in which there are several primitive 3d models of people, animals and other objects or buildings, as well as upload your low poly 3d models. which will create a framework for generation. This is the only way it will be possible, for example, to make a comic or a storyboard with actors who are guaranteed to have their hands and feet in order, as well as, for example, to make a crowd of pedestrians in a city with the correct geometry of the body.

2. Stable diffusion should identify the characters in the composition and give them an ID number in order to be able to use this character or many characters in the future when changing locations.

3. The 3D program should be able to edit the parameters of the character's body.

4. As well as the ability to generate faces with a variety of parameters, mixing, and loading parameters from photos.

5. It should be possible to upload a photo of a face and generate its 3d model as a meta-human.

I do not know where to stir ideas, so I put it here.

[peak dove]

Reinstall Python?

[tropic turret]

[sharp robin]

Idk comfy is easy to install, reinstall?

[trim orbit]

i don't know how comfy ui would've messed up your python folder. don't think that's possible

[peak dove]

Reinstall everything?

[trim orbit]

but that's whats wrong. your python folder got haggard

[tropic turret]

😦

[deft coral]

@trim orbit ill just take your soul then

[trim orbit]

you probably did something else and that broke python install

[peak dove]

Does deleting VENV work in ComfyUI?

[tropic turret]

ComfyUI worked fine until my PC crashed

[trim orbit]

i'm happy to provide. i've got plenty of soul always

[azure oxide]

stable diffusion is only the image generating parts, all of those things that you mentioned are to be developed by 3rd party developers

[floral island]

not sure why i got the flowers, but eh... i'll take it ❤️

[peak dove]

The PC crashed and took ComfyUI with it? OR ComfyUI crashed and took the computr down?

[trim orbit]

that's a very differetn program. what stable diffusion is is more like an open sourced rendering component. blender uses a raytracing component as it's renderer. blender is the visual suite that uses that renderer as part of it's architecture

[azure oxide]

id post this into #🤝｜tech-support , hopefully areytis and lone are active

[orchid lava]

how long the generation takes for you guys with euler a, 80 setps and batch of 4? Its around 20 minutes for me and I feel somethings wrong. (edit 8gb ram rtx)

[trim orbit]

https://www.cycles-renderer.org/ this is the renderer i think that blender uses

[tropic turret]

PC crashed and took ComfyUI with it.

I think?

[trim orbit]

python being borked speaks more to me that something else happened

[dense chasm]

does anyone has a clue how generating mythological symbols with appropriate prompts like this

[trim orbit]

the symbols are likely being added after

[azure oxide]

that looks like a gryphon so you can just prompt for that

[trim orbit]

and that's a 1.5 render. you knew that already. you're posting that for an audience to see. not a real inquiry. but yeah, reminder, this is the sdxl chan

[amber pecan]

I know how the blender works, but there is an addition for the blender stable diffusion, why not develop this promising direction, the SDXL version has not solved the problem of crooked hands and only a 3d program can solve it

[trim orbit]

good i'm glad you know, then you'll understand what i mean by a smaller component of a larger project

[dense chasm]

how about the feature extraction?i mean the major design style, can we make it thr sdxl?

[orchid lava]

How long the generation takes for you guys with euler a, 80 setps and batch of 4? Its around 20 minutes for me and I feel somethings wrong, in sd1.5 its 30 seconds at max (8gb ram rtx 3070)

[azure oxide]

developed by a 3rd party. SAI is just here to provide the image generation part

this is why the devs were hyped about the community because its on us to go fucking crazy with it

[trim orbit]

with stable swarm and comfyui and stable studio all being developed alongside stable diffusion, there will be convergence into a fully featured studio app like blender at some point. your suggestions are good and clear. it's a good direction and where theyr'e headed already i believe @amber pecan

[floral island]

awesome ^^

[viscid girder]

Hi! What model do I download to get this look

Or something similar to it

[trim orbit]

controlnet will help a lot for getting hands out of sdxl

[azure oxide]

i think thats just a simple style that you can change to in the settings

[deft coral]

okay

[viscid girder]

I’m super new to SDXL is it the github to get it?

[deft coral]

i need to redownload 1.0 since the model weights got released

[trim orbit]

pretty sure thats just dark mode

[deft coral]

[azure oxide]

no, thats a picture of automatic1111, its a setting there

settings -> User Interface

youll see this

[viscid girder]

Great!

[viscid girder]

I just download the GitHub?

To get the automatic1111?

[azure oxide]

do you have automatic1111 already?

yes

[viscid girder]

Ok

[viscid girder]

I haven’t downloaded anything

[amber pecan]

Can possibly at least implement this function 2) ?, stable diffusion should be more stable and not random, the work should be controlled by the creator of the pictures, otherwise it's just an endless empty game of generating random pictures

2. Stable diffusion should identify the characters in the composition and give them an ID number in order to be able to use this character or many characters in the future when changing locations.

[trim orbit]

i love the git clone url command that's so popularized with this realm.

[azure oxide]

i think there might be more steps, the github has very simple installation instructions so you can just follow those

[trim orbit]

i know git commands aren't the most user friendly, but i lovem

[viscid girder]

Great thanks!

[azure oxide]

2. ?, stable diffusion should be more stable and not random, the work should be controlled by the creator of the pictures, otherwise it's just an endless empty game of generating random pictures
   if you can figure out how to do this then im sure many companies would love to hire you lol

[worthy bear]

so control net is currently not compatible in SDXL?

[azure oxide]

correct there are no controlnet models for sdxl currently

[trim orbit]

controlnet gives artistic compositional control. a studio set up for building thhat controlnet input would be really nice. its likely already underway by a few teams.

it wont be too long before there is controlnet for sdxl. it is inevitable /echo

people are already using blender as an open pose creator

[azure oxide]

ideally we dont even use a gui to generate images, we could just generate images from scans of our brain activity like from that one research paper a few months back

[wary bay]

the second setting is set to "fixed" which will make your seed static, its overlapped so its hard to read

just set that to randomize

[lethal hedge]

imagining SAO worlds with a simple thought

[lethal hedge]

if you trust yourself to the elon neuralink tho XD

[amber pecan]

Unfortunately, I am not a programmer, but the solution to this problem lies in the fact that the AI has to explain what is in the picture and Meta has succeeded in this. And then it will be a new level of generation, through dialogue with AI.

[trim orbit]

https://toyxyz.gumroad.com/l/ciojz?layout=profile here's a decent looking one

[lethal hedge]

for some reason my answer to that is just better prompting lol

[azure oxide]

have they? i havent kept up with them

[lethal hedge]

no amount of systems on top can solve lazyness

[azure oxide]

ya currently we're limited to just trying our* best to get the AI to generate what we want. ideally tho we dont need to

[trim orbit]

soon, public ads will use focused microwave beams to scan brains of people in the area and target diffusion generated ads to display specifically towards them

[lethal hedge]

so they'll try to sell me anime figures whenever i'm buying stuff from the supermarket huh

[lethal hedge]

i hope i don't have to learn to use comfyui to use sdxl, my pc can't handle it on auto11, it takes too long T__t 3050 8gb user <-

[half ivy]

a boxer boxer

[azure oxide]

no need to learn how to use nodes when you can just be lazy and steal workflows like i did 🙂

[steady chasm]

It's already light-years beyond developing genuine artistic ability to paint or draw or take a photograph yourself- further development is always underway but as it is the simple leap from "no generative ai" to "nearly indistinguishable from photographs when properly processed" is astronomical, and likely not to be reproduced in scale for a good while if ever in terms of the technology

[lethal hedge]

i can never steal workflows because what i make is niche

[trim orbit]

i use auto for an hour long session, and images start taking 10min on the vae step. there's a huge memory leak somewhere. closing and opening it again helps. sometimes its only 10min of work before the vae step starts getting longer and longer to cook

[steady chasm]

If you want ai to be better sooner though, feel free to collaborate on the open source nature of many of them

[steady chasm]

Auto is famous for memory leak sadly. In my experience comfy is too, with the new sdxl base model at least

[lethal hedge]

at first it was 10mins per image on my end, had to used medvram for auto11 to generate properly

[trim orbit]

maybe it's just pytorch

[lethal hedge]

gonna be checking 1.5.1 now tho

[uncut steeple]

Finally got wildcards to work in comfy

[steady chasm]

Could be . I'm not a programmer, nor am I particularly versed in the technology of generative ai. I just keep restarting it lol

[azure oxide]

well by workflow i mean just simple stuff like getting it to generate an image and using the refiner. with sytans workflow for example, you can load it up and just get straight to prompting, very similar to a1111. assuming you simply prompt and generate as your normal workflow of course.

