✨｜sdxl

[hardy cipher]

I've considered that myself actually

but not sure how it'd work within the graphical interface

[ionic gulch]

i have a variation of those in my pack:
difference is that off is left and on is right, which is more in line with many other on/off switches. comfyroll nodes do some things i don't like. for example the switches have a=1 and b=0 which always confused me.

[hardy cipher]

I just stuff 10 loras and switches into one node

I don't think I've ever used 10 loras in one workflow

[wide crane]

max ive ever used at once is 3

after 3 i usually get conflicts

[ionic gulch]

if i need more space in my menu section i would do that too. for now i like the flexibility of 4 separate nodes, even if they overall take a little bit more space. for example i scan spread them horizontally.

[wide crane]

it gets less and less organized

but for now

its still good

[ionic gulch]

pipe-like nodes are nice for menus. as they only have one "pipe" output and take less space in the menu section. other than that i don't see much use and using too many pipe-type nodes was part of the performance problem i had a week ago.
example - menu node + output node in a different part of the workflow:

with normal outputs the menu part would be twice as high - wasting menu space and looking bad

[hardy cipher]

I'd really like if they offered the pipe option but didn't require it's usage

[ionic gulch]

would require complicated "real-time" code, to hide the outputs you don't use. that is a thing that can also cause the lost wires problem i had. had to replace all the impact pack xin1 switches for that reason. made simple 5 in 1 switches for almost all types for that reason.

[hardy cipher]

well image loaders have mask outputs they don't usually use and they don't appear to be very complicated

[thorny frost]

hi! having a lot of ai generated png images it would be useful to be able to convert them to jpg to save space. Problem is losing the prompt. Is there some util to batch extract prompts from png images?

[hardy cipher]

gotta be, if not it wouldn't be that hard

[uncut fiber]

@thorny frost there is at least Save Prompt To File, which can save a prompt.

[hardy cipher]

extremely cursed

[uncut fiber]

yes not sure it can work with already generated images, but there is chance it somehow can.

[thorny frost]

I know. The interesting thing would be to apply it to 1000 or 2000 images. Generate a txt file with the same name as the image, convert the imag to png and delete the png. All done automatically. I can program some simple scripts in python (with help from chtgpt) but this would be beyond my capacities i think.

[floral island]

uhhh...

[hardy cipher]

well it's all in the metadata

[uncut fiber]

i read it already 🙂 Converting images to jpg is for xnview question of few minutels. You can as well keep metadata, but i think it is not in those one can keep.

[uncut fiber]

why uhhh... seems o.k.

[floral island]

@uncut fiber well,tried absurdism, it's totally awesome 😮

[shy kelp]

formulaxl workflow

[uncut fiber]

i love gertrude abercombie work, can work together probably.

[floral island]

art by <random artist>, absurdism, cute

[floral island]

lemme run 4 on that 🙂

[uncut fiber]

🙂 we all have different tastes 🙂 dont be dissapointed!

[floral island]

it for sure is something!

[uncut fiber]

But you can notice, it is always subscribed

[floral island]

horror warning

[floral island]

lel

[uncut fiber]

i got similar 🙂

not much, memory enhance past...

[uncut fiber]

he looks like beholder is next in pot 🙂

[floral island]

imma eat you after this.

[thorny frost]

done! chatgpt is fantastic.

[uncut fiber]

congrats!

[hardy cipher]

did you have gpt write that?

[thorny frost]

is there some way to paste code here?

so it reads nicely

[hardy cipher]

because if so, and you don't go over it, you're going to have a very bad time

in a block you mean?

like this?

[thorny frost]

yep

[hardy cipher]

the little hash thing to the left of 1 on the keyboard `

[thorny frost]

alredy tested

[hardy cipher]

repeat 3 times above and below

[floral island]

fat zombie hamster... very underwhelming

[thorny frost]

from PIL import Image

# Directorio que contiene los archivos PNG
input_dir = "D:/SD-outputs_storage/convertojpg/aa"

# Itera sobre cada archivo en el directorio
for filename in os.listdir(input_dir):
    if filename.endswith(".png"):
        file_path = os.path.join(input_dir, filename)
        try:
            # Abre el archivo de imagen
            image = Image.open(file_path)

            # Obtiene los metadatos (EXIF) de la imagen
            metadata = image.info

            # Crea un archivo de texto con el mismo nombre que el archivo de imagen
            txt_filename = os.path.splitext(filename)[0] + ".txt"
            txt_path = os.path.join(input_dir, txt_filename)

            # Escribe los metadatos en el archivo de texto
            with open(txt_path, "w") as txt_file:
                txt_file.write(f"File: {filename}\n\n")
                txt_file.write("Metadata:\n")
                for key, value in metadata.items():
                    txt_file.write(f"{key}: {value}\n")
                txt_file.write("---\n")

            # Cierra el archivo de imagen
            image.close()

        except Exception as e:
            print(f"Error procesando el archivo: {filename}")
            print(e)```

input_dir change inputdir to the folder where your png images are obviously

[hardy cipher]

all the LLMs are psychopaths. I wanted gpt-4 to give me advice on some code and help me figure something out and it kept messing up my code

they'll do little subtle things you might not even realize until way later

[thorny frost]

you need to explain it very clearly

[hardy cipher]

well it kept messing up the math

I said

do not touch the math do not touch the math

[uncut fiber]

Yes 59% of code it do is wrong, some recent tests, but for ppl that know language properly it can be very helpful

[hardy cipher]

over and over

and it repeatedly broke the math

[thorny frost]

it seldom do it at the first attempt

[hardy cipher]

it would apologize, say it understood now, and do it again

[uncut fiber]

Probably somebody already did it before?

[hardy cipher]

no no, they will never get it right because their entire existence is a calculated hallucination. they have no idea when they're lying or messing something up

I'm not saying that's a bad thing per se

just that I understand it's limitations, at least for me

[thorny frost]

in this case i used a code from reddit from a guy that also used chatgpt to write a code to extract all metadata in a single txt file. Copied it, pasted it in chatgpt and asked for a modified code, so that each metadata would go in its own txt file with the img name

so it was a AI-human1-AI-human2 interaction. 😄

[uncut fiber]

i just dont know. In xnview it showing metadata, but not those you are looking for.

[floral island]

redbull clown mecha... uhhh

[thorny frost]

Chatgpt is better at modifying code than making it from scratch i think

[hardy cipher]

well there are approaches that can work, but they require human interaction to a degree. run it by bing every so often. but don't bother with bard. he's a bit tarded these days

well, just lies about EVERYTHING

[thorny frost]

bard is pretty dumb

LOL

has not a clue about anything

[hardy cipher]

hey bard, can you help me with this thing?

sure I know about that thing and I'd be happy to help

cool, so here's my issue with the thing we just discussed

an an AI language model I'm unable to assist you with that

[thorny frost]

have you tried to run llama model locally?

there are some that code acceptably well, not so good as chatgpt though

[hardy cipher]

well if you want actual programming assistance just get something like github copilot. but then, that thing doesn't have any clue about anything. it knows code real good, but it definitely misses a lot of other things

people want these things to be binary or like switches. but that's not how it happens

[thorny frost]

for amateurs like myself chatgpt is just enough.

