✨｜sdxl

[hardy cipher]

longboi

[supple knot]

he would rule the basketball court

[shy kelp]

https://tenor.com/view/forrest-gump-running-me-when-im-late-tom-hanks-gif-5144739

couple of besties

[uncut fiber]

Rebecca Kit, Baloo.

[supple knot]

that yarn lora seems cool

[hardy cipher]

I have my very own lora now

[shy kelp]

nice

[supple knot]

I saw that post congratulations

[uncut fiber]

🙂

[hardy cipher]

https://tenor.com/view/yes-nod-agree-proud-jeremiah-johnson-gif-13827796

[uncut fiber]

how they are called?

[supple knot]

https://youtu.be/SC4xMk98Pdc?si=KpICUpk2Be6vqdRQ

[hardy cipher]

you mean the lora?

[uncut fiber]

i mean those dwarfes

[hardy cipher]

they started off as creepy rainbow goblins

it's a book I had when I was little

well, tried to make those, lol

then kept recycling them

[floral island]

ai chip

[scarlet atlas]

i want to use stable diffusion for free

how do i do that

[glad grove]

get a gpu

[supple knot]

@hardy cipher can you post a link to your lora

[scarlet atlas]

how about google colab

[hardy cipher]

he just pmed me a link to it on his google drive. I'll ask about uploading it or sharing it. didn't even think about that

but soon

[supple knot]

oh

[glad grove]

they banned use of webui there

[shy kelp]

lol when?

[hardy cipher]

well there you go

[floral island]

kek

[supple knot]

I was searching rainbow goblin on civitai

[shy kelp]

oh wait, actual webui or the lora notebook?

[scarlet atlas]

how i still see many people use it

[glad grove]

actual webui

[scarlet atlas]

idc wat google think

[shy kelp]

[scarlet atlas]

give me Free COLAb

[glad grove]

well theres prob some ways to bypass it but idk anythin about that

[noble shoal]

@hardy cipher You might share the drive link. But also share the prompting instructions (sorry currently at work)

[scarlet atlas]

i got this one stable diffusion colab but it keep disconnecting everytime its finished coding

[glad grove]

prob banned

[hardy cipher]

I can do that. just didn't want to link anything without running it by you.

[scarlet atlas]

how about you, are you using local ???

[glad grove]

yea

[floral island]

i think most of us here are

[hardy cipher]

you know, there are lots of alternatives to colab that are actually better and some of them even have free options

[shy kelp]

i am local

[hardy cipher]

or you can pay a few cents an hour to use a gpu

[supple knot]

Im using Loco

[shy kelp]

close!

[noble shoal]

After work I throw another 50 pictures in the dataset and get it cooked for a bit longer for V2. 😉

[floral island]

do you want them closer, or...?

[shy kelp]

just one head on another shoulder, need to get rid of hands too

close though

[hardy cipher]

https://tenor.com/view/gigachad-chad-gif-20773266

[shy kelp]

NO

[hardy cipher]

https://drive.google.com/file/d/1QjP1--cCOykNpLKo4HGsPsizrmJZpvd1/view?usp=sharing

[floral island]

lol, i just prompted the song title i'm listening to The last attack

[hardy cipher]

Example Prompt: p0p, a weird man wearing a meat hat eating ice, in the style of p0p
It reacts a lot on the word "weird" and "group of people". Best at strength 1.5 on the base checkpoint

[tender timber]

I've updated the nodes, here's that is in them:

Six Text Input Switch: Allows the user to select between six text inputs and uses a slider to make the selection. Useful for multiple inputs for prompt creation

Six Integer Input to Six Integer Output: I've seen a fair number of 3-, 4-, or more way text input and outputs, I wanted to do something for numbers as well. Use it as you wish.

Six Integer Widget: As above, but with widgets for entry instead of connectors

Endless Node Parameterizer: This node has a collection of inputs for the CLIP text Encoder and Refiners for SDXL based workflows

Parameterizer with prompt: As above, but with TEXT_G and TEXT_L outputs

COMBO Parameterizer with and without prompt: As above, but with two slots for aesthetic scoring, one for positive, the other for negative. Now you can use one box for both positive and negative refiners, but keep separate aesthetic scorer. Also comes in a variant that has the prompt boxes for you.

[noble shoal]

Try to prompt "a weird 80's boyband"

[floral island]

now, i add in my randomizer too. glorious!

[hardy cipher]

on it

[supple knot]

@hardy cipher idk they said they were rainbow goblins

[floral island]

holy shit, these are amazing 😮

[supple knot]

fighting contact Scenes are difficult for AI

[floral island]

it's very close though

aight, new songtitle, lets see how it does ^^ -> more next level randomness, instead of lyrics, now we take a song title and randomizerv2 to add spice

[hardy cipher]

is that pete davidson?

[supple knot]

yeah

[floral island]

don't question my methods

[supple knot]

so I look at the mess ups and try to describe what they could be

[hardy cipher]

alright, syntax error fixed so I can load things again

accidentally an S

don't question my methods

[floral island]

its. a thing. not sure what i'm looking at. but i'm looking at it!

[stuck bobcat]

I created a bot that automatically posts my SD images for me on FB, so any new SD image I produce it automatically posts on facebook without needing a user to upload over and over again their AI

https://www.facebook.com/profile.php?id=61551343954166

[noble shoal]

Wow, is that some alternate version of one of my heart black hole galaxy picture?

[stuck bobcat]

dope

lil bit from today

[floral island]

people are stillon facebook?

[stuck bobcat]

@floral island I'm planning to get my AI bot to work on Instagram/threads/twitter, but for now I did facebook because my elderly father & mother only know how to use that platform XD dang boomers

[floral island]

just keep having it on generation, flood the threads!

[nimble heart]

that your output folder?

[stuck bobcat]

eface can you tell me if you see my images being generated on the fb bot?

[floral island]

fb bot?

[floral island]

one of my output folders

[stuck bobcat]

https://www.facebook.com/profile.php?id=61551343954166

[floral island]

i think that's the sdxl output folder

[noble shoal]

Godverdomme

[floral island]

JE MOEDER!

😄 ❤️

[stuck bobcat]

if anyone can verify they see the images being automatically generated and posted that'd be dope

[nimble heart]

do you really need 30k images?

[hardy cipher]

man, I will just put random images into this workflow when I leave to do things. come back to all sorts of strange things

[floral island]

my old sd1.5 output folder

and got another 1.5 folder somewhere

so uhh... i might have an issue

[noble shoal]

Make a printed photo book only with the thumbnails.

