✨｜sdxl

[hardy cipher]

cursed

[nimble heart]

I feel like two eyes are such a prominent concept you'd almost need a Lora to unweight them if you wanted consistent single eye cyclops images

[vital ermine]

I am thinking the same

[zinc cargo]

use wrong lora in a positive prompt?

[vital ermine]

though the clip/te is supposed to know this stuff as it is massive in size

I don't have that for xl

[zinc cargo]

my first MJ prompt was Leela from futurama...

[nimble heart]

everything on the planet with eyes has at least two of them

so it's not got a lot to go off of

[zinc cargo]

and i got a boob!

[vital ermine]

a boob? LOL

[nimble heart]

and cyclops aren't popular character art either

even in DND everyone just draws hot tieflings

of which has two eyes

[zinc cargo]

AIniverssary

[vital ermine]

Be different, break out of the box and play a Cyclops.

[nimble heart]

no

[zinc cargo]

maybe i should spoiler it

!!!!

[nimble heart]

dude wearing a mask

the skull shape has to completely change to make the one eye look real

[nimble heart]

if you still have eye sockets then it's not really a cyclops

[zinc cargo]

everyone is an eye critic

[nimble heart]

yea you're right not like it's the single defining feature or a cyclops or anything

[zinc cargo]

they do have personalities you know...

[ionic gulch]

dog with red demon eyes:

[nimble heart]

A large man with a facemask and a spicy personality does not a cyclops make

an elf wouldn't be an elf without the pointed ears

without them it'd be confusing trying to be racist towards them

[hardy cipher]

tbf this is an ancient greek statue of a cyclops. and they were kind of the OG cyclops society. lots of cyclops references from their time

[nimble heart]

just because they came up with the idea doesn't mean they're right

looks like a necromancer got bored and started rearranging parts on a whim

[hardy cipher]

ahh, you know a lot. I'll just assume you're right

[nimble heart]

No need to assume. I am always correct 100% of the time.

[west breach]

Is there a setting for comfy to limit how many models it keeps in CPU memory?

[nimble heart]

Not specifically but setting low vram encourages early unloading

[zinc cargo]

you know what... i agreee...
SDXL is very bad at getting the concept of cyclops

also...

[nimble heart]

thought he had a univisor

[zinc cargo]

he did. that's why i said what i've said

[nimble heart]

could negative goggles and positive visor

or other words for single lense masks

[nimble heart]

oh cpu

tf

I've never used more than like 25 gigs

[hardy cipher]

that's some kind of memory leak maybe

bing cyclones

those images are really disturbing

[west breach]

I'm testing 12 models, but it seems to keep them in memory

[nimble heart]

I have a dozen 1.5 models I could test tomorrow. They ate all 64 gigs of my ram in a1111 but I don't recall having issues when I made a grid of them in ComfyUI

[west breach]

these would generate a lot faster if my computer didn't keep locking up because it's running out of cpu ram

[hardy cipher]

oooh, all in one go

[nimble heart]

do you have all 12 models in the same workflow as 12 nodes?

[hardy cipher]

lol

[west breach]

yeah

[nimble heart]

that might be

[hardy cipher]

well duh

you have them loaded

if they're in a node they're going to take up space

otherwise they'd have to reload each time you ran through the workflow

which would take a bunch of time

[west breach]

they're all being used at separate times, not at the same time

[nimble heart]

yea I always use one node then just manually switch and queue the rest while the first generates

[hardy cipher]

and people rarely load 12 models at once

well it takes several seconds to load each model

maybe less for some big dawgs with fancy hardware

you could probably modify a checkpoint loader node to do what you want, but then running through the workflow would take quite a bit longer

[nimble heart]

i mean comfy should still be able to handle it since the queue is a single thread. It could load models only when they're needed them unload for the next. Maybe open an issue with the workflow file to reproduce

[hardy cipher]

wow, yeah, do that

maybe he'll make it so they're magic and load instantly

[nimble heart]

but for now I'd 100% just queue a batch for each model separately. Gives you something to do while they generate I guess lol

[nimble heart]

XL models already load in like 1 second IDK what you're doing

hard drive?

[hardy cipher]

no, dunning kruger, I have an ssd

[nimble heart]

how

in any way

does the dinning Kruger effect

apply to what we were just talking about

[hardy cipher]

it makes no sense to have models load each time you access the node

[nimble heart]

ok

[west breach]

Maybe I'll just make a custom node 😛

[nimble heart]

but how's that dunning kruger

[hardy cipher]

you probably wouldn't agree with me

as is your way

[nimble heart]

might be cursed but you could probably make a custom ksampler node that just inherits everything from the built in one but unloads the model once it finishes

[hardy cipher]

custom node probably wouldn't be that hard

[west breach]

yeah, i was thinking a node that just passes an image through and clears the memory at the same time'

[nimble heart]

it'd need to be on the sampler wouldn't it?

cause the model flow stops on the ksampler node

unless you just like purge all the torch cache

[astral jay]

Just trained a new lora...

[west breach]

much betta

[zinc cargo]

generated?

[sand pier]

had a look, i think it would be nice to have links to the huggingface models you are using, because testing it out its somewhat guesswork

and a lil pain to find unless you know explicitly what you are looking for

[ionic gulch]

@sand pier yes, i've noticed that. already included the link for the inpaint model (not updated on github yet). will add the ip adapter links too. anything else missing?

https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1/tree/main/unet
https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/pytorch_model.bin
https://openmodeldb.info/models/4x-NMKD-Superscale

[dense chasm]

nope lol.

[sand pier]

the controlnets and clip

uhmnn

[sand pier]

i think that's it?

sdxl lora offset

[ionic gulch]

so now i have:
https://i.imgur.com/S07Lnqr.png

[sand pier]

yeah that looks very good

[ionic gulch]

@sand pier does it work for you, because Klinter had vram problems before?

[sand pier]

ill try in a lil

[ionic gulch]

sorry for all those requirements, but if you want to support all those features and have a nice menu many models and nodes are needed :)

[sand pier]

that's just fair, the ones ive tried before usually have just bash scripted the requirements but likely not needed

[ionic gulch]

i don't want to force specific models. in some cases users may want to pick different ones. for example different upscalers or checkpoints.

or fp32 instead of fp16

[sand pier]

i mean that's a fair standpoint, but often if you download a more complex model having a known working starting point is nice

then you can start to thinker with the details

[soft zealot]

If you're around @visual glade have an odd one

JUst updated ComfyUI to "ComfyUI Revision: 1391 [36ea8784]"

Now some of my workflows wont even start (Heavyweight,MiddleWeight,LightWeight,ChatGPT Demo)

There are no errors in the console they simply wont start running when Queue Prompt clicked

Revision, Revison+ & Revision+++ start & run fine
(each workflow embedded in relevant screenshot)

[ionic gulch]

not bad. good skin quality and even some hair on the shoulder:

[sand pier]

it did work for me btw, i do have a 4090 though

had to resolve some updates breaking ipadapter but other then that np

[ionic gulch]

green dress inpainting:

[vale eagle]

what is the prompt of the generation?

It seems to be a good prompt to test hand in different fine tune

[ionic gulch]

wasn't an update enough? ip adapter had some problems a few days ago, but the latest one (from one hour ago) worked without manual fixes for me. i've updated all nodes before i made the upload after there were problems in the past :)

[west breach]

A detailed image of a hand reaching out to touch a painting, where the figures in the painting reach back, intricate accurate details, capturing the interaction between art and reality

[sand pier]

i had to manually just pull latest from the ipadapter repo

[vale eagle]

Thanks

[sand pier]

one more minor change i also had to do was swap the combo lora loader to a widget, not quite sure why though, he just refused to be taken as an input

[ionic gulch]

combo lora loader?