[hardy cipher]

if you say so. but as I said, if you don't go over everything it does you're going to have a bad time at some point. it'll do something so subtle you won't recognize it. use a deprecated library, forget a comma in a place that rarely comes up. who knows what sot of hidden frustration mines it will leave

maybe?

like, I'm in no way trying to discourage you from using it

just want to let you know what I've found it's shortcomings to be. it's still a huge help to almost anyone

[hardy cipher]

whoa

are those dentures made of stacks of teeth?

[thorny frost]

i know. There is some times it gets exasperating. But for using it ocasionally is nice. It has solved me more of an issue.

i wonder what the promt was for this. Sadly i converted it to jpg before having my brand new metadata extractor script.

[hardy cipher]

it's just a new kind of tool, like a hammer that thinks real good. but any tool has it's limitations

is that a butt to butt shrimp centipede?

[thorny frost]

dont know. All i know shrimps are less appetizing now

[thorny frost]

i use one button prompt script to generate a bunch of images with random subject, i only stablish the style and loras to use

[hardy cipher]

man, I have something like that sort of

except instead of random prompts I use images for the prompts

[thorny frost]

it generate garbage but sometimes some pearl appears, si it is interesting to keep prompts somewhere

[thorny frost]

you mean clip interrogator?

[hardy cipher]

this one rarely makes garbage. just doesn't always turn out eactly the way expected

it's a combination of a few things

which ui are you on?

[thorny frost]

ah, i know what you mean

ipa

[hardy cipher]

well it's clip, blip, ipa, controlnet, etc

[thorny frost]

i use both a1111 and comfy

[hardy cipher]

all balanced delicately

and I don't think there's a one size fits all for the settings and parameters, but I have done some goofy things with conditioning data whch seems to have made it less susceptible to the black boxes I was getting at fiirst

[thorny frost]

have been disconnected for a couple of weeks and didnt remembered IPA

[thorny frost]

comfy is limitless

[hardy cipher]

dude, this combo I have going never gets old. just need 3 images and it's on

[thorny frost]

is it embedded in the image

?

[hardy cipher]

yeah, but it's a bit sloppy as I've been moving things around

[thorny frost]

will try it anyway. :p

[hardy cipher]

love the space monsters chasing children images

the monsters always look so full of joy

I have some examples of what the same images look like in different orders too

each slot has a different role

first image is ipadapter with interrogater running to clip g

second is two different types of clipvisions plus clip l

[peak dove]

SDXL Bot

[hardy cipher]

third image is latent base and controlnet

might have to install some nodes if you want to use it properly

[thorny frost]

dont get any workflow when loading the images in comfy. 😦

[hardy cipher]

ahh

give me a second. I know why

I was just copying and pasting rather than dragging them from the folder

I'll make a new one and send it your way

[thorny frost]

ok ty

[hardy cipher]

let's see what happens here

[shy kelp]

deconstructing

[hardy cipher]

let me know if you see anything that's obviously not the way it should be

[thorny frost]

got the workflow 👍

[hardy cipher]

excellent

[thorny frost]

has to update everything after two weeks not using comfy zzz

[hardy cipher]

these thigns happen. and that ipadapter is so finicky and needs updating constantly

[peak dove]

AUTOMATIC1111 using Unstable Diffusers v6 model (Civit.ai) and SDXL Refiner

[shy kelp]

(haven't added the lora stack but it's working )

[hardy cipher]

which flow is this?

[shy kelp]

frankenstein'd the formulaxl workflow

stuff on far left is from first image on this page: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

what are these switches? if i had like, 3 loras, do i have to move the switch to on? hmm

or even 1 lora, does switch need to be on?

[shy kelp]

Hey can i use stable diffusion 1.0 using api

[lusty wolf]

You need apply Lora stack.

[zinc cargo]

so many jumps and previews..

[shy kelp]

maybe mine is old or something, idk, just confused on where it says switch_1, switch_2

[shy kelp]

[zinc cargo]

what goes on with the face?
is this a public workflow?

[shy kelp]

face fixing, i just deconstructed formulaxl workflow #✨｜sdxl message, the facefix, hires 1 & 2

[uncut gull]

Can your prompt do that? 😁
Practical applications of the overlay prompts I've been working on. Her face is actually inside the helmet!

Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 3784295722, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, VAE hash: 551eac7037, VAE: sdxl_vae.safetensors, Version: v1.6.0```

[shy kelp]

nice.
i remember the other day, we were trying to get this woman's face behind some glass, pretty tough

[uncut gull]

Yeah that was me. But check out how short my prompt is now, and the results are much more consistent. 🙂

[shy kelp]

turned out great!

[uncut gull]

Veils too.

Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 2151636245, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, VAE hash: 551eac7037, VAE: sdxl_vae.safetensors, Version: v1.6.0```

[uncut gull]

And of course... the holy grail of fine-structure overlays... 🥁 🥁 🥁
Portrait rainfall coverage! 🥳 🎉 🌧️

...With a pretty convoluted prompt though...

Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 1538189682, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, VAE hash: 551eac7037, VAE: sdxl_vae.safetensors, Version: v1.6.0```

Basically, SDXL sets up an internal 3D scene at the very beginning of the render, in the first 20% of steps or so. Within that 3D scene, it has the concept of a background, a foreground subject, and a screen-level overlay. Using prompt-switching lets us prime that screen-level overlay with large-scale structures that resemble the final effect we want. Once the 3D scene is primed like this, we can switch to a description of the prompt we actually want, and SDXL will convert those large-scale structures into fine details.

The screen-level overlay in general is very fragile. SDXL will just discard the entire structure in most renders. For that reason, we have to prime a special kind of scene: One with a background and screen-level overlay, but no foreground subject. Then, only after 20% of steps (once the 3D scene is well-established), we insert a foreground subject. At this point, the screen-level overlay's structure is so solid that it will not be discarded in favor of the subject.

[zinc cargo]

how does it goes on about known faces? or Lora

[shy kelp]

honestly, not entirely sure

im not sure with a lot of things lol

it's weird, with formulaxl, my gen was failing due to memory issues but then i set it up exactly how it was and no memory issues now

[zinc cargo]

did you share the formulaxl workflow?

i only find its' model

ok, just got it that i can copy paste a workflow off civit.ai website

[uncut gull]

It's the ComfyUI bottle! 🙂

[uncut gull]

This is the Tooth Furry. He sneaks into your room while you're sleeping and takes the teeth of children who don't brush their teeth.

Really though, SDXL struggles so much with teeth...

[uncut fiber]

He will eat your fat daemon! 😄

[uncut fiber]

going to rest, will try open his mouth later 🙂

[uncut fiber]

Gremlin?

[shy kelp]

idk what is going on

[noble shoal]

Nah, just some alien creature

[uncut fiber]

Fatal battle of humankind vs evil beasts. Last part 🙂

[shy kelp]

it found a face in there

[noble shoal]

I am starting a new series called "Sports done wrong"

[stone fossil]

Id like to request soccer icehockey.

[stone fossil]

Id like to add deap sea tennis diving.

[uncut gull]

Would play. 👍

[noble shoal]

Heck yes! Monstertruck Polo

[uncut gull]

Prompt a hand with 6 fingers, six fingers, polydactyl. I guess you just can't have it both ways. 😮‍💨

Well, I mean, if SDXL knew how to COUNT, now that would be AMAZING! Seriously, Stability, text is awesome, but please teach it how to count next!

[noble shoal]

@uncut gull That was the request, right?