[floral island]

aight, seems the song made a zombie baby. if not suitable, lemme know, i'll remove

[hardy cipher]

he looks happy

[nimble heart]

just output to RAM so it automatically clears every time you power cycle

[floral island]

nah, i love looking at old outputs, seeing how far we came, both the sd models and the prompters

[nimble heart]

I have one or two 1.4 renders lying around lol

let me find them

[floral island]

lemme get one of my first prompts, vs what we do nowadays on sdxl

(first need to generate the bunch tho)

[nimble heart]

one of my first widescreen SD 1.4 images

also the latest SDXL gen I saved. Crazy its only been a bit over a year

[floral island]

gpu still chugging lol

oh i have this "illness" that i always render in batches of 4+

[nimble heart]

i mean

for 1.5 based models I did batch 16 for everything

[noble shoal]

Supposedly generated on 17.09.2022

[floral island]

i mean, you do get the point , it's a dude in front of a caravan

[nimble heart]

for XL I do 4 mid res but only 1 high res. High res takes so long to render its better to just watch the live preview for each image so you can cancel it early

[floral island]

spot the difference

[nimble heart]

esp. if you're going above 1080p. 4k takes 7s/it on my system...

[nimble heart]

she grew up

[floral island]

and she dumped her ugly ass robe

[nimble heart]

that the same prompt or did you tweak it for the new clip

[floral island]

i did tweak it

i'll do acopule with the old prompt

see what happens XD

it shouldn't be too bad, except i'll get a christmas elf

[nimble heart]

it does OK but not nearly as well.

[noble shoal]

Same prompt a month ago, or so.

[nimble heart]

even just rewriting it into a sentence instead of comma separated terms helps on the new clip

[floral island]

small improvement ❤️

anime elf in forest -> this is the reason if you prompt for elves without further weights, you have to neg christmas

still, huge improvement

[nimble heart]

?

[hardy cipher]

used it to make that last image

[nimble heart]

ohh its the ip vision workflow you're doing

wait the cabin's on fire

[hardy cipher]

lol, yeah. it makes some strange things

[floral island]

lol, somehow it can't get hte scale right

[nimble heart]

she's a fair elf in a village of christmas elves

[hardy cipher]

so I've been low key feeling crazy for about the last hour because the lora I was trying to use wasn't showing up in the list. and turns out I put it in the wrong folder

so that's cool

[floral island]

i prompted 'lowkey'

it went about as well as i expected

[nimble heart]

not Pagan enough

damn you can make a 16x16 latent now with the lastest comfy

16x16 batch size 128 lets go

oh the max batch size is 64

[hardy cipher]

beautiful

[nimble heart]

Lowkey Loki

[uncut fiber]

could be interesting to play with it 🙂 to resize it to higher values.
@nimble heart try to colorize it 🙂

[nimble heart]

i adjusted my image math node to infer some colors based on lightness

not colorized per se but cool enough

[uncut fiber]

great, it at least do impression of old colors.

[nimble heart]

its a pretty small formula

could probably write a more advanced one to match something like those old psuedo-color photos

[upbeat summit]

nice - new node? looks interesting

[nimble heart]

That's been on my repo for a while

should be installable through the manager too

[upbeat summit]

yeah I got your nodes installed of course. I might have missed the commit on your repo

[uncut fiber]

@nimble heart there is bot on discord colorizing B&W images, just not sure if it is safe or not.

[uncut fiber]

@nimble heart 🎉
https://cdn.discordapp.com/attachments/1141988096943079464/1153615430276087968/deoldify.png

[uncut fiber]

It is lamborghini? i think

[hardy cipher]

I forget exactly what I used there, it was my ipadapter setup

[nimble heart]

looks like you took "make colorful" in a bit of a different direction lol

[hardy cipher]

these are prompt only, btw

no ipadapter magic

[nimble heart]

how do you plot a spiral if x and y are normalized 0..1

[hardy cipher]

import numpy as np
import matplotlib.pyplot as plt

# Generate theta values from 0 to 10*pi
theta = np.linspace(0, 10*np.pi, 1000)

# Calculate r values using theta^2
r = theta**2

# Normalize r values to the range [0,1]
r = (r - r.min()) / (r.max() - r.min())

# Convert polar coordinates to Cartesian coordinates
x = r * np.cos(theta)
y = r * np.sin(theta)

# Plot the spiral
plt.plot(x, y)
plt.show()

[uncut fiber]

i think it is realistic colors, yes not much colorful.

[hardy cipher]

that pig is high af

[nimble heart]

yea no idea how to translate that to absolute values lol.

rasterized I guess

[hardy cipher]

I don't know. I just guessed tbh

[nimble heart]

how do I re-blur an image after i click on it

[hardy cipher]

I've never figured it out

[noble shoal]

Wow, impressive.

[nimble heart]

well i drew a polar gradient so I guess that's like half way to a spiral

[floral island]

imma go ahead and be a bitch, but sorry, those are genitals visible -> i know it's artistic nude, but we can't have that around

i'd love it if we could keep it, but stability.ai don't like that stuff around here

[uncut fiber]

@nimble heart i think just move to different channel?

[nimble heart]

?

[uncut fiber]

to reblur spoiler

[nimble heart]

Oh yea

[hardy cipher]

don't worry, everyone, you're safe now. Eface saved you

[noble shoal]

Oh my god. I forgot that I am actually at work 😱. Thank you Eface.

[leaden fable]

Hey ! anyone know a control model capable of achieving similar results as this model for SDXL ? control_v1p_sd15_qrcode_monster

https://www.reddit.com/r/StableDiffusion/comments/16jtfpn/spirals_no_gigachad_yes/

[hardy cipher]

hmm

it'd take some finessing

[floral island]

hottest girl in town

[shy kelp]

lol

[noble shoal]

Sus but not subtle. 🥸

[shy kelp]

yeah im having a bit of trouble getting the subtle part right atm lol

[shy kelp]

then again, the inside of amogus is same color as bg hmm

[uncut fiber]

nice

Tried some stickers 😄

need some modification but doesnt look bad

[floral island]

how much psychedelics do you want in your shroom? -> yes

[uncut fiber]

it reminds me something 🙂

@floral island what do you think?

[floral island]

mona lisa looks a bit burned 😦

rest pretty damn good ❤️

[uncut fiber]

yes should be sticker

thank you for your response!

[leaden fable]

wow verry good, what did you used ? simply canny ?

[shy kelp]

the model? not sure, i used an xy plot and picked the good looking ones
not sure if it's prompts or what but man, my models aren't working the way i want them to

i should look into controlnet workflow for comfy

[floral island]

a council, of cats

[uncut fiber]

Jefferson airplane

[uncut fiber]

have you count them?

[uncut steeple]

Anyone running open pose with SDXL in Comfy? Did you manage to get the hand/face detection work?

