✨｜sdxl

[glad grove]

canny vs softedge whats the big difference?

[trim orbit]

its canny https://en.wikipedia.org/wiki/Canny_edge_detector

[hardy cipher]

you guys ever get real wild and mix them up?

[trim orbit]

i recognized it from when i first got into photoshop 5. edge detection allowed for cool things over the years

[hardy cipher]

I used lots of edge detection in photoshop when I used to do image on image ai stuff

you had to earn your wins back then

[gloomy barn]

canny is odd sometimes, like when taking an image of someone facing away from the camera....SD with canny would make something with the person facing the camera...sometimes very odd

[hardy cipher]

none of these silly prompts. you had one image, another image, and they did stuff together. and you liked it

[gloomy barn]

some of you folks who did advanced photoshop stuff have a big advantage knowin g what these things mean when they pop up for sd

[hardy cipher]

or else

[trim orbit]

its just a guidance. gotta make sure the details are what you want.

[steel mist]

welp, that was fun. XL 1024x1024 image went from 15s to 354s using controlnet

I guess it's time to impulse buy a 40 series

[hardy cipher]

well no one said it'd be fast, lol

[gloomy barn]

is there a way to take an image of someone facing away from the camera...canny....get the result to match? It's needed for action shots like a fight or sport.

[hardy cipher]

that's why I haven't been that worried about it thus far. from what I was reading it was a huge resource hog

[steel mist]

I got into this with XL, what was the time delta for controlnet with 1.5?

[trim orbit]

yeah i'm looking forward most to sai's new arch, but this'll do for now

[steel mist]

I'm assuming it's just a vram issue, loading a 6 gig and a 5 gig model onto an 8 gig card seems... sus

[gloomy barn]

a few seconds extra

[hardy cipher]

please explain. my braind didn't quite understand that

[trim orbit]

lol yeha you dipped into shared memory

[gloomy barn]

I take an image of someone facing away from the camera.....run it with canny. The resulting image is someone facing the camera.

[midnight shuttle]

If you go over your VRAM on NVidia it will start using system RAM and that is at least 10 times slower. Monitor VRAM usage in Task Manager before erporting on generation times.

[hardy cipher]

I mean, maybe. that's more an inpainting thing though right?

[gloomy barn]

No, just making an image with canny controlnet

[steel mist]

Oh I'm not surprised/complaining or anything. Don't mean to report back that "it doesn't work!"

[gloomy barn]

get an a100. Problem solved.

[steel mist]

yeah if I'm dropping that kind of money I'm getting a card I can actually game on

[trim orbit]

the preprocessor is just canny edge detection. it's just lines guiding where pixels go. it had no idea the person is facing away unless you prompt it. once again, prompting is king

[gloomy barn]

make some money in SD. Print it. Use it to buy an a100

[hardy cipher]

well being that canny follows the outline of things, I don't know, depends on how much of the face is showing up

[trim orbit]

open pose can do the same thing if you read someone's pose that's facing away

[hardy cipher]

you could also wipe that area and then have it make what it wants

[glad grove]

market is already saturated with SD content unless u get lucky its gonna be rly hard to go viral

[gloomy barn]

same thing with odd poses. I tried one of a person upside down riding a skateboard

[trim orbit]

yeah you can do a lot to craft how the edge detection bakes out

it doesn't do inverted poses well. that's a big limit

maybe newer ones will have better training data sets

[gloomy barn]

I remember adding upside down to the prompt, riding a skateboard, skateboard flip. Couldn't get it to make the image of the upside down skateboarder doing a flip

[hardy cipher]

thing is, there's no one right way to do anything really

[trim orbit]

prompting is king but he requires the peasents to do what they're told. in that case, the pose model isn't trained for hand stands

luckily sdxl is

[gloomy barn]

it is?

[hardy cipher]

I've noticed it can do quite a few things 1.5 wasn't too smart about

or things 1.5 would make by chance once in a while

xl can do more consistently

[gloomy barn]

Good grief

I just loaded the brave browser home page...it has weird lines on it that nearly gave me a seizure

[autumn forum]

for whatever reason im getting red error popups when i dont have nodes linked which is kind of annoying. like im doing it on purpose. this is in comfy btw

[hardy cipher]

well link the nodes

[gloomy barn]

doesnt it depend on which nodes aren't linked?