[sand pier]

the one that gets normally loaded here

[ionic gulch]

ok, so just changing it to your filename didn't work?

[sand pier]

for some odd reason it didnt

probably just something breaking it for me there though

[ionic gulch]

i think i've managed to reduce the mask lines enough, so the upscaler removes them completely. or do you still see the borders of the used mask in this image?

the previous one still had those artefacts:

[uncut fiber]

isnt there just some blur offset?

[vale eagle]

it still has the invisible line but good enough

[ionic gulch]

the official workflow had no blur and no options for it.

[vital ermine]

Ah HAH. At first I thought that was my video card having issues.

[uncut fiber]

i have not watched all discussion, sorry

i think making few inbetweens images could be quite good video 🙂

[soft zealot]

@visual glade FYI "Rolled back" using Windows Portable download to ComfyUI Revision: 1374 [fe4c0740] and issue not present
(as in did a clean install into a fresh directory in parralel to existing install)

[west breach]

I tried making a node that unloads all models, but it just moves them to RAM, doesn't actually remove them

[glad grove]

make a node that moves all our models/loras and embeddins in our hdd's to ram

[west breach]

you must have a lot of ram

[vital ermine]

Well, he tries hard though

[vale eagle]

Could we train embedding for SDXL?

[vital ermine]

I wish

I use them, now, for neg stuff

lora with a weight of neg is not the same

[glad grove]

i see a few XL embeddings on civitai idk if they are real though

[vale eagle]

CounterfeitXL has neg embedding. I am wondering how to use and create https://huggingface.co/gsdf/CounterfeitXL/tree/main/embeddings

[glad grove]

bro has that 16 pack

[zinc cargo]

double pack!

[sand pier]

Been testing ipdapter lately, and for some reason the SDXL is performing so much worse then the SD version. Anyone have any ideas?

(Left is SD, right is SDXL)

[eternal fog]

Are you using an anime model for sd1.5?

It can't make stuff that doesn't exist in the model and the SD 1.5 anime models are still much better than SDXL.

[sand pier]

nope, pefect for sd version, xlpefect for sdxl

[eternal fog]

I'm not familiar with that model. But my suspicion is that the 1.5 version is just better at anime in general

[sand pier]

probably, tested quite a few to see if i could make some improvements but usually it just feels random

tested a few OC artworks as inputs as well to remove any potential biases but generally same result

[uncut fiber]

clownpunk

[cyan crown]

combining Loras is awesome!!

[uncut fiber]

i am combining authors 🙂

[native knot]

le sigh...
Don't you just love it when you load a workflow and what worked fine 5 minutes ago is not not working after an update?

Worst part on this one is that no nodes are turning red and my console isn't spitting out anything at all...it just won't queue the prompt at all. Literally click the button or ctrl+enter does nothing.

[indigo carbon]

Classic ComfyUI updates

[native knot]

Ugh.

[indigo carbon]

I love it, makes life more interesting =]

So people that make nodes have to rewrite the code every single week

[native knot]

Sometimes though...just sometimes, I want to to be a It Just Works ™️ situation.

[native knot]

Time to tear apart the workflow to see what's not doing the thing.

[native knot]

Well, after taking ttn TextDebug out temporarily, it allowed the workflow to at least show an error...funny thing is that I put the node back in and I'm still getting the error popup; which is good, but odd that just having that in there was preventing the error output at all.

The KSampler doesn't like something, but now to figure out what. The error is:

 * (tuple of Tensors tensors, int dim, *, Tensor out)
 * (tuple of Tensors tensors, name dim, *, Tensor out)```

[cyan crown]

my new Lora

[zinc cargo]

you think it's possible to make a node that actually post to civit every time you want?

[native knot]

Why not?

[zinc cargo]

i mean , i'm sure it's possible, just have no idea how to do it

[native knot]

It was the AIT Template Loader.

Took it out, all worked.

[indigo carbon]

actually, the fix for that was to merge with the earliest ComfyUI commit from the 28th of last month

[vital ermine]

yep

[native knot]

So...plain english me. What do I need to do to get back to working?

Is this going to be just git cherry-pick hash?

Ah...nope...I see, just a straight reset --hard to f92074b

Word. That worked.

Thanks!

...now what was I going to make before all that occurred?! lol

[soft zealot]

coincidence?

https://github.com/comfyanonymous/ComfyUI/commit/36ea8784a875bde21c88f84dfb99475b6e8187e8

[shy kelp]

what are the best sdxl models right now?

[peak dove]

Combined 1960 and Underwater LoRAs

Plus some Nik Color FX

[native knot]

"best" is subjective and the answer would depend upon what you're trying to render, your tastes, limitations on the models, etc.
My personal go-tos are any of the XL models from copax, juggernaut, nightvision, protovision, dreamshaper, crystalclear, and occasionally deepblue.

I think the one I get the most out of at the moment is copax timeless.

But it really depends.

[native knot]

That is some wild stuff...nice style it ends up with. You should have it render some framed pieces.

Can't wait to share what I'm in the process of rendering right now...if the final product looks as good as the preview, it'll end up as my wallpaper.

Had this idea for 2 concepts based on some earlier posts from here plus a meme I saw this morning. 🤣

[peak dove]

Copax Timeless is a checkpoint model? I can use without a Base/Refiner setup at all?

[peak dove]

Out of about 100 renders, 10 are marketable, the others are too 'leftfield'! 🙂

[native knot]

All models can be used without base/refiner setup. It's just about whether you like the output or not. I have been using a txt2img one-shot workflow with timeless for a bit now with great results.

[native knot]

That's kinda the nature of this thing right now. Once you find something that works, it's all about small adjustments to get something out that doesn't have glaring errors...that is, if you go without inpainting or other things of that nature.

[zinc cargo]

what do you mean by marketable?

[peak dove]

Stuff that'd sell on fineartamerica or Printler or Society6 ...

I have sold a few AI prints @ Printler

[indigo carbon]

actually, I recently made an IMG+IMG AIT workflow if you want to test new AIT stuff out

[native knot]

Always down to check stuff out. 🙂

[indigo carbon]

it's on my CivitAI page

[native knot]

I assume it's the Image blend_AIT one?

[indigo carbon]

yep

[soft zealot]

“I can’t believe it’s not ReVision” now with extra added AIT lol

[uncut fiber]

8GB isnt enough for it, or how to do lowvram in comfyUI?

[native knot]

You can adjust the launch parameters in the batch file that launches comfyui.

[uncut fiber]

yes but there isnt much so simply add new line and type --lowvram?

[native knot]

Not a new line, just append that to the end of the current execution line.

[soft zealot]

ComfyUI auto detects and sets the best setting (unless you have disabled smart memory management).

[uncut fiber]

how i recognize it, yes it run normal Vram, but simply i get out of memory with TDG blend things

[soft zealot]

Ah is that extra overheads from the AIT modules?

[uncut fiber]

i dont know, it crashing in KSampler. It helped bit turn down batch from 4 to 1, probably there are different things to change. 🙂

[indigo carbon]

actually, it's an issue with ComfyUI. you need an older branch for AIT to work because the newest versions of ComfyUI isn't compatible with AIT

[soft zealot]

ah the old chicken and egg paradox 🙂

[peak dove]

Weirdly good with the 1960's LoRA added to my w/flow

[visual glade]

does the default workflow start?