[somber hill]

https://twitter.com/GozukaraFurkan/status/1704148923714056310

Furkan Gözükara (@GozukaraFurkan)

These images are generated with the SDXL DreamBooth workflow that I have discovered after 66 full training experiments

Used training dataset is not even good and I don't use captioning and only rare token approach "ohwx man"

Hopefully a full tutorial coming soon stay tuned

Twitter · Sep 19, 2023, 3:02 PM

[crisp owl]

Yeah I think my works as expected, it at least shows it did hand and face detection in the openpose preview, but I've only used it a single time, so would have to verify that

[uncut steeple]

For me it never shows hands and the face in the open pose image

[crisp owl]

which openpose node are you using?

[crisp owl]

looks like it didn't quite catch the hands in this one either, but for sure the face

[uncut steeple]

Why does your node look different

[crisp owl]

when di you last update?

well let me first see if it's part of a custom pack

you see here?

hmm actually I think I'm still using the old fannovel update before he started the new github for it

[shy kelp]

(this is not sdxl though)

[zinc cargo]

is there a reccomended python version to run comfy?

is it unrealated?

[peak dove]

How to instal OpenPose for A1111 and ComfyUI?

[uncut steeple]

I'll have to check tomorrow, everything should be up to date tho

[crisp owl]

@uncut steeple I just updated my folder to the newer release for that node, so now I have the "openpose pose recognition" node, and no longer does mine detect the face in that last image

yeah doens't seem to work

[uncut steeple]

so it's a bug with the new release

[crisp owl]

This is the repo I was using before where it was working
https://github.com/Fannovel16/comfy_controlnet_preprocessors

[uncut steeple]

The bane of always updating

[shy kelp]

i hit the queue prompt button, but it not registered, any solution?

[crisp owl]

did you use someone elses flow?

[shy kelp]

no, i create my own, is it the flow?

[crisp owl]

often it's due to using someone elses flow and there's a node with a model you do not have.

Check that on yours even though you made it. Make sure they're all filled in.
First easiest thing to check is that

Otherwise, would have to look at it to see what's the issue

[uncut steeple]

What does the console say

Also any errors? Or red circles etc

[steady grove]

you can edit the original image a bit too. increasing the contrast so that the open pose has a better sillohuette to read

[shy kelp]

nothing, but i just figure it out, it was facedetailer, i ctrl+b on it, but when i enable it, it works

[uncut steeple]

Yeah but it doesn't detect hands or faces with the new version at all

Tried multiple images

[crisp owl]

yup

[steady grove]

ahh i see. well you can always load up the entire automatic 1111 webui with the extension there, and use one of the many openpose readers installed there

then save that preprocessed pose and inject it directly in the comfy workflow

[uncut steeple]

Yeah but I have pretty much abandoned a1111

[steady grove]

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

i got the image loaded in the extension and none of the preprocessors are doing a great job at detecting the hadns on that either

i think its one of these edge cases where the preprocessor can't read the image well. nothing to do with an update. same ol preprocessor edge cases.

[uncut steeple]

Yeah the DW one did alright, but it generated me hands over the face

[uncut steeple]

Nope the update still fucked something, pretty sure

[steady grove]

i'm using old preprocessors on the extension and they're buggered on that pic too. sometimes the preprocessors just can't read the source very well

they're trained classifiers just as well and have their own limitations

[uncut steeple]

He had it working and then updated and it stopped detecting the face

[crisp owl]

toss open an issue on the repo

[uncut steeple]

Maybe tomorrow, I'm omw home now

[crisp owl]

I may do it since I can compare to the other install I have as well

[uncut fiber]

@shy kelp have you solved it?
dont you generate image with same setting already generated image?

[lusty wolf]

Very much related. "Comfyui has this standalone beta build which runs on python 3.11. "

[shy kelp]

bratz taxi

https://tenor.com/view/bratz-dolls-gif-25419272

[forest snow]

i am not quite getting used to just how good SDXL is, it is a model still full of surprises. great job. can't wait for SDXXL

[shy kelp]

NOW HERE

[indigo carbon]

now I made an image+text input version of the workflow

[hardy cipher]

nice

I've found that, with the right configuration of things, ipadapter can do miracles to lower quaity images

interesting how it works

plugging some of my old 1.5 images into it in different ways and combining them. it does impressive things

[indigo carbon]

again, this seems to beat MJ's version most of the time. also didn't have to do any streaching again to use it with AIT

man, BLiP is so much entertainment value haha

[hardy cipher]

I've been looking into the data and what not more, how it modifies things. I have some ideas but I'll wait to elaborate on them since not sure what's viable and what's not

[shadow condor]

Hello everyone

[hardy cipher]

oh yeah, it just repeats itself over and over sometimes. not sure why

[hardy cipher]

greetings

[indigo carbon]

this isn't much of a problem; this doesn't bother SDXL too much judging by the outputs

[shadow condor]

I used to play with stable diffusion 3-4 month back - I didnt had that much gpu but I have purchased 4090 and I have done my setup

so whats new ?

(I want to play with sdxl and new controlnets and so more

I even want to learn about comfyui and so more

any way to get started?

[indigo carbon]

if you would use AIT you might be able to get close to instant outputs on a 4090

[shadow condor]

what is AIT?

[indigo carbon]

this is what a normal AIT workflow looks like

[indigo carbon]

new optimization capable of almost tripling the speed on newer cards

[shadow condor]

this look so cool

how to get started -

I am noob with comfyui

I used to play with auto1111

[indigo carbon]

there is also something called "OneFlow", apparently it can be even faster than AIT; but we haven't figured out how to implement it. I heard that's what SAI used for training SDXL and for the bot and ClipDrop

[shadow condor]

is this comfyui ?

[indigo carbon]

If you're going for using AIT you might need to get an older commit of ComfyUI, but it might not matter

[indigo carbon]

yeah

with a bunch of extensions and modifications though

sadly, there is no solid frontend for Comfy yet, but there will certainly be one in the future

[shadow condor]

I have auto1111 installeed and I am installing sd_xl_base_1.0.safetensors
I am not familor with comfyui -

I used chainer with ebsynth before for video but I am noob with comfyui

any help or guide docs will be appriciated

[indigo carbon]

also I'd say if you're moving to Comfy, it will be smart to make a new VENV then use pytorch nightly and build Xformers from source for even extra speed (this is 9-12% on top of AIT (in my experience at least))

[shadow condor]

I am confused :}

[shadow condor]

let me see some tuts

[hardy cipher]

I realize it's not a precise representation of what was there, but look what this does to a photo that was essentially not even worth keeping. first two essentially help it come into focus here

[hardy cipher]

not like those are perfect images on the right, but kind of neat how it works

[slender coral]

Hey guys, I'm using comfy with sdxl 1.0 with Searge workflow (https://github.com/SeargeDP/SeargeSDXL), but no matter what I do I'm not able to use other checkpoints other than the sdxl1.0 or loras (https://civitai.com/models/112287/head-texture-map). Can someone help me understand what I'm doing wrong?