[uncut gull]

LOL maybe. But are you saying they're "female" just because of the way they look? Do you want to get cancelled? 😱

[noble shoal]

Female was part of the request. Where do i need to sign the cancelation form?

[noble shoal]

I have tried "Hand with a huge amount of fingers" and it doesn't look so healthy

[uncut gull]

It's interesting. I'm trying right now to work out an experiment that will investigate how well SDXL can count. E.g., "four cars", "eight cookies", "three houses". How often does it score a hit on the number? It seems to be doing better than chance from my early tests. Does it have a generalized concept of counting, or is it memorizing??? I suddenly have so many questions.

[noble shoal]

Oh, try "thousands"

[uncut gull]

You can't visualize a "thousand" in your head, neither can it. Also, human counting doesn't generalize perfectly to all scenarios. Just because you can count change in a cash register doesn't mean you can count beats in a musical measure. The question is: Can it count at all, or does it only have some numbered-pictures memorized?
(But it'll take a while to work out how to set this up and then run it.)

[noble shoal]

If i would take a guess, than i would go for "only have some numbered-pictures memorized"

[rustic garnet]

I guess so, too ^^°

[uncut gull]

LOL that's why it's one of my two hypotheses. Obviously that would be the boring answer. The other answer would be awesome.

[noble shoal]

Most Finger picture i can get so far

Most of them need a manicure, though

[uncut gull]

Can SDXL count?
To get a real answer, you would need to do latent saliency probing.
That would mean: SDXL would need to show an embedding that activates on the prompt "four". But, when you prompt for "cookies" and SDXL generates an image with four cookies, that "four"-embedding activates even though "four" wasn't in the prompt.

What I can do... Is at best, just check how it scores across a random sampling of subjects. I'll have Llama2 get me a huge list of countable things, then pair them with random numbers between 2 and 10, and measure the hit-rate... Oof. That's one long experiment.

[hardy cipher]

gpt 4 can barely count

and sdxl doesn't understand human anatomy

[uncut gull]

Humans can barely count. Have you tried working a cash register?

The question is not "can it count well", but "can it count at all?"

[noble shoal]

What the hell are you talking about? My finger picture is flawless.

[hardy cipher]

well I'm saying, their training didn't necessarily give them an understanding of the concepts we take for granted as simple

[hardy cipher]

that's because you have a special relationship with sdxl

[uncut gull]

But they do learn real concepts. We have multiple papers demonstrating that now. Everything from board game rules to theory of mind.

[hardy cipher]

they learn them, but not flawlessly

[noble shoal]

You might be onto something

[hardy cipher]

hae you ever tried having chat gpt give you a list of x number of something?

gpt 3 will fall off the rails quickly

3.5, whatever

4 does better, but it starts breaking down soon as well

they can do insanely advanced things in an instant, but struggle with computational tasks that most 5 year olds could work out

[uncut gull]

I honestly only use Llama2 on my local PC. It can count lists fine as long as you don't make them too long. Same for humans. You can count chairs or coins or whatever fine if there's just five or six. You can group them into stacks if there's a few hundred, but tens of thousands? Forget it. Humans can't count that without messing up. (We use technology, as old as wax and styli, to make up for it. 🙂 )

But... Can SDXL count at all? I will have evidence one way or another soon!

[rustic garnet]

neural networks in general can barely generalize mathematical functions

you can train a DNN to interpolate a sinus function, but it will still struggle to extrapolate the very same function no matter how many neurons you use for training

same for large language models. There is a funny anecdote with large language models couldn't get finetuned on larger context windows, because the model was initially trained on a fixed size context window. The problem was not that it could not be generalized on the larger window, it could not learn the positional embeddings because it has never seen such large numbers.

same is for SDXL. We know that it struggles if we generate images with unusual aspect ratios

the problem is not that the underlying unet wouldn't be able to generate these aspect ratios

it just doesn't understand the numbers

768x1344 is fine cause it was trained on these numbers. 1080x1480 is weird, cause it never seen these numbers in its training data

[noble shoal]

That's some interesting information. Thank you.

[indigo carbon]

SDXL likes to only make things that look right. if you ask for 4 fingers, you will likely get 5. same goes for other logistic things

[uncut gull]

It's been over-trained on five-fingered hands IMO.
Like the scene from 1984, "How many fingers on on this hand?" 🙂

[indigo carbon]

yeah, I hope that NeXT-GPT gets released soon, I can't wait to have instructive editing and such using SDXL

[hardy cipher]

it just replicates the patterns that it learns. and sometimes when denoising it'll see an opening for 7 or 8 fingers on the one hand. so that's what it'll give you.

I'd imagine a lot of it's logic also comes from the clip models right?

[uncut gull]

The repo you found said they'd release the weights. Hopefully works out!

[indigo carbon]

I think they said the weights already exists. it's like gluing expert models of certain things together, right? it says it uses SDXL for the diffusion and LLaMa 2 for interpretation (I think?)

[hardy cipher]

an ensemble of experts, interesting

[indigo carbon]

it's just how to make all the weights work together in a single inference

[indigo carbon]

this

[uncut gull]

They're using ImageBind, a 1.2B parameter multimodel encoder. https://arxiv.org/pdf/2309.05519.pdf

[hardy cipher]

ooh, an all in one LLM team

[indigo carbon]

so idk if they really train anything. it seems they just glue together diffusion and LLM, etc..

[hardy cipher]

this tech is moving at warp speed

well that's what I think is going to keep happening

train specialized models that are really good at their one specific task or few tasks

then link them up with other models that are experts in other tasks

[uncut gull]

They did, they trained two things they call "linear projection layers". They have little 🔥 emojis on that pic you posted.

The 🔥 means these weights are live / updated parameters during training. The ❄️ icon means these weights are frozen and just used for forward propagation during training.

[indigo carbon]

ah, yeah. I think that kind of inference can allow image blending and instructive editing, right? that's likely just the tip of the capabilities when using SDXL and LLaMa2 together and such..

[uncut gull]

But also, hold your horses there. They've been working on this for a long time, and they're using SD 1.5 and Llama 1 (Vicuna). That's what their weights are trained on, so you can't just drop in SDXL and Llama 2. 😕

[hardy cipher]

well once you get the first then you move to the second

[indigo carbon]

I don't see why that won't be compatible with stuff like SDXL and MythoMax. LLM is LLM and Diffusion is still Diffusion

they have a demo
https://989c787409ca0e66e1.gradio.live

[uncut gull]

And you might be right. I doubt it though. Even prompting for SDXL is not the same as prompting for SD 1.5.

[hardy cipher]

well yeah, but the framework is there

basic inputs and outputs are comparable. they might differ in some of the specifics but they are largely the same

[uncut gull]

It didn't mention that she's feeling "wet".

[half ivy]

Got no solid workflows for sdxl (control'd), Auto1111 cant produce (for me) = sad bois club, heart broke island, waterfall of tears, depression cave vibes

[uncut gull]

I feel your pain. I love that my A1111 install works. Good luck.

[half ivy]

so cool 😂

[half ivy]

bro 👌

[indigo carbon]

that's still crazy.. can you get it to blend images?

[uncut gull]

😮‍💨 It really doesn't know what it's looking at. The LLM just has a very vague understanding of what's in the image, and zero understanding of the images it creates. (It never "sees" its own output.)

[indigo carbon]

damn, so I guess we aren't there yet, huh?