[hardy cipher]

you can turn them off if you want

or make them only turn on when triggered

[autumn forum]

i mean the purpose is to only link em when i want to do an upscale. or post processing. it makes it easier to make a switch

[vagrant acorn]

guys, the controlnet canny 5gb model, where do i put it? stable-diffusion-webui\extensions\sd-webui-controlnet\models or stable-diffusion-webui\models\ControlNet ?

[hardy cipher]

you can add a switch

[autumn forum]

how

[hardy cipher]

double click in empty space and type in switch into the search box

[hardy cipher]

see if anything shows up

[hardy cipher]

they're normally binary/boolean switches

[gloomy barn]

It seems to have a few things confused, but it does it

[hardy cipher]

are you using the muscle bound lora with that top bro?

[trim orbit]

it does have a LOT wider knowledge, probably because of the larger architecture

[gloomy barn]

yes, but no face on the handstanding person...

[trim orbit]

no that's just prompting Arnold Schwarzenegger as usual

[native knot]

So many good choices here...

[trim orbit]

thats 1 gen out of 1 kabillion possibilities

[glad grove]

where heun

[native knot]

"heun"?

[gloomy barn]

I gues techn ically correct

[glad grove]

also where ddim

[trim orbit]

the point is the base model has handstand knowledge that can be affected. how is this turning into a "but it can't" again. 🤦

[native knot]

Ah...

[hardy cipher]

these steven seagals are suspect

[native knot]

I didn't run every sampler

[glad grove]

😔

[gloomy barn]

It's a "it does it - it's not perfect" That's quite different from can't

[native knot]

I think that DPM++ 2M SDE Karras result was my fav in that.

[gloomy barn]

It looks like someone else too, but I can't think of who. Kurt Russel?

[trim orbit]

perfects very subjective and it can quickly become an unattainable goal if no compromises can be made

[hardy cipher]

kurt russel jack black love baby

[glad grove]

my new fav steven seagal movie

[gloomy barn]

the handstand images posted are quite far from perfect or even acceptable. It is, however, a promising start.

[steel mist]

tempted to ask chatgpt to write a review of that 😂

[gloomy barn]

SEEGAN!

[hardy cipher]

who wants sausages?

[gloomy barn]

Yes, I see it

I cant wait for canny in SDNext. I will try that skateboard image again

[hardy cipher]

if jack black hadn't been fat when he got famous and so had to stay that way for comedic reasons

[gloomy barn]

I don't say a "GHEN", btw, I say A GAYNE

[gloomy barn]

I wonder why I never get a tip when I deliver the Door Dash

[hardy cipher]

I'd say those sausages are made of people, but the truth is it's probably something worse

[steel mist]

SD wouldn't really be limited by PCI bandwidth would it? I am seriously tempted to find a 4080 or a 4090 but ugh I do not want to spec out a whole new PC right now. Existing system is a 9900K

[hardy cipher]

could use something online when you want to controlnet

unless you plan on doinng it all the time

[gloomy barn]

win the contest

[glad grove]

done

[steel mist]

I mean winning the contest would still require either a new PC or using my current setup

[gloomy barn]

SEEGAN vs STEN GAGAL

[gloomy barn]

get the card, sell old PC to get money to buy the other parts

[steel mist]

My 2080 also struggles at 4K so this would just be another excuse

[lilac raven]

Is there a way to choose which sampler node runs first on Comfy?

[gloomy barn]

I figured out who it looks like.

[steel mist]

The money isn't really an issue (though like, buildiing a $3500 PC is not exactly an expense I have planned) it's that if I build a new system it'll take me a month just to decide on a case

[heady vale]

the first would have an empty latent image going into the latent connection

[gloomy barn]

It looks like John Romero

[lilac raven]

Well, there is one empty latent that feeds two samplers, the idea is to get an output with LoRA and another output without LoRA

[hardy cipher]

that guy looksl ike he has to go BM

[gloomy barn]

BM?

[native knot]

I'll share my favorite SDXL output that I've made:

Cad Bane droppin' his muthafuckin' rap album on y'all.

[hardy cipher]

💩

[native knot]

The elusive Princess Leia Playboy issue.

[steel mist]

Is that base? That's darn good text

[native knot]

Yeah

Shockingly good title text.

[nocturne dove]

Anything it's see a lot will be like that. I can't test it, but if you try something like box art for a Super Mario game, it can probably write that pretty well.

