🏞｜general-with-images

[late sorrel]

now that looks kinda cool

[true frost]

miku is tired 😮‍💨

poor girl is so exhausted 😖

[analog goblet]

(this was a weird batch, so weird that i post three images of it 🙂 )

this's some funny idea:

cannot recognize the model though 🙂

and i guess this model is called "the overkill" 🙂

and "the overkill 2" 🙂

[languid pebble]

1963 version of Apple Vision Pro

[vague cloak]

Doing a dreambooth on beautifulRealisticv7 with like 1000 photos of an idol, I am at 39% now and this is a sample image now

do I continue 😐

[languid pebble]

Sorry! No idea about that ...

[languid pebble]

I should take another look at deforum ^^

[crisp stream]

well 😄

[vague osprey]

#artisan-faq

[languid pebble]

But I have more cables 😛

[minor glen]

https://tenor.com/bofEK.gif

[crisp stream]

Not if you count them all 😄

[analog goblet]

it's tesla-coil-man !

[true frost]

dragon mommy

[clever oar]

hello

[clever oar]

mojo

Do you punish migrants who make doner from rotten chicken?

[languid pebble]

Yeah 🙂

Not my job, police and justice will do ...

[clever oar]

they do this often or in your country quality control>?

[languid pebble]

I hope so 🙂

[clever oar]

😎

[languid pebble]

But you can ask what they found ...

[clever oar]

in your city many sirian guys?

[languid pebble]

Hamburg has always been pretty mixed cause of the port ...

[clever oar]

what they found?

[languid pebble]

When they test a restaurant ... you can ask for the result. Not sure whether they are already online ...

[clever oar]

Can you get cancer and AIDS from a bad doner?

[languid pebble]

AIDS ... no ... cancer ... maybe if they are burned?

I buy my Döner an Central Station ... they sell hundrets or more every day. So they have to take care a lot.

[clever oar]

what is included in the doner you take, I wonder how different it is from what is sold here

[languid pebble]

I prefer a Döner Box: fries, salad, beef, Tzazik sauce

[clever oar]

its like this?

[languid pebble]

That's a usual Döner. I buy a box without bread but with fries

It's funny that it's a German invention 🙂

[clever oar]

i buy with bread but in not together

[languid pebble]

That's a good point in having many different cultures. So many different foods to discover!

[clever oar]

sometimes I make dishes from different countries

[late sorrel]

let me drink it :3

[languid pebble]

Yummy!

Islamists are pretty loud here at the moment ...

Having cake 🙂

[late sorrel]

you have a nice cake 🙂

[steep sleet]

PiXart-Sigma+SDXL+LoRA

[clever oar]

ok it for you

and it for mojo

cake inside!

[vague cloak]

I think I know what's wrong, I way over trained the model 😭

[languid pebble]

Didn't you save in between?

[late sorrel]

❤️

[final dagger]

greetings... ai visual production bots have been removed, anyone know?

[steep sleet]

Go to #1237459938901491852

[clever oar]

alien its not monster is woman

[languid pebble]

How are the Aliens called? Facehugger? This is more a brainhugger ^^

[clever oar]

huggingface.com

[languid pebble]

lol

I should buy huggingbrain.com 🙂

[clever oar]

you can i think

not work now

[languid pebble]

I already have 2 domains 🙂

Already sold ^^

Hi Moofi 🙂

[clever oar]

mojo

I made 4 cheese sandwiches in the oven for dinner

I wanted to treat you and mofi but I ate it myself

[languid pebble]

Hope you've enjoyd them.

[clever oar]

👍

next time you give me true kebab

[languid pebble]

If I win the creativity contest 😄

[clever oar]

100% win

[languid pebble]

12 days left ...

[clever oar]

oh

[languid pebble]

At the moment I am the only participant in the category short movie

[clever oar]

then you win it!

the only participant at all 🤭

[languid pebble]

I prefer contests without other participants 🙂

[late sorrel]

what about a pillow fight contest? :3

[languid pebble]

🙂

[late sorrel]

haha

[languid pebble]

I'll visit my pillow now. Good nite!

[late sorrel]

awww gn

what is the german version idk, guten tag is like good morning, but idk good night :3

[languid pebble]

Gute Nacht 🙂

[late sorrel]

ah yes i even knew that but forgot damn

Gute Nacht 🙂

but isnt Nacht also naked in german? LOL

[languid pebble]

Not really ... but in Hamburg you can just say: "Moin!" the whole day ... I'm off! C U

[late sorrel]

i think it's spelled Nakt maybe? anyway, gn haha

[broken moth]

this is all i came here to say

[late sorrel]

nice

[late sorrel]

shark car :3

[nimble mason]

shark rat 😄

[late sorrel]

haha nice 🙂

i cant wait to try some creepy pics

i wonder if sd3 will also have like cascade, the image remixing, cause that was cool

[nimble mason]

we have even better stuff now

new IPA nodes blow that out of the water

[late sorrel]

ah yes of course, i did some IPA stuff, really crazy

[nimble mason]

yeahhhhhh

[late sorrel]

imagine IPA with sd3 😮

greatness is coming baby

[nimble mason]

thisi s all with a lora i trained on some of my outputs

sure hope so 😄

[late sorrel]

cool, i wanna try training a lora soon, but im currently gathering a dataset, then i have to tag

[nimble mason]

use onetrainer, it's easier and uses less vram etc

[late sorrel]

yea i have onetrainer, heck even @junior sky did a whole tutorial on it over DM haha

[nimble mason]

awesome

[vague cloak]

I been trying to use automatic1111 dreambooth plugin to train on a face, is onetrainer better?

[languid pebble]

Good morning!

[nimble mason]

prompt = good morning

[languid pebble]

With a cat every morning is good ... 😄

[nimble mason]

also prompt = good morning

(red spots = residual clown makeup from lora training, not blood)

[languid pebble]

In my early days with A.I. I would have asked myself what the A.I. has been thinking creating these pictures for the prompt 😄

[nimble mason]

hahah

wow, had glowing eyes in the prompt... they are indeed there............

[languid pebble]

Did you use "realistic reflections"?

[nimble mason]

is that a lora?

[languid pebble]

Just a prompt ... found out A.I. often inserts water in the middle of the picture if you add this to a prompt

[nimble mason]

ahh, interesting

[languid pebble]

If you create a city with realistic reflections the street becomes wet or kinda canal

[languid pebble]

tried with SD3 and a desert:

[languid pebble]

SD3 maybe better ...