[supple knot]

This LorA says its for SD1.5

[slender coral]

Download the checkpoint? https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main

[supple knot]

@slender coral Yeah any 1.5 checkpoint will work, for the general rule SD 1.5 is trained on 512x512 images and SDXL is trained on 1024 x 1024,So if you don't use the right one they don't line up

[slender coral]

okay understood, I can continue to use the vae and everything else as normal (Just find a 1.5 workflow)?

\

[supple knot]

@slender coral you need the 1.5 refiner also

and vae

[slender coral]

Could you point me in the direction where I can get those, I can only find https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main

[crisp owl]

1.5 does not have a refiner as SDXL does

[slender coral]

Uh... 😄 could someone just point me in a tutorial or some other resource to get 1.5 working with a workflow?

[crisp owl]

here's my 1.5 workflow if ya wanna check it out.
I haven't touched it since SDXL came out, so some nodes may potentially be out of date

it's not super user friendly

I need to update it

[glad grove]

Ackchyually u can use sdxl refiner on 1.5 models 🤓

[indigo carbon]

it might actually help unironically; the refiner is meant to take the mistakes and fix them. it's now not as needed for modern XL models due to them being way more than enough. when it doesn't have anything to fix it kinda degrades the latents

the world's comfyiest pond:

[slender coral]

So I guess my question is what do I do here?

[slender coral]

Basically how do I get a scene like the couches, that's exactly what I'm looking for.

[indigo carbon]

I didn't have a couch in the prompt. I just wrote "the worlds most comfy pond, relaxing"

[slender coral]

No I mean I'm asking about the workflow / sdxl 1.5

[indigo carbon]

1.5 doesn't have a refiner

[slender coral]

I think for the sake of geetting confused on people suggesting things that I don't or cant use, do you or anyone here know of a good resource on how to set up 1.5? I have looked and unable to find anything. All resources I found are 10+ months old.

[indigo carbon]

1.5 is old tech and this isn't the channel you should be talking about it in. this channel is for SDXL, not 1.5

[slender coral]

You can clearly see I don't know what's going on so why don't you suggest something instead of telling me where I should be talking about it?

People here suggested 1.5, they also said I needed a refiner, so what do I do?

[slender coral]

If 1.5 isn't a good solution can anyone suggest anything.

[glad grove]

https://www.youtube.com/watch?v=14W__bxStMc second google result

[slender coral]

Thanks but that's for Auto1111.

[crisp owl]

Choose a LORA that is meant for SDXL, not 1.5.

[slender coral]

#✨｜sdxl message

[crisp owl]

that's 1.5

[slender coral]

Yes I understands, thats already been established. I'm getting compleetly wrong advice on what to do, the first suggestion said I need the 1.5 refiner. Now I'm getting links to Auto1111.

So all I'm asking for now is how to set up 1.5 or other suggestions / resources so I can learn on my own, because it seems you guys are just being rude instead of actually offering a simple solution (as I've said I couldn't find anything on setting up 1.5)

[glad grove]

because this is not the tech support channel #🤝｜tech-support

[slender coral]

If 1.5 is old tech, great can you offer a suggestion on what to use then?

[slender coral]

i'm not looking for tech support.

I'm also not looking for advice in auto1111 but you sent that anyways...

[glad grove]

because im not obliged to anything here tbh

[crisp owl]

If you want to use A1111, then you can use any 1.5 model in A1111, load up that lora, and go.

If you want to use ConfyUI, there are probably plenty of workflows on CivitAI for 1.5 you can find there, just search for em and include that LORA in the model.

Otherwise, as has been stated, this is getting into the woods of 1.5 talk, you can bring it to #1072238304042438758

[slender coral]

I agree, and I don't expect you to, but you acted pretty rudly like I wasn't able to find something when you couldn't either so what was the point of sending me a link to something I wasnt asking about.

[glad grove]

I'm sorry, but as an AI language model, I may not always be perfect and can make mistakes or provide inaccurate information. Please verify important details from reliable sources first before asking or you could try asking again by providing more details.

[slender coral]

I'll jump over there, but one final question to see if this fixes what I need. Is 1.5 the only solution I have in this regard? Or is there something that I may be missing or could use that would offer what I'm looking for?

[slender coral]

Sounds about right.

[crisp owl]

If you want to use that LORA, yes. It is trained for 1.5, and thus will only work for 1.5.

If someone makes a similar LORA for SDXL, then you could alternatively use that in SDXL.

With SDXL being just a couple months released, there's still much missing, and much being added.

[slender coral]

Thank you, is SDXL 1.0 the best option in your opinion?

I'm asking since I assuemed 1.5 would have been newer.

[crisp owl]

SDXL is substantially larger, and will be able to produce much nicer end results. But as stated, it's new, and more and more fine-tune models/LORA's are being released daily by the community

[glad grove]

depends on what you want to generate if you want realism,some cartoon styles or abstract then SDXL is the best out there

[slender coral]

kk, understood, I assumed 1.0 was older than 1.5 so that had me confused. Makes sense, thanks for your time.

[vital ermine]

Bear seems happy

[next plume]

It's glorious

[cedar tusk]

Hey @hardy cipher , are you using my wlsh nodes? Or do you have them installed, at least? Some people are reporting a bug where several of the I/O nodes just don't show up, but there's no output to explain it and it works on my system without issue so I'm totally lost. Was hoping to get some active feedback from someone with the issue on their system

[hardy cipher]

I've been using the upscale by factor node, no issues there. Which nodes specifically?

[cedar tusk]

The saving ones it seems. Image Save with Prompt Data (under IO) for example

They say nothing happens when you click

It SEEMS like it may be related to the new INFO input on it and a few other nodes

[hardy cipher]

Hmm. Well I'm actually not home so on my phone currently. But I can check it out shortly. Could try to load comfy from my phone on one of the online services. But wew, that is a chore

It is not mobile friendly

[cedar tusk]

Yeah no rush, haha

[hardy cipher]

So it's a prompt saving node? Haven't actually used that one, but I'll check it out

[cedar tusk]

(If it works)

[hardy cipher]

I always have things to post because leave things to render and link the folder to my Google drive

[cedar tusk]

There's three broken nodes: Image Save with Prompt Data (basically embed metadata like Auto1111 does), Save Prompt Info (writes metadata to txt file, no image save), and Image Save with Prompt File (the two combined). I added an INFO input to those that comes from a modified KSampler with an INFO output, which is basically just a dict with things like step count and sampler.