[uncut gull]

LOL no wait. After you post the next message, it gets to see what image it outputted. Not super helpful. 🙂

We're not there yet though. Not with this exact setup, anyway.

[indigo carbon]

oh, so I guess that's a limitation of the diffusion model then. I bet SDXL would follow the bot's description

[uncut gull]

Yeah. SD 1.5 here is like a "filter" on the LLM's imagination. The filter is bad, even though the imagination is fine. SDXL would be better, but still not perfect. After all, I'm not even sure it can count.

[indigo carbon]

when it comes to diffusion, SDXL is by far the best.. the only issues is it's lack of capabilities such as instructive stuff and blending. honastly idk where this is going

I don't know what MJ does for blending, but it's not IPA for sure.. I think when it comes to pure txt2img SDXL destroys stuff like MJ but again, the lack of capabilities

[peak dove]

AUTOMATIC1111 DynaVision XL for a model

[peak dove]

MJ has a finish to die for. Comparing SDXL and MJ is like comparing salt and sugar imho

[zinc cargo]

android to apple

[peak dove]

Chalk2Cheez

[zinc cargo]

rounded corners and censorship

only thing left is for SD to do pistachios and im good.

[uncut gull]

Bad news guys... SDXL can't count! 😭 At least that's what all my tests are showing so far.

[hardy cipher]

best image made by sdxl thus far?

[indigo carbon]

Idk what do you mean by "finish" but the quality of the images made with SDXL is miles ahead of MJ. the only advantage MJ has is the image input capability

[uncut gull]

Woah there. I wouldn't defend it that strongly. (I've never used MJ though, so I can't weigh in here.)

[zinc cargo]

MJ has an esthetic flavor that is very... MJish...

[uncut gull]

SDXL can't count at all. It doesn't even have a few numbered-pictures memorized. 😐
The prompt was {two|three|four|five|six} against a white background, and only the two ever succeeded.

(I'm testing now whether it understands the difference between 1 and 2.)

[indigo carbon]

idk.. I never saw MJ get to this quality, when it comes to the images themselves I don't think MJ stands that much of a chance tbh

[uncut gull]

Yeah, I really have no idea. I feel like if MJ was actually worse than SDXL, couldn't they just switch to SDXL behind the scenes and not tell anyone? But people still use Yahoo, so...

[peak dove]

We can agree to disagree 🙂

[peak dove]

IMHO these are the equal of MJ

[indigo carbon]

why don't you prove it then? mj won't get to that kind of detail if its life was dependent on it. it would be able to blend the images though

the blend thingy is impressive, but txt2img is not as good as SDXL

[peak dove]

As any opinion is subjective, there is no proving 🙂

[hardy cipher]

well some opinions hold more weight than others

[rustic garnet]

dunno, I can't try MJ myself, but I sometimes see images other people generated with MJ that look extremely good. At least it seems that MJ is better in doing human anatomy right than SDXL. What's noticeable, though, is that all MJ images I saw so far look all "the same", as if MJ can only do one style (or all people who use MJ and post images all do the same style?=

[peak dove]

Objectivity is the democracy of subjectivities.

All subjective opinions are pooled and the objective emerges ...

[hardy cipher]

mj seems like it produces amazing results with minimal effort. which is great. but in that same vein it also feelss like the user is in no way controlling the real outcome

it's like taking a train somewhere

[uncut gull]

Somebody's using SD for captchas.

[hardy cipher]

or having a driver

[peak dove]

That tendency to sameness a wise and careful prompt-wrangler finds as a challenge to surmount

[lusty wolf]

Anyone have a problem with the just updated comfyui? My nodes sticks to the cursor and you cant put them down. Even when you click on screen starts draging the screen and does not let go....

[hardy cipher]

stable diffusion give you a pwerful engine but you need to know what to do wtih it

[indigo carbon]

I laterally wrote "cube made out of water in rainy cyberpunk city" and it made this is 30 seconds. if this isn't minimal effort idk what is

[peak dove]

I am back to AUTOMATIC1111 - just for the sake of a "tidy screen!" 😄

[peak dove]

This is cool!

[hardy cipher]

yeah, and what accompanying tools and fine tuning did you use?

[rustic garnet]

to be honest: this looks totally like a MJ image 😂

[shy kelp]

https://tenor.com/view/barney-the-dinosaur-nothing-like-a-good-pistachio-pistachio-pistachios-healthy-food-gif-27567503

[indigo carbon]

bruh

[hardy cipher]

I'm not putting it down. what I'm saying is that a lot of powerful tools really aren't plug and play

[peak dove]

I agree - I couldn't tell MJ or SDXL?!

[rustic garnet]

hm, SDXL generates great images without fine-tuning imo. I sometimes use other models, but in many cases the base model still produces best results

[floral island]

someone gimme dumb prompt quick, wanna test my new merge

[peak dove]

A camel eating a banana

[indigo carbon]

just loaded a workflow, then wrote those words. that's it

[peak dove]

A serpent in silk pyjamas

[shy kelp]

person observing extremely muscular monolithic rubber duck

[floral island]

why do you have to make it THAT hard?!

[zinc cargo]

naaaa, "just" loaded a workflow 🙂

[peak dove]

This is a good prompt - (To a girl: If You Were A Booger, I'd Pick You!)

[shy kelp]

lol

[indigo carbon]

I mean, I did use the workflow I made; but that's all I did, really..

[shy kelp]

that's a nice looking camel
and spock

[peak dove]

A snake in silk pyjamas

[zinc cargo]

well General, we havent spoke in a while now

[lusty wolf]

No turning back, Comfy is so much faster, and versatile...

[zinc cargo]

man i saw your workflow about 2 weeks ago and my butt still hurts since

[shy kelp]

ngl, i picked up comfy a few days ago and love it so far

been on a1 for like 1+yr or something, too long

[shy kelp]

#1 and #3

[peak dove]

I tire of expanding Comfy just to see the text, and then reducing again so as to see the whole thing!!!

[zinc cargo]

had to put it in midj... it's not so bad 😛
before upscaling etc

[floral island]

looking at the previews i can only say "what is wrong with you"

OMFG

actual prompt: This is a good prompt - (To a girl: If You Were A Booger, I'd Pick You!)

[uncut gull]

Okay! In some circumstances, SDXL can tell the difference between one and two with a 100% hit rate. That counts as counting! (Because I said so.) However! It can probably only count animals and flowers???
Prompt {one|two} against a white background. 100% accuracy, and it is obviously not pasting in pictures that originally had two subjects. (Some of these are clearly not pasted pictures.)

[indigo carbon]

not as good. plus, MJ isn't free like SDXL is

[zinc cargo]

@indigo carbon you're preachign to the quire, we are all team SD here...

[rustic garnet]

of course mj is not free. It also does not have this amazing ecosystem of tools, plugins, extensions, controlneta and workflows. There are so many reasons why SDXL is better than MJ. But if you just look at text2image I would say MJ is not worse than SDXL

[zinc cargo]

but, MJ has it's strong points, it's a wonderful tool with pleanty of stuff that doesnt require the GPUs and serious knowledge.
it simplifes alot of the stuff we actually enjoy - so like SD, it's a tool. not a way of life. you can still use both

[peak dove]

L8r

[floral island]

this reply was used as the actual prompt. I like this new merge

[indigo carbon]

idk what y'all want. even though MJ is paid, it's still not as good when it comes to txt2img. that's all I'll say.