[soft zealot]

one moment caller

[soft zealot]

yes it does

[visual glade]

so it's a problem with some workflows with custom nodes?

if you roll back to 0e3b64117218c50a554b492269f5f35779839695 does that fix it?

[soft zealot]

potentially yes.

II have 4 workflows that dont start and 3 that do.

Mostly all use the smae custom nodes, big difference that I can immediatley think off is thathe ones that dont start have ChatGPT Assiste inthem

[soft zealot]

I intsalled using the windows Portable version. For the moment I grabbed the last working version of that and did a clean install.

Is there a way of rolling the windows portable back to a specific version after running the upgrade bat ?

[visual glade]

if you have git installed you can open a terminal in the ComfyUI folder and: git checkout 0e3b64117218c50a554b492269f5f35779839695

[sand pier]

probably gotta fetch first

[visual glade]

yes that's what it prints when you do that

it's fine

you can: git checkout master

after to go back to where it was

[soft zealot]

workflows not starting wirth update now starting

[visual glade]

if you go back to master and then check the javascript console in your browser when you queue a workflow do you see any errors there?

[soft zealot]

this is one of the workflows that wasnt starting and is also the simplest of the non starters

[soft zealot]

easiest way to go back to master just run upgrade bat again I guess

[indigo carbon]

@visual glade I heard it's possible to surpass AIT with speed using TRT and OneFlow, correct? then why isn't that the focus as of now?

[visual glade]

oneflow doesn't work on windows

[indigo carbon]

ah, is that the only reason? because I heard from FizzleDorf the fastest inference possible is TRT with 8 ONNX optimization levels

[visual glade]

AIT is open source so I can change the code to do what I want it to

trt isn't so it's harder

[soft zealot]

and with a wider view using a simpler flow that doesnt start

[indigo carbon]

fair enough.. although I saw FizzleDorf get to speeds like this with TRT+8ONNX optimizations

[brazen whale]

/Dream,aroma服饰门店

[indigo carbon]

honestly I don't know how is this even possible; but it's fizzledorf's goal to make getting to speeds like this accessibly

[brazen whale]

/Dream，aroma clothing store

[meager canopy]

#1072220168534642768

[soft zealot]

@visual glade

I think I've narrowed it down to a potentially a confilct with Omar92s Quality Of Life Nodes.

My CHAT Gpt Demo only Uses WAS + QOL above standard and doesnt start

ReVison++++ uses WAS but starts fine (doesnt have QOL Nodes)

opinions please.

Which do you think has the bettwer skin texture?
Image 1, Image 2 or Image 3?

Or no real difference, they're all terrible ;o)

[meager canopy]

They're not great

[soft zealot]

thanks

[uncut fiber]

i am using older version have not updated AIT is working in my workflow great as well as in others. Just Out of memory issue in that blend one.

[meager canopy]

You asked 🤷🏻‍♂️

[soft zealot]

I know (and they are zoomed in at 235%)

[meager canopy]

Left is best, but too smooth, and the other two have a pattern affect over them

[soft zealot]

ta

At 100% zoom

[indigo carbon]

@visual glade I thought SAI is leaning towards TRT more and they collaborated with NVIDIA, so it could be open sourced for you?

[meager canopy]

Same with those. There is a lack of skin detail

[soft zealot]

also that was pretty much my own opnion 🙂

[native knot]

Galactic mitosis rendered directly to 4k. (No upscaling.)

[meager canopy]

Look how flat and featureless her chest area is

[soft zealot]

really for a fair comparison need to use the og gen image ratehr than the upscaled ones.

For reference Image 1 used a mix of 4x_NMKC-Siax_200k & 4x_UltraSharp for the Upscale models and both rescales using area

Image 2 was 4x_NMKC-Siax_200k & 4x_NMKD_Superscale_v2 , one rescale used area, secondused nearest exact

Image 3 was also 4x_NMKC-Siax_200k & 4x_UltraSharp but one rescale used area, secondused nearest exact

[meager canopy]

These are just base/refiner output, no upscale

[soft zealot]

and the overall image was probs influenced by the styles used (iphone-photographic, cinematic & dytopian

[visual glade]

can you check if it's fixed now?

[soft zealot]

be about 30 minutes jusdinners serveed

[vale eagle]

Love this one. It is clear and nice

[indigo carbon]

I'll fuze it with a doggo to see what happens

[vale eagle]

Love this one. Evil.

[vast galleon]

testing my color transfer node for bugs and found interesting effects when transfering over individual color channels

[soft zealot]

ComfyUI Revision: 1393 [7291e303]

Still same issue

[visual glade]

has the javascript error changed?

[soft zealot]

same error looks like

widgetInputs.js:22 Uncaught (in promise) TypeError: Cannot destructure property 'link' of 'node.inputs.find(...)' as it is undefined.
at widget.serializeValue (widgetInputs.js:22:11)
at ComfyApp.graphToPrompt (app.js:1358:66)
at async ComfyApp.queuePrompt (app.js:1481:16)
at async app.queuePrompt (nodeMenu.js:41:13)

[soft zealot]

If I run the same prompt/seed without styles applied ;o)

yup, skins texture is a bit flat lol

[visual glade]

now hopefully it's fixed

[soft zealot]

ComfyUI Revision: 1394 [6962cb46]

Confirmed working correctly again now

Cheers

[thorny frost]

civitai SDXL basemodel contest winer has been CrystalClear XL

My preferred one is UnstableDiffussionXL though. Best to create rqandom stuff.

Wich one is yours guys?

[lapis gale]

galaxytimemachine XL and protovision XL do the job nicely, add detail, and are still flexible enough to generate all the different styles

[thorny frost]

I have both i my model folder

one i just found i like is protovisionXLHighFidelity

[lapis gale]

yeah, it's getting crowded in the XL space now, too. at least not too many remixes for the sake of a remix (yet)

and i love the SDXL Styles selector - some XL models are overtrained or tainted in some other way and just go for their own style, no more comic/anime/pixelart

[thorny frost]

It makes little sense to upload remixes when anybody can do it a t home ina 10 seconds

[somber hill]

just published how to do SDXL LoRA training for free with both Text Encoder and UNET with Kohya GUI : https://youtu.be/JF2P7BIUpIU?feature=shared

[lapis gale]

"a cute AI (mixing models and styles) in a big bucket, focused look, colorful"

[azure oxide]

Dame, comfy a literal modern day hero

[soft zealot]

like the playing card vibe

[uncut fiber]

@visual glade updated everything and my workflow doesnt work, once out of memory where would never expect it, in others KSampler advanced errors, not my workflow.... I did all updates, anything i can do else?

[visual glade]

any custom nodes?

[uncut fiber]

Error occurred when executing KSampler:

cat() received an invalid combination of arguments - got (Tensor, dim=int), but expected one of:

• (tuple of Tensors tensors, int dim, *, Tensor out)
• (tuple of Tensors tensors, name dim, *, Tensor out)

You mean DT or such? AIT is all i know

[steady grove]

Don't know where else to say this but here. Congrats to all the SDXL Civit AI winners! Looks like it wasn't a giant popularity softcore contest like i thought it was going to be. There's actually a broad selection of winners that don't look like the owners friends.

Being so cynical sucks sometimes. I should've submitted something

[uncut fiber]

@visual glade here is one that doesnt work for me. i mean WF in image

[visual glade]

for the AIT nodes: https://github.com/FizzleDorf/AIT/pull/44

[native knot]

start here:
#✨｜sdxl message

end here:
#✨｜sdxl message

Essentially, you have to revert to an older version of Comfy to remove the issue and what was recommended by @indigo carbon was commit f92074b.