The modified sampler seems to not give errors though...

And on my system they all work!

[fringe barn]

What models are you guys using for upscaling? My upscales have been looking like trash.

[hardy cipher]

Hmm, it's always something small. At least for me

[cedar tusk]

I've recently been using DAT (DAT_light_x4 to be specific)

[hardy cipher]

Last two node errors I was informed about took me about 2 minutes each

[cedar tusk]

Yeah I expect it's something super minor

[hardy cipher]

Deprecated antialiasing in pillow. Now its lanczos

Also realized I forgot to map two nodes, lol. So they worked, just weren't there

[cedar tusk]

@visual glade If a node just doesn't load upon clicking the entry in the list (no error message), is there any place to find out why that happened? I have some nodes that work on my system but have user reports of the above failure happening

[hardy cipher]

That is something I've never encountered

Have you tried flooding it with print statements and trying to pinpoint where it goes awry?

[cedar tusk]

Ah, but it works on my system, which is the issue. One user commented out the INFO input on those nodes which allowed them to be placed BUT they didn't work (becaue of the internal logic being broken due to the missing INFO)

[visual glade]

check the console in your browser

[cedar tusk]

Forgot that was a thing, thanks

[hardy cipher]

Well don't leave me hanging if you figure it out. I'm curious now

Why did it make me these collage images for seemingly no reason?

Not complaining, but I do not understand

[cedar tusk]

It's trying to tell you something

Do you use refiner on your workflow or no?

[hardy cipher]

Sometimes. But don't think I was there

[cedar tusk]

been generating phone wallpapers with that wild aspect ratio (19.5:9)

[hardy cipher]

Nice. I've messed with some weird aspect ratios myself. Made a node that'll input any 2 numbers for aspect ratio and rounds to 64 pixels on each side. Not sure why 64 is the number rather than 8. But don't know enough to question it

[cedar tusk]

One of my nodes is Empty Latent by Ratio, where you pick from a list of aspect ratios and set the short side resolution and whether you want landscape or portrait. It also rounds. Made that for 1.5 but it works on SDXL too theoretically. Probably going to make a new node where you give apsect ratio and it spits out a 1MP for SDXL

[hardy cipher]

I made one that loads images and automatically resizes them to 1 mega pixel. Well by default but it can be adjusted

I wonder if anyone uses that one besides me

I'm really trying to learn more about the data flow now. It was a complete mystery to me until Ioved from a1111 to comfy

[uncut fiber]

i cant find my with extreme AR. pitty i think they were o.k.

[cedar tusk]

Isn't that one in the default nodes now? ImageScaleToTotalPixels?

[heady vale]

"watermark" included 😜

[uncut fiber]

@heady vale where

[heady vale]

see the water drops

[hardy cipher]

Yeah, but at least as far as I know it's a second node

[uncut fiber]

yes

[hardy cipher]

Not a huge deal, but I like to have all the images in using for controlnet, ipadapter, clipvision etc to either be uniform or at least a known size. So adding that to a leader just made sense to me

I pretty much just make things I want to use. Never sure if anyone else will find a use for them

[uncut fiber]

ive been using your workflow i think, and i likes it. So ppl using it, going try something right now 🙂 @hardy cipher

[hardy cipher]

They workflow was an accident, lol. Never really intended for it to become what it did. But I'm also a fan of it. Looking into ways of making it tighter. Not sure everything in it is necessary

The key is the pieces of flair I added to conditioning

Tried it with and without the changes. And they definitely change how the image renders

Cloud pigs

How does ipadapter make everything so detailed?

[uncut fiber]

yes very detailed. very sharp

[hardy cipher]

Yeah. I've been using it to overhaul old 1.5 images

Example. This image was trash.

And like magic it morphs it into something extremely detailed

Sure, not exactly the same content, but that could be fine tuned if you wanted

Structure is intact

[uncut fiber]

woah

it is result of image blend, just not sure if your wf or Tg wf

i only made one workflow, with very questionable usefulness... Not sure i ever can do anything new, sensible.

[hardy cipher]

I just try things to see if they work. Well to start at least. Worst that can happen is it doesn't work and I try something else

Reay trying to figure out ipadapter though. It's so new, it's like exploring uncharted land or something

[uncut fiber]

yes. It can be adventure, but i dont know where to start, and i am very slow. I am lucky i can use yours wfs

this is mickey mouse with evil nero like portrait

[hardy cipher]

Lol. Nice. Have you been adjusting parameters at all in the workflow? That's one thing I need to organize a lot better.

But actually wasn't planning on anyone else using it

Normally I build them up until I'm sick of them and then start over

[uncut fiber]

i only deleted i think resize and some models, because have not them and can resize later. I got not much space where i would love to have and dl here another SDXL model wouldnt help much

nice glasses 🙂

going to sleep, see you! @hardy cipher

[pure crystal]

https://www.mystic.ai/paulh/open-journey-xl:latest/play

[shy kelp]

wurstchen is a model or training?

[urban fjord]

I finally got to training my own Controlnet LoRA and it is much simpler than I expected. So yeah, it was really a good decision to scrap the old chunky Controlnet models in favour of this.

[hardy cipher]

Could you elaborate on the training? I haven't really looked into yet. And curious about how it works

[urban fjord]

You find the instructions here: https://github.com/kohya-ss/sd-scripts/blob/sdxl/docs/train_lllite_README.md

You need a beefy GPU, I don't know if 24 GB VRAM is even enough, I trained it in the cloud using an A40 with 48 GB VRAM.

[hardy cipher]

Cool. Saw that but haven't went over it yet

[low hill]

Made with the new model of 3HM, WowXL

[hardy cipher]

For how long?

[low hill]

WowXL

[urban fjord]

Basically you have two folders, one with input images and one with the normal training images where each input is paired up with a training image by giving them the same name.

Apart from that, the training process is basically identical.

[hardy cipher]

It's one of those things I've wanted to learn more about, but there are so many different things

I'm just curious what people train it on. Something similar to what's established like depth or canny? Or can they come up with novel new concepts?

[urban fjord]

What I tried was to teach it to colour black and white images as a simple proof of concept. Yes it is done before but I wanted to see what was required to make it work.

[vital ermine]

Hands 😦

[urban fjord]

I used only 6 training images and got a satisfactory results in just 77 epochs. I don't expect every task to be this easy, but it gives me confidence that I might be able to manually create the necessary input images if I want to try something fancier.

[shy kelp]

hmmm, "dragon scale armor"
i dont want a dragon!

smug bastard

[hardy cipher]

Right on. Yeah, loras are the way to go. So efficient

[urban fjord]

Here is an example output. With input on the right and output on the left. My training images were just 6 simple drawings so this is a completely new example for it. All in less than 10 minutes of Controlnet LoRA training.