[steep sleet]

Pixart-Sigma+SDXL+LoRA

[jovial tiger]

These are great

Inspired by the yellow screaming guy

[jovial tiger]

Good morning

These guys are just having the best time

[normal wagon]

https://youtu.be/I8hZyJhhIEU?si=BCnamxpvbM1gTEUB

[steep sleet]

Thanks 😄

[sterile kiln]

when does SD3 arrive?

[late sorrel]

May 20

[steep sleet]

May the MaY-BE™

[languid pebble]

[languid pebble]

Working on my robots instead ^^

[unique condor]

Unless there has been an actual statement from SAI, this is just wishful thinking. Everyone swore up and down it was going to be may 10th until it was May 11th lol...

In reality, it will likely be some time within the next month or two. Basically, whenever they see a large enough dropoff from API revenue. I'd guess more like early June.

[analog goblet]

experimenting with my new workflow 🙂

[languid pebble]

I like to have an eye on it ...

[late sorrel]

yea i started the whole May 10 thingy i think :3 haha

[late sorrel]

is there a shower cam too?

[languid pebble]

Not sure Aliens need a shower ^^

[maiden hawk]

the gap between midjourney and stable diffusion is crazy

[analog goblet]

is it ?

[maiden hawk]

@maiden hawkhi

[analog goblet]

i havent looked into mj for a while, you got some good examples ?

[late sorrel]

the gap is crazy indeed my friend, for example midjourney doesnt have controlnet, ipadapter, total control of the inference pipeline, i mean i could continue, but the gap is already wide 🙂

[languid pebble]

But looks like they don't need ...

[analog goblet]

any of you made some great mj images lately ?

[languid pebble]

No 🙂

[analog goblet]

sad, i would like to know if quality really is different

[pallid ruin]

It depends, sdxl can have specifig loras so the gap isn´t that much

[analog goblet]

can you recommend some loras ?

[languid pebble]

They manage to give their costumers something they like. Maybe even if it's not 100% what they wanted ...

[pallid ruin]

I have hundreds, idk which one XD

[analog goblet]

abstract/surrealistic art and photography/portraits if you have 🙂

would be cool if you could send me some

[late sorrel]

what's the point of quality anyway if you cant have some sort of control over the image, like you can with stable diffusion models. i personally never understood why people go for those other models instead of stable diffusion models. i guess mj and dalle are "sufficient" for whatever those people are using image generators for... but i personally find it very limiting idk

[maiden hawk]

[analog goblet]

i find stable diffusion very limiting, too, to be honest

[languid pebble]

You can join their Discord Channel and watch ....

[maiden hawk]

ai isnt meant to be understood 🙂

[late sorrel]

well sure, but i feel like you have way less options with those other ones, no?

[analog goblet]

i would say about the same

i'm trying really hard to produce new kinds of art, but its complicated/impossible

you can tell by a simple look at an image if it's a photography, mj or sd image

[late sorrel]

not to mention those other ones cost money, so idk how you juggle with that too

[empty jolt]

I recently reinstalled stable diffusion and it's giving me this error. Before formatting the PC and reinstalling it, it generated images normally, can anyone help me?

[pallid ruin]

That´s correct but mostly because everyone tries to do "perfect" images with perfect lighting and such things

[analog goblet]

no, it's a style thing, too

[pallid ruin]

This image has errors, of course, but look at the lighting

[analog goblet]

like this here screams "i m stable diffusion"

[pallid ruin]

Well that´s correct

[pallid ruin]

That says mj

wait no

[analog goblet]

to me it says sd

it actually screams sd to me, too

the kind of how the faces/proportions are distorted

[pallid ruin]

I looked withouth opening the image, yeah its sd

[analog goblet]

the flowers, the colors, everything

[pallid ruin]

But midjourney has that type of blurry around the corners

in all the images

[late sorrel]

i actually never used mj and dalle, do they also produce bokeh pics?

[languid pebble]

Mine was SD3 🙂

[pallid ruin]

👀

[pallid ruin]

I hope it gets released locally @languid pebble 😭

[analog goblet]

if this was sd3 then i m disappointed by sd3

[languid pebble]

Learned to like stuff like that cause of working with A.I. 🙂

[pallid ruin]

You can´t just compare a base model of sd against midjourney or dalle

Community make the best finetunes

[languid pebble]

Beta ...

[analog goblet]

screams sd, too (at least to me)

[languid pebble]

But cool!

[analog goblet]

i sometimes cant even really put into words why, but after seeing 100k+ images it kinda is obvious

as hard as i try to break out of it

[late sorrel]

does mj or dalle have better human hands?

[analog goblet]

those skulls are 100% sd, too, and it's exactly what you get if you prompt "surrealistic"

i'm pushing sd as hard as i can, but it's staying within some boundaries

[languid pebble]

MJ 🙂

[pallid ruin]

That 4 fingers 🤣

[languid pebble]

I don't know the prompt ... have been searching for the first hand i could find 🙂

[languid pebble]

Must be a Simpson ^^

[late sorrel]

kek

[languid pebble]

That's ComfyUI 😄

[pallid ruin]

Do you have the prompt of that image? It looks nice

And yes sd has some limitations, lets hope that with SD3 we can get closer to the closed source alternatives

[unique condor]

Yeah that's the boring reality lora

[south temple]

Just….. out in the hood…

Thump thuuuuuuMP

[wispy nest]

@wispy nest

Does that make sense?

or no?

[opal vessel]

L'image pourrait représenter une personne en train de s'entraîner, peut-être en train de courir, de faire des étirements ou de soulever des poids, avec une expression faciale montrant qu'elle ressent de la douleur. Derrière cette personne, on pourrait voir deux chemins divergents : l'un représentant la continuation de l'entraînement malgré la douleur, avec des panneaux indiquant des stratégies d'adaptation et de gestion de la douleur, comme "Écoute ton corps", "Modifie ton mouvement", "Consulte un professionnel", etc. L'autre chemin pourrait représenter l'arrêt de l'entraînement, avec des panneaux indiquant des raisons de repos obligatoire, comme "Blessure nécessitant du repos", "Cicatrisation nécessaire", etc.

[past pelican]

Lmao 🤣

[analog goblet]

@pallid ruin i dont really work with prompts anymore

[clever oar]

you save your money in hood?

i have not ideal english this man try buy gpu in hood?