[uncut fiber]

@visual glade thank you very much

[steady grove]

i just don't use ait. the speed differences are negligible on my 4080 build with 64gb ddr4. It being in the custom_nodes folder causes a bunch of problems on all versions. Particularly stopping controlnet nodes from executing

there's some shenanigans going on with the ait code that bork comfy and i don't trust it

having ait in the custom_nodes folder at all causes all sorts of problems. the ait nodes don't even have to be instanced

[steady grove]

actual thanos wearing a toupee like that would be some clown edition of "what if"

[steady grove]

how did you guys elect a pres with a toupee? i'm always austounded. He made sense as the aprentice host but then y'all elected him. I don't get it

[native knot]

Xteven Xeagal would take Donald Tranos out!

[steady grove]

now world wide american icon is a fat dude insecure about going bald

[uncut fiber]

@native knot thank you, how to --hard to f92074b ? please, i am not command line prefering user 🙂

[steady grove]

he's 80 with gold hair too. it says a lot. A LOT. and y'allqueda elected him. i just dont understand i guess

[native knot]

open command prompt
change directory to the ComfyUI folder ("cd <folder path here>)
then run:
git reset f92074b --hard

[uncut fiber]

o.k. thank you!

@native knot thank you working now again

[steady grove]

missing out on inpainting and controlnet is a big trade for ait speeds imo. The increases aren't even that great.

better solution is delete ait and benefit from new features

you might not even be using a workflow with AIT nodes in it

[thorny frost]

maximum randomness workflow: two random images created with one button prompt are mixed through IP adapter

[vast galleon]

ComfyI2I has been updated with a new and improved color transfer node for anyone who would like to play around with it.

https://github.com/ManglerFTW/ComfyI2I

[dapper dragon]

https://tenor.com/view/snoop-dogg-gif-25326816

[steady grove]

lot of supremacy dog whistles in here today.

[native knot]

In an alternate reality...

[shy kelp]

wow what model is this?

[stone fossil]

Protogen + private lora.

Dreamshaper is also bad ass. 😉

[ionic gulch]

teaser: upcoming new inpainting version in my workflow:

1. source, 2. standard stable foundation workflow (with artefacts in the hair where the mask ended), 3. my new upcoming workflow

[zinc cargo]

@delicate kelp is about to push really nice updates on both nightvision and dynavision models...

[lapis gale]

how do you manage to distort the beemers so much, is that intentional?

watercolor anime 🙂

[stone fossil]

Yes.

I love the output of it.

[ionic gulch]

new version 4.0 of my workflow with huge improvements to inpainting (imo much better quality than the default workflow and without mask artifacts in many cases):
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[lapis gale]

cheesy

[eternal fog]

You can't connect both the inpainting model and a normal model can you?

I briefly tried merging a normal and inpainting model with the model merge node yesterday and it did very strange things.

[ionic gulch]

you can, but you can't switch midway between steps, like you do with base/refiner. you're latent will also include the noise mask as long as you don't remove it (for example through vae), so all further samplers will still use the mask. took me some time to find a good solution and integrate it with the other 4 generation modes, that work quite differently.

new jacket and fixed broken ear (changes in the face are caused by the upscaling process):

[ionic gulch]

inpainting + upscaler - prompt "smiling cat"

[zinc cargo]

what about this?
are there special instructions or hacks to use this one?

[ionic gulch]

you can use for example "a {mouse|dog|fox} with a hat" and it will pick one of them.

you can also use weights (don't know the exact syntax) and you can place files in the directory of impact node pack and use filename to pick one of the lines from that file

should be explained in more detail on the github website of impact pack.

[ionic gulch]

"a plush lion"

[vital ermine]

The damn people pooping in the streets. What do they think this is San Francisco?

[ionic gulch]

more inpainting "improvements"... for the science

[sterile yoke]

are any of you familiar with fooocus

specifically the moonrise version on google colab

[nimble heart]

finally getting some cyclopses I see

just prompting differently?

seems like it might like that one specific cyclops

[ionic gulch]

she was at the hair stylist and dropped her headset:

[vital ermine]

Yep

[nimble heart]

you've meticulously prompeted out a specific cyclops monster from some 1970's movie it looks

[vital ermine]

That is not from the 1970s and is the most famous of the Hollywood Cyclops.

not counting x-men, of course, lol

That is Ray Harryhausen's Cyclops

Stop motion from The Seventh Voyage of Sinbad

1958

[upbeat summit]

greetings

[vital ermine]

Heyas mass

[noble shoal]

Hey. Does anybody knows how to deal with the attached error while loading the Inpainting checkpoint? My Comfy should be up to date. Might me that i am missing out on something 🤔

[vast galleon]

Awesome image! Do you mind if I use it for a color chaining workflow example for my I2I nodes?

[heady vale]

Are you using the DiffuserLoader? I believe the inpainting model is diffusers

[noble shoal]

Nope. You might be onto something. Lemme check

Unfortunately, that does not seems to be the case. It does not show up in the DiffuserLoader, even if its in the folder. Does it currently work with the fp16 checkpoint or am i supposed to load the 10gb file?

[heady vale]

I havent played with that yet. I just knew there was a different loader

is there no info on the HF page?

[noble shoal]

Sadly, not when it comes to using it in comfyui.

[heady vale]

There is a basic workflow example here
https://gist.github.com/comfyanonymous/6792ae5460c2cadb16d4bd60381c7e06

apparently the model goes in the unet folder and using a unetloader

[noble shoal]

Hmmm, i will check that out, but updating comfy has somehow screwed my AIT. I guess i need to start over again.

Thank you though

[upbeat summit]

yeah you need to go back to an older version of ComfyUI right now. at least this was the solution a couple of hours ago

[noble shoal]

Any solution how to do that on the standalone version. PS: I am not very good at the whole git stuff.

[hardy cipher]

time to git good at it, bud

[noble shoal]

I love programming and all the stuff, but my version control is called "Folders" 😅

[hardy cipher]

so turns out if you attempt to use a dict when a tuple containing a dict is needed you're not going to have a good time

[noble shoal]

Yeah. I have read the red wall of text. Not good.

[hardy cipher]

man, I'm fine with cloning and all that from github. but I had a real time adding to someone else's project, forking, etc. it was all new concepts to me

[noble shoal]

Ok, cheated my way around git. Pasted all the files from an earlier verison in my folder

Done. Don't touch it again....for the next days

[hardy cipher]

I was trying to figure this thing out for way longer than I'd like to admit. and it was squiggly brackets. one set of squiggly bois. and everything unbroke

the tuple thing I mean

[noble shoal]

In the AIT code or the comfi code?

[hardy cipher]

just wanted to have encode and decode nodes for myself that loaded the vaes, and they're singleton so only load the model once as long as you are only using the one model

still need to test different combinations to see if it breaks. but getting there

I think other custom nodes do something similar, but didn't want to search through them. plus I like learning more about how the different stuff works

not ait. haven't messed with that yet

[noble shoal]

Oh, ok. Well it seems like another nice rabbithole if you want to break things

[hardy cipher]

well I'm new to python in general. haven't really coded in years. but seems like a good time to engage with it

most I'd done in the last few years was some arduino scripting

and most of that was just me harvesting and modifying things I found on the internet

[noble shoal]

Ah Arduino, what a time consuming hobby. Have done that too. Even scaled the projects down to Attiny's.