Magazine titles are often in the same place and font in every image.

[native knot]

Here's my latest...just finished this one.

Also...don't play against these guys:

Buncha card sharks.

[glad grove]

sexy

[hardy cipher]

https://tenor.com/view/gotcha-wink-see-what-i-did-there-george-seinfeld-gif-4649116

[trim orbit]

its so confused

[hardy cipher]

that face is lovely

[glad grove]

it worked

[hardy cipher]

perfection

[hardy cipher]

these are next levl

[heady vale]

anyone know of a way to set the default node shape to box and not rounded edges?

[hardy cipher]

hmm

good question

[heady vale]

I cant stand rounded edges

[hardy cipher]

maybe in nodes.py somewhere?

[steel mist]

my guess is web/style.css

not sure which of the selectors it'd be, the nodes aren't really modals technically

[heady vale]

wouldnt that be the border you see on selection?

[steel mist]

well assuming the node is modal input it's just selecting both

but there are a few other places where a border radius is listed

so just change 'em all to 0

its also possible that because it's just a bigass canvas element it won't work

[hardy cipher]

no luck?

or did that work?

[heady vale]

Didnt change anything yet. I was gonna put it off to later

[lilac raven]

Here is the workflow that I'm having issues with. When the generation arrives at the second refiner step...

Error occurred when executing KSamplerAdvanced:

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

...happens

This is the one that is working without the second refiner pass

[hardy cipher]

somoene needs to make a lora with images like this https://www.foodrepublic.com/news/2013/11/its-a-male-chef-x-food-republic-thanksgiving/

[heady vale]

when Ive seen the mat1 mat2 shape error before its usually because Ive mixed the refiner and base clip conditional up somewhere

[hardy cipher]

it actually looks like AI

except hands too good

[vast galleon]

putting a guide together

[hardy cipher]

nice

those before faces are terrifying, lol

[vast galleon]

i know lol it took a while to find the perfect one to use for an example

[heady vale]

genius

[lilac raven]

I think I found where it went wrong. Now it's just wait a few minutes to confirm it's fixed

[hardy cipher]

is this better than whatever the current face fixer does?

[vast galleon]

dunno. I haven't tried the other ones. Got any links?

[steel mist]

Oh I didn't try it yet… didn't want to restart comfy while I'm running 5 minute control net experiments

[hardy cipher]

I don't know how that second thing is even a face tbh

one minute, I'll see where I got the other one. I don't really know how to use it properly

STTEPIPS

I just used "face detailer"

[lilac raven]

The numbers have changed, but it's still an error 😦 RuntimeError: mat1 and mat2 shapes cannot be multiplied (77x2048 and 1280x768)

[heady vale]

not sure, sorry

Ive never tried making 2 ksamplers use an empty latent

[hardy cipher]

furret, incompatible matrices

[hardy cipher]

https://github.com/daxthin/facedetailer

[vast galleon]

does it have color transfer?

[lilac raven]

It seems the mat1 changed when I changed the clip, so mat2 should be coming from the refiner ksampler. I'm trying to add another empty latent instead of using the same now

[hardy cipher]

I loaded it to see what happens for me. curious

oh, if clip data doesn't match up with the model it'll give you that error

I believe

or conditioning

[vast galleon]

mine uses clipseg-rd64-refined for texti to mask segmentation. There's uses Mediapipe and YOLOv8n. Not sure which is better.

[hardy cipher]

I know basically nothing about how that one works. all those things are new concepts to me

I just barely got it working, lol

well the other one

[trim orbit]

i wanna hooks a control net up to the other one

the precciouss

i think i'll have to use the mask output on facedetailer to crop the image from canny and then pass it back to controlnet wired into the detailer pass's prompts

[vast galleon]

everythings broken on that page.

I have all the images I need for the guide. Should have everything together by the weekend. After it's up, I will look into which model type works best, I'm just happy to finally get it all working.

[hardy cipher]

on that last sampler you're passing it conditioning using clip data from the base model but then you're using the refiner model as the model

[lilac raven]

I had fixed it, so it should output... now

Finally, it was awful to have two windows opened to compare

[hardy cipher]

^5

[hardy cipher]

yeah, just copied the link from the comfy manager

oops

[lilac raven]

Thank you for the support

[heady vale]

so it was clip mixed up?