[loud loom]

Dune Sci-Fi inspired blonde woman

[wispy nest]

I be rolling up on the ops with my gat, bro

[clever oar]

but you may not be the only one with a gun there...

[wispy nest]

Damn straight all my soldiers strapped

[clever oar]

count the money so that there is enough for the GPU

[wispy nest]

I just gave that n---- six dollars

(this is an American meme)

[clever oar]

😎

[still rose]

@dry crow I can make these without it going "unable to allocate vram" or whatever

[bold briar]

Hello! Could someone please help me figure out how to create a comic for my game NPC in ComfyUI, ensuring it has a consistent style and the characters resemble those in the game? For example, I've used ControlNET TTPlanet SDXL, which effectively creates similar NPC (see the second photo). However, I'm struggling with adjusting their poses, compositions, and changing the background. Does anyone have ideas on how I could accomplish this?

[dry crow]

Great to hear 🙂

[clever oar]

cs10 changed room 😄

[still rose]

I will also do the cat prompt soon

[dry crow]

Np, take your time

[tough flame]

@wispy nest The middle house (red brick with white accents)

this is the base

So I want to replace the middel house with the following reference

So I want to place this steel-perforated facade in the base image

Therefore my question, can I inpaint the steel-facade and link it to my base image where I inpaint the building I'd like to replace.

Photoshop can do the trick, but was wondering if SD had the capabilities

Yea it does the trick

Might be a good plugin for SD

[shut sinew]

Please use #🤝｜tech-support for any tech related questions and for help

[wispy nest]

Ok black guy

[shut sinew]

Also please dont upload random videos here, but if you have to, use #🌶｜off-topic

[shut sinew]

No, you dont

[junior sky]

Uhm....

@shut sinew He's gone, right?

[languid pebble]

Gone with the wind ...

[clever oar]

Mojo

I bought croissants

Do they eat this in Germany?

[atomic nymph]

Hi guys, where can I ask about recognizing art styles? 😄

[languid pebble]

Moin discordos! Sure we do! 😄

[languid pebble]

Working on a musicvideo again 😄

Katzengras Song 😄

[steel apex]

Skyrim in the style of Breath Of The Wild

Link!

[pallid ruin]

Nice!

[pallid ruin]

Are you using an inpainting model? 👀

You should use juggernaut X inpainting version for that 😁 (or any other inpainting model, but there are not many)

[late sorrel]

there is also something called lama remover, its like an inpainting model that can replace/remove objects in a pic and it usually replaces them with the background logically, but i started using it recently, so there are prob ways to make it better idk. it's not a bad tool to have as well in some situations.

oh im not saying not to use jugg haha, i was just saying there is another tool that could help maybe in some scenarios idk :3 i personally like to have as many cool tools as i can get for all sorts of things

and then experiment

there is the classic thing you know, the student surpasses the master haha

[gaunt ember]

Technology background picture

[late sorrel]

so wait you were not using the inpainting version of jugg X ?

[nimble mason]

the inpainting models with sdxl drive me kinda nuts

they're almost great... but not imo

they still leave weird seams etc, even if it's in the form of structure of lighting etc

the sd15 ones are really good

the sdxl ones... almost... but just a bit short

[late sorrel]

i wanna see if that brushnet improves results tomorrrow

[nimble mason]

ooo, very interesting, i will try tomorrow too

[late sorrel]

because based on the pics i saw, it seems to give nice results

[nimble mango]

Good morning,
I tried to install reactor on stable diffusion but it doesn't work, can someone help me please?

[noble zephyr]

well, for one, uninstall roop. you don't need it if you're going to be using reactor. secondly, if i recall correctly (i haven't used reactor in a while), it should be under the "scripts" dropdown.

ah, i see you're already getting help in tech-support.

[unique condor]

Just use differential diffusion inpainting and get amazing results with any sdxl model. Make sure to blur the masks when you do it. Since the mask gradients are where the magic is done and use enough steps.

[wispy nest]

guys which is best local SD or Paid SD?

i mean generative SD or midjourney or DALL.E?

[unique condor]

Here's a super quick example of using differential diffusion with a regular model. You really don't need inpainting models anymore.

I generated the image with the same model, so that's why you see the set latent noise mask node there, vs the usual inpaintmodelconditioning node. Just tweaked the same prompt a little to turn the face into a robot. And you don't need the PAG node, I was just testing it out to see if it broke inpainting or not, seemed to still work great

[nimble mason]

Yeah, that's what I've settled on doing

Sd15 there's no need as the inpainting is ph.nomenal

Sdxl... Just not quite there

[wispy nest]

guys can u use SD in mac or ipad pro m4?

[unique condor]

dif dif works with 1.5 as well and is as good, if not better than 1.5 inpainting models. as far as i know, regular inpainting models still only do binary masking under the hood. there are some hacky ways around it, but they aren't perfect and you'll still usually end up with seams or blurred seams. meanwhile dif dif slowly grows/spreads the mask based on the gradient and applies more/less denoise based on mask intensity, so it can feather itself in almost seamlessly into any image.

[nimble mason]

yeah it is really good at that

i haven't tried it with sd15 tbh

been pretty satisfied with it with sdxl

using ipadapter really helps with it with sdxl too... give it some context of what's going on in the image as a whole

[unique condor]

within the first day of using differential diffusion, i deleted all my inpainting models... and yeah, things like ipa can help even more, but the same is true with any workflow really

[late sorrel]

can you send me a very simple workflow that has the differential diffusion inpainting stuff? i would love to try it out :3

[unique condor]

Yeah sure, give me a few minutes, I'm walking the dog

[late sorrel]

cool ty

[unique condor]

And I just realized that in my example above, I had denoise at a full 100% and it still handled the transitions seamlessly

[nimble mason]

literally all you need is the set latent noise mask mode with a gradient mask of some kind

and to put the differential difusion node in the middle of the model pipeline, so checkpoint loader unet -> diff diff -> ksampler

that's the simplest implementation

[unique condor]

But having a mask blur is absolutely required

[nimble mason]

but diff diff does mess with the image as a whole, it's subtle but it will add up, it's a good idea to use latent composite mask to then take what you changed and paste it back onto the original latent

[unique condor]

The mask has to have gradients otherwise, it's just doing regular inpainting

[nimble mason]

even unmasked areas are affected so i do that every time

i'll drop a workflow on that in a min here

[unique condor]