[lyric helm]

Is there a checkpoint model somewhere I can drop into DiffusionBee or other SD installations?

[sharp robin]

What nodes? Also nice.

[lethal hedge]

still taking too long with 1.5.1 without medvram on my gpu, i guess i'll be using comfyui for sometime ;s

GL with generations guys ._./

[rancid junco]

i updated automatic1111, downloaded the "refiner" model and put it in.. and it sucks. i feel like i skipped a step, what am i missing?

[tawdry trellis]

Are you generating at least 1024p?

[rancid junco]

yes, im trying 1024x1024

[azure oxide]

the steps go: create image with base model -> img2img with refiner. the refiner itself shouldnt be creating images

[rancid junco]

ohhh

[azure oxide]

a1111 is a 2 step process right now, comfy's is better imo since you can just do 1 click

[uncut steeple]

Using Impact node now, had issues getting it to run before
But its acting weird for me still

[scarlet pine]

Can't wait for controlnet, gonna be game changing again

[hard fractal]

Also takes a ton of VRAM

Whereas comfy runs on 8

[sharp robin]

Thanks!

[trim orbit]

after clicking 100x to set up a workflow

[echo pine]

just press the arrows on the control after gen to the right until u see randomize

[trim orbit]

another ux thing i hate about comfy, go to pan the screen and it doesn't catch my space bar, so i click to drag but accidentally click a node and drag it way off messing the whole layout

soemtimes i just get fed up and load auto

[magic dawn]

When dragging image into comfy is it possible to change only prompts and seed and keep the workflow? @visual glade

[wary bay]

I just want to see if anyone's used the SD Ultimate Upscale script with sdxl, or if it is compatible

[trim orbit]

going to try to get stable studio running with comfy tonight again. i just have to symlink it all up with my existing folders i think

[cyan crown]

Sometimes with this method images are flattened in my case

[sharp robin]

It works

[wary bay]

is it worth using without controlnet tile?

[heavy stump]

hello everyone .... so i have a 3060 8GB vram band i downloaded dreamshaper SDXL1.0 and i tried to run it at 1024/1024 resolution on atomatic 1111 and a basic prompt no lora no nothing but every time it keeps saying Out of cuda memory ... why is that ? like other SD1.5 models works fine in that resolotion

[paper phoenix]

because cuda is out of memory. just is. the model is bigger

try medvram though

[heavy stump]

already on

[deft coral]

i forced my self to koyha ss

[paper phoenix]

a1111 holds too much stuff in memory. i have 6gb vram and comfy is at least running stuff out at 1024 for me (without refiner)

[rancid junco]

okay fine i guess i have to install comfy and get used to it

[south horizon]

I gotta give this comfy a try some time...

[rancid junco]

i installed it a while back but it didnt do anything a1111 didnt so i skipped it

[paper phoenix]

yeah its kinda like a dog walking on hind legs for a while but

[deft coral]

installing it now

[paper phoenix]

well "work" is something it can add to the list 😉

comfy deffo for middle end machines though. im sure a1111 will work tearfree eventually

[deft coral]

rip torch 1

[south horizon]

is comfy the one with the weird nodes thing?

[wet nacelle]

https://tenor.com/view/jerma-scowl-angry-iron-lung-gif-25675354

[paper phoenix]

yeah. its painless if you already know blender or use music DAWS but hard to wrap your head around a pegboard UI at the same time as grasp AI art generation

[wet nacelle]

What DAWs use a node graph?

[south horizon]

I've used the setup with Blender and in Fusion with Davinci Resolve, but haven't tried Comfy yet.

[paper phoenix]

not node graphs per say, but you use your ins and outs and ducking and tracks and stuff, handy concepts for comfy

[stray lake]

change to randomized and enter 0 as value

[paper phoenix]

ITC: we complain about comfyui like children going for their first haircut.

i dont wanna! what if it hurts?

[stray lake]

comfy rules but i still can't figure out masks + compositing

[wet nacelle]

Ah I see what you mean. Stuff like digital modular synthesizers

[wet nacelle]

tes

eys

yes

[paper phoenix]

in fact, it'd kinda like if comfy could be controlled with mixers or make macro surfaces for it.

[wet nacelle]

what the fuck!

[merry jasper]

How do I find the negative_prompts used in the bot generation?

[dense chasm]

diffusers is awesome for pretrain models

[paper phoenix]

something about that image makes me expect it to move suddenly

[wet nacelle]

https://tenor.com/view/stelalris-jerma-gif-23276652

[vocal rapids]

so much fun

[wet nacelle]

https://tenor.com/view/yoinky-sploinky-the-rock-blackchan-jerma-gif-26461974

[bold osprey]

or drag and dropping an image that was already generated with comfy

[worthy bear]

where is the best place to learn how to use comfy? for a complete beginner

[bold osprey]

https://github.com/comfyanonymous/ComfyUI_examples and olivios videos

[worthy bear]

thank you sir

[zealous horizon]

Just finished my first LoRA based on SDXL 1.0 (Trigger J3nn4) (using Jenna Ortega as a Test here)

[bold osprey]

well damn

[trim orbit]

has some weird videogame engine aesthetic to it i love it

[wet nacelle]

like early dierectx

[zealous horizon]

This is without the refiner or VAE.

[wet nacelle]

directx

[sharp robin]

Is it private or public? Good work

Looks nice

[bold osprey]

now that you said it we need a direct comparison :P

[sharp robin]

What’s the question tho?

[sharp robin]

How are we supposed to know?

[wet nacelle]

teh fuck?

the

[bold osprey]

just use clip and try to work it out

[sharp robin]

This is new clip

[zealous horizon]

Since I used random highres images from the internet I will keep it private. I just used Jenna as a test subject

[wet nacelle]

the fuck! no

[bold osprey]

should get close enough to a working prompt

[wet nacelle]

should I ping staff?! the fuck!

[half ivy]

any1 got any face detailer workflows going?

[trim orbit]

good luck genning today generators. Stay classy. I'm out

[young mantle]

hey

can anyone tell me how I can use SDXL?

[wet nacelle]

did you look it up?

[bold osprey]

drag this into comfy UI

[uneven swallow]

Why do you mention 6 billions of images?

[wet nacelle]

"What's that?"

[kind pewter]

moments before a FHRITP

[light stag]

Got dammnit give me the promt

[inner ruin]

which parameters did you use?

[bold osprey]

it is included in the metadata

[light stag]

What is metadata

[fierce hollow]

because that's the amount of images on danbooru and also the size of the nai dataset according to nai at least
//oh sorry I meant 6 million

[wet nacelle]

too harsh of me

[bold osprey]

the workflow and prompt stored in the image itself, stuff you don't actually see visually

[shy kelp]

comy doggo vs. sdnext doggo

[wet nacelle]

Do they know what comfy is though? that is the question.

[light stag]

How do i see it?

[shy kelp]

upscaled comfy doggo

[zealous horizon]

But the interesting thing here are the spec, right?
20 images 20 repeats
40 regulisation images 1 repeats
stopped at epoch 8
used epoch 7
Training time ~1.5h on a RTX 4090 (64GB RAM)

[wet nacelle]

see what I mean.

[bold osprey]

yep

[zealous horizon]

see above

[inner ruin]

what about LR, did you train text encoder?

[bold osprey]

https://github.com/comfyanonymous/ComfyUI
download this and read what is written on there

[wet nacelle]

This is why I don't understand why new people join the server and ask these quesions. It's like as if they haven't tried to search Youtube for it. @bold osprey

[light stag]

okay

[shy kelp]

can't believe i'm a comfy shill now

[bold osprey]

the legs look like they don't have bone damn

otherwise the color grading is insane

[zealous horizon]

yes, TE and UNET too

[scarlet pine]

https://tenor.com/view/leg-day-gym-girl-giraffe-limp-gif-12296316

[bold osprey]

accurate

they look like plushies

[wet nacelle]

https://tenor.com/view/spin-spinny-spinning-rotate-rotating-gif-22920564

[bold osprey]

are you guys using any icons for SDXL / Comfy? I want to tuck it neatly into my taskbar

[fast moat]

Anybody using it in A1111?

[shy kelp]

not recommended

[wet nacelle]

shart

[clever sky]

The arm with two hands is curious hahaha

[slender raft]

If I make a large batch of images in comfyui, how can I find the seeds of the individual images from the batch? They don't seem to increment and I can't find them recorded.