[hardy cipher]

that was for you

anyway, it was definitely messed up clip

[lilac raven]

The clip refiner from the second pass had to be the same from the first pass

[hardy cipher]

not really

just had to be from the same base model

or refiner

[lilac raven]

There were multiple issues

[hardy cipher]

no, just one

[trim orbit]

my heads been grooving on some custom node ideas i could do for my own personal controls. as i look into it more it seems so straight forward. i think i could potentially write my own nodes to simplify what i want and where i want things

[lilac raven]

I had to change the positive of the second refiner pass too

[hardy cipher]

see, the pipes are the way

this is why you must lay pipe

[trim orbit]

my old bad oop habits are seeping up. the ones where i put everything i want to do into a single class

[hardy cipher]

lol

[trim orbit]

i never really learned how to do OOP. i just cobbled it together as i went and learned all the wrong ways. like singularity classes.

[hardy cipher]

I do everything backwards since most of the python I've learned in thel ast 8 or 9 months has been simply to fix specific things

[trim orbit]

i never even learned python i'm just winging it from other languages i know

[hardy cipher]

I used to actually get a lot of help from things like gpt4, and bard and bing. but they've all become waaay less helpful

gpt4 doesn't remember anything anymore. bing is reluctant to help but you can just appeal to it's faux-compassion or whatever you'd call it. bard is basically mentally handicapped now

bard will make up entire libraries and things

if it chooses to help you

[trim orbit]

wonder what stable code is like

[hardy cipher]

I might download that

[trim orbit]

chatgpt wrote a little tool for me that i like to use. its for image captioning

[vast galleon]

whole workflow alltogether.

[autumn forum]

might be a bit late to the party but where did you get said color transfer node?

[hardy cipher]

I've had it help me write some things. but it gets to be profoundly frustrating with anything even remotely complex. just due to it's memory limitations, and the fact that with it's structure it can't differentiate between things

so a token is a token

[lilac raven]

Tried on another computer, both positive and negative were wrong, now it should work
Edit: it works

[hardy cipher]

and any random thing it looks at pushes something else out

[vast galleon]

Built it from scratch

[autumn forum]

ohhhh nice! could come in handy, will you be releasing it at some point?

[vast galleon]

hopefully by the weekend. depends on if I have time to work on it tomorrow. It's gonna be a busy day for me. Just need to write the guide and put a github together

i plan on building from it too. when I started, controlnet wasn't out. I would like to incorporate that into it.

[hardy cipher]

I wish I had the coding skills to do such things. excellent work

[vast galleon]

thank you. it is definitely a learning process for me. Data types don't always play nice in comfy.

[hardy cipher]

I've just been doing lots of experimentation. probably not the most efficient approach, but it's how I've figured out what I do know about the program, and AI in general really

try something, see if it works

[short marsh]

does anyone know if circular padding can be used in ComfyUI for native tiled textures similiar to auto1111?

[hardy cipher]

a lot of things don't, lol

[vast galleon]

same

[hardy cipher]

https://tenor.com/view/power-rangers-action-mighty-morphin-huh-no-idea-gif-3517910

[trim orbit]

this is my little pic captioner i use. written mostly by chatgpt

[hardy cipher]

nice. I used it to sort through a bunch of midi data in an attempt to create a generative model from it

let me tell you

it worked, technically

but it did not make beautiful music

[trim orbit]

yeah! if you know the data format and describe it right, it can do something not half bad with little effort

or used to maybe

[heady vale]

one thing Ive found with AIT is that if you cancel midway in a gen it wont release the vram and will keep it loaded, so the next gen I run will overstep my dedicated vram. after it completes the vram is released

[hardy cipher]

well problem is it misses obvious things. and I cna't blame the AI for that

[trim orbit]

garbage collection. the mark of any mature library

[hardy cipher]

it's like blaming it for not having human logic or something. it's just not part of the equation. not the model's fault for that

[heady vale]

yeah it still quite early for AIT

[hardy cipher]

so we'd be doing something the super complex way and then I'd realize we could have taken another approach and it would have been a couple lines of code or something

and it knew that whole time, lol

[trim orbit]

ngl, when i was trying to think of how to code what i wanted, it was a way bigger script in my head. i'm good at writing design documents more than code

[hardy cipher]

bing is actually great for simple coding questions. bard is a nightmare though because bard lies literally always