It's minor and not much different from when you vae encode and then decode an image

[unique condor]

But yeah if you're doing a multipass workflow, you want it to stay as a latent the whole time and avoid repeated img>latent>image>latent stuff, but that goes for ALL multiples workflows

[late sorrel]

btw i have a question if anyone tried/knows how to do this: lets say i want to inpaint or imprint a tattoo on a person's body somewhere, but i dont want the tattoo to be random (text), i want it to be based on maybe like a picture of something, is that possible? but the pic has to be i guess transformed to "look like" an actual tattoo on the body, not just like a pic on the body, if you know what i mean? is there any way to do that, would be kinda cool

[nimble mason]

yep, it is indeed minor, but the degradation is real so i just do it outta habit

[dense nova]

Got to add apiflow for this. (But comfy's api is so evil... 😭)

[dense nova]

I have theories I've been pondering, but never tested. Basically you need depth -> mesh, take it into blender, uv unwrap preserving edge lengths, paste the tattoo onto the texture, render w/ a transparent background, photoshop that render onto the original image, then inpaint that with "tattoo" prompt and low denoise to get lighting / color right.

I should be able to automate this into a tool in ParrotLUX, but not anytime soon. Few months.

[nimble mason]

i know how to do this

doesn't even need SD tbh to get pretty close

you want to do frequency separation

[late sorrel]

teach me sensei 🙂

[dense nova]

No in order to wrap the tattoo along the shape of muscles / arm curvature. Not easy.

[nimble mason]

ohh, yeah for that you'd need more for sure

[dense nova]

Forehead tattoo? Chest tattoo? Those could work.

[nimble mason]

but freq sep could at least get you started, maybe a few iterative light denoise cycles could do it

the linear light/gaussian blur method is okay, that's what everyone uses

much better in most cases imo is my hard light/median blur method

[late sorrel]

well i dont necessarily want a huge tattoo, just something i can put on a random area of the body, doesnt have to be big at all

[dense nova]

As long as there's no curve changing, yeah just edit the skin color in photoshop / gimp / krita.
Select area, feather 1.25 pixels, lower saturation, darken, and shift hue toward green (seems most common black tattoo ink).

[late sorrel]

but i wanted to see if there is a way to do it all inside of comfy :3

[dense nova]

Sure, just grab yourself a python ide and start writing a custom node. 😛 (Might be stuff out there idk.)

[late sorrel]

damn it

maybe someone out there has something for this case already, just have to find it

[dense nova]

If google actually worked, the internet would be amazing.

[late sorrel]

yea google sometimes either doesnt find what im looking for at all, or it gives very weird results

heck even youtube doesnt give the results im searching

[unique condor]

Here, try this. I tried to comment it out a little, but it should get you started

[late sorrel]

ty il try it later, busy now ❤️

[unique condor]

for real though, i need to make a custom node to replace the preview node so that it has an image and mask output to it. one of the only annoying things about doing inpainting where you stay in latent space. you have to copy clipspace from preview images, paste it into a load image node and then mask in that, even though you don't use the image output pin

or i need to make a load image node with image and mask input pins, probably easier to do it there

scratch all that, im a dumbass, impact pack has a preview bridge node... and even a latent preview bridge node...

[unique condor]

I made a few changes, use this one instead

[late sorrel]

kk

is the frog gonna eat that ball? :3

[unique condor]

it was originally a cat, then i turned it into a frog

[late sorrel]

haha

[unique condor]

this was the cat that i generated first, then masked out and turned into the frog. so as you can see, differential diffusion did a really good job of filling in the blanks behind

[late sorrel]

indeed

[orchid schooner]

https://tenor.com/view/multipass-freepass-gif-12760477

[late sorrel]

that is some inpainted hair 🙂

[unique condor]

an just as an fyi for those who do a lot of multipasses and that want to keep things in latent space, the impact pack's preview bridge is a godsend. you can do your masking there and it outputs the latent and mask. i really do feel like an idiot for doing things the pain in the ass preview copy/paste clipspace way for so long

[languid pebble]

I need a multipass ^^

[clever oar]

mojo

today I bought fish with pepper

[languid pebble]

Jummy ... I went to my fav. Restaurant from my past ...

[clever oar]

I need to cook something for dinner and I'm lazy

[languid pebble]

I know that 🙂 Like to see Beta of my last video?

[clever oar]

ok

[languid pebble]

I'll add you as friend ... second ...

[clever oar]

yesterday a man with a suitcase of money wandered into a bad area and disappeared 😃

[languid pebble]

And today you are rich with a 5090?

[late sorrel]

i just tried to import, it doesnt give me the workflow, can you send the json?

[unique condor]

open it in browser

then click and drag the image into the comfyui tab

when you click on the image in discord, click the open in browser button*

[late sorrel]

hmm no it doesnt want to load even that way

[unique condor]

lemme check

[languid pebble]

Good nite!

Not sure whether Discord deletes the Metainformation ...

[late sorrel]

yea maybe

but then again i remember some workflows were available from here, so idk

[unique condor]

it usually doesnt, weird that its doing it right now. maybe i only pasted in the copied preview or something, well crap, lemme see if i can find the workflow again

[late sorrel]

i even checked on an online image metadata viewer, the pic doesnt have the metadata indeed, it was stripped

[unique condor]

yeah i tried it as well and it didnt work

[late sorrel]

😦

[unique condor]

ill whip it up real quick, but without comments and being all tidy and stuff lol

should only take like 2 min

[late sorrel]

it's ok haha

ty

i have to eat, but just ping me with it

[unique condor]

try this, and jfc trying to use the basic example image that comes with comfyui is kind of a nightmare lol (the bunny eared mspaint looking one with the pink dress and huge yellow ears)

but i put two workflows in there, one for if you're going from ksampler to ksampler and one for if you just want to inpaint on an already saved image

play with the kernel and sigma values on the blur nodes and keep your step count high for inpainting. i usually just use 50 steps because diminishing returns starts to kick in

[late sorrel]

ok i got it working, ty 🙂

it does give me better results

now im gonna try some BrushNet stuff

[unique condor]

Also remember that you don't want pixel perfect masks. Be somewhat sloppy with it. It needs some local context to know what to blend into

[devout barn]

guys I'm new at this and making a good bit of progress but.

I think my genius idea of just taking a photo of my own hands for the poses did not work...