[hardy cipher]

Whoa. 10 minutes?

[urban fjord]

I didn't measure, it might be closer to 5.

[hardy cipher]

Got me curious what sort if novel things it could be trained on

[urban fjord]

Though be warned, anything novel might be significantly harder to train.

[upbeat summit]

5-minute-fine-tunings - could be perfect for single image projects

[hardy cipher]

For sure, I've just considered it

[urban fjord]

As I mentioned, this is ControlNet LoRA modules and not normal LoRA models so for the most common tasks there should already be a Controlnet model for what you want to do. The hardest part for anything advanced is probably getting the necessary training data.

[urban fjord]

But this might be of use for certain style training actually as there isn't one-size fits all when it comes to colouring and it might be simpler to teach a style this way.

[native knot]

Thought some of you might find this an interesting model comparison. Nothing super special, but ran a prompt for The Statue of Liberty on Liberty Island. Interesting to see what models excel at various pieces of the image. Some do better with the statue, some do better with the base, some omit an island aside from the base altogether. Some have a very interesting angle.

[crisp owl]

If he does the Vulcan sign with 5 fingers, is that flipping someone off?

[heady vale]

[worthy orbit]

Stable Diffusion generation with less fingers

[hardy cipher]

thicc hands

[worthy orbit]

relaxing

[hardy cipher]

that's one of my old 1.5 images that I overhauled using ipadapter, clipvision, and controlnet

[indigo vine]

diffusers_xl_canny_full.safetensors
kohya_controllllite_xl_canny.safetensors
sai_xl_canny_128lora.safetensors
t2i-adapter_diffusers_xl_canny.safetensors
t2i-adapter_xl_canny.safetensors

Hi, I am a bit out of the loop those days with CN. What is the difference between all those models available?

[worthy orbit]

any advice on ip adapter? I don't get good results (it is very good when it does), so I don't use much

clipvision is a Comfy node thing I guess, I use Automatic1111, tried Comfy but got tired of scrolling the screen

[hardy cipher]

are you familiar with blip?

[worthy orbit]

yes

I mean the thing that interprets images

[hardy cipher]

one thing I've been doing is running blip on the image, sending it to clip g with xl, and then running it through a styler. also have a text box to add things, but makes the prompt sort of an after thought. but anyway, that seems to help. before I was getting lots of black box images

[worthy orbit]

a styler is like a magic prompt thing?

[hardy cipher]

well it's not required by any means. and the thing I'm referring to is pretty basic. it just adds some words to your prompt to give it a different theme

I'm actually going to start a new ipadapter workflows from scratch. not giving up on the other one, but I feel like it could be made a lot cleaner

[worthy orbit]

oh I understand

[mild birch]

Anyone know why auto1111 reloads model everytime I change prompt?

[crisp owl]

Probably just the clip portion

[hardy cipher]

did you see that workflow I've bee messing with. it's pretty cool actually

[worthy orbit]

I haven't checked it out, I've seem some nodes around xD and some crazy images. I've uninstalled Comfy so I wasn't into it

yes it seems like in Comfy you can add some crazy pipelines like a guitar pedal line

[hardy cipher]

well with this one I'm not really referring to the complexity as much as the simplicity

well under the hood maybe complex

but I can essentially just add 3 images and start it running and it'll combine them

not like perfectly 33 percent of each

lots of ways to adjust all that

[worthy orbit]

yes that's what drives me crazy of using "wild" things

it is like the casino xD

[hardy cipher]

well it's more like they do't all have the same kind of impact. one has more impact on structure, another one is more thematic. but anyway, not for everyone

not really like a casino for me. I can adjust things when I want to and then it changes

[worthy orbit]

For example I tried Magic prompt that adds more prompts according to the base prompts, and usually delivers unrelated meh images

[hardy cipher]

well I normally pay lots of attention to my prompts and didn't use those tools much

but just got sick of it and wanted to try something different

[worthy orbit]

yes I guess you have more control. And I don't know if I'm using Magic Prompt wrong or it sucked xD

[hardy cipher]

I've actually never used that oen

[worthy orbit]

Yes I want to start trying things like that, but for me it is like a goose chase, but I just need time to test

[hardy cipher]

well I genuinely don't understand why so many people are scared to connect things wrong. or it seems like it

if it doesn't work ijust do it differently

when I got it I had absolutely no idea what I wsa doing. and thought the tuturials were just sarcastic mockery or something because didn't realize you could load the images as workflows

but going back to your original question, I've done some silly complicated things that may or may not have helped

[uncut fiber]

do you know about tag "color" theme?

[hardy cipher]

hmm

what's that"

[uncut fiber]

yellow them red theme blue theme

[hardy cipher]

you just add that color to the prompt or something?

[uncut fiber]

yes just tag {yellow|red|blue} theme in my case

so yellow theme should do it

[hardy cipher]

well I definitely have color themed images

hadn't tried what you're doing though

[uncut fiber]

i am doing long dynamic like prompt bit limited random
one resulted in weird image

[hardy cipher]

lol, wow

I like this one

[uncut fiber]

yes it has as well limited colors

[hardy cipher]

that's oen of my old 1.5 images I think. but overhauled it

I like that this one

[uncut fiber]

nice

[hardy cipher]

looks like straight up clothe

[uncut fiber]

nobody has hands 🙂

tried to play with stickers, but lack of quality

[hardy cipher]

what's the story on those stickers?

[uncut fiber]

no story, there is some #1152260531982434434 thing

cow jumping over the Moon 😄

[hardy cipher]

gotta be some tricks to that

[uncut fiber]

yes just stickers itself.

[uncut fiber]

and higher cfg

[uncut fiber]

this is picture i dont understand very 🙂

[hardy cipher]

lol

[uncut fiber]

this last i tried to play with edges in krita...

[hardy cipher]

those look like quality stickers to me

[worthy orbit]

he got stuck on his foot a giant condom full of fluorescent material

[uncut fiber]

@hardy cipher not sure
@worthy orbit definitely something weird 🙂 looks like somebody was trying to inpaint his foot

[worthy orbit]

looks pretty nice

[worthy orbit]

last two looks good

[uncut fiber]

Thought rainbow theme will work like colors used, but it insert rainbow in picture. Will try some non colors themes

motorcycle and car theme

[upbeat rivet]

I'd like you to make some cute chibi pictures for me, please.

So I want to request a workflow?

[uncut fiber]

what UI are you using?

[upbeat rivet]

comfyUI

[uncut fiber]

so just drag image you like in it. If it is not reconverted, it has workflow in it.
Or ask for specific picture probably. I mean what is posted. If author can share with you workflow.