[kind pewter]

drag and drop the image you want the seed from in comfy

[slender raft]

That just sets up the workflow for the whole batch.

[kind pewter]

(in another tab if you don't want to lose your workflow state)

AH

[fierce hollow]

you have to use the index from batch node, or something like that, forgot the exact name

[polar epoch]

do you lads know of a youtube video that explains all the nodes in comfyUI in good detail? And some of the addons? Really wanna learn this stuff to make full use of it

[fast moat]

oh...

[fierce hollow]

the seeds are somehow between the numbers so just incrementing by 1 doesn't work like with auto

[high escarp]

In automatic 1111 discarding next to last Sigma seems to help

[bold osprey]

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

[polar epoch]

Thanks

[shy kelp]

a1111 and its related forks do function, but the results i've seen haven't been great. even if you manage to get things running smoothly enough, an overnight update could screw it all up

[slender raft]

Is it doing a random seed each time and forgetting them?

[uneven swallow]

Yeah, my jaw dropped, when I saw "6b"
I know that waifu diffusion project has 15m dataset

[polar epoch]

eyy! Olivio!

[bold osprey]

yeah you can also just set a seed in the sampler if you want to

[inner ruin]

what did you use for the text encoder learning rate?

[fierce hollow]

it's not random, just that the seed counts for the whole batch, not 1 image

[slender raft]

Perhaps I can use increment, but that's far less conveniant than random.

[prisma wyvern]

Has anyone tried this combination sd1.5 512px for base and sdxl 1024 for refiner, will it work????

[fast moat]

could you please explain in simpler way. it went over my head 🙈

[bold osprey]

thank you for the delivery

[candid pond]

What did you use as your reg images?

[fierce hollow]

yeah that would be like every anime image on the internet ever made probably

[shy kelp]

just don't use a1111

till things stabilize

[scarlet pine]

So many wheels for a light load

[slender raft]

I would love to be able to recreate a single image without having to regenerate the wole batch. I've got a generation I really like, but it's part way through a batch.

[fast moat]

oh... i was thinking about using it in deforum 😕

[noble shoal]

I do it the other way around. SDXL Base -> Juggernaut as "refiner" after upscaling.

[fierce hollow]

0 will be the first batch image, 1 - second, etc

[shy kelp]

BRO this video literally explains how all this shit works in less than 3 minutes

[slender raft]

Thanks.

[polar epoch]

So, someone yesterday made this with 1.0 but, when i installed necessary node packs from the extension manager, i still keep getting these errors. Any idea?

[obtuse berry]

HELP

[bold osprey]

helped me getting into comfy as well

[shy kelp]

use comfy manager to install missing nodes

[glacial lodge]

anyone get a fine tune / LoRA / etc. working yet?

[obtuse berry]

Anyone Plz Help me. After i generate one or two images. it shows me that out of memory error. I am Using 1660 Super 6Gb vRam.

[fast moat]

ok

[zealous horizon]

40 images generated with Photon (sd 1.5) that were using "Jenna Ortega" in the prompt (2048x2048) from various views and angles

[bold osprey]

https://github.com/ltdrdata/ComfyUI-Manager
is it this one?

[shy kelp]

that's the one

[fast moat]

on A1111? same issue with me

[polar epoch]

Read the whole message :P Already got the missing nodes installed, even clicked "install missing ones" and still no go :S

[uneven swallow]

But, how many of these images are pages from manga... Because it can be good to have not only arts

[hollow halo]

LoRA, yes

[polar epoch]

So it's weird how missing node addons are installed, but image/comfy still errors saying it's missing

[shy kelp]

hahahah yes bro I DID IT

I literally just made this workflow

[glacial lodge]

that looks great

[bold osprey]

let's gooo

[glacial lodge]

and cute!

[polar epoch]

1.0 trained? Or can one use lora from below SDXL?

[bold osprey]

what happens after the restart

[hollow halo]

Running some testing on it and getting images, but will be releasing it later 🙂

[hollow halo]

1.0 trained, took about an hour

You can use 0.9 LoRA's but they don't work as well as retraining

[bold osprey]

getting the refiner in there will do magic to you result

[zealous horizon]

"text_encoder_lr": 0.0004,
"unet_lr": 0.0004,
"learning_rate": 0.0004,

[hazy moat]

can someone help me understand how swarmui multithreading works when it is using a backend that, presumably, doesn't have that?

[polar epoch]

same thing. For every addon i installed myself/added that was missing/sdxl related, i closed the CMD to restart, and it started to install dependencies, but once running again, it still complaned

[polar epoch]

Oh damn! With kohya?

[hollow halo]

yep!

I trained both TE and Unet, even though I'm not sure how it handles the TE at this time, being that there are 2

[polar epoch]

Wow, so just regular kohya training? Cause i tried it before, but it constantly complained about stuff in the wrong folder lol. So gave up back then

[shy kelp]

Yeah thats the next step :D

[bold osprey]

I need to try out UNET

[pearl shoal]

I am 16 Gb ram and 12 GB vram, using sdxl on comfy ui and it keeps killing itself, ```got prompt
model_type EPS
adm 2816
making attention of type 'vanilla-pytorch' with 512 in_channels
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-pytorch' with 512 in_channels
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}
torch.Size([1, 1280]) 512 512 0 0 512 512
torch.Size([1, 1280]) 512 512 0 0 512 512
100%|███████████████████████████████████████████| 20/20 [00:05<00:00, 3.69it/s]
Killed

[hollow halo]

I use Derrian Distro's Easy_LoRA_Training scripts, the Gui is a little better for LoRA's in my opinion than Kohya_SS from bmaltais

[pearl shoal]

Can someone give a solution?

[prisma wyvern]

I thought it only worked with 1024px So it now works with 512p??? What about 256px??

[bold osprey]

anything below 512px with the base results in noisy and overdone images

[hollow halo]

honestly I have no idea, but hope someone helps you. It shouldn't be killing itself after a gen

[earnest anvil]

@high skiff trying your workflow for the first time. you're a legend, thank you

[shy kelp]

That's the base model. The base model is supposed to make a 128x128 latent image

[pearl shoal]

there is no generation,

[shy kelp]

if im understanding right

[bold osprey]

just don't

[slender raft]

Thanks, that worked great.

[hollow halo]

Did you check your output folder? It looks like it ran 20 steps and then died

[fierce hollow]

using different conditioning dimensions and target below 1024 I think you can get images that look normal but 1024 is still the default

[shy kelp]

?

[bold osprey]

the results I got were not usable at all

just noise

[prisma wyvern]

On my system take 1,5hr I want to speed it up 512>1024

[hollow halo]

128 Latent = 1024 pixels, just to be clear

[fleet veldt]

yes works fine

[shy kelp]

base makes 128x128 latent

(or is optimized for that, I suppose you can make images from it just fine)

[bold osprey]

not with only the base and without the separate vae

[fast moat]

Mine just stops after final step of generation

[pearl shoal]

I did no outpul

[fierce hollow]

that's kind of a strange graph since the image decoded from base latent will also be 1024, likewise the latent from refiner will be 128...

[prisma wyvern]

Which conditions do you used?

[fierce hollow]

makes it sound like there's some sort of upscaling going on when in reality there isn't

[glad fulcrum]

you are what you eat.

[hollow halo]

That is why there is no image, and possibly could be a bug killing it

[fierce hollow]

comfy has nodes for text encoder where you can specify width/height and target width/height

[fierce hollow]

if you specify target width and height lower than 1024 but width and height higher than it, the result should in theory be pretty decent as well

[fleet harness]

SDXL Aspect Ratio is nicer because it has all the trained ratios ready for selection

[rich coral]

@hard fractal how goes control net? 🙂

[fleet veldt]

did you double check your VAE is not hardcoded in the command args of the startup script

[bold osprey]

possibly, instead I set the resolution separately for the latent image

[fierce hollow]

that's for latent but from my understanding the target dims should also match the latent size

might be bullshit though

[bold osprey]

Can't say for sure what works better yet

[fast moat]

It prompted to use --no-half-vae , so I did

[bold osprey]

makes sense though could also just be downscaled

[fleet harness]

this is the way

[bold osprey]

now would you look at that

[lusty raptor]

ooh, with a multiplier now. still need a proper math'd out node though

[fleet harness]

you know more than i do! I just like the settings being accessible 🙂

[fleet veldt]

gotcha, i put the 0.9 vae in the VAE folder and then selected it under settings and saved it. did a full restart of auto1111 and then it worked for me. search my prior messages to see what i was getting before and if it's just stopping, check the command window for an error you can google

[lusty raptor]

i learned from the devs last night that target width/height should be the resolution you're generating at

[fleet harness]

Are you saying that upscaling isnt good idea?