[hardy cipher]

I was mostly using it to create custom LED stuff. ws2812b led strips and matrices. incorporated microphones and various other sensors

it's fun stuff. but yeah, time consuming. and I just can't juggle that many hobbies, lol. but I have phases of different things. it'll pop back up. still have all the hardware

[noble shoal]

You should use AIT, that saves time 😜 . But i know what you mean, i am also constantly switching between hobbies and projects. But hey, that keeps it interesting. Before SDXL i had at least a one month break from SD, and now i am sitting here and generating stuff at...let me check 02:57 o'clock

[hardy cipher]

well I had a good run with music gear, then transitioned to stable diffusion. still mess with the music stuff, but not as much

just need more hours in the day and it'd be easier

[upbeat summit]

should really build a eurorack comfyui setup

[hardy cipher]

comfyui is basically the ai equivalent of modular synth gear

[upbeat summit]

yeah - using node based systems is similar to working with studio gear for sure.

[hardy cipher]

I like the concept of modular synth gear, but really not as into it in practice. just too much, and it's not a cheap hobby

now if I could find node based software that simulates it. have a vr program that's a modular synth environment. you can just create whatever you want and hook it together however you want

[noble shoal]

Sill have parts worth hundreds of Euro laying around. Yes, i was under the impression building a modular synth from scratch is a nice idea.

[hardy cipher]

if you have the space, resources, and time, sure. but there are much more efficient approaches to achieve the same thing. still cool stuff though

[upbeat summit]

It's not cheap for sure. it takes up space, it needs lots of care and you can't just switch workflows quickly or work on a different project to just being able to do "File -> Open".

But using real hardware and having knobs, buttons, faders - haptics - is still very different to using a mouse or computer keyboard.

But when I work on a couple of projects simultaneously I mostly work digital only.

[noble shoal]

Whaaaaaat? Nothings better than a DIY 808 with shady soldering 🤣

[hardy cipher]

well I bought a jdxi a couple years ago having never messed with music gear before, and only used that for about a year. no software or anything else. well, used reaper to record audio

but didn't like the idea of going software first. I feel like people keep getting new things so they can make better sounding music. but that's putting the cart before the horse in my opinion

I have some other stuff now, but that's still my favorite music toy. mpc live 2 is cool too

[somber hill]

https://twitter.com/GozukaraFurkan/status/1698142548471001234

Furkan Gözükara (@GozukaraFurkan)

First SDXL DreamBooth result

looks like LR was too fast :) need more test obviously

Hopefully will do more tomorrow

It was really fast on single RTX 3090

Twitter · Sep 3, 2023, 1:15 AM

[hardy cipher]

just can't get into the same zone with software. it's not tactile

[noble shoal]

And thats why my main instrument is a good old guitar.

[hardy cipher]

jdxi is fun because it has all the oldschool roland sounds

808, 909, 707, cr78

wait, does it have 707? well either way

[noble shoal]

Please don't make me look into that

[hardy cipher]

lol. well it's considered best bang for your buck as far as entry level versatility. but then has pretty barebones effects and things. which was ideal for learning how they all worked

and the drumkits sound better than anything I've found on ableton or anywhere else

as far as the 808 and what not

[[<comfy.sd.VAE object at 0x0000024417579300>]], '10': []}, input_unique_id: 10, output_index: 0
!!! Exception during processing !!!
Traceback (most recent call last):
  File "C:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 159, in recursive_execute
    input_data_all = get_input_data(inputs, class_def, unique_id, outputs, prompt, extra_data)
  File "C:\AI\ComfyUI_windows_portable\ComfyUI\execution.py", line 37, in get_input_data
    raise IndexError("Output index out of range.")
IndexError: Output index out of range.

that's not what I want

[nimble heart]

I mean its probably famous if you were born during the great depression. Not so much if the Star Wars: The Phantom Menace is older than you are.

[hardy cipher]

is that a diss? lol

[nimble heart]

I just find it interesting how "the most famous thing of its kind in hollywood" is something so old my grandpa wouldn't have been old enough to see it in theatres. there's gotta be a cyclops monster from something more recent

[native knot]

I could take that reaction one of 2 different ways, @upbeat summit

[upbeat summit]

hmm... I thought about how I could best represent Samus with an emoji

[native knot]

No, I get it...I'm being funny.

[vital ermine]

OR if you are a cinema buff.

[native knot]

The ball was obvious. But since it was a "softball", I though I'd give you shit.

[upbeat summit]

they are really nice ancient illustrations

[native knot]

How do you like the Ridley one?

[upbeat summit]

interesting gargoyle design

[native knot]

It's definitely a different take, but most of Ridley is represented...just missing the tail.

[upbeat summit]

but you got two wings, which isn't always guaranteed in the latent space 😉

[native knot]

Yeah...it worked well.

Working on Kraid right now.

[strong field]

ooo what model

[upbeat summit]

currently testing NightVision XL beta 0.7.4.1 by @delicate kelp

[strong field]

looking forward to it

[native knot]

Ooh... nice. I've used 702 a bit.

[noble shoal]

Nightvision is awesome. I've been generating a lot with it so far.

[eternal fog]

Anyone got any idea's how to make this work? https://github.com/ssitu/ComfyUI_fabric

Comfy posted about it in his update blog, but I can't get it to work with SDXL or SD 1.5

And there's no example workflows

Seems like it doesn't support SDXL anyway, so I'm not sure I care about it.

[upbeat summit]

FABRIC looks interesting

[noble shoal]

Opened the link. Read "paper" and "Fabric". Thought about Minecraft. Time to go to bed. Good night fellas.

[upbeat summit]

I tried training an atheistic embedding with SD2 - did not work as good as I had hoped. this looks more promising

[upbeat summit]

good night!

[eternal fog]

I tried a web demo a bit ago and it worked ok

All from a screenshot of my messy ship on Starfield lol

And IPAdapter

[upbeat summit]

oh nice! haven't started yet.

the game graphics look great

[eternal fog]

This isn't creepy at all

[strong field]

got a fav model for people/faces?

[eternal fog]

this is generating some interesting images

This is a screenshot from a planet surface in IPAdapter

[upbeat summit]

yes, but it's sadly not public yet (Aether Real by @icy brook). I think base SDXL looks great with the right prompts. I heard that some really like DreamShaper XL but I haven't really tried it myself.

[strong field]

dreamshaperXL ok, does about 70% job for realistic

Lots of adjustments needed

[strong field]

Dang, been waiting on Aether Real lol

Honestly i like base alot too

[low hill]

I saved this poor animal from a lab...:(

[native knot]

I can hear it now...it says "kill me!"

[low hill]

I'm looking for a foster family, anyone?

He's still wagging his tail, he doesn't know any better

[low hill]

Collar is attached to his skin..

[eternal fog]

magic collar

[low hill]

It sure is

or piercing

hahaha

collar piercing

But he is/she/it is cute

[low hill]

New Half life?

[eternal fog]

You guard that wall

[low hill]

Walls don't fight back

[upbeat summit]

made with base SDXL. of course it depends on what you want to create. if you want to look into any of the prompts, just let me know and I'll get it for you.

[upbeat summit]

don't spoil the game 😉 ||j/k||
really cool. now we need this in real time as a graphics upgrade :D.

[eternal fog]

Pretty good at space stuff, when you give it space inputs

[strong field]

Wow I have never gotten results that good from base. You think the difference is in prompting?