[upbeat rivet]

I want pictures in this style. Do you have any workflow?

[uncut fiber]

it is lora, so yes ask somebody with WF with lora support

i am not using loras

[upbeat rivet]

I can't create a workflow

[uncut fiber]

i cant as well, dont know how to add lora.

[upbeat rivet]

Okay, no problem, wait for the person with lora.

[uncut fiber]

7th mad max 1 2001 space oddysea @ionic gulch

[uncut fiber]

love this, hope i found way to generate it with more steps

[supple knot]

thats fake?

looks like a real photo

[uncut fiber]

normal pure only prompt

two hulks?

[supple knot]

Wasnt there a dc guy that looks like that

[uncut fiber]

A nobody is at sunrise in bad weather. BREAK A action movie theme. A Blue theme,telephoto lens art by famous author

[supple knot]

martian manhunter

[uncut fiber]

you can drag drop image in ComfUI to generate it, it is model Jugernaut

[mild birch]

anyone know why my automatic1111 reloads the model every time I change my prompt?

[supple knot]

I like how hulks now Capt America

[uncut fiber]

yes 🙂 and that on left can be turtle ninja

[mild birch]

I made my cat into a prince

He's thrilled

[uncut fiber]

@mild birch i dont know why it reloaded each time, definitely seems like bug.

Giants

[supple knot]

looking out for those giants

[ionic gulch]

Yes - 2 + 4 and maybe 3 should also be easy if you know the movie. 5+6 are probably the most popular movie in that style. for 9 the black/white style is a hint.

[uncut fiber]

9th is Blade runner? 4th i think is AI? realy dont know 🙂

[ionic gulch]

9 = metropolis, 4 = ex machina (great movie!)

[supple knot]

I like the madmax one

[ionic gulch]

ex machina real scene:
https://www.maenner.media/downloads/39833/download/Ex-Machina-Download-Wallpapers.jpg?cb=a0baeb0f32d481e58a4b190a99b09727

2 = 300, 3 = black swan, 5=eyes wide shut, 6=in the mood for love

the hardest is imo 8 = melachonia - you can see lars van trier style, but not so much the movie

[supple knot]

Giant Spotted

I'm going try and get closer

[uncut fiber]

@ionic gulch most of movies i haven't seems realy. But know by names

[ionic gulch]

i can highly recommend "ex machina". in the mood for love and black swan are also very good.

from lars van trier i would perfer "Dogville" over "Melancholia".

[uncut fiber]

will try it. thank you @ionic gulch

[uncut fiber]

Not sure, probably properly hold umbrella, haven't see this too much 🙂

[stone fossil]

Testing the steps / epochs now will release the printed circuit board lora tonight / tomorrow. 🙂

[wet nacelle]

Nevermind

[worthy orbit]

very nice, what model?

[wet nacelle]

ChimeraMi(XL)

[worthy orbit]

I'll look it up, thanks!

[wet nacelle]

yuppers

[noble shoal]

Cozy

[indigo carbon]

which one?

[uncut fiber]

Number 6 is living

[uncut fiber]

@noble shoal try rainbow theme 🙂

[noble shoal]

Lora or generating with that theme?

[uncut fiber]

just tag "rainbow theme" 🙂 if it can be more rainbowish

[noble shoal]

Lemme try

[uncut fiber]

xxxxx theme is very powerful tag

[noble shoal]

10/10

[uncut fiber]

lol cats 🙂 and happy some 🙂

It is repost today my best image. only 24 steps, doing more it making it worse.

[stone fossil]

Can u make her face the cam? 🙂

[uncut fiber]

i cant, i just show what future glasses will looks like 🙂
I will see what i can do, going to run Comfui

[stone fossil]

Just type facing the canera in the prompt.

[uncut fiber]

it was generated randomly from prompt

A {man|woman|nobody|robot|animals} is {walking|staying|playing|watching|staring|roaring} at {sunrise|sunset|night|midday} in {beautiful|bad} weather, in the {seaside, village, town, city, desert, space, landscape, sky-land, kingdom, realm}. BREAK A {action-|adventure-|romantic-|mystery-|war-}movie theme. A {Red|Yellow|Blue|Green|Colorful} theme,{picture|painting|photo|telephoto lens}, art by famous author

[uncut fiber]

camera is triggering she is holding camera. So i added viewer.

[stone fossil]

Ah nice. 🙂

[hardy cipher]

that's awesome

[stone fossil]

I will release the LoRa soon.

Testing between A B and C.

For our own good. 🙂

[uncut fiber]

Vermeer is AMD ZEN3 🙂

[stone fossil]

ZEN is good hehe.

[uncut fiber]

yes i have only zen2

it is Matisse

[ionic gulch]

next thing for my sdxl syler node (and workflow) after movie styles, will be artist styles:

[hardy cipher]

true art

[uncut fiber]

gogh in red period

[uncut fiber]

it is real photo? @wet nacelle

[wet nacelle]

no

[uncut fiber]

only thing there are two same thing for cooking

[wet nacelle]

This is your house. bro. I started the fire.

[crisp owl]

prompt: "the result of generating images for 24 hours straight"

[stone fossil]

I cant wait for PoP to combine this and a van Gogh lol.

[wet nacelle]

Wrong reply?

oh wait

I see what you mean

The actual guy

[stone fossil]

Ofc. 🙂

[wet nacelle]

https://tenor.com/view/catshakira-cat-dance-catdance-catdancing-gif-22503034

[ionic gulch]

really promising first results for the artist styles - prompt is only "detailed photo of a woman" besides the artist style:

[peak dove]

Can do hybrid prompts in A1111, which is fun, but they kinda take a long time for A1111 to braid up each one 🙂

[peak dove]

Superb!

[uncut fiber]

you mean combinator and all combination?

iam using comfy and it can repeat some combination, i think it is posible in A1111 as well @peak dove

[floral island]

https://civitai.com/models/148871?modelVersionId=166165 self-advertisement spam, i uploaded my model ^^

[native knot]

Meat.

[floral island]

now time to relax ^^

[native knot]

Be sure to pick up a copy of The Game 3: Game.

[hardy cipher]

messing with different ipadapter configurations. needs some work, lol

I guess a square wasn't the best choice

[uncut fiber]

reminds me some sort of scary fairy tales.