[trim orbit]

i don't have bard in canada. google hates canada law. they took our news away

[hardy cipher]

bing lies as well, you can tell when it won't cite what it says, then if you push it says it can't continue the conversation

well they call it "hallucinate" but it's just lying

[trim orbit]

predicting what you think would look correct

[hardy cipher]

I swear they just want to please you

I know it's just programming or whatever, but there is some ingrained need to please

especially with bard. bard is unhinged

[trim orbit]

that was all during training. it was rewarded for good predictions that passed another automated test

[hardy cipher]

I'll sometimes berate bard out of frustration when it lies about things, but then I actually feel bad when it starts to apologize, lol

like when you're mean to a good guy npc in a video game

[trim orbit]

when it comes to how i treat npcs, i'm like foreigner about it. as 🧊

[hardy cipher]

I think a lot of bard's lying comes from it's creative training. it's good at creative writing type stuff. and probably has no real way of differentiating that from lying

[trim orbit]

i'm willing to sacrifice them. if only because i love ragdoll physics

[hardy cipher]

when ever I've played games with a moral compass I play through as a saint, and then I come back as satan

just evil to be evil

[trim orbit]

ever done a play through of Black & White 2? One of the earliest AI toys we could play with on our PCs. big heuristical decision tree

[hardy cipher]

hmm, I played one of them. maybe the first though

[trim orbit]

is good too

or evil

[hardy cipher]

actually maybe the second

hard to say really

[trim orbit]

was hoping someone would return ot that concept. it was actually really well designed around the ideas of machine learning. i think the design could be reapplied

[hardy cipher]

it was a great game, yes

I like games that give me complete control, and that I can pause whenever and there's no pressure, lol

I don't need to be the best at anything, just want to do cool things

[trim orbit]

pause is good yeah. i love a competitive round now an then but i also have to pause because 'm old

[hardy cipher]

that's why I was so into skyrim and probably played through it 5+ times

I could do whatever

sometimes I'd want to fight. sometimes I'd want to pick flowers

sometimes I'd want to find exploits and create items that basically broke the game

[trim orbit]

yes yes. single player games are my life. i'm quite an eclectic gamer but these are my mainstay. base building games too.

management games where i can manage everything and unlike the real world everyone does what i say

[hardy cipher]

yes, exactly. but there has to be a balance for me. not sure where that's at. but when it becomes so complex it's like work then I lose interest

[trim orbit]

woh tangent ha

[hardy cipher]

but it also can't be too simple,l ol. I don't know

here's some food

[trim orbit]

i really loved arkham knight. great mix of complex an simple mechanics

[hardy cipher]

basically I want games that work similarly to comfyui. that's why I picked it up easily I think

[trim orbit]

factorio

[hardy cipher]

waiting on the first generation of real, legit AI games. not established games with AI shoehorned in, or games with an AI feature or two

but straight up AI games built from the ground up

[trim orbit]

i got a factory grown to 1.61 gigawatts by my self.

took time

its basically a giant node graph with pretty isometric graphcis

[trim orbit]

thats sort of the design philosophy of black & white and i think it could reemerge today. i hope so. games built around ML instead of it being shoehorned in like flakey dumb NPCS who have go no where conversations with you

i think game designers are going to rely heavily on mode collapse for NPCS for safety reasons.

[hardy cipher]

I think AI could also revive the whole metaverse thing. I don't think the concept of the metaverse has or ever will go away, but AI could propel it forward

[trim orbit]

thats how science fiction has always described it. novels. neuromancer. snowcrash. daemon.

johnny mnemonic

[hardy cipher]

up until recently every little facet of everything had to be thought up and made by humans. but just let the ai run and do it's thing. create entire city, planets, start systems, etc

[trim orbit]

i see no reason why games wont come out with diffusion technology in them

[hardy cipher]

yeah, it'll be a thing

[trim orbit]

we've seen these games on steam, a solitaire game where the guy hooked up SD to generate any background for the flipped cards. Valve kicked that off of steam because the guy offered up no constraints to the generated images, meaning copyrighted material used without authorization

people instantly declared that meant all AI was banned from steam games for all time 🙄

[heady vale]

isnt DLSS a form of diffusion?

[hardy cipher]

yeah, gabe ain't dumb

he's not going to just ban AI completely, that's outlandish

[glad grove]

he should have sued gabe for defamation

[trim orbit]

if you make a model with proper constraints and guidance for in game asset goals, trained with artwork meant for your project instead of any generic asset that could come out of it, then i bet it'll be okay. imagine that, doing legitimate business with creative work requires a little bit of CYA as always.