What are my mistakes

my first thoughts are less opacity on the green & more outlines of the hands

but I'm sure there's more

Also in the first image, the stage left hand is mostly perfect but I want its extended straight out, not going forward towards the camera

[nimble mason]

what are you trying to do?

your hands are absurdly huge for that image

[late sorrel]

yea that makes sense

[nimble mason]

and not even at a sensible angle

[devout barn]

I figured that because the original image started with that the pose was based in iwas so low quality and it did a pretty ok job, that doing the same thing with higher quality camera would be better

And I only made the hands as good as I could perfect in other iterations too

Lemme find

[nimble mason]

if your hands are ever that size relative to your arms, you need an ambulance

also, the left hand is attached to the wrist by the knuckle

and the thumb is on the far side, like if your wrist attached to your hand from the middle of your pinkie

[late sorrel]

are you using open pose controlnet?

[devout barn]

Ye Ill try making the hands smaller but I feel like im missing more

Idk what that is so no

I looked in citivai but could not find what im looking for

[late sorrel]

so what are you currently doing then, just img2img?

[nimble mason]

yeah seriously think about the angles you have there

first, you gotta think about anatomy and geometry

[devout barn]

inpainting at this stage. Started with img2img with prompts

[nimble mason]

if you want to go with this approach, doing it manually, try seeing if you can make a photoshop that doesn't look like the guy fell out of a plane and landed on his wrists

if you can get it to look reasonable

[devout barn]

I get it lol

[nimble mason]

then you have something you could take into openpose

lol

sorry it's just hilarious looking haha

[brazen sand]

I think the problem there in the final image I think is the denoising strength or the inpainting, that's why you are getting something like extra stuff

[late sorrel]

first and foremost use a higher quality pic :3

[brazen sand]

also as some stated, the hands are posed kind of off, but it shoul work in some way

[devout barn]

The hell. How did you do that? Just ai upscaling?

[late sorrel]

no i did reverse image search and found it

[devout barn]

O cool

[late sorrel]

so use that instead

[polar shadow]

#👥｜roles

[sinful moss]

I want to make some realistic photos I have tried several times but can't make hyper-realistic images but true-AI images

[languid pebble]

On Katzengras? 😄

[clever oar]

dont know what is it😃

[languid pebble]

It's just kinda grass some cats react like they are on drugs. They start to purr and meow, roll around in it and lick it up.

[clever oar]

cat mint

[languid pebble]

Yes! I think that's how you call it. Fun fact: we call it grass just like weed ...

[past pelican]

Catnip

[clever oar]

wanna people mint?

[languid pebble]

Not really a fan 🙂

[languid pebble]

Albino frogs should stay out of the sun ^^

[clever oar]

beneficial symbiosis

[languid pebble]

And it doesn't look like you wanna eat it. So protection from enemies, too ^^

[clever oar]

after 100 years symbiosis

[languid pebble]

lol

Thanks for making me laugh ... lost 3 seconds upscaling videos and now nothing fits any longer 😄

[clever oar]

evolution shouldn't be funny!

[languid pebble]

Sure ... blondes will extinct ... 😦

[clever oar]

oh, now I wish I could meet a natural blonde

[languid pebble]

Not natural ones can even be cooler 🙂

[clever oar]

girls always need to paint something, they just love sd but don’t know about it

[languid pebble]

They can paint my apartement 😄

What have I done? 😄

[languid pebble]

Looking for a new car?

[sterile kiln]

I can currently afford to buy a wheelbarrow, maybe...

[languid pebble]

Always good to have 🙂

[sterile kiln]

not in an apartment xD

[languid pebble]

You could use it as TV chair ^^

[sterile kiln]

already done on a building site, during work

[languid pebble]

😄

That's what you call work? So so ... 🙂

[sterile kiln]

the life of a slave sometimes has its little moments of joy

[languid pebble]

So better enjoy them ...

[clever oar]

upscale
breast😎

[languid pebble]

😛

[nimble mason]

hahah what the hell was that

[clever oar]

is special guest for you😃

[clever oar]

stop it i need sleep soon

[nimble mason]

i find it helps me sleep

[clever oar]

sharks will be offended

[nimble mason]

will u marry me

[clever oar]

bruxa

from witcher

gift for children

[nimble mason]

Oil painting style, a detailed black kitchen, set against a dark, textured background with subtle brush strokes, soft shadows highlighting the delicate curves of the counters, and deep black granite with hints of dark veins, creating a dramatic and moody atmosphere with rich contrasts between the appliances and the background.

high quality, sharp

base on the left, base + multi style lora i just trained on the right

[nimble mason]

fire ain't enough, we need a lil mushroom cloud emoji

[nimble mason]

cgrazy cool shit

[nimble mason]

just another day at the office

[pallid ruin]

👏 amazing images

im stealing your prompt btw 🤭

[nimble mason]

all is meant to be shared 🙂

[nimble mason]

black hole

[opal heart]

hi

I have some images gene bu ai

but it is not available

[nimble mason]

loving this lora

very versatile

[past pelican]

Porny Burger

[languid pebble]

😂

[fallow marlin]

Hi guys I trained LORA for consistent character and after training and using.
After it Lora works, it returns my character, but performance of juggernautXL with applied LORA worse. Why so?
I trained with 28 img (1024x1024) with 10 epochs and 10 repeats. LR 5e-4, I used LoRA-LierLa with network_dim = 128 and network_alpha = 96.
Is it because I used wrong params or what?

[crisp stream]

The contours are sweet

[wispy nest]

pony XL

[upper sierra]

\image ago

[crisp stream]

ty 🙂

[idle spruce]

Does anyone know why aa1111 always puts my eyes out with this tone?

[crisp stream]

I don't really think it's A1111 😉

[idle spruce]

what?

[crisp stream]

no

[idle spruce]

Just ask if anyone knows what it could be due to, I don't understand what you said.

[crisp stream]

what didn't you understand?

[idle spruce]

It wasn't a1111, if it isn't the program I don't know what it could be, it happened to me on all the models

[crisp stream]

then I at least assume it's prompt related.

[late sorrel]

i like the colors

[nimble mason]

upscaled

[slate sedge]

Anyone have any workflows to make GPT4o images more realistic looking

[tacit cipher]

Is stable diffusion automatic 1111 web ui still able to be used as an api? For some reason when I’m hitting the docs route I can’t see the option for inpainting or im2img

[crisp stream]

@lost valve would you mind demonstrating sd 1.5 being superior to SDXL in terms of photorealism?