[lusty raptor]

it's handy, i just threw it together as a quickie until something better comes along.

[fast moat]

Ok. It's not showing any error. It just sort of hangs

[lusty raptor]

idk, does it work? if it works, then great! (i haven't tried)

[fleet harness]

its hit and miss honestly

[fierce hollow]

I just use this for multiplying since it ouputs both before and after

[bold osprey]

trying something out with the CLIPTextEncodeSDXL node

[fleet veldt]

did you go to the output folder to confirm if it generated anything but just didn't render it in the web.. sounds like you have a specific issue here with your install but web1111 works

[pearl shoal]

No i mean that I checked the output folder, there is no output image

[candid pond]

Hey Annuvin. Can you share the setting you use in your kohya gui lora training?

The Json file?

[hollow halo]

Maybe the saving is crashing it, that is an odd issue. Have you checked if there is an open issue about if on ComfyUI github?

[polar epoch]

If you're using an upscaler, wouldn't you want to use the same seed to generate the higher res upscale with the exact same details, just more of it in the higer res? Watching Olivio's video

[fierce hollow]

I think I pasted them earlier? not really sure I have any jsons with them, I use my own script with variables and stuff

[candid pond]

Oh you make the training using the script? I wanted to use the gui for the training. (

Thank you though.

[fierce hollow]

oh I don't use any guis

that said if you're using bmaltais' repo or something most of the arg names should match up

[fleet harness]

for those struggling with MISSING NODES:

https://github.com/ltdrdata/ComfyUI-Manager

[quasi remnant]

don't forget to load the SDXL Vae https://www.huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors otherwise you get chromatic aberration & noise over your generations.

[wet nacelle]

No you won't

it's jpg not png

[quasi remnant]

i see it over your two gens

lol

[shy kelp]

Error occurred when executing KSampler:

mat1 and mat2 shapes cannot be multiplied (154x2048 and 1280x768)

File "C:\Users\emerg\Desktop\webui\StableSwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 144, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\emerg\Desktop\webui\StableSwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 74, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\emerg\Desktop\webui\StableSwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 67, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

[quasi remnant]

if you use a vae loader with the sdxl vae it'll be fixed

[autumn forum]

I saw you were having trouble with the workflow from my image, did you try the one that I uploaded as a file? If it’s from the image yes I use custom nodes.

[fierce hollow]

you're not supposed to set latent size to 128 but idk if that's the problem

that's the dimensions for the image in pixel space, not encoded

[autumn forum]

@polar epoch if your having trouble using the workflow I used for that image use this instead. I took out the custom nodes for you. Or does it still give errors? #✨｜sdxl message

[polar epoch]

Aye, the file worked, just oddly not when dragging over that old man pic

[wet nacelle]

The fuck are you talking about.

yes there are faint lines but it's not noticeable on surface level.

[quasi remnant]

and all I'm saying is you will get rid of it if you use the official vae they posted.

[autumn forum]

Yeah the old man pic is my private workflow haha. It has custom nodes and stuff. That I use for ease of use and to make pretty pictures like bloom and stuffs

[quasi remnant]

if you don't find it necessary then no problem.

[lyric grove]

It's super noticeable even at 2 monitor height-distance view. /shrug

[quasi remnant]

yup. 100% very noticeable

i will never not notice it looking at a pic

[lusty raptor]

looks like the refiner doesn't like working at 1024x1024 but being passed a 128x128 latent?

[fierce hollow]

I can see it from the couch 4 meters away

[lyric grove]

I just use the other VAE for it, no issues there

[wet nacelle]

I'll be autistic just for you sweetie.

[quasi remnant]

they reverted it so it's the official vae version now

[polar epoch]

Also, is there a setting/way/addon that has a "PNG info" where it notes what nodes were used, settings for nodes and such? Similar to png info in automatic, that way i can add nodes where due if it's needed without shifting my entire current workflow :P

[quasi remnant]

the sdxl_vae on the main branch on huggingface is the proper one now

[hollow halo]

Damn dude, really pulling out that card.

[fierce hollow]

the official vae still has fp16 issues afaik, the one from madebyollin doesn't

[lusty raptor]

#✨｜sdxl message

[polar epoch]

Ah ok :P Could you take a screenshot of the workflow so i can see what you used? As i wanna achieve the same level of details you guys get :P

[shy kelp]

It doesn't seem to like any resolutions I set in that configuration and spits out seemingly random matricies multiplication missmatche dimensions

[quasi remnant]

i've had no issues with using the one with this hash from the official upload

[fleet harness]

https://tenor.com/view/suzaku-kururugi-gun-point-aim-gif-13318800

Kindly asking for all workflows...

[polar epoch]

That's what i used to get VIA's workflow, but they seem to use even more custom ones the manager doesn't have :P

[candid pond]

Does anyone have a the SDXL LORA Fine-tuning parametres for KohyaSSGUI?

[fierce hollow]

it might be working for inference or you're using fp32 unkowingly

[autumn forum]

For the best images sytan will be releasing a workflow with all stock nodes that will achieve great results! Hopefully he will release by this weekend. I’d just way for that.

[fierce hollow]

for training this is the best in any case https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors

[wet nacelle]

that moment when bitches think that giving help to the people that don't want it is a good idea.

[quasi remnant]

my workflow hasn't changed much, are you asking for inference workflows to be posted or...?

[fleet harness]

just messing around :p

[quasi remnant]

lol

[polar epoch]

Aye. I can do addon nodes as well, just need to learn the basics of nodes and what goes where to achieve X :P

[autumn forum]

Probably a custom sdxl aspect ratio helper that I got from someone. You don’t have to use it.

[fleet harness]

I hunt workflows for sure but i just check to see if meta is intact, if not, i figure its TOP SECRET stuff

[polar epoch]

I'd love to use it actually :P Would be perfect for ultrawide generations :P

[fleet harness]

Ajani?

[autumn forum]

Yeah I may end up making a YouTube video lmao. For those struggling with comfy or something for 2.0 when sytan comes out with his workflow idk

[fleet harness]

:p

[autumn forum]

I’d send you the file but I’m not home right now 😦

[polar epoch]

Aye. I'm smart, but not so clever sadly. So takes me fairly long time to get the basics of anything new and foreign

[shy kelp]

What is the point of including stuff like this upscaling in comfyui nodes?

the blurry one is the upscaled one lol

[lyric grove]

Are you latent upscaling with a low denoise?

[shy kelp]

its just a bilinear

[polar epoch]

No worries, i'm gonna watch youtube videos that explains what is what, why place X there and what it achieves hehe

[fleet harness]

maybe you forgot to disable add noise?

[fierce hollow]

you're supposed to use that upscaling with a second pass

[fierce hollow]

for a normal upscaler check out hat-l/esrgan ones

[wet nacelle]

So your saying that if I use this VAE is should have lesser issues than I already don't have?

I'd be using a VAE without having to use it to begin with correct?

Instead of using the official one I should instead be using a fork that supposedly fixes the issues I don't see nor have.

AM I understanding this correctly?

[fierce hollow]

huh? I wasn't even responding to you

[autumn forum]

No worries. You can do it! I believe in you

[lyric grove]

He's in need of attention

[eternal fog]

Is that the hacked together one I put on here lol?

[lyric grove]

Just ignore him lmao

[fleet harness]

possible? I forgot which workflow it came from.

[daring smelt]

You've been timed out for 1hr
See if you can play nice with everyone without being so rude when you come back

[fierce hollow]

if you have to know the official sai vae produces nans in fp16 precision which makes your loras and stuff basically broken

[quasi remnant]

I linked the official vae from the official fork https://www.huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors

[polar epoch]

Ohh. There we go, think i found out why certain nodes didn't wanna show. Some of them didn't install dependencies

[fleet harness]

I just like the easy selction lol

[quasi remnant]

This vae is meant to be used alongside SDXL

If you don't trust the fp16 fix then trust the official vae

[eternal fog]

It looks like mine, I just took throttlekittys and added a multiplier to it

[fleet harness]

https://tenor.com/view/ben-stiller-confused-owen-wilson-dont-know-how-to-use-computer-zoolander-gif-8968881

[lusty raptor]

it could be mine. i never intended it to actually be used. I guess i could make a gist and register it with the manager

[fleet harness]

i difer to you guys on that. I like it. I'm lazy and it helps things along

[polar epoch]

As i downloaded 1 click install/portable comfy, how would i pip install stuff to it? Just simply open CMD in it's folder and install missing module? or what's the way for venv ones outsude of anaconda?