[floral island]

MJ really making a case for the face

[floral island]

what's the prompt? i'd love to try on my model mix

[lusty wolf]

Do you know what I'm missing with this error? Using AItemplate nodes on this workflow?

[indigo carbon]

cube made out of fire in rainy cyberpunk city

[indigo carbon]

to use AIT you need either the latest Comfy and latest AIT node or a specific Comfy commit and a specific node commit

I use commits from a week or so ago because it's faster on my 4070ti for some reason

[lusty wolf]

Then it must be specific, as I just updated everything to the latest.

[indigo carbon]

including the nodes?

[lusty wolf]

Yeah...

[indigo carbon]

I see. then you need the specifc commits

@lusty wolf
https://github.com/FizzleDorf/AIT/tree/ca8f0627992649eb2211bbdc2f269484b054e62b
https://github.com/comfyanonymous/ComfyUI/tree/bc76b3829f5fbba7c5a439c7833d313a3ca87398

[lusty wolf]

Been trying to use the AI templates for the last week, (that is why I updated) even on the old it gave me the same error. So I give up.

[indigo carbon]

this is what I use

[lusty wolf]

Thanks.

Can one see the commits in the manager, or where do one see that info?

[shy kelp]

lol

[indigo carbon]

git rev-parse HEAD in directories with repos

[noble shoal]

A tribute to Tdg8uU

[indigo carbon]

banana cubes

[hardy cipher]

that's pretty good tbh

[uncut gull]

Okay... SDXL can count "birds" specifically! Why birds? Why not.
It can count "one" and "two" birds with 100% accuracy, and "three" birds with 75% accuracy.
Prompt {one|two|three} birds against a white background

How far does its concept of "one" and "two" generalize? Stay tuned.

[hardy cipher]

well think of it like this, they can count, sorta kinda, but not in the same context that we can

they don't understand what it all means, probably don't recognize larger numbers, etc

they are simply trained to recognize patterns. and sometimes things have numbers attached to them

[shy kelp]

couple loras attached and a custom model

[uncut fiber]

how about tags 1girl and 1boy ??? @uncut gull

[uncut gull]

Oy. I'm talking about real results here, you know? Does SDXL have a baseline concept of "one" and "two"?

But here's OthelloGPT. I point people to this when they say NNs just "recognize patterns". https://arxiv.org/abs/2210.13382
Trained only to predict the next move in an Othello game, a GPT learns to predict the next move in an Othello game. No surprises there.
How does it do this, under the hood? It develops the concept of an 8x8 board, and learns the rules of Othello. Why? Because that's how you predict the next move in an Othello game.

How do you count? By understanding numbers. And I, as a human, do not understand the number 1.23 * 10^60 in the same way that I understand the number 5.

[uncut gull]

Lets keep it SFW. 😛
SDXL can count cats more accurately than it can count birds.

[hardy cipher]

I don't know if these results are very scientific

[uncut fiber]

depends if 3 cats are where 3 cats realy should be.

[vital ermine]

the main reason I jumped to 2.0 is that it could count up to 5

got it to count up to 5 knives and people were shocked. Never looked back

[uncut gull]

Yeah, I'm checking every image one by one. So it takes a while.

[hardy cipher]

well what I'm saying is that it's not a binary concept

[uncut fiber]

o.k.

[hardy cipher]

can it count, can't it count

it's not an on off yes no thing

[uncut fiber]

tried 355/113 today, it changed theme 😄

[uncut gull]

And I'm saying: "How far do experiments show its ability to generalize 'one' and 'two' to different subjects?"

It cannot count cookies. Flat out.

[uncut gull]

It can count Kermits as accurately as birds and cats! 🙂
(Each one of these is tagged "one" "two" or "three", and all of them are accurate.)

[hardy cipher]

how does it know what defines a "cookie" from our perspective. those are weird angles. how could it know what one cookie is visually?

[hardy cipher]

it has no baseline of knowledge of human concepts

[uncut fiber]

cookies are probably difficult know where it starts and ends, if it is one or two cookies.

[hardy cipher]

yes

I don't even know sometimes

it's a fuzzy concept

[uncut fiber]

@hardy cipher i read it now, didnt noticed before 🙂

[uncut gull]

I wonder if it could count something like Oreos though? I'm starting to suspect it can only count things with faces.

[uncut gull]

It can count "two" and "three" Sailor Moons with 100% accuracy, but not "one"? Or is it because I prompted "one Sailor Moons" (plural)?

[floral island]

sorry, mom called, that's a call i'll never not take (this on my model merge)

[indigo carbon]

I kinda still like Copax_colorful better, but nice!

[uncut gull]

Fixing the prompt for plural / non-plural, we get {one Sailor Moon|two Sailor Moons|three Sailor Moons} against a white background, and now it can count Sailor Moon with 100% accuracy. 🙂

[floral island]

i think i had it in my initial model tests, but decided to drop it -> it doesn't work well with nonsensical prompts, and too many artifacts

[floral island]

all 3 prompts combined!

[uncut gull]

It is 100% accurate with "one Oreo" and "two Oreos", but definitely not "three Oreos". It is generalizing at least a little bit to things without faces.

[shy kelp]

a matte grey cube in rainy cyberpunk city

[floral island]

that's not a fiery cube prompt tho

[shy kelp]

1 cookie

[indigo carbon]

I wrote water here, but it still nails it

[floral island]

made of wwater?

[uncut gull]

Yeah, my prompt was "one cookies" technically.

It is counting 1 and 2 Excaliburs, and 3 with better-than-chance.

[floral island]

@uncut gull it all depends on model, which models "learned to count"

usually 1, 2, 3, 5 and multiple are the things you can rely on

[indigo carbon]

this is with fire in the prompt

[floral island]

for example elysium anime model learned what 'punks' are

[uncut gull]

True. I'm using SDXL CrystalClear. These tests do not necessarily apply to any other models, but it would be straight-forward to repeat these tests on another model.

[floral island]

cube. matte. rainy

[floral island]

don't have to test it too much, usually 1 and 2 are decently understood

anything above will be 'guesstimate levels'

[oblique tartan]

anyone using rtx3060, how long does it take for you to render a 1024px image? kindly include sampler use and step count.

[uncut gull]

That in itself would be amazing to me. If SDXL had a concept of "one" and "two" under the hood that it could generalize to any subject. (Including subjects that had never shown up with "one" or "two" in the training data.)

[floral island]

sdxl doesn't have the concept

the checkpoint needs to associate the tag "one/two" with similar elements

that's a training thing

[shy kelp]

why its elevated, idk

[floral island]

have you ever seen non-flying cubes in cyberpunk cities, duh?!

[indigo carbon]

mine's aren't floating =\

[floral island]

obviously, worst timeline to live in.

[uncut gull]

So you're saying it associates the tag of "one" or "two" with its internalized concept of "one" or "two" elements? 😛

[floral island]

yes, this is the 2 cubes concept

it TRIES to make 2 similar elements

[lusty wolf]

Depends on which side of the globe you are... floating here too.

[floral island]

but one image failed

three, will probably have even worse results

why?

because one and two are the most common numbers used in tagging images

three, lot less common

the rest, even less common

[shy kelp]

run

[floral island]

you'd have 'thousands' or 'a group of'

[uncut gull]

Yep. That's how number concepts work. That's why you can glance at three chairs and know it's three chairs. But glance at twenty, and you have no idea. Then you need an algorithm and a process.