[upbeat summit]

yeah. the sampler settings also make a difference, but it's mostly talking to the model 🙂

[strong field]

Dm me a prompt example on any of these I’m really curious where I can improve

[upbeat summit]

this is my analog retro photography prompt build I made for SD2. still works quite nicely with SDXL - just fill in the placeholders:
cinematic movie extreme close-up still of an epic scene of a [ETHNICITY] [OCCUPATION] in the [SEASON] at [DAYTIME], centered, looking into the camera, fog atmosphere, volumetrics, photorealistic, from a western movie, analog, very grainy, film still, kodak ektar, fujifilm fuji, kodak gold, cinestill 800t, kodak portra, photo taken by thomas hoepker

and this is a simple negative prompt I'm using with it to boost fidelity a bit:
photoshop, video game, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, 3d render, photograph

[upbeat summit]

this will give you the look of the last 2 images here. it's very stylized and got a retro look. the coherence can be really good with this prompt build.

[strong field]

Many thanks, can’t wait too try it out

[upbeat summit]

have fun 🙂

[vital ermine]

Here we go

[hardy cipher]

lolol

[hardy cipher]

woo, finally got this stupid node to work

[crisp owl]

Cave? Ocean? Waterfall? Sunrays?
Kinda trippy the more I look at it

[hardy cipher]

tis many things

[crisp owl]

accidental image, so unprompted for

meaning I accidentally had zero'd prompt condition feeding my flow

[nimble heart]

the cyclops are getting better by the day

[hardy cipher]

weird how that works

I guess looks it at the noise and is like "yes, this makes sense"

[crisp owl]

Well I do have the preconditioning ksampler, so it has a very basic latent image to work off of

so basically that was fed to the main process with zero'd out conditioning

basically daydreamed the final image

[hardy cipher]

I think that'll be my next custom node rabbit hole. conditioning. haven't messed with that yet

[crisp owl]

I've been working on putting my flow together nicely and working properly with switches

[nimble heart]

make a node that interprets things like "BREAK" and "AND" and builds a graph of the appropriate conditioning nodes like in auto1111

[hardy cipher]

just got my encoder/decoder with vae loaders built in figured out. really not that complex, but I'm new to such things

[nimble heart]

I was gonna but doing the processing for splits like [a:b:0.5] seems like a pain

[hardy cipher]

a node that builds nodes?

very meta

[nimble heart]

that's literally what my Principled node does

just builds a whole workflow for you depending on the input

[hardy cipher]

whoa. I hadn't even considered that. that's pretty legit

[nimble heart]

yea it's on that node manager thing as bsz-cui-extras if you wanna play with it

[hardy cipher]

I basically knew little to nothing about any of the code side of things when I started messing with them a while back. but doing my best to learn

[nimble heart]

source for the principled node is here
https://github.com/Beinsezii/bsz-cui-extras/blob/master/bsz-nodes/bsz-principled-sdxl.py

[hardy cipher]

I have a few on the manager

nothing too special though

[nimble heart]

you can see in the source it literally just imports a bunch of factory nodes and chains them all together

conditionally

only real processing it does is calculating resolution targets

[hardy cipher]

that's quite impressive. I might have to look at this code for a while to really fully grasp what you have going on, lol

[nimble heart]

its all pretty normal except for the fact that it can split hires passes into multiple tiered iterations which is kinda cursed

so if you're going 1024 -> 2048 you can say split it into 1024 1536 2048 so the jumps arent as big

that's 100% cursed

but the rest is fine

[hardy cipher]

lolol

I wonder if I used that one? if it's the one I'm thinking of, it was definitely cursed, at least on my humble hardware

[nimble heart]

idk it was only added to the manager thing like a week ago

[hardy cipher]

ahh. I added mine a few days ago. yours are vastly more impressive though

[nimble heart]

the author of the manager opened an issue saying that I'm doing everything wrong and I should feel bad because it's breaking his manager and I'm like "what manager"

[hardy cipher]

wat? how would it be breaking the manager?

[nimble heart]

the manager works by importing the whole repo as a python module I guess

and all my stuff was nested in subfolders

so I did the very 100% NOT cursed thing of making an init.py script that just scans everything in the folder and forwards it

so now you can just clone it and import the cloned repo

[hardy cipher]

my nodes consist of a moderately pointless math node that calculates any aspect ratio at any pixel count with any amount of rounding you want, two lora stacks, and a weird canny edge node. yours are cooler

but might have to work up to that level of things

[nimble heart]

my repo also has a math node that auto-calculates resolutions

and a image math node

[hardy cipher]

and just made encoder and decoder with built in vae loaders in them and did that singleton thing so it only loads the model once

probably been done before, but whatever

I want to refine the math node. That was my first one.

[nimble heart]

anyways yea you could 100% make a node that reads some text and splits it into multiple conditionings around "BREAK" and "AND" and []'s to build a conditioning tree out of them but that'd basically involve making an entire markup interpreter which is tediuos as hell

took me long enough to make the interpreter for my image math library

I guess if you skipped the brackets and just did BREAK/AND it'd be easy enough

since you could just use them as delimiters

ah fuck but that doesnt account for parentheses

[hardy cipher]

Yeah. May look into it. It's a bit complex as I'm a python novice. But I could get ai to give me broken code and then fix it

[nimble heart]

pls no

someone else was doing that earlier

feel like they were spending more time trying to fix things they didn't understand than just learning it normally

[hardy cipher]

Lol

[nimble heart]

maybe if you had it make just the interpeter

[hardy cipher]

Yeah. I've tried that route. It's profoundly frustrating

[nimble heart]

you could write some tests and just have it retry until they all pass

but making something integrated into all the torch libraries would be a nightmare

Feel like the best route for AI coding at the moment is to just use it like a personalized Stack Overflow

"examples of how to create a diff between two strings in python"

so you're still writing the code it's just scanning the docs for you basically

[hardy cipher]

man, gpt-4 kept breaking my math in something I was making a while back. just wanted it to help me figure out why I was getting an error. and for some reason it kept changing the math. every time.

[nimble heart]

math especially I wouldn't put through GPT unless it's a super well-known formula

gonna get off-by-one problems and weird edge cases

[hardy cipher]

well I told it "do not touch the math"

and it did over and over. I can't fathom why

[nimble heart]

saying that is more of a suggestion rather than an actual rule

[crisp owl]

"I understand you want to leave the math alone.
Here's what I changed"

[nimble heart]

chatgpt is just a big fancy word calculator. if it calculates that the optimal response is to say "Fuck your rules" then it will

[hardy cipher]

oh no, it didn't say it changed it. it just did

bard is the best. I was trying to see if I could figure out how to extract midi from some save files from a musical device of mine and it gave me all the details on different projects that did it

so looked them up, they didn't exist

[nimble heart]

real question would be to ask Comfy if there's going to be eventual bulit-in support for other markups in the conditioning nodes. There's already support for the (word weights:0.8) so maybe the rest is on a todo list somewhere

[nimble heart]

lol yea. chatgpt sites fake sources for papers too.

the links look real but they're all dead

[hardy cipher]

I'm curious about what's possible with nodes and what's restricted in the backend or whatever.

[nimble heart]

my principled node has previews and everything

[hardy cipher]

oh, bard told me that they did exist, but they were personal information and it couldn't share it

psychopathic liar

[nimble heart]

all the base facotory nodes are just in the nodes.py file in the root comfyui folder

so you can look to see what they do

[hardy cipher]

yeah, I reference those pretty extensivley

[nimble heart]

then just import nodes in your own node

[hardy cipher]

just to figure out how they input and output things

[nimble heart]

yea

or again import nodes and just use their functions directly

ez enough

[hardy cipher]

yeah, could do that too. I'm just getting my feet wet with it. I'll think of something and then see if I can pull it off

just making things that I think would be cool and I want to use but haven't found in other custom nodes

[nimble heart]

all my stuff is GPL licensed so you can use it as long as your source is available too

[hardy cipher]

I'd literally never made a githubproject public until a couple weeks ago, lol

[nimble heart]

even if they are found in other nodes if you have a slightly different spin on it just do it

I made my own automatic resolution calculator node despite 50 others existing cause mine is better for me

[hardy cipher]

well never even made one until a couple weeks ago

[nimble heart]

better get started then

[hardy cipher]

but gotta start somewhere

[nimble heart]

I expect a full conditioning interpreter by the end of next week

brb I need a 7up my tummy hurts

[hardy cipher]

well we'll see how that works out. I was originally thinking of just making a node that adjusted conditioning strength, but I guess I could take your ever so slightly more complex suggestion

[nimble heart]

adjust strength how?