[hardy cipher]

ipadapter magic

[uncut fiber]

Tim Burton? Not sure it is style 🙂

[dim steeple]

Is there a controlnet model for simply giving it an image to use as a inspiration?

sort of like img2img

but in text2img

[ionic gulch]

bottom middle is tim burton.

at the moment i have ~60 (quite different) artists and i'm only though letters A to C. this will be a huge list. maybe i need a different way of selecting the artist.

don't know what's the limit for the dropdown selection in comfy, but i guess it can't handle hundereds of artists very well

i'm testing and filtering out artists at the moment. to check i do 3-5 images per artist. this could take days for the complete list i have

[uncut fiber]

yes have you seen wait a mmnt

https://sdxl.parrotzone.art/alphabetical/S

https://sdxl.parrotzone.art/

better

[ionic gulch]

yes, this is my main source. the first part of the selection phase is through the images they offer - this helps me to come down to 10%. second phase is my own check with a low strength ip adapter image, a simple prompt and 3-5 test images, but most of them work if the preview on their site looked good. so i still keep 75% after my own tests. so in the end there will still be hundreds of good artists.

[uncut fiber]

o.k. 🙂

[hardy cipher]

JPS do you know what the deal with ipadaper moving things to the right? I'm guessing one of the resolution settings?

[ionic gulch]

new style for the red carpet:

[uncut fiber]

egyptian?

[crisp owl]

Better than most of the crap that goes down it

[ionic gulch]

you just never have to shake or move your head in the wrong way :)

[hardy cipher]

n

[floral island]

the dancing dead ^^

[hardy cipher]

pure ipadapter

[slender coral]

Does inpainting work like photoshops content fill? And how well would you say it works for you?

[hardy cipher]

I don't even use it, lol

that's why all my images suck though

[ionic gulch]

got a better test setup now, that adds artist caption to the image, so i can increment through the list and check them later:

[ionic gulch]

will try with 4 different input images and prompts and only keep the ones that got at least 2 good or 1 outstanding results (or adjust those numbers if they kill to many/few artists from the list)

[cyan crown]

If anyone is using Auto1111 I found that changing VAE with fp16 version speed up a little generation

[lusty wolf]

ComfyUI: Can the default path setting for output be changed, without having to use the node path?

[bright valley]

Can we be friends maybe

[bright valley]

We have similar interests

[cyan crown]

both eating pizza

[ionic gulch]

he really is a master:

[hardy cipher]

Strange forehead on this girl

[livid cradle]

Finally LIVE - AI Product Photography App (SDX) based - Full trianing here: https://youtu.be/kUdJKNFQS-w

[hardy cipher]

Can it make this image?

[native knot]

Rhianna would definitely be a blue.

[somber hill]

https://twitter.com/GozukaraFurkan/status/1704625590437814424

Furkan Gözükara (@GozukaraFurkan)

Here SDXL LoRA vs SDXL DreamBooth

Both trained on RTX 3090 TI - 24 GB

Same epoch, same dataset, same training settings (except different LR for each one), same prompt and seed

LoRA has xFormers enabled & Rank 32

Will post previous epoch too next tweet

LoRA settings correct

Twitter · Sep 20, 2023, 10:36 PM

[uncut gull]

Trying to make stuff out of weird materials. Very hard to control.

[hardy cipher]

Perfection

[crisp owl]

picturesonpictures likes cursed images, so here ya go. One for you lol

[hardy cipher]

Hairy muscles or muscular hairs

[urban fjord]

Turns out that ControlNet LoRAs makes for decent style models too. Making these of Seagal alot more photographic.
Left 4 is no Controlnet while right 4 is with controlnet as the labels suggests. No normal LoRA applied. No Seagal in my training set as it was originally a failed pose Controlnet I made.

[hardy cipher]

do you happen to know much about the width and target width input parameters in xl? trying to understand what they're doing. all these images have identical parameters other than those two things.

*width/height

ipadapter has a thing for moving the subject of images halfway out of the image

[urban fjord]

Nope, I've got no idea. You probably need to ask someone who worked on SDXL.

[urban fjord]

1200 epochs with 9 images at learning-rate 0.0002. So seems hard to overcook.

[hardy cipher]

damn, I like those

now we need a lora based off this

well not a cute little mini white tiger kitten, but it's fire paper fur

[supple knot]

Mine is like 15th gen Warriors

[supple knot]

The Frogs are back!

[hardy cipher]

the frogs never left, my friend

they were just hanging out, waiting

[supple knot]

for the perfect time to make there return

[hardy cipher]

best frog girls

should be interesting in my new ipadapter flow that I'm trying to get rolling

[supple knot]

It looks like Chloë Sevigny on the left

and Mia Wallis

[somber hill]

https://twitter.com/GozukaraFurkan/status/1704649928138514565

Furkan Gözükara (@GozukaraFurkan)

DALL-E 3 vs SDXL 1.0 base

Hats off
Prompt: The inauguration of a wormhole portal between Shanghai and New York. New York is inside the portal and Shanghai outside of it

First 2 are DALL-E 3

Twitter · Sep 21, 2023, 12:13 AM

[urban fjord]

Did you get early access to DALL-E 3?

[somber hill]

someone posted dall e results. i generated SDXL myself

dont have access yet

[urban fjord]

Alright, DALL-E 3 seems more and more impressive for more examples I see. I hope SAI won't fall too far behind here.

[hardy cipher]

all those stupid resolutions

[eternal fog]

Hey @visual glade Have you changed something regarding groups on the latest comfyui update?

All the titles from my groups have vanished. If I right click > Title, it's still there, but it's not actually displaying.

Oh, it's changed all the font sizes to "Undefined", guessing because of this - https://github.com/comfyanonymous/ComfyUI/commit/b92a86d7370b28af6777c3859f7d486191f6379a

[keen nimbus]

screenshot from a new openworld stranger things MMO above

[hardy cipher]

decent

[visual glade]

should be fixed now

[eternal fog]

ok cool. I ended up just going into the JSON and adding in the "font_size" option to them all as it was missing.

[urban fjord]

I've been playing around some more with Controlnet training...
(Left is input, right is output) Using just txt2img and Controlnet LoRA.

[hardy cipher]

ipadapter on 3 images

[urban fjord]

What exactly is ipadapter?

[hardy cipher]

https://github.com/tencent-ailab/IP-Adapter

they explain it with better words than I can use

basically these were my 3 input images. use blip to interrogate each of the images and don't even touch a propmpt. just load images and adjust a few parameters as needed

results from some things I left to render

[urban fjord]

I don't really understand much of what's written on the github

[hardy cipher]

it looks at the model using vit g currently, does it's best to understand what's going on, then adds itself to the model data. sophisticated image prompt basically

also works with clip vision

not sure what this was but it was a mistake

[crisp owl]

I've been testing different images at different width/height & target width/height. I do everything fixed except for those variables for each set. I do one where I match my latent size, then one at each 1024x, 2048x, 4096x. (3076 was no good).
I honestly can't decide which of those 4 I prefer, still plan to test it out more.