[nocturne oak]

lul

[crisp stream]

Could you please simply post an SD 1.5 image that you think is superior to SDXL in photorealistic terms?

[tacit cipher]

Is stable diffusion automatic 1111 web ui still able to be used as an api? For some reason when I’m hitting the docs route I can’t see the option for inpainting or im2img

[crisp stream]

@lost valve ⬆️ ⬆️

[rain gazelle]

dang, can't get a preview to show up... hmm

will try a rescale...

there we go.... which one of these is the best? 🙂

[cedar perch]

Hey everyone I am new to stable diffusion and would like to create a sort of comic strip maintaining the same character throught. Is this possible? For example if the initial prompt was 2 dogs walking on the street could I make those exact same 2 dogs in a different setting and pose like eating inside?

thanks for the help

[rain gazelle]

I think I like 0.3, 0.35, or 0.4.

[solemn pier]

could someone generate a shirt for this stock image for me?

[late sorrel]

kek

[nimble mason]

a witch and her alligator go for a walk

and bat

[nimble mason]

that makes no sense

you can just delete metadata yourself lol and really, is this like some kind of mob trial? jurors with paper bags over their heads to protect their identities, holding up AI generated images with garbled text to make statements without giving away the sound of their voice

[past pelican]

All your art has a very similar artstyle. Is there a specific style or artist you are prompting for?

[nimble mason]

nope

trained my own lora on my fav outputs over the last month or two

so it's CSBW_Style pretty much

[past pelican]

I should mess around with making my own Loras more.

[nimble mason]

onetrainer is pretty easy to use

and the captions i'm getting with gpt4o are crazy good

Act as a creative agent who generates a pair of highly creative detailed 70 word image prompts derived from each uploaded image. Begin each prompt by specifying an artistic style (such as a type of painting or illustration or drawing, etc., for example, oil or acrylic or watercolor painting, or charcoal drawing or engraving or sculpture, etc.) or camera settings, using descriptive elements that explain what makes the artistic style or camera settings unique. Then, continue with descriptive visual elements of the subject, surroundings, and background. Please do not omit any details, especially objects, lighting, colors, and textures, even subtle ones hidden in the background! It is imperative that you carefully consider every aspect of the image, and provide satisfactory descriptions! Also, please be very careful when considering whether the image really is a photograph or actually a detailed painting or illustration. Include elements presented as discrete descriptors in one long 70 word sentence without story. Put the most important descriptive elements at the beginning of the sentence. Only respond with the 70 word prompt, don't preface the prompt with anything else. Use simple language. Limit your response to one sentence. Each prompt should be 70 words long. Please create a pair of 70 word image prompts for each uploaded image.

prompt it with that and drop 6 images in there (maybe more, that's the most i've tried so far)

via the web interface

i was able to smash through over 200 images yesterday before it stopped me for a few hours

[blissful cairn]

The transition.

[nimble mason]

the photorealism is not better than sdxl................

there's dead giveaways in all of those lol

[unique condor]

Bro you posted some other pics with that Asian woman's face like a week ago and seeing these two just now made me immediately recognize them. I don't even need to search to verify

But I'll double check for sanity's sake

#🆕｜sd3 message

https://cdn.discordapp.com/attachments/1230206273451069540/1238532428738596935/Blown_away_by_her_features.png?ex=66483275&is=6646e0f5&hm=d37a61907eb8c5994b22f7ce4ba04e3a995011a2ce73770d18d09549fb9ed1e5&

those ones

and no, they aren't really all that realistic, they look pretty deep in the uncanny valley

most 1.5 models are overtrained on limited datasets of whoever or whatever the trainer was simping over at the time. and the law of averages wouldn't converge an "asian woman's" face into that shape. you'd be surprised how little variety they actually use to train these kinds of models.

[strange mortar]

is this automatic 1111 or comfyui run ?

[unique condor]

Like this?

i'll make a bunch more with other ethnicities

not even using controlnets or PAG either lol

and the lora is clip strength only

not even using the model portion of it

[strange mortar]

@unique condor any idea

[unique condor]

keeping the seed the same while I change ethnicities

here's another

and a nutha

want me to keep going or do you get the point now that SDXL models are far superior in detail now. Again, I'm not even using the lora model, only the clip portion(doesn't really change much anyways, just makes things a little more reliably boring). No special sampling tricks, 30 steps and a basic profile for automaticcfg that helps a tiny bit.

[strange mortar]

can you pls see the above tag and let me know if its done on comfyui or automatic 1111 like the site he is using

[unique condor]

not watching a video at this hour

[unique condor]

Bro I didn't even save them because I have no interest in making people. I guess I could have added in some tags for washed out image or lowered the cfg to drop the contrast more, to make them match your 1.5 generations more. But you can replicate the workflow yourself

And spend time around real women, or people for that matter, you'll see that most people have natural oil to their skin that reflects light.

It's not a straw man when I'm literally showing the proof lol

Reproducible proof at that

Minus my custom automaticcfg profile

Oh and it's not just reproducible proof, it's reproducible one-shot proof with zero cherry picking and zero editing. Just a single sampler workflow with nothing extra done to it after generation.

[terse fulcrum]

fellas I need sollution for this

[nimble mason]

i'm talking about the images themselves

[foggy dagger]

I found this pic randomly on Google, i want to modify and make his hair short so eyebrows are visible and remove the lightning on left 100% , iam new to this and my normal text promt dosent seem to work, pls help

[junior sky]

Is this photoreal? Never mind his lazy eye, he has a health condition.

[terse fulcrum]

Got the solution

Downloaded the CUDA from NVIDIA

and Bam it worked

[frank wolf]

Yooooooooo anyone else getting to try out chatgpt4??

[frank wolf]

both eyes are very sus

[junior sky]

Yeah, i think he is up to no good

[cedar perch]

Hey everyone I am new to stable diffusion and would like to create a sort of comic strip maintaining the same character throught. Is this possible? For example if the initial prompt was 2 dogs walking on the street could I make those exact same 2 dogs in a different setting and pose like eating inside?