[fleet harness]

do it. I like it.

[autumn forum]

Yeah yours is the one I use. I love it! Thanks for making that. At least I think it is lmao

[hollow halo]

Besides them holding the cup on the wrong side with the force, I like this image 🙂

[fleet harness]

force sip

[hard fractal]

I'm using it

what does width and height do?

[fleet harness]

seemingly nothing lol

[lusty raptor]

nothing. I just wasn't sure if the node needed values; I had a massive headache trying to get the node to run in the first place because of windows(?), and was too annoyed to fix it further

literally copy/pasted to a new file and it worked

[fleet harness]

https://tenor.com/view/ted-frank-stevens-that-took-guts-we-need-guts-courage-gif-26327150

[lusty raptor]

https://github.com/throttlekitty/SDXLCustomAspectRatio

[slate jewel]

Does the newest version of Easy Diffusion contain SDXL 1.0 ?

[potent summit]

Which folder do I put the refiner for ComfyUI?

[lusty raptor]

with your other stable diffusion models

[hollow halo]

If you like Easy Diffusion, check out Arthemy. They are making a super cool application imo, I like their evolution workflow (just img2img but super sleek) I don't think Easy Diffusion supports SDXL yet though, but I could be wrong

[soft zealot]

Dont know if this helps but I've added a routine to my workflow that writes the prompt & seed used (want to add some other parameters in when I get 5) to a .txt tfile in the output directory with the same filename as the image generated from it.

NB full info on any custom nodes, styles sheets etc needed to work ar in the Credit & Notes textbox in the workflow

[slate jewel]

Alright i will look into that. Thank you for your quick response 👍

[primal nimbus]

where is the ''trainer ready'' now?

[mental pecan]

just tried sdxl

thoughts?

png info for promt btw

[fleet veldt]

looks like it's working 👍

[mental pecan]

took me a while lol

[hollow halo]

How much coke was on the table before you started vs after? Jk lol but nice images

[mental pecan]

hahahah nice joke (5kg)

[shy kelp]

kinda boring

[mental pecan]

reccomendations?

im all ears

[polar epoch]

Thanks! :D

And is there a addon that has like a png info node, and would read from current workflow you have, state "node missing", list nodes, and highlighting with a red outline between the nodes image originally used and where rest is fine to note what's missing?

[mental pecan]

oi you taking long to reccomend its scary

[edgy sierra]

how to config sdxl aspect ration?

[shy kelp]

less cartoony, bigger dynaminc range of colour, but I feel its more a a problem with sdxl

[upbeat summit]

hey, that looks cool. I've build something similar - a metadata .txt generator with just nodes. It helped getting to know comfyui and it's fun building components. I should probably start learning to build my own nodes 😉

[mental pecan]

i use an extension (a1111) sd-webui-aspect-ratio-helper

[fast moat]

Ya.. no img in output folder. It was generating earlier but then stopped. Found a similar issue on A1111's GitHub. It's an issue with vae it seems.

[mental pecan]

nah nah i used an artist name for that

its actually fun mixing artists yall should try it

[soft zealot]

Im just tinkering with this at the moment from WAS Suite

[shy kelp]

im uncultured I dont know any artists xD

[strange mist]

looks nice. which UI are you using? 🙂

[mental pecan]

a1111

why, cant you run it on a1111?

[strange mist]

oh, is refiner working already?

[warm hazel]

SDXL can run on auto1111.

[mental pecan]

yes but you cant use the vae-ft-mse-840000-ema-pruned.ckpt vae

or actually any vae i tried

[fierce hollow]

are you using some theme to get the connections so straight?

[mental pecan]

because it will mess up the image agt the last step

[jagged cove]

How'd you get those connections to be square?

[mental pecan]

i updated webui to 1.5.1 for it to work too

[soft zealot]

@fierce hollow @jagged cove

Link Tidier. Save in \ComfyUI\web\extensions (0=straight 1= slight curves 3 = Standard) https://drive.google.com/file/d/11pAM5HW3S72DQv4Qw6cXQXTwCrLFlN2o/view?usp=sharing

[strange mist]

you need sdxl vae for that

[mental pecan]

oh really

well

T0T

[fierce hollow]

thanks

[upbeat summit]

ah - great idea! looking good 🙂

[potent summit]

Has anyone figured out how to get text right?

[glass wolf]

wow, comfyui is a game changer for sdxl; kept getting oom with sdnext (i'd get maybe 10 images generated before crashing, but was always at 95% gpu vram utilization). sitting at around 75% vram used with comfyui. rtx 3070, 8 GB ram.

[urban fjord]

If you're using Auto1111 I recommend using the fp16 VAE for SDXL that you can find on huggingface as that has a better performance.

[lusty raptor]

gawsh, haven't done stuff in a long time, like... pull requests. but i did it! 😮

[jagged cove]

Ah, SO much cleaner. Thanks!

[strange mist]

you mean the third party vae?

[mental pecan]

ill try that thanks

just got the sdxl vae

litterally no difference than no vae

that i can see

[soft zealot]

you want clean ?

This is my daily driver workspace in ComfyUI ;o)

Managed to hide most of the lines altogether

[late marsh]

Few random SDXL images made in Comfy today

[urban fjord]

Yes, then you do not need to use the no half VAE param and I pretty sure you get better performance.

[shy kelp]

workflow.json PLZ

[jagged cove]

I love that.

[late marsh]

They're in the images

[fierce hollow]

you might be interested in https://github.com/space-nuko/ComfyBox it's like a cross between comfy nodes and normal ui

[hearty ginkgo]

How do you get that image save node?

[hearty ginkgo]

This one

[strange mist]

thanks, will try that tomorrow 👍

[soft zealot]

Treid it, liked the idea, my workflow refuses to load, move on 🙂

[fierce hollow]

fair enough

[jagged cove]

I was so hesitant to move from A1111 but i get 10x better speed with comfy

[soft zealot]

its part of WAS Nodes Suite

[late marsh]

@shy kelp Just throw the image in Comfy. There's the workflow

[shy kelp]

How do you do that?

[late marsh]

Save the image from here first and then drag&drop it in Comfy

[jagged cove]

Has anyone tried making Loras yet?

[soft zealot]

open image in browser , save to PC, dragndrop into comfy

[urban fjord]

mfish: Yes, LoRAs work great

[fleet harness]

unfortunately some of the nodes do not show up in the WAS nodes when you update using manager. I don't know what to do about that lol

[jagged cove]

Did you use Kohya?

[shy kelp]

Interesting how does it do that??

I did get a bug though When loading the graph, the following node types were not found:

Text Multiline
Text Concatenate

[soft zealot]

update manually using git clone

[shy kelp]

xif data or something Im assuming

[fleet harness]

did that too and still missing nodes. Not really shure what to do.

[fierce hollow]

I did a bunch with kohya's scripts earlier, they work quite well for the most part, still playing around with the settings

[fleet harness]

usually i can figure it out lol

[jagged cove]

Do you need the sdxl dev branch still?

[soft zealot]

delete completly from custom_nodes and then do a fresh git clone

[peak wraith]

I swear nobody knows this lol

[fierce hollow]

afaik yes, the main branch is outdated by a couple weeks

[upbeat summit]

I have so many QoL ideas. I guess I need to build some front-end helpers. I think it would be really nice, especially when you build compact builds, when you hover over a pin that it shows the source / destination node - maybe even highlighting the corresponding source / destination node with a border.

[fleet harness]

will try that as well

[late marsh]

Yeah I didn't believe it at first either 😄

[urban fjord]

Yes, here is one I made to get cleaner white backgrounds.
Top is SDXL 1.0, bottom is SDXL 0.9.
Left is no LoRA, then 0.9 LoRA and then 1.0 LoRA to the right.

[fierce hollow]

you also need to use sdxl_train_network.py, not train_network.py

[jagged cove]

Cool cool. Gonna get the branch now. Thanks! So it wont work with the GUI?

[soft zealot]

oh and dont forget to restart Comfy after updating

[urban fjord]

I used the bmaltais/kohya_ss version.

[fierce hollow]

with bmaltais' gui? I don't think so unless they also have an sdxl branch, too much stuff to modify otherwise

[shy kelp]

I'm seemingly missing Text Multiline & Concatenate nodes, those sound like pretty basic node names, should I just reinstall my swarmui?