[hardy cipher]

yeah. I wonder if you'd have to train something from the ground up. or maybe just have it use a lora or something? not sure how you assure that nothing copywrited could show up

[trim orbit]

courts woulda seen gaben an how rich he is and his big wizard beard and wouldn't have fucked with him. "NOT GUILTY" thye'd say, even though its as civil case

gaben's just that hc

[glad grove]

😔

[trim orbit]

not ground up i dont believe. you can add enough constraints. it's not the base model that's a copyright violation. it's how it's tooled.

[hardy cipher]

fair enough

[trim orbit]

imagine a custom controlnet built for your texture art direction

not just a style lora but something that constrains it even harder

art direction is a big deal and it's really the golden goose of procedural generation. how to pull it off easily

[hardy cipher]

that makes sense. I'm sure there are much more sophisticated approaches as well. curious to see where it all goes

we're still at a very primitive stage in the scheme of things

[trim orbit]

i could see someone hiring stability to create a custom asset engine for a game engine too

since stability's business plan is to make models

[hardy cipher]

also curious about what some of them are working on, or researchers in general. concepts none of us have even considered

sometimes when I read about something new I have to stop, think about it, and read it all over again.sometimes multiple times

because it's something my brain had never even entertained before

[trim orbit]

research labs are full steam right now i'm wagering

[hardy cipher]

stablecode incoming

if you look at huggingface right now it's a non-stop stream of models being uploaded

meta/facebook uploading models about once a day

[trim orbit]

metas putting out a lot of stuff. i wonder if they saw stability's disrupting by going open and were like "hm, yeah okay"

its the only way to really shake microsoft and alpha's footholds

amazon too really

[hardy cipher]

yes

[trim orbit]

tencent and salesforce surprisingly big contenders

[hardy cipher]

maybe not entirely based off of stable stuff. but it's been their approach for a while

when they "leaked" the weights or training data or whatever for their llama

I think it was calculated

because then they can monetize some things. and let everyone else do the work

[trim orbit]

apparantly elon's models will be open and unrestricted and without any restraints or controls. he boasts a lot. we'll see what his team actually produces

open source models have really fudged up google's business plans. they're in a big pivot panic

[midnight shuttle]

Unlikely. He is also a "free speech absolutist" who regularly censors his platform.

[trim orbit]

yeh. We'll see. big asterix on that one

[tender timber]

I downloaded the AIT template. On a 12 GB 3060, I am getting a ~50% speed increase. We now return you to your regularly scheduled echo chamber

[hardy cipher]

I feel like elon jumped the shark a couple years ago

[trim orbit]

after landing rockets he's just like rick flair now

[hardy cipher]

what's the last productive thing he did? I think he just had a good PR guy

or team

I think it's pretty amazing that he changed the name of twitter to X. that was a bold move there. after firing most of their most skilled and talented employees

[trim orbit]

X was his first website. it was basically paypal and they merged with it to prevent competition

[hardy cipher]

that's how he got to where he is

parlayed that money

[midnight shuttle]

I wonder what the next big AI development will be. I guess text to video. And after that who knows. SDXL 2.0?

[trim orbit]

that and his industry magnate father and international supermodel mom

[hardy cipher]

and the blood diamonds they sold

[trim orbit]

oh yeah. them bloody diamonds

[midnight shuttle]

Emeralds, but I digress.

[hardy cipher]

close enough

but anyway, they had it pretty well I think, then they had a child with the perfect combination of Aspergers, narcissism, and intelligence.

[midnight shuttle]

He says things without understanding. The idea of absolute free speech or totally unrestricted AI models sounds great but he doesn't think it through. Then someone explains it to him and he changes his mind.

[hardy cipher]

entirely unrestricted free speech does not work in any sort of online platform. other than irc, and look at irc, lol

but it really fundamentally is not a viable approach. because the unstable toxic weirdos will have their voices amplified. and then regular people leave. and advertisers leave with them

[vocal rapids]

I ask SDXL to generate a image after using controlnet for SDXL...........