[jovial tiger]

https://youtu.be/6i417F-g37s?si=8tEJRfqpWmpPBaH4

[cedar perch]

oh ok thank you so much I was trying to do it on dream studio ai lol

[nova copper]

Hi, trying to replicate something like that "Stack of timber 4x4 on white background" and all I getting is close up of wood texture. I am using SDXL with Refiner. Did anyone try something similar and it is willing to help me out?

[crisp stream]

Like said, not into photorealism, simply wanted to see what you can do, so I got an idea and be able to compare with SDXL stuff I can find (including loRA/checkpoints)

Thank you for posting btw

[crisp stream]

Are you referring to the glossy 3Dish style as well? Asking ecause that's resulting from the use of IPAdapter+, regardless of what you prompt or use modelwise

[junior sky]

Not necessarily. Here is a stack of sandwiches in the style of the timber blocks from above. Made with IPAdapter+

[crisp stream]

Yes, true, forgot to mention it´s either the 3d-ish and/or the photoreal look, then I noticed I wouldn´t produce anything flat looking but at least inhabiting a 3D appearance. If you got tricks up your sleeve going around that issue, I´m all ears 😄
Btw I was about to generate examples of what I am referring to exactly, then IPA+ seems to not work anymore here (possible have to update?). Gonna check and come back with examples if I manage to get it working again

ah, it´s working, just thought it wouldn´t, took a while to load the model/nodes (?)

[junior sky]

Do you mean this with flat?

[crisp stream]

currently experimenting, just a moment 🙂

[junior sky]

No problem

[wispy nest]

#▶｜stable-video-diffusion message

[dark robin]

In A1111, is there a way to give the AI a reference photo to pull the art style from? I know there's controlnet but I'm not sure that's what I'm looking for. I just want it to look at a photo and generate something similar to that photo. Thanks.

[crisp stream]

Just to give an idea, what I am talking about:

Original:

IPA+:

From what I was told it´s because of the VAE-thing of SDXL and it would change with SD3. Then I´m no expert, so it´s simply what I read

@junior sky Oh and to say it isn´t producing anything flat looking isn´t really correct in that sense, it´s more like it´s adding its touch on top that can still result in flat looking stuff yet often results in the effect shown

[junior sky]

I get what you meant now. You are right with that, yes.

[junior sky]

Solved it 😅

[crisp stream]

😄

What did you do btw?

[junior sky]

To get this?

[crisp stream]

yep

[junior sky]

I am feeding 2 images into a node i coded. The new image goes into the IPA as embedding

[crisp stream]

Aaaaah 😄

[mild thicket]

From what I was told it´s because of the VAE-thing of SDXL and it would change with SD3. Then I´m no expert, so it´s simply what I read

[nimble mason]

would this be anything like what you have in mind?

[languid pebble]

Ahoi 🙂

My first released A.I. song ;-D https://youtu.be/vxLnqgTCloY?si=nur0Pem3XH-j6kfP

[crisp stream]

yes, that´s sort of the typical look I was referring to

[crisp stream]

like in the sense of what IPA+ tends to look like (not what I want)+

[nimble mason]

i have a lora i've been training on selections of my outputs i could share with ya if that's a look you like

[late sorrel]

is there a clownshark lora? :3

[crisp stream]

like said, it´s basically a look I try to avoid because of the stereotypical appearance. Not a bad look, just occuring too much when using IPA+. Thank you anyway 😉

[nimble mason]

Yeah, not published on civitai or anything but been working on it

[nimble mason]

Gotcha yeah I trained for that as it was a 3D paint look most models struggle on afaik

[nimble mason]

Get me some examples of what you're looking for and maybe I can make a lora

[crisp stream]

I'm fine, thank you 🙂

[nimble mason]

Got an example anyway? Just curious what you have in mind

[crisp stream]

No, I don´t have anything in mind 🙂 Possibly you understood it that way, yet I was merely demonstrating the effect IPA+ tends to have 🙂

[late sorrel]

i have no idea what im looking at, but its cool

[crisp stream]

ty

[rain gazelle]

did we ever get controlnet stuff like posing with xl models?

[late sorrel]

yes

[clever oar]

robot feminist🤭

[crisp stream]

how? 😀

[clever oar]

color😃

[soft meadow]

Hi I have a probelm here

Im using qr code monster , the qr code I make I can scan by using the camera

But when I save the picture, I can't scan anymore when I input the qr code in Instagram as picture

[unique condor]

Try editing it in Photoshop of krita. Boosting the contrast or playing with the midpoint of the levels would probably be the easiest way. You need the dark and light spots to be more distinct. Oh and maybe try adding a bit of sharpening to it as well if that doesn't work

[soft meadow]

Ok tQ

[languid pebble]

Burn mother****** burn!

[jovial tiger]

Your images are always great style templates

[rich badger]

Burn mother** burn!

[crisp stream]

Thank you 😄

[crisp stream]

The first one I like

[clever oar]

its he?

[crisp stream]

could be 🙂

didn´t prompt for anything alike, just came out 🙂

[clever oar]

Trigger words are needed to activate lora if i have <lora name> in promt?

i know its works

but many example prompt have it

[languid pebble]

Moin! 🙂

[clever oar]

[pallid ruin]

cute!

[clever oar]

leob how are you 😎

[pallid ruin]

Im good! 🤗

Its cold here so I left he pc at night training loras 🤣

But idk which one should I train now 🤔

[clever oar]

where cold?

you live in North Pole?

[pallid ruin]

Nah its not that cold, I live in Argentina

about 4°c sometimes

[clever oar]

how is possible argentina is very warm

[pallid ruin]

Its winter now 😁

this is rn

[clever oar]

is not cold 😄

i have 11 C

[pallid ruin]

🥶

[clever oar]

just walked in a T-shirt

It's very hot at home, I needed to cool down

🔥

where you train lora?

[pallid ruin]

On my PC using kohya_ss

[clever oar]

i try install it too but fail

[nimble mason]

OneTrainer

[humble oracle]

I train a SD-related model on generating the MRI images in 15 hrs atm

I think just finish 30% of the training section

[tacit haven]

Oil painting with influences from Byzantine frescoes illustrating a cheerful monkey expertly surfing a massive wave. In the background, an old fishing boat

[lavish latch]

Campus Graduation Song Party

[golden pebble]

hey guys, how is it possible to do this exact quality?

i guess it's made with comfy-ui

[pale bison]

sun

[fierce ravine]

po

[late sorrel]

quality? i see two great qualities

[scenic mulch]

draw a logo which named River Tree，it is a company name which operating nonwoven machines, use green and black color

[eternal arch]

When your promt gives you this body.