[hearty ginkgo]

I have thay installed I can't find it thougj

[urban fjord]

bmaltais/kohya_ss version has SDXL support on the main branch.

[fierce hollow]

ah I see

[soft zealot]

did you restart comfy after uodating?

[candid pond]

Can you please share your Lora config file?

[fleet harness]

of course, still no dice, trying more things

[candid pond]

I have been using 2e-6 for checkpoint fine-tuning on 1.5 based models and have no idea what le etc to use for xl lora.

[late marsh]

I'm sorry I don't know how to install the missing custom nodes in SwarmUI. Didn't get it to work at all so I don't know how that works.

[soft zealot]

If in doubt ask the maintainer on his git page

[prisma wyvern]

Look awesome please share image with this workflow 🙏👍

[fleet harness]

ok!

[fierce hollow]

if you use adafactor it will set the lr for you based on the initial one which is like 2e-6, it's pretty cool

it gets lower/higher depending on loss

[soft zealot]

#✨｜sdxl message

[candid pond]

Oooh, I see!

[digital shoal]

does kohya_ss work on AMD gpus?

[shy kelp]

do you need the vae or is it built in? I mean I know there is a vae, so whats the advantage to not having it?

[candid pond]

I also have no idea about how to set the steps on Kohya. What exactly is epoch and repeat?

[fierce hollow]

epoch is how many times all the datasets repeat, repeat is per dataset within an epoch

not sure that's understandable...

[candid pond]

Oooh

[fierce hollow]

if you're using a single dataset epoch and repeats are basically interchangeable, some settings like checkpoint saving rely on epochs though iirc

[candid pond]

So for checkpoint training I usually do number of images x 100

So for LORA I should, lets say I have 20 images

[urban fjord]

Config for that LoRA with paths removed.

[candid pond]

I should do 40 epochs and 50 repeats?

[urban fjord]

There is no need to do repeats unless you need to balance your dataset.

[fierce hollow]

depends on how many steps you want, that's probably going to be an overkill though

[vocal stream]

if you have one dataset that sounds equivalent to just having 200 epochs

[urban fjord]

20 epochs 1 repeat is probably what you want to aim for.

But do 40 epochs and save every 10 to be sure.

[candid pond]

But then 20x20 would give 400 steps.

Wouldnt that be too low?

[pearl shoal]

I reinstalled everything and I got this now
It did not even pretent to create an image unlike the last time.

Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is None and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is 2048 and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is None and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is 2048 and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is None and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is 2048 and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is None and using 10 heads.
Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is 2048 and using 10 heads.
model_type EPS
adm 2816
making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
left over keys: dict_keys(['cond_stage_model.clip_l.transformer.text_model.embeddings.position_ids'])
Killed

[bronze kestrel]

Does anyone know if SDXL support tileable / seamless generations natively? If not I'm looking for that and can hire someone too if anyone is interested. Thanks in advance.

[vocal stream]

it's not steps, it's how many times you train over the whole dataset, even 50 sounds overfitty at least naively looking at that

[fierce hollow]

probably, yeah, 1000 is what people usually recommend as a good starting point

stuff like flip aug also factors into the number

[urban fjord]

That's why I said do 40 epoch, save every 10 and see when it works th ebest.

[pearl shoal]

Searching the github it looking like a memory problem

[urban fjord]

And this is why 90% of Civitai LoRA are deep-fried.

[pearl shoal]

But I have 12 gb Vram and 16 GB ram

[candid pond]

Oh I see. I understand now.

[fierce hollow]

I mean, if you have 1 image that's going to fry, sure

[candid pond]

So for the config, I should live it as default and chose adafactor with 2e-6 lr.

[urban fjord]

This one was 20 steps.

[vocal stream]

huh I dont see anything sdxl 1 on civitai yet, Id have thought someone wouldve uploaded something by now

[lyric grove]

dreamshaper

[quartz yew]

found the problem! Your Refiner strength is way to high! Saw a video of some testing Refiner strength and the more it goes from 0 to 1 the longer the faces get!

[hot dagger]

How does one use the refiner on a1111

???

Merging safetensors?

Img2img?

[vocal stream]

oh yeah i can see it from his profile but not from search

[hot dagger]

It’s doesn’t seem to load the refiner standalone

[frozen inlet]

the option of tiling should be available still in automatic1111 am I right? still didn't test Im downloading now

[candid pond]

I think you might have set 1.5 in the filtering.

You can change it by clicking the filtering icon on the right and choosing sdxl

[vocal stream]

no I specifically set it to sdxl 1.0 and I get 0 results (and I get some when I set it to 0.9)

[fierce hollow]

does it crash when loading the model into ram or vram?

[vocal stream]

ah maybe some otherr setting actually because now I see it, probably just had a style on or something

[pearl shoal]

It crashes I add to queue, not when I selected the model

[bronze kestrel]

thanks I really don't know. I have been using a 2.1 model available on replicate to date. I don't know what I am doing at dev level so just need to find the right help.

[sweet abyss]

will there be an inpaint model released aswell ?

[fierce hollow]

that's rather weird, but I remember people had problems with memory when their page file was too low, might want to check that

[pearl shoal]

I am using linux btw and I dont have problems running Sd 1.5, what is min spek of sdxl? I am also using xformers

[fierce hollow]

you should be able to load the base model with 8gb of vram, but it will require around twice that to load into ram first

as in, around 16gb of cpu ram, but I'm not sure if it actually crosses that

that's why issues with page file/swap in any case

[pearl shoal]

Maybe I am running out of ram? I have 16 GB ram, couple of GB must be occupied by my OS

[urban fjord]

If you can afford it without any issue I do recommend upgrading from 16 GB RAM to 32 GB or higher.

It will make the SDXL experience so much nicer.

[pearl shoal]

Is there any tricks I can use to reduce ram usage for now? I just want to experiment.

[fierce hollow]

I just checked and I'm only using around 6gb for loading, strange

[high skiff]

Very interesting

[fierce hollow]

one thing you might want to try is --dont-upcast-attention when starting comfy

[high skiff]

@upbeat summit are you here by chance?

[fierce hollow]

also --normalvram if that's not set already

[urban fjord]

A problem with the low VRAM settings is that it is kind of leveraging your RAM and if that's also relatively low you will get nowhere, but sure try.

[fierce hollow]

--use-pytorch-cross-attention (sdp) might or might not work better than xformers too, worth a try

[hearty ginkgo]

Yes

[shy kelp]

what folder do lora's go in in comfy ui's directory?

[soft zealot]

as per response to someone else with the same problem, might be worth asking the question at the maintainers git page

[polar epoch]

Only trick would be to quit anything you don't need right now.

Got anything here? Quit anything running

[hearty ginkgo]

Quickbooks uses alot of computer resources or is the qb not quick books?

[lusty raptor]

@visual glade I've got a request: Can we get a UI function to store/load values across multiple nodes for the current graph?
Say, sometimes I like the base model only, and sometimes I like base+refiner, and I have my graph set up to do either- but I have to manually click around to convert from one to the other. If I had a dropdown to do that at will, that would be ...comfy.
It should mostly concern itself with values on nodes, and their muted state, not position or anything extra.

[polar epoch]

Oh no, that's a torrent program lol

[hearty ginkgo]

Oh

Use a vpn

[polar epoch]

These are my tasks atm, then you can see what uses the most and quit anything using significant amount

[tall lion]

Just use it like aspect 3:4 in positive prompt?

[lusty raptor]

you select what you want from the dropdown.

[azure oxide]

requests probly should go in the discussions/idea section of comfyui's github, theyre tracked and responded to better there

[cyan crown]

what is a good nuber of steps for model and fopr refiner ?

[hearty ginkgo]

Why ur letting Google use 5gb of ram?

[wet nacelle]

@fleet veldt I swear this is a bot. They've only messaged once in the server and spammed reacted my reply.

It's as if it knew the last message I sent before I was timed out.

[fleet veldt]

am not a bot..and try 839 times back before this even went public

[fierce hollow]

[wet nacelle]

then you were somehow very tentative.