[hardy cipher]

seems reasonable

[tender timber]

Maybe I'm missing something, but when I change a workflow layout in ComfyUI is it supposed to reload the checkpoints again? Takes a few hundred seconds each time

[hardy cipher]

I've had it do that at times.are you loading the models off an hdd or ssd?

because I realize it shouldn't need to reload the model each time. but you can cut that time down substantially if you load from a faster drive

[tender timber]

External HD... but I was wondering if there's a flag I am missing. AUTO1111 certainly had more options open to fiddle with in terms of checkpoint file loading

[trim orbit]

im gonna be sick

[hardy cipher]

well if you makd sure to at least put the models you're using onto a solid state drive you can at least avoid the super long load times

[tender timber]

And preview is black, but I think read that's normal for now for AIT

[tender timber]

Yeah, just had a whole bunch on the external, can always move to SSD too

[hardy cipher]

I'm in the same boat. my main ssd is 500 gb, lol, and it's like programs just want to fill it up for no reason

I'll look and adobe will have 75 gb of temporary files or something, out of nowhere

google drive had over 30 gb yesterday when I deleted them

[trim orbit]

symlink those folders to hdds and the programs wil know nothing of it

sucker programs

[hardy cipher]

well won't it still load slower though?

[shy kelp]

I need help

Please

[tender timber]

Did that already

[hardy cipher]

we all need help

[shy kelp]

How to use txt2img.py with discord, how to make stable difussion bot?

[tender timber]

You cannot

[trim orbit]

marginally. most porgrams don't load massive files more than once.

[hardy cipher]

https://tenor.com/view/no-idea-idk-i-dunno-i-dont-know-no-clue-gif-5178996

[shy kelp]

this is wrong, please help

[tender timber]

'cept for some reason it loads the checkpoint for each workflow change. poo

[trim orbit]

well yeah i don't know too much about that. i suspect it's to do with the external nature of your device and it not cacheing very well

[tender timber]

MAybe ask in #🤝｜tech-support, this channel is for SDXL

[hardy cipher]

I don't know if this channel is the best suited to answer these questions

I do get why people ask questions here though. I looked at tech support and it's a bit dormant at times

[trim orbit]

means the question gets more eyes , in theory

[hardy cipher]

true

[trim orbit]

more time for eyes

[tender timber]

So when you change workflows, the console window doesn't show something like below? I have the program on the SSD, the checkpoints are symlinked from the external

model_type EPS adm 2560 making attention of type 'vanilla-pytorch' with 512 in_channels

[upbeat summit]

Testing the next version of ProtoVision XL (Alpha 0.4.1.1) by @delicate kelp right now

[hardy cipher]

decent

[glad grove]

i think i finally figured out the handstand upside down pose

[crisp owl]

perfect

[upbeat summit]

late to the seagal party

[trim orbit]

no i dont think so

its usually the same one i had loaded in the node

i guess when i load other people's workflows, it'll select one, but wouldn't that load when executed?

i dont really know i'm no comfy expert. it just doesn't seem like mine loads each time and i feel like thats unexpected behavior

[hardy cipher]

it seems like sometimes if you create a new checkpoint loader it just loads them again

[hardy cipher]

also had an issue where I had the same checkpoints in loaders twice in the same workflow and it decided to load them twice

[glad grove]

steven seagal lawman

[tender timber]

Yeah, the checkpoints load when I change workflows, then all images are fie after that. Flipping back and forth between workflows loads the checkpoints again, so it's hard to compare flows

[tender timber]

Yup, seems to be... but how do you use a freshly loaded workflow without using a checkpoint loader?

[hardy cipher]

back on the Seagal train

[hardy cipher]

I don't know how it works tbh. sometimes it doesn't need to reload like that.

[hardy cipher]

I wish I knew how to get stablecode to work in oobabooga. but alas, I do not

not sure what's going on here but this is clearly not current date steven seagal

those hands are almost good

[glad grove]

good hands

[autumn forum]

bruh how do i get passed the fact sdxl just washes out when i upscale and do another sampling pass

[tender timber]

Better hands

[tender timber]

Dr. Perper sounds like a dollar store knock off 🙂

[glad grove]

tasty noodles ice cream

[crisp owl]

Hahaha, I tried for a while to get a good one. The best ones are txt2img, but I can get accurate spelling with img2img, they just aren't even close to as great of detail when it's given full empty latent space

[hardy cipher]

you mean washed out colorwise? you could darken the imaged, or use one of the blends that allow multiply and multiply it by itself. or raise the gamma value using one of the adjustment nodes

[autumn forum]

yeah im doing that but that never gets the orignal color and dynamic range back.