[floral island]

and that's how the models work too

only they're dumber than us

[floral island]

so after 2, they already start to fail

[uncut gull]

Yes! But I promise, what you just said has society-ending implications, and there are philosophers who profoundly disagree.

[floral island]

2 out of 3, three cubes. my model is pretty smart, but i like it that way

can it do. 5 cubes?!

[hardy cipher]

have you ever tried.... 5 cubes?

[shy kelp]

exactly 20 chairs lol

[floral island]

well, color me impressed, again 2/3

[hardy cipher]

this is what you guys need to be doing

[floral island]

now. ipadapter merge that shit, and show us what it becomes!

[floral island]

uhh... what am i looking at XD

[shy kelp]

here is a guard tower

[floral island]

this is glorious!

[indigo carbon]

IPA doesn't really blend that well. when used on SDXL, SDXL looses a lot of its aesthetics and coherency, sadly.

[hardy cipher]

just gotta treat it right

[floral island]

@uncut gull this is 'millions of cubes' -> the model tries to resolve the request and 'decided' that it sub-cubes would suffice

[noble shoal]

Mine are also not floating. But they are made out of skin

[uncut gull]

It can count Einsteins as well as Kermits and cats. It can consistently make "one" or "two" {things with faces}. "Three" is ~75% accurate across the board for things with faces.

[uncut gull]

Mmm. Pork rinds.

[shy kelp]

i have not been able to recreate that skull in the sky

[floral island]

either model preference, or flavour text doing you favors

meat. cubes. (technically prompt was : cubes made of meat)

[indigo carbon]

was BLIP the solution? I don't see how we would achieve instructive image editing and blending without SDXL loosing its qualities though. I'm thinking the ultimate solve might be to make SDXL work with something like LLaMA to give it image input abilities like MJ has

[floral island]

'a few meat cubes'

what's with the laser dome?!

[hardy cipher]

tbh I think the blip thing really does help. just kind of stumbled onto the idea. might not be ideal but it works

[native knot]

They're made out of meat.

[floral island]

oh. i still have bismuth in my prompt

[uncut gull]

Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 1920749068, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, VAE hash: 551eac7037, VAE: sdxl_vae.safetensors, Version: v1.6.0```

[floral island]

meat cube, but it also has neon glass bismuth cube in the prompt. no wonder i get these packaged meats

[hardy cipher]

plus I kind of stopped trying to go pure ipadapter because that seemed to fizzle out a lot. but using it with accompanying tools seems to work alright

[zinc cargo]

this is epic

[shy kelp]

the skull has to be made out of smoke!

[native knot]

The Impossible Burger supplier had a container fall off the ship, basically.

[floral island]

sigh. i hate it when prompts listen too close.

[indigo carbon]

I think the solution for image inputs might not be IPA tbh, SDXL's txt2img destroys MJ but IPA makes SDXL loose quality in my experience. I think going forward the goal should be achieving multimodal inference that doesn't make SDXL loose quality. idk if IPA is the way to do it, but it sure is progress

[floral island]

steak block, now farmable in minecraft

[hardy cipher]

not sure what's going to happen, but I bet it will look delicious

oh my

good idea, eface

[indigo carbon]

when doing stuff like IPA it makes a whole lot of new issues like speed decrease and compatibility issues with optimizations, so multimodal might be a solution

[shy kelp]

war

[hardy cipher]

I believe the concept will go through a lot of refinement over time. when it doesn't work out the way I want I have to remind myself the paper on it was just released a month ago

[floral island]

he angy

[floral island]

lol "how do you want your steak done -> blocky"

[shy kelp]

steak

[indigo carbon]

yeah, I heard about something Facebook is working on that's called "CM3LEON", but they said it's closed source, unfortunately.

[zinc cargo]

is there a prompt for this? it's soo nice

[floral island]

this looks as raw as possible xD

[native knot]

Add bone.

[floral island]

adding "minecraft block" help make it a defined block, i just noticed

[hardy cipher]

mostly non-text prompting stuff. I have blip doing most of my prompting, and then have style overlays

[floral island]

ugh. i hate you xD

it's awesome!

[native knot]

Every good meat cube needs a little bone.

[floral island]

reminds me of my last spareribs

you look at the ribs -> they just fall out

good spareribs ❤️

[native knot]

Too small.

[floral island]

those blocks would cost $10000 a piece

[hardy cipher]

think of the cow they'd have to be cut from

[native knot]

lol

[floral island]

1000 feet cow

[native knot]

https://giphy.com/gifs/mrw-video-finally-fukS8yBmpvGGk

[floral island]

333 non-freedom units

[native knot]

Imagine the size of Salt Bae you'd need.

[zinc cargo]

https://www.instagram.com/p/CvRxqfroVMn/

gilibenshahar

Salt a-AI
@nusr_et and that marbled-marbled meat 😁

#saltbae #aivideo
@robosapien_renaissance

I wonder if he'll ever see and like this.

Instagram

[floral island]

asked for electronic pork chops -> am dissapointed in my model

[hardy cipher]

they've evolved

[native knot]

Too much minecraft influence there.

[hardy cipher]

that's the economy beef cube

[noble shoal]

This even has vents

[floral island]

here is where you get that 1m cubes of beef get from

[native knot]

Massive unit, that one.

[hardy cipher]

the size of that lad. geez

[floral island]

that would feed a nation for months

but, i'm not done yet!

[floral island]

YEARS OF BEEF STEAK!

i mean, a single leg would keep a family fed for years

[floral island]

but. i'm sure i can make bigger cows!

still not big enough.

[hardy cipher]

full meat

[floral island]

look skinny

[noble shoal]

Ate already all recources

[floral island]

oh he just donates 0.1% of his body mass for all the world so they can eat for year

universe scaled cow, your move atheists

[hardy cipher]

that cow actually dictates the physical laws of the known universe.

[floral island]

meatball star lol xD

[floral island]

meatball star. capable of feeding a civilisation for a gazillion years

or 5 years, because nestle be like 'yeah it's gone to waste'

been there for 1000000000 years, but went bad in 3 months

[zinc cargo]

meat ball star gave me this lol

[hardy cipher]

tried to make a lovely steak island

almost

[floral island]

meat gundam?

[floral island]

meat gundam.

[hardy cipher]

that comic blew my mind tbh

[floral island]

oh yeah!

[hardy cipher]

those are unironically good

[floral island]

more meat robots!

[hardy cipher]

going through things I made recently, lol. don't always look at all of them initially

[floral island]

comes with integrated salad. lol this is a walking meal xD

[shy kelp]

space cow

[floral island]

"i'm ur dinner"

[hardy cipher]

wet meat gang

[shy kelp]

finding out the hard way that using a photorealistic lora on a photorealistic model is just overkill

[shy kelp]

meatpocalypse

[hardy cipher]

I love it and hate it

that base image is a keeper though

[shy kelp]

wardog

[floral island]

somehow, i'm not giving this guy much of a chance...

[zinc cargo]

look into my eyes

[shy kelp]

https://tenor.com/view/steve-buscemi-weirdface-gif-10492567

[native knot]

My eyes are up here, buddy!

[shy kelp]

dog is hungry

[shy kelp]

not res'd but i liked this gen

[floral island]

uhh.. whatever this is.

oooh... this mech might wiwn against the dogs!