[solemn scroll]

Will I really buy a $500 GPU to generate cute women? ||kekking hard rn, I'm not even running it at the correct 1024x1024 size yet|| I am thinking about switching to comfy, but also worried about loosing all those cross attention and lowvram optimisation :/ (RTX 2060, 6GB)

[soft zealot]

I know a man whos using sdxl comfyui with OOTB settings plus one of those no issues

[fleet veldt]

last 3090 i got was a used\refurb just needed fresh paste. +$750

[azure oxide]

comfy should be way more optimized than a1111's from what ive read

[soft zealot]

my 1080ti with OEAM cooler plus fitted with an EK Waterblaock was <£200 and runs cjust fine. A1111 or comfy, SD1.5,SD2.1 & SDXL

sure it s a 1/4 f the speed of a 3090 but who cares

[solemn scroll]

Wow, I'll probably install it this night then. So no optimization needed, it'll run better just by installing comfy's UI?

[fleet veldt]

I do a bit more than just inference. and run huge batches during inference.. but gotta do what works for you. M40 wasn't enough so I kept throwing money at the problem 🙂

[soft zealot]

Just Playing

[pearl shoal]

python3 main.py --dont-upcast-attention --use-pytorch-cross-attention and closing discord worked. Generration in 147 second. Thanks @polar epoch @fierce hollow

[cyan crown]

any of you is using upscaler with comfyui and sdcl?

[honest flint]

i really hope the next gen nvidia cards finally ship with more memory, last time they increased memory on the cards was like what Titan? Titan RTX half a decade ago?

[shy kelp]

UpscaleModelLoader:
    - Value not in list: model_name: '4x_NMKD-Siax_200k.pth' not in []
UpscaleModelLoader:
    - Value not in list: model_name: '4x-UltraSharp.pth' not in []
VAELoader:
    - Value not in list: vae_name: 'sdxl_vae_1.0.safetensors' not in []```

Anyone know how to fix this? I have the files but I can't seem to get it to locate them, I don't know which folder its trying to read them from but I've tried 'upscale' and 'upscaler' in the models folder

[high skiff]

ok, due to shockingly popular demand, I am gonna be starting a patreon with hopes to further fund my AI research and provide even better things in the future

[azure oxide]

takemymoney.gif

[polar epoch]

Yep.

Don't go too wild on upscalers, or it'll chug on 64GB ram and 24GB video memory as a afternoon snack

[rigid prawn]

where can I put "CUDA_LAUNCH_BLOCKING=1" for comfyui?

[cyan crown]

well have only 12GB of GPU RAM (4070) and 128GB of RAM. BTW wich upscaling method are u using?

because with 4k ultrasharp it comes out too sharpened

[quasi remnant]

Pretty small filesize for 16kX16k

8k x 8k is already like 350mb lossless lol

[zealous horizon]

I've had great results on foolhardy. My go-to upscaler for basically everything. definitely worth trying

[flint copper]

Will your upcoming SDXL workflow be locked behind a paywall then?

[high skiff]

no, that will be live and for free on patreon

sorry

on github

[soft zealot]

anime prompt through an analog photo style filter

[high skiff]

but more advanced documentation and future research will likely be behind the patreon. At least to some extent to try and support myself, as I am fully unemployed

[boreal bough]

I have to use .exr at work XD that size would go into the multiple gb

[quasi remnant]

when ur just starting out it's better to do Ko-fi FYI

[cyan crown]

which one? Remacri Remacri Backup Mirror ?

[quasi remnant]

Patreon is better when u have an established base of people following you

[flint copper]

Wonderful to hear, looking forward to it. Thanks for your efforts!

[quasi remnant]

and Ko-fi takes no cut of the donations

[vernal cloak]

Is the base SDXL model capable of image2image? I assume so? I'm getting blurry results right now tho.

[flint copper]

Will join in once your page is setup. All the best.

[shy kelp]

kofi is better

[soft zealot]

why not use somewhere like https://www.buymeacoffee.com/ rather than Patreon (which takes 8%)

[quasi remnant]

@high skiff seriously consider Ko-fi if you're just starting out on a platform for donations; use Patreon once you're more well-known and have an established base of users willing to subscribe.

[urban fjord]

Yes the base model is capable of img2img

[soft zealot]

snap (almost) lol

[high skiff]

I have had buy me a coffee up for a while, but I assume that people either don't see it, or its not a common platform that people have access to

[vernal cloak]

Ahh thx. I must be doing something wrong then. All good. Thx

[quasi remnant]

ko-fi is moar well known than that too xD

[high skiff]

I am setting up a kofi right now

[shy kelp]

just use both lmao

got on as many platforms as you can if you want money

[fierce hollow]

cool

[soft zealot]

you do ? Its not obvious on your github ;o)

[polar epoch]

Curious what workflow you use to achieve that :P

[soft zealot]

scrolls to bottom and walks away shame faced

[quasi remnant]

i don't upscale myself cuz no point. the current settings with AI produce an image that's the size I'd want to use. my workflow is only for 1024x1024

[zealous horizon]

LoRA: I just tested 8 low quality (iPhone front camera) images with 20 repeats over 10 epochs. The results are insane for that input

[polar epoch]

I set one half up, but turns out they want your phone number, and as it's my personal and only number, i chose something else instead :P

[high skiff]

I put it at the end to not seem desperate

[quasi remnant]

Patreon's gonna want it too, ain't it?

[polar epoch]

Ah, so it's not a upscale, it's simply just resizing image

[polar epoch]

Not that i can remember when i set mine up.

[high skiff]

here it is in the read me 😅

[quasi remnant]

once @ the beginning and once @ the end lol and use a visual for it

heck, SDXL can spell "Donate"

you could make a donate visual/logo with it

transparent it out and make it fit nicely

[polar epoch]

I upscale with the "highres fix" equivalent where it resamples the image a bit. But i wanna use a proper sampler and redo the same steps for more details

[high skiff]

alright, I will update things to be a little more forefront, now that I have more community support and legitimacy behind my name now

[high skiff]

I have a new upscale workflow coming out

2048x reliable

[elfin flare]

Anyone know if there's a DML method of running SDXL?

[flint copper]

Would appreciate any recommendations for ComfyUI arguments and settings to utilize for a 3060 with 6GB VRAM and 16GB RAM? The refiner is a no go in most of the workflows I've tried in ComfyUI.

[quasi remnant]

Ye, don't just use a small link nobody will notice, consider making a DONATE logo

[sharp robin]

breh sick af prompt

[high skiff]

alright, will do haha

[shy kelp]

what does this mean File "D:\stable-diffusion-webui-1.5.1-RC\stable-diffusion-webui-1.5.1-RC\modules\textual_inversion\textual_inversion.py", line 283, in create_embedding
cond_model([""]) # will send cond model to GPU if lowvram/medvram is active
File "D:\stable-diffusion-webui-1.5.1-RC\stable-diffusion-webui-1.5.1-RC\venv\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "D:\stable-diffusion-webui-1.5.1-RC\stable-diffusion-webui-1.5.1-RC\repositories\generative-models\sgm\modules\encoders\modules.py", line 141, in forward
emb_out = embedder(batch[embedder.input_key])
TypeError: list indices must be integers or slices, not str

im trying to create a Textual Inversion

[austere comet]

Подскажите ,как с телефона это сделать ?

[quasi remnant]

like zero shot just now lol

but obviously you wanna mke it in the style you want

[elfin flare]

On your phone?

[quasi remnant]

xD

[elfin flare]

uh.

[quasi remnant]

it can spell DONATE fine kekw

[austere comet]

Да

[elfin flare]

Скорее всего, вам нужно будет запустить его через Google Colab.

[soft zealot]

enjoy a small pizza ;o)

[austere comet]

Это приложение?

[high skiff]

thank youuuuuuu

[elfin flare]

Это не приложение, а сайт.

[austere ferry]

Made with xl 1.0?

[elfin flare]

Однако я не уверен, была ли еще создана коллаборация Google для SDXL.

[sharp robin]

yep

[elfin flare]

Hey guys.

The guy speaking in russian is trying to figure out how to run it on a phone.

Is there a google colab link for SDXL yet?

[quasi remnant]

lol more prompt tinkering

dun forget sdxl can spell

like

[austere comet]

Я через другой делаю,в компе ок,а с телефона никак

[high skiff]

I bet I could make a dope DONATE logo lol

[quasi remnant]

up to 3 words

xD

[boreal bough]

I mean - the bots are still working
a lot of people in the chat here didn't know that when they just wanted to mess around

[quasi remnant]

hells yeah

it can even pull stuff off like this

lol

[austere ferry]

With lora? Mine are all mutated with both versions 😅🥲

[quasi remnant]

splitting the word up real stylized

[boreal bough]

corgi flowerfield donation? 🤣

[high skiff]

mmmm perhaps

[elfin flare]

Если на вашем ПК настроен репозиторий Automatic1111, вы можете добавить --share в командную строку, чтобы сделать его общедоступным URL-адресом для доступа на вашем телефоне.