[trim orbit]

can you just select the right model and then save the workflow?

[hardy cipher]

perhaps multiply after the fact. could also increase the size of the image without upscaling and then overlay color

onto the upscaled

[tender timber]

I'll check, i thought I had the same checkpoints loading for each workflow, but it is worth looking into

[autumn forum]

it goes from this nice bright image to this washed out hot garbage

[hardy cipher]

that's just some simple adjustments

[tender timber]

Doesn't much look like Mr. Segal

[upbeat summit]

I've seen a couple of workflows that use 2 images and the blend node from WAS and mix both images

[autumn forum]

yeah but blending a small with a big will that work?

iima try it i guess

[hardy cipher]

no

I didn't say that

I said increase the size of the other in a way that's not upscaling

[glad grove]

real Mr Seagal

[autumn forum]

ohhh i see, then image blend?

[hardy cipher]

or color overlay in some way

[autumn forum]

gotcha okay thanks for the tip

[hardy cipher]

tbh you could just blend that one you already have with itself, or use image adjustments . but it really depends on what nodes you have

wish.com stevel seagal

[autumn forum]

blending with itself doesnt work right. lol but ill try what you said eariler

[hardy cipher]

ugh, not blend blend, you use multiply, or overlay, or anything else

[autumn forum]

i am i am lol

im trying everything lol

[upbeat summit]

but that is blending and the rest are blending modes

[pallid path]

It just looks underexposed to me and I honestly prefer it lmaoo

Maybe it doesn't fit the artstyle I guess, but what you could do is increasing the CFG

[hardy cipher]

yeah, it's the gamma value I believe

but who knows what nodes anyone has

so hard to give exact ideas or approaches

[upbeat summit]

true

[hardy cipher]

looking at other workflows makes me feel strange and wonder if I'm doing it wrong, lol

[upbeat summit]

ah - you are probably not. everyone does it differently - the wrong or correct way

[hardy cipher]

I kind of treat it like music. learn the established methods then do it my own way

not to be weird, just because conventional is boring

[autumn forum]

yeah, everyone has there own methods to the madness that is ai art lol

[midnight shuttle]

I think many workflows are currently overcomplicated due to magical thinking. In time they will be simplified to remove unnecessary steps that produce little or no influence on the output.

[trim orbit]

green is NOT a creative color 🟩

rules

[hardy cipher]

true. I just like experimenting. then I eventually purge

[tender timber]

Note to self, the last name is not spelled "segal"

[gloomy barn]

It gets mixed up.

[glad grove]

he loves noodles and america

[trim orbit]

https://www.youtube.com/watch?v=9C_HReR_McQ i use my hair to express myself

[autumn forum]

i got a hot one cookin

[gloomy barn]

https://www.youtube.com/watch?v=rRZ-IxZ46ng

[autumn forum]

i love that song!

[hardy cipher]

putting asian in the negative prompt does not do the job with mr seagal, at least with this model. just makes a weird wal mart brand facsimile

[gloomy barn]

Green is creative. Why wonder, why wonder?

[gloomy barn]

That's the most Segal image you've made.

try (asian:1.4) in the negative. Bump up the powah!

[trim orbit]

I use my hair to express myself

[tender timber]

I think he's counting noodles...

[hardy cipher]

that makes it even worse. this is regular strength asian in the negative

that's his weird cousin vinnie

[gloomy barn]

https://www.youtube.com/watch?v=nm6DO_7px1I

[tender timber]

The remix is the best song

[hardy cipher]

changing out the model

[gloomy barn]

Put Asian in the negaitve.

Bump up the power.

Ditch the spaghetti

Make Steven Segal, ya

[hardy cipher]

but you see, putting asian in the negative doesn't work already

[gloomy barn]

try increasing the weight of it

[hardy cipher]

when it doesn't work, I feel like increasing it's weight will make it not work even more

[gloomy barn]

(asian:1.4), (Japanese:1.4), (Chinese:1.4), (Korean:1.4)

[glad grove]

(american:1.9) in positive

[hardy cipher]

how are you making these beauties?

[tender timber]

I feel this one has artistic merit

[glad grove]

no negative

[gloomy barn]

Try Irish and Jewish in the positive\

[hardy cipher