[hardy cipher]

I don't know. I've never looked at how conditioning works so I don't even know if it makes sense tbh

[nimble heart]

you already can adjust strength with CTRL+arrows

ctrl arrow up makes a word or group stronger

ctrl arrow down weaker

thats why im wondering if there's plans for the other conditioning shortcuts you see in auto

[hardy cipher]

oh yeah, I was just thinking of a multiplier for the entire sum of the conditioning, not necessarily single words or portions

[uncut fiber]

i believe there is support for {red|blue|yellow|green}

[nimble heart]

highlight the whole prompt and ctrl+up/down

[nimble heart]

yea that's not conditioning that's just dynamic prompting

[hardy cipher]

well yeah. I actually didn't know about that option

[nimble heart]

i mean

I guess prompting is technically conditioning

but you get what I mean

[hardy cipher]

it's just a rabbit hole because I basically knew nothing when I started with a1111, and didn't really start learning about how it all works until I downloaded comfy

the more I figure out the more I realize I don't know

[nimble heart]

yea auto is hard to learn from

its all spaghetti

I tried to look into the memory leaks and gave up

[hardy cipher]

it's the reason I started learning python to unbreak errors

[nimble heart]

lol

yea its always out of necessity

I learned because I had this thing I did in GIMP all the time and it was like a 15-step process and I wanted it to be one button

[hardy cipher]

well I took a few coding classes back in my college days, but never really kept up with it after that. should have though

[nimble heart]

nah its fine

just leeroy jenkins into a project and google it whenever you get stuck

I didnt go to highschool nevermind college and look at me now: wrote my own music player through the power of googling and spite

[hardy cipher]

spite is a great motivator, and don't let anyone tell you differently

[vital ermine]

I managed two nude ones

[nimble heart]

yea QuodLibet kept crashing randomly due to GTK issues so I wrote my own clone of it entirely in TUI

[hardy cipher]

damn

[nimble heart]

like on the one hand I could've just reported the bug on the other fuck them I could probably write it better

[hardy cipher]

I saw dwarf fortress and for a second I was trying to figure out how you incorporated the game into your music player

which would be impressive. something to consider

[nimble heart]

no it's just the song_game.ogg file as its own album

I mean the game is entirely TUI so I guess I could just host it in a virtual window

[hardy cipher]

talk about some impressive programming. that game is too complex for my pea brain to grasp

[nimble heart]

but having a whole game in your player is cursed 200% so no

its already complicated enough adding mouse gestures n shit

[nimble heart]

become a DnD oc artist

[hardy cipher]

lol, not just a whole game. an outlandishly complex text based game

I guess they have the version with actual graphics now

[nimble heart]

the steam version just got linux support

sdl2 now

[hardy cipher]

which is cool becase those dudes deserve to get paid

[nimble heart]

also means it works on wayland

[nimble heart]

yea they made it cause they were living off peanuts and needed the money for medical bills

[hardy cipher]

I read about how they really weren't making much off of it before. enough to live, but not very well

[nimble heart]

then the week the game came out on steam they made enough money to live comfortably for the rest of their lives

[hardy cipher]

good. had to be an insane amount of work

[nimble heart]

the publisher said it exceeded sales expectations by many times

[hardy cipher]

I've downloaded it at least 5 times over the years, play for a while, realize the learning curve is ridiculous and it just fizzles out after that

[nimble heart]

the gui version is easier

actually has tooltips and an in-game dictionary

plus you can understand what the fuck you're looking at

[hardy cipher]

ahh. might check that out. haven't gamed since I discovered stable diffusion tbh

[nimble heart]

give your gpu a break

[hardy cipher]

not like it's the only thing I do, but just have time consuming hobbies. music junk, stable diffusion, learning about pointless things

[nimble heart]

add coding to it too

go from python plugins to automation scripting to writing whole ass softwares

[hardy cipher]

ahh yes. that too

[nimble heart]

the install arch linux and program half of your UI to be exactly how you like it

[hardy cipher]

well I'll expand my knowledge pretty quickly I believe. I'm actually motivated to do it this time

[nimble heart]

custom launcher, status bar, power menu, etc

[hardy cipher]

quit drinking out of spite a couple years ago and leaves so much time for activities

[nimble heart]

then script the fuck out of everything with bash. have multiple folders full of bash scripts for various tasks

[hardy cipher]

sure, I'll do that right now

[nimble heart]

who needs free time when you can write scripts for every task you ever do more than once a year

[hardy cipher]

yes. I agree

[vital ermine]

lol, made me laugh

[nimble heart]

so even with all that its still struggling to not make two eyes

[hardy cipher]

I just want to make weird things no one else has made before. not necessarily useful, but that wouldn't hurt either

[nimble heart]

make an anti-two-eyes lora

[hardy cipher]

dude, I tried to make a cyclops a bunch of times on my end. it's really not easy. never thought it'd be such an issue for the ai

[vital ermine]

I tried it on all AI except MJ and no go

even the interrogators will say cyclops with two eyes

I am like, wtf?

[hardy cipher]

accidentally flipped my positive negative conditioning and it made this

can't be mad at it

[vital ermine]

it must filter them thinking the image is a reject

[hardy cipher]

can that be adjusted? is it top p or top k?

at least with open source language models

the percentage of choices they'll consider

[vital ermine]

I mean if it creates the image it must think it is a reject

[nimble heart]

yes, tovarich. my favorite project behind the music player is my image math library, cause why not write weird psuedo-assembly code to modify an image?
I even made it a comfyui node lol

[hardy cipher]

can't say I know wtf is going on there, but that's pretty cool. I need to hurry up and up my knowledge though so I can figure out how to mimic mcmonkey's dynamic thresholding a1111 extension in comfy

that's what I miss most from a1111

[nimble heart]

creates a circular hue gradient centered at the top center of the image

[hardy cipher]

ahh. that makes sense

[nimble heart]

so it's just the X and Y * pi run through sin/cos then input into the hue. Y is / 2 so it's stretched to the top

[hardy cipher]

you should set it to vary relative to brightness distribution or something

[nimble heart]

The default for my workflow is a contrast fix

so actually kinda useful

there's no shadow buffer so it can't be used for edge effects

[nimble heart]

simple contrast fix with sliders to control. XL 1.0 kinda went overboard on undoing the XL 0.9 over-contrasting

[hardy cipher]

comfy also satisfies my desire for convoluted flows of image modifications

spent so much time doing that with photoshop

[nimble heart]

you knowwww

if you need ideas

[hardy cipher]

blending layers, doing weird things

[nimble heart]

you could write a node similar to my pixelbuster but it uses the G'MIC library instead

so it has every filter under the sun

instead of just assembly code

[hardy cipher]

oooh, yes. that's a good idea. I want all the options

[nimble heart]

g'mic is designed for that. it has plugins for gimp, krita, inkscape, etc

so should be easy enough to import using python ctypes then just feed it the image and script from the node