God google translate has gotten a lot better

Lol

[sharp robin]

u need to gen at a minimum of 1024x1024 or a ratio of similar size

[high skiff]

I am trying to come up with a new 1.0 release prompt that will be even more prolific than my corgi prompt

[elfin flare]

Language gaps are starting to slowly not exist.

Can't wait for real-time voice to voice translation.

[abstract comet]

new SDXL is 🔥 amazing stuff, anyone know if inpainting works yet? (on comfyui)

[sharp robin]

Do those sizes

[boreal bough]

damn that armor is close to overwatch O:

what artist/style is it?

[high skiff]

man, I could only imagine

[sharp robin]

This basically just a little bit of modification #✨｜sdxl message

[high skiff]

train an RVC on your own voice, feed a real time TTS translator into it

[elfin flare]

We already have RVC, which is an open-sourced voice to voice deepfake software.

[quasi remnant]

Use GPT-4 or DeepL for translation

[elfin flare]

Yeah.

[quasi remnant]

accept no less.

Even GPT-4, not even connected to a translator

[high skiff]

oh yeah, I love RVC haha

[quasi remnant]

is better than GT

[high skiff]

I hang out with their staff when I am not here lmao

[elfin flare]

I cannot afford GPT-4, so what I usually use

[quasi remnant]

GPT-4 is free

[elfin flare]

is a locally-run LLM.

[quasi remnant]

it's on Microsoft Edge

[high skiff]

they were showing me a new thing they implemented just a few days ago

[glad fulcrum]

chatGPT is free, not 4

[elfin flare]

Oh you mean Bing-GPT4.

[quasi remnant]

bing AI is better than gpt-4

[elfin flare]

I have access to it of course

[elfin flare]

That doesn't make sense, since they're quite literally the same model.

[quasi remnant]

if you didn't know

[elfin flare]

I do know.

[quasi remnant]

They also add another model on top of it, prometheus.

prometheus+gpt-4

[elfin flare]

Ah.

[quasi remnant]

it's a much better stack than just gpt-4 alone

[elfin flare]

That I did not know.

[quasi remnant]

I've literally used this for all my tl needs for JP->en en->JP etc

it works for any language practically

and it's waaaay more fluent than GT

you can also ask for specific tones

like casual/friendly/for a discord message/etc

since some languages use diff words for diff tones

like JP you don't wanna use formal speak for discord xD

[boreal bough]

bing makes terrible custom dnd goblins!
I don't know why that has become my turing test XD but that's the golden standard I now judge LLMs by

[quasi remnant]

like what kind of custom gobby do u want? actually this is sdxl not general, getting offtopic xD

[boreal bough]

good point 🤣

[cyan crown]

and what about step count?

50 or more? or less?

[austere ferry]

Base 0.9?

[sharp robin]

1.0 base, 1.0 refiner, 0.9 vae

[high skiff]

as intended

[austere ferry]

Oooh

[elfin flare]

me with intel arc

Gotta wait

[austere ferry]

Wait in automatic 11111 checkpoint what did u use? The base model or the refiner model?

[quasi remnant]

i'll send u a dm for what i got with something i composed just now

[polar epoch]

How do i properly install missing modules for the 1 click AIO version of comfy? Added a few more addons to it

[quasi remnant]

if it doesn't come with a requirements.txt

can just keep pip installing every time it fails?

like pip install numba

[fierce hollow]

was suite should have a requirements file

[quasi remnant]

if there is a requirements.txt, get to that and open that folder in terminal/git bash/your preferred cmdline

[fierce hollow]

but idk if you can just pip install it with the portable version, might need to activate the venv

[quasi remnant]

pip install -r requirements.txt

[quasi remnant]

we are way past this but i left this genning for a while (and this is waaaay too big to be a logo tbh) but i thought this was badass lmao

[high skiff]

that is pretty sick haha

[zealous horizon]

results from my very low quality image input LoRA (8 images, 20 repeats, 10 epochs) All training images were similar to the photo of me, same background, same shirt. (my goal was to test the results on basic input that anyone can create on the fly. I used my iPhone front camera)

[upbeat summit]

that looks great 🙂

[zealous horizon]

here are some more

[cyan crown]

with kohja?

[zealous horizon]

Yes.

[hoary saddle]

wen Pixelass - Idiotproof Easy Step by Step Guide for Simple Lora Training release?

[cyan crown]

parameters are the same that I use with 1.5?

[fair crow]

Interesting insight -> "{token} man" >> "{token} male person". If you use male person SDXL won't put you in imakes like "baking cookies"

[polar epoch]

It would add numba to anaconda's location on my C drive instead of within the venv, so as it's contained, it can't see it

[soft bone]

@boreal bough what are your thoughts on 1.0 lora params? 1e-3 worked well for me on 0.9, but on 1.0 it looks way worse

[hoary saddle]

or SDXL Lora's for Dummies manual would work too 😉

[soft bone]

speed?

[zealous horizon]

he makes way too bloated videos, but this is basically what I used as a base: https://www.youtube.com/watch?v=AY6DMBCIZ3A

[fair crow]

You can port attention masking from lora_diffusion on github and use mediapipe selfi segmentation to create a attention mask to ignore the background. That works incedibly well

[zealous horizon]

I don't rememebr but fast compared to 20 images + 40 reg images.
My own face loRA did not use any reg images.

[soft bone]

LR?

[soft bone]

how complicated is this lol

[solemn scroll]

Hi, has anyone tried to copy paste A1111 venv to comfyUI ? I'm an a very slow internet (around 300kB/s) so I'd prefer not to have to worry about it

[zealous horizon]

LR, TE, UNET all at 0.0004

[tall lion]

Base+Refiner+Upscale SDXL1.0

[fair crow]

Not too complicated, essentially just this bit: https://github.com/cloneofsimo/lora/blob/bdd51b04c49fa90a88919a19850ec3b4cf3c5ecd/lora_diffusion/cli_lora_pti.py#L339C1-L363C1

[glad fulcrum]

I believe ComfyUI is not using too much space, but I could be wrong, I think I just unzipped the 1.5 GB file and that was it.

[quasi remnant]

gotta make a venv & conda install & activate all the stuff then iirc

i'm not gud with portable packages/using python stuff on diff drives xD

i usually just keep all the stuff installed on my main drive and use any python stuff on my main ssd

[shy kelp]

HI there. Should I be waiting for some sort of official update to A1111 before I can use sdxl 1.0?

[zealous horizon]

would that create transparent background? basically waht rembg would do?

[glad fulcrum]

update, its already supported

[jagged cove]

Comfy creates its own venv I believe

[fair crow]

Sort of - but better. It's more like "don't learn this part of the image" instead of "this part of the image is black/white/whatever" which might be picked up by the model

[zealous horizon]

thx, will play around with it.

[boreal bough]

still messing around with it. I can confirm that 1/16, 5e-4 will take about 200 hours to train. so no need to go there XD

[abstract comet]

what do you use for upscaling?

[boreal bough]

might be time to change scheduler

[eternal fog]

@high skiff Did you finalise your upscale workflow yet?

[high skiff]

still working, goal release is in a few days along with 1.0 workflow, my kofi, and my documentation

got a lot going on ATM

[eternal fog]

Could you send me a screenshot of your ksampler setup? I want to try it on mine and see what's so different. Because the way I'm trying to do it is seemingly ass.

[high skiff]

yeah, i can share in DM's, just please keep it between us cause its not finalized just yet

[viral narwhal]

SDXL+Refiner

[soft bone]

wait then how did he do this with 4e-4? #✨｜sdxl message

[placid edge]

RTX 5000 series rumored to be 32GB of memory for the flagship, we could be seeing 64gb professional cards 😮

[thin nova]

where'd ya hear that?

[autumn forum]

To bad they prolly won’t come out until 2025 lol

[random crystal]

Hmpf. Installed A1111 yesterday, and was able to generate 512x512 images super quickly (eg ~10 seconds). Today, it's more like ~10 minutes. I'm trying to figure out what the hell I did, or messed up, to make it slow down so much.

[placid edge]

yup but still xD

[quasi remnant]

after the 4000 rumors being proven totally false do you really trust some random rumor?

remember when 4090 was supposed to be 48GB

and make all the AI enthusiasts happy

lol

it's still like 1.66x faster which makes me jelly