[shy kelp]

[floral island]

meat gundam. GO

uhh....

[zinc cargo]

wonderful!

[native knot]

Delicious

[zinc cargo]

and glad you didnt put sausage into the prompt

[floral island]

MEAT. GUNDAM.

imma keep going till i get the ultimate meal gundam

did i just invent the ultimate meal?

[shy kelp]

what is that green stuff?!

[floral island]

the stuff you don't eat.

[shy kelp]

whats in its hands??

[floral island]

the the thing which either makes him, or you, dinner

[shy kelp]

i feel close

[floral island]

fern gully!!!

[shy kelp]

will always remind me of sea of thieves

[floral island]

this is how i feel...

the fabled eye of the storm

"so uhh... what are we going to do about that?"

"bullshit inspirational quote"

[glad grove]

she speakin fax tho

[floral island]

"space italian"

"british with a heart"

[uncut gull]

It's still not a cloud though.

[native knot]

Hey fellas...any of you done something with comfy where you point it at a directory full of images (or in my case, one with subdirectories full of images) and then have that run through the WD14 Tagger? Trying to get a bunch of images tagged and I figured comfy might actually be useful for that since I can otherwise build the workflow how I need. I just can't seem to find a way (yet) to load a directory of images in for processing.

[uncut gull]

No idea, sorry.

[stone fossil]

You want to use WD14 to caption?

[uncut gull]

Card backs by faction.

[stone fossil]

Use Kohya.

[native knot]

I'm using pythonssss WD14 tagger and it does a good job pulling them in. I'm dropping them off into a custom script that then adds the tags into my Lychee server when the images are scraped in.

[stone fossil]

I here for my stuff both BLIP and WD14 and then go over them by hand and my love. 🙂

[native knot]

I only need to do it once through for all the images that were generated prior to now. I've got new ones saving the tags off and being processed, but I've got thousands of images I want to pipe through the WD14 tagger and it'll be trivial once I can figure that part out.

[stone fossil]

Oh well this is for when I make a lora.

The both free caption methods give me a free start to edit each caption and merge them.

[uncut gull]

I give up. It's not consistent, and it's always just somewhere between bone and cloud, never really a cloud shaped like a skull.

Steps: 100, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 1629730332, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, Version: v1.6.0```

[native knot]

If you want a new challenge, go do the most recent in #🔆｜dailies

🙂

[uncut gull]

I tried one, but it worked on the first try... and it wasn't as interesting either. But I'm gonna sleep now.

[native knot]

New from Burger King?

[hardy cipher]

Yes, with authentic blue cheese

[native knot]

Your way, right away, as you die, from Burger King now.

[slender coral]

Could someone reccomend a comfy workflow for generating with upscale and use of loras?

[hoary saddle]

no addon made with oob nodes, very basic

[slender coral]

Thanks, could you advisae me what to put in the clip?

[hoary saddle]

or open this image

[hoary saddle]

download again, i added clip node 🙂

[hoary saddle]

this one

[slender coral]

Thank you prompting now, could I ask, why is there no refiner in this workflow? trying to get a better understanding of nodes.

[slender coral]

Am I to assume generated images contain meta for the nodes?

[hoary saddle]

just the default comfy workflow

some use refiner, some don't, was more needed with sdxl 0.9 i think

loras good a good job in my opinion

[hoary saddle]

yes, if they're saved correctly, sometimes discord will remove it, but yeah they have workflow embedded into the image

if you save image it will retain it, if you right click copy image it won't

[slender coral]

I needed this info, thank you so much.

I was wondering why people dropped images here.

[slender coral]

You've been extreamly helpful so I don't want to take up more of your time, but final question (as I clearly know nothing about comphy). I currently don't have this checkpoint after testing the image load, do I need to search HF for the model or is there a built in way to get any dependencies such as this in the UI?

[hoary saddle]

dunno what that one is either, i use civitai website mostly for checkpoints and lora's (make sure to check sdxl 1.0 in the filter), download and copy to your comfyui/models/checkpoints and loras folders and refresh webui to load them in dropdown

[slender coral]

perfect, yea hwas just wondering if there was a similar thing I may be missing to

[hoary saddle]

as you load in others workflows youll come across tons of errors as most are using custom nodes, that button will help you figure out which addons to install

[shy kelp]

yeah its a bit tough! 😭

[vital ermine]

Vulcan need no poles

[vital ermine]

I give up on people as I can't train them

[hardy cipher]

Man, these are legit.

[native knot]

Got that tagging workflow figured out. Took a bit of ingenuity, but:

[native knot]

Basically spit an output from find into a text file, used that with an incremental loader paired with load image from file path, then a bit of string manipulation to extract the file name so that I can re-concat that with where the tag file needed to be saved and boom. Tested with a small batch of .png files to start, then thousands of .webp files are next.

[upbeat summit]

nice work 🙂

[native knot]

4675 webp files to get tags for, then a quick script to inject those into the mysql database for lychee. I've already got most of that written in a different way for images moving forward. Once all these get injected, finding an image on my site will be so much easier.

[hardy cipher]

Look like mf doom album covers

[shy kelp]

ghostly hershey kisses

[native knot]

Nearly 2000 images have tag files created for future integration into the database already. This will do just fine.

[hardy cipher]

Nice

[vital ermine]

dang

[uncut lichen]

if i combine a pre-existing sd 1.5 (i think) model with sdxl, will it gain anything from it?

or should i just not bother until someone trains an sdxl model that has roughly the same style that i want

[hardy cipher]

You mean into one checkpoint?

[uncut lichen]

yeah

[hardy cipher]

Don't think you can

[uncut lichen]

unfortunate

[hardy cipher]

They're trained on different sized images, have different parameters. Just wouldn't work.

But believe me. Sdxl models are quickly surpassing the best the peak for what could be done with 1.5

[uncut lichen]

i know, its just that there aren't any available in the style that i like

[hardy cipher]

What style is that?

[vital ermine]

Big dudes are back

[shy kelp]

https://tenor.com/view/ok-star-trek-leonard-nimoy-beam-me-up-gif-15983748

[ionic gulch]

info: my nodes (and workflow) are fixed for linux users now. someone reported it on github and i've noticed that my paths didn't work for linux users:
https://github.com/JPS-GER/ComfyUI_JPS-Nodes
also included is the current version (WIP) of the movie styler.

[vital ermine]

XL is fighting me for training a person yet every vid makes it seem so easy

[shy kelp]

is the process still the same as with 1.5 but just a bigger resolution src image?

[vital ermine]

pretty much

for styles I have that down pat

funny, as styles is supposed to be harder, lol

2 days of about 30 trainings

[shy kelp]

with 1.5, i didn't even bother tagging and got pretty good results
wonder if i could do that with xl

[vital ermine]

various ways etc...

Tried that too and no

[shy kelp]

:(

[vital ermine]

they say yes but damn if I could pull it off

if you don't tag it should make a hard coded image cause tagging/captioning is just telling it what you want to be able to change.

brown hair then I can change the hair colour

[slender coral]

Anyone have any suggestions on more realistic?