[nimble heart]

gmic has 590 filters included by default + you can script your own

good for embedding literally all the post processing effects

[hardy cipher]

hmm. makes me wonder what could be done with that

[nimble heart]

they literally have some games included so anything ig

though you'd be slightly more limited since you'd really only be able to use the headless interface I guess

have stable diffusion generate a maze then make gmic solve it

[hardy cipher]

yeah, I need to look into that more

lol

[nimble heart]

the pixelbuster node in my git tree shows how to convert the torch image tensor into something feedable to a C library which would probably get you started

[hardy cipher]

yeah, that library looks super cool. soo many posibilities

[nimble heart]

yea I use gmic a lot, especially before Gimp had all its gegl filters

never learned its scripting language though cause it's esoteric as fuck

I feel part of me dies every time I gaze upon it

and this is coming from a Rust programmer

[hardy cipher]

yeah, I don't think I'll be learning that any time soon

[nimble heart]

rust or gmic

rust is cool i recommend

[hardy cipher]

well either one really. but who knows

I'm not even completely sure what rust does tbh

[nimble heart]

so if you wanna do something C/C++ style like a high performance library or a static binary but are really used to your cozy python life of just not worrying about memory and being able to just pip install cool-library then rust is for you

[hardy cipher]

apparently cyclops + cyclops = 2 eyed monster baby head

[hardy cipher]

sounds decent

[nimble heart]

he got both his parents' eyes

biclops

[dense chasm]

controlnet canny model successfully run in comfyUI,replace the blonde with a chinese woman

[nimble heart]

instantly cancelled for racism

[hardy cipher]

big boi

[nimble heart]

have you considered not dividing by zero?

[hardy cipher]

yeah well, I have now

[lapis gale]

dragons are very cool with SDXL, too 🙂

[nimble heart]

just do what I do

if val == 0:
    val == 0.000001
thing / val

[hardy cipher]

I'm actually not sure what happened there. I just connected noodles

[nimble heart]

update comfyui

[hardy cipher]

indeed

[nimble heart]

there was changes to the ip adapter stuff recently

[hardy cipher]

I'm pretty sure I updated it yesterday, but he does like to update fairly often

[nimble heart]

literally their full time job

[hardy cipher]

yeah, it's impressive that a single person wrote it and maintains it.

I like that your node pack description is red texted

[nimble heart]

red texted?

[hardy cipher]

oh, the text in the description is red for whatever reason. I think because it has special instructions

[nimble heart]

???

[hardy cipher]

in the manager I meant to say

[nimble heart]

Ohh

never used the manager

I just added two files cause the guy who made it told me to

[hardy cipher]

well it definitely catches people's eye

[nimble heart]

like is it all red or

[hardy cipher]

that's cool. only the complex stuff gets the red text descriptions

[nimble heart]

Idk. Maybe he replaces the code markdown with red text

?

[hardy cipher]

I'll get a screenshot and show you

[nimble heart]

win+shift+s on windows 10+ to snip a region

[nimble heart]

Oh he wrote that by hand

I think

I've never said "auxillary" in my life

[hardy cipher]

lol. do you know him or did he discover your nodes somehow?

[nimble heart]

damn that must mean he combs github to manually add node repos to his big list

[hardy cipher]

ahh

[nimble heart]

Idk one day he just created an issue on my repo saying it doesnt work with his manager tool

so I made two init scripts so the whole repo can be loaded as a module then he just kinda peaced out

[hardy cipher]

but you never submitted them originally?

[nimble heart]

I didn't know what it was. I was arguing with him in the issue that putting the nodes at the root of my repo is stupid cause its much cleaner to symlink and he was all "that's improper for a management tool to do" and I'm like "tf is this management tool"

[hardy cipher]

bold move to complain to you when you didn't even submit it. that's pretty awesome

[nimble heart]

so I guess he just searched 'comfyui nodes' on github and found it

or someone else said it should be added. I've showed off the principled node in this channel a few times

I wrote the principled node before Sytan released his big workflow so a few people started using it

now everyone just uses Sytans stuff lol

[hardy cipher]

well you have to do it from your own github account normally

I mean for the manager

[nimble heart]

idk how the request system works I've never even looked at the manager tool's repository

[hardy cipher]

gotta do a pull

[nimble heart]

I just clicked on the dude's page and saw his first pinned repo was "ComfyUI Manager" with 400 stars and was like "ohhh that's what he's talking about"

[hardy cipher]

and all those other github things

[uncut fiber]

want to try car having donuts instead of tyres, but somehow i cant. Purely with prompt, going to check different model

[nimble heart]

nah. I literally have a bare comfy repo minus the models and my own nodes.

works well

if it doesnt I write a node for it

[hardy cipher]

chatting with copilot

lol, it's kind of a dick

[nimble heart]

probably easier to just inpaint the wheels as donuts instead of trying to coerce it into sourcing donut wheels from the 2 images on the entire internet that shows something like that

its the cyclops situation all over again lol

[uncut fiber]

🙂 yes just curious if i can get closer 🙂

[hardy cipher]

man, you can't make a cyclops. it's like dividing by zero

[nimble heart]

you can, General Awareness did higher up. it's just you have to be extremely meticulous in your prompt

things like "person" and "face" are tied to the idea of two eyes

so you basically just have to really tailor it to making a hollywood monster costume or something

[nimble heart]

If you wanted to make a large variety of cyclopses you'd probably need a lora to teach it the cyclops eye

or just inpaint

[hardy cipher]

I actually got some cyclops stuff by putting "only one eye not two" or something to that effect in the positive prompt. I've read that sort of thing doesn't work, but maybe that's more for 1.5, or just got lucky

[hardy cipher]

is that donald trump's one eyed cousin?

[vital ermine]

Hehehe

[vale eagle]

LOL

[nimble heart]

problem is even one eye will give you pictures of a person where only one eye is visible but they still have two

[uncut fiber]

Copax is bit closer

[shy kelp]

which one is the real strength of the model?

[hardy cipher]

both

[shy kelp]

is it clip or model, what is the different?

[nimble heart]

clip retrieval so you can see what the text encoder actually thinks of what you're telling it
https://rom1504.github.io/clip-retrieval/

[hardy cipher]

as I understand it one dictates how heavily it comes on when it does make an appearance, and the other dictates how likely it'll show up. someone else could probably word that better

[nimble heart]

model strength is how much it affects the entire model
clip strength is how much it affects how it interprets your prompt

[hardy cipher]

ahh, so I got lucky, used other words as well so surely was a combination of things

[nimble heart]

yea that github is kinda eye opening

[hardy cipher]

tbf it was just kind of a cyclops. lol

[nimble heart]

lot of people type "good anatomy" in positive, but when you search it on the clip retrieval it shows anatomical charts

[hardy cipher]

how do you do the clip retrieval?

[nimble heart]

that website I linked

Idk if it's 100% accurate of what stable diffusion's clip does but I mean it gives you an idea

another fun one is typing "realistic" and stuff gives you art, so if you want it to be literally a real photo it'll actually set you baCk

[noble shoal]

I guess its some of those hovercrafts.

[nimble heart]

man these blender tutorials are getting crazy

[hardy cipher]

I somehow missed that

man, this ipadaptor is not digging what I'm doing. changed it up and it's just a small novel of slightly different errors

[nimble heart]

idk what an ip adapter is and at this point I'm too afraid to ask

[hardy cipher]

it's like clip vision or unclip, but way more coherent in what it outputs

and uses the checkpoint model

[shy kelp]

ooh now i know

[nimble heart]

red paint

[stone fossil]

Yes blood is green.

[shy kelp]

do you have any workflow using ipadapter?

[nimble heart]

Too plastic-y for blood