But then gives you this face

@dry crow Got any recommendations for face fixing stuff? xD

[vagrant dust]

FaceRestore can be achived on different ways. Either use a restore model like codeformer etc., try inpainting or use a upscale with face restoration capabilities

[eternal arch]

Ran ADetailer through a few times, and then upscaled the img by 2. Holy crap, the change and improvement

[vagrant dust]

Used the image and added a tile controlnet on it. Guess the strength is a bit high 🙂

[eternal arch]

Nah man. Me like. What stuff u used for that?

Hair looks nicer too

[vagrant dust]

Well i loaded you image, added a generic prompt "anime woman standing in ruins" then used the controlnet tile model with 0.7 strength and animerge model.

[eternal arch]

Wat is dis controlnet u r talking boot?

[vagrant dust]

This btw. is the same setup but with the euler Scheduler instead of the dpm2m it keeps the soft rendering

[eternal arch]

OOh

I had one particular model i was working on, where Euler made it more... orange, so idk. I've stuck to dpm for sum reason

[vagrant dust]

Control-Nets are a set of models to take more "control" of the generation of the image. You could use depth maps to give informations about how near and how far items are. Or OpenPose how the person should stand, or give a sketch ...

[eternal arch]

Also. Euler Scheduler?

[vagrant dust]

Second one was euler. First one dpmpp_2m

[eternal arch]

Or, Sampling methods

Oh*

I.. keyboard.. halp

Can't bloody type

[vagrant dust]

Yes sampling for the second pass i made. So he goes thru the image and tries to add details but the rendering is decided again by model sampling etc.

[eternal arch]

Toss me some links? Me being new to this still, has no idea of sourcing stuff, except civit

[vagrant dust]

https://github.com/lllyasviel/ControlNet-v1-1-nightly
and then scroll a bit down, they explain all the control net models with examples.

[eternal arch]

Standby. Reloading UI

OOooh. Is in beta for Auto1111 UI (https://github.com/Mikubill/sd-webui-controlnet.git)

Was wondering why i couldn't find it

Well fudge. I already copied your link in xD

[dry crow]

Adetailer extension

[eternal arch]

Got that one ❤️

#🏞｜general-with-images message

[crisp stream]

You can as well simply take the face only, scale it up (in Photoshop for example) and then use it for an input image along the original prompt. Paste the result onto the original in PS and scale it down to the original size. @eternal arch

[eternal arch]

I ain't got Photoshop OR gimp, nor do i understand any of those programs 🙂

[crisp stream]

http://photopea.com

Basically PS CS6 online.

It´s actually pretty easy

[eternal arch]

Upscaling is fine via the webui. Less complex

[crisp stream]

for the given task

[eternal arch]

And less steps

[crisp stream]

yes, sure, that could be done as well

then you would have to mix them afterwards nonetheless

and scale the result down that for

[eternal arch]

I mean. atm i can't even launch the webui, soo. Not doing any getting for some time 😦

[crisp stream]

well 🙂

[eternal arch]

The resident Tech Support Guru is afk atm, it seems too just waiting for him to come back 😄

Soi kinda just been scrolling down the lora/check/embed list on civit for the past 2 hrs

Think I've looked at 3000 packls to far

[crisp stream]

;-D

[eternal arch]

RIGHT BOYO! I got my UI back. Where did we leave off?

[eternal arch]

What was the settings you used to obtain this?

Don't say u went to bed 😐

Is only 1 am at my local timezone

[surreal heart]

I'm experimenting with stacking Ksamplers in comfyui ( first day using it)

I am having a lot of fun!

[nimble mason]

great! many of us (myself included) embed our workflows in the images we post here, feel free to borrow workflows 🙂

[lusty moon]

https://tenor.com/bjFCx.gif

[vagrant dust]

Well borrow your workflows.... Let me think about it - no 🙂

i usally try to keep the amount of nodes below the mark of 1.000

[languid pebble]

😄

[unique condor]

Bruh... I know that workflow... I don't even have to see the name in your reply to immediately spot it. Reminds me of my workspaces, which is why I rarely share them unless I make a new specific one for an example

His sigma cluster in the top right is unmistakable lol

[limber tide]

/dream: large orthodox cross in black outline, white background, the inscription “ ТАРТЮФ» in a cross in thin black letters, minimalism, purple shade

[sterile kiln]

SD 3 is out?

https://tenor.com/view/happy-excited-dog-dance-sitting-gif-8923590

[surreal heart]

You mean the end resulting image? 😮

[vagrant dust]

No he mean the workflow to create the image. Confyui by default keeps the workflow stored within the png file. So you can just drop it into the confyui browser tab an see the complete workflow, seeds, models, ….

[surreal heart]

That's cool! Having the information directly in the png preview of the flow

[nimble mason]

You have to click open in browser just fyi

The preview from discord doesn't have the info, yo u need the full image

[surreal heart]

10-4!

[nimble mason]

yeah there's some custom nodes of my own on there

including the modified res_momentumized sampler that takes itas as a scheduled input

glad to share with anyone that wants them

have been too lazy to set up a github

unzip in ComfyUI/custom_nodes and you're good to go

same thing here, python file included in case you already have that node pack, just overwrite with it if you do... both zips contain a backup of the originals

[late sorrel]

that is an AI generated workflow 🙂

[junior sky]

The workflow from Clownsharks image has only 273 nodes. Here is the top 10:
Node: CLIPTextEncode, Count: 42
Node: KarrasScheduler, Count: 27
Node: Sigmas Concat, Count: 21
Node: Multiply sigmas, Count: 19
Node: ImpactInt, Count: 16
Node: LoadImage, Count: 14
Node: Sigmas Truncate, Count: 12
Node: PreviewImage, Count: 10
Node: PolyexponentialScheduler, Count: 10
Node: FlipSigmas, Count: 9

[nimble mason]

my standards are slipping

[nimble mason]

what's the full list?

perhaps my work will look more complete with variety, not repetition

[junior sky]

You lowered your Load Image count and ramped up the CLIPTextEncode nodes

[nimble mason]

hahaha

only 37 scheduler nodes... i'm disappointed in myself

[junior sky]

You have 52 unique node types in the wf

[nimble mason]

excellent

[junior sky]

I think you could get rid of around 40 CLIPTextEncode nodes 🤣