[half ivy]

@glad grove
Found it

https://reddit.com/r/therewasanattempt/s/vKb11HQGXS

[glad grove]

sensei 🙇‍♂️

[willow bane]

hlo anyone here

i want to create a clothing graphics for a website which software should i use

[hardy cipher]

this one kind of terrifies me

[hardy cipher]

cursed

[supple knot]

there all wearing one outfit

[hardy cipher]

their cousins

[supple knot]

Must be fashion Week

[hardy cipher]

good call

[tender timber]

I've tried my hand at some very, very, very basic nodes for ComfyUI. Have a look if you want:

https://github.com/tusharbhutt/Endless-Nodes

[hardy cipher]

not sure who this is exactly

[glad grove]

thats a lot of mikus 😔

[hardy cipher]

not sure wtf is going on here

[ionic gulch]

Version 12 of my workflow - now with real "on/off" switches for Base, Refiner, Base Upscaled and Refiner Upscaled:
https://i.imgur.com/h98OyZA.jpg
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

Features:

• Easy-to-use menu area that fits on a single screen (if your screen is big enough)
• Generate images with 100% Base model and/or Base + Refiner (selectable percentage for base and refiner, default: 80%)
• Generate upscaled versions (2x)
• On/Off switches for enabling/disabling generated images and/or upscaled versions
• Easy selection of resolutions recommended for SDXL (aspect ratio between square and up to 21:9 / 9:21)
• Switch between your own resolution and the resolution of the input image
• Automatically adjust input images to the nearest recommended SDXL resolution
• 5 Generation Modes: TXT2IMG, IMG2IMG, ControlNet Canny, ControlNet Depth, Inpainting
• Use the VAE included in your model or provide a separate VAE (switchable)
• 4x Lora - use up to 4 Lora models, menu area offers an on/off switch and all necessary parameters for fine tuning
• 2x IP Adapter - use up to two input images, menu area offers an on/off switch and all necessary parameters for fine tuning
• 2x Revision- use up to two input images, menu area offers an on/off switch and all necessary parameters for fine tuning

[hardy cipher]

version 12? prolific workflow production

[ionic gulch]

usually each "real" update gets a new major version. only bugfixes get smaller steps :)

so next week i have the workflow with the highest version number available :)

finally having a solution for turning on/off parts of the workflow with real switches may result in some other updates in the next days

[hardy cipher]

do you have it on civitai? or just github?

[ionic gulch]

only github

[hardy cipher]

what do you mean by real switches though? I have switches on my things. too bad I goofed on some of them, lol. but they're fixed now

they looked like they worked, but when you don't use matching variables things tend to not work right

[ionic gulch]

before you had to cut wires. now you can choose the four output images with a switch (shown in the first screenshot) and really turn off that part of the workflow

[hardy cipher]

ahhh, an entire part of the workflow

gotcha

I've had to change four words total so far. well in the last 2 or 3 weeks. and those four words broke 2 nodes

I guess it wasn't that bad. but don't like being sloppy like that

[hardy cipher]

he looks like some of his ancestors were cats

one of his grandparents was a khajiit

[supple knot]

@hardy cipher it was this

[hardy cipher]

lol

wow

brilliant

[hardy cipher]

I just keep throwing things back in the pot. it's kind of like the upton sinclair book "the jungle" where butchers used to throw the old sausages back in the mix when they would make new sausages

gets weird after a while

[ionic gulch]

the usual xin1 switches with for example 4 inputs, 1 output and a value to select the input will still run code for all 4 inputs. the new switch used for the 4 output images will set one node to mute which prevents unneeded code from being executed. you also couldn't turn off some things, because missing inputs throw errors that stop the workflow. so for that use case there was no solution other than cutting wires (that were rerouted to the menu section). now as it mutes the vae-decoder you get no error and everything "left" of the muted node is not run.

[hardy cipher]

ugh, that missing inputs thing is annoying

[woeful patio]

is there a qr code control net for xl?

[crisp owl]

impact any switch with execute on prompt option

[supple knot]

@hardy cipher what the root of the frog people?

[hardy cipher]

told a guy in another channel to give me a random picture and he gave me hypnotoad. and then hypnotoad's frog babies started showing up in other images

and it just went from there

[supple knot]

it looks like Wes Anderson, mixed with like Vogue Magazine

[hardy cipher]

well I added some style of course. I really liked the frog girls. this is where it all began

the patriarch

[supple knot]

Gen 1 ha

you think it matter the size of the photos for this process?

[hardy cipher]

I don't know about size per se, but I do like to have them similar sizes. I know that the shape can matter for some things

that's one of the reasons things get cut off with ipadapter I believe

I just try to make sure they have similar pixel counts though

[supple knot]

So I worked on Giant Spock for way to longer then I wanted to, even editing in photoshop and rerunning 10 total maybe

[hardy cipher]

lol. well got em at least

[supple knot]

Ever generation he would lose the Vulcan salute more

Now hes just saying Hi

[vernal spire]

[noble shoal]

More like: stop right there!!!

[vernal spire]

honestly if 2005 fashion 80 foot tall spock told me to stop while i was on the freeway, i would

[hardy cipher]

what about these guys?

[noble shoal]

I love the meat helmets

[hardy cipher]

wet meat gang

[supple knot]

Those are MeatBoys I think they fought the Warriors

[hardy cipher]

they look so comfortable in their own meat skin

no worries or strife

this guy is a different story

[nimble heart]

Hypnotoad frog girls you say?
https://youtu.be/TiFlMspJvN0?feature=shared

[hardy cipher]

this will create top tier fashion

[hardy cipher]

chipper tune

nice

brought them from circa 1978 to circa 1984

[supple knot]

Woah

give them some toad

[worthy orbit]

great The Warriors, really looks like the film

[supple knot]

Ha, its to generate

[hardy cipher]

they're hiding, one moment

[supple knot]

As cast of SNL ? kinda

[hardy cipher]

alright, still trying to remember what folder they're in. in the meantime tried to bring some kool keith fashion into the mix

aged them a bit

[hardy cipher]

lol, this is genuinely perplexing. I tucked that whole folder of images into some weird corner somewhere and cant' find it. I'll find it as soon as I stop looking. but that's unacceptable

[supple knot]

its all good

[hardy cipher]

but I must find them

to feel at peace

[supple knot]

its the folder labeled frogs

isn't in your comfy outputs folder too anyhow

[hardy cipher]

nah, it was something smart like "new folder 3"

[supple knot]

what about looking at the thumb nails

Prompt : Gang in Christmas Sweaters

[hardy cipher]

Internet decided to go down right in the middle of my search

[hardy cipher]

hmm

[hardy cipher]

almost frog

if you squint your eyes it's frog

[supple knot]

FrogoWarriors

[hardy cipher]

whoa

[supple knot]

Were from France

[hardy cipher]

those are body paint vests

[supple knot]

Because you cant lose a body paint vest

[hardy cipher]

we've taken a wrong turn

turning back

[supple knot]

yeah Im done before I unearth a demon

[hardy cipher]

think I did already

I was just mixing them with spinoff rainbow goblins

their shoe game is on point though

[supple knot]

if they can get those goblin faces ,is that them too

[hardy cipher]

you want the goblin faces?

[supple knot]

Yes

[hardy cipher]

these are my og rainbow goblins

a bit of variation in them

[supple knot]

There really Good , all of them .A+

[hardy cipher]

that time they found the end of the rainbow

these ones are so grimy

little skittles colored street urchins

no demons here

[supple knot]

3rd maybe demon, ha

[hardy cipher]

might have to work this one in too

[supple knot]

Give them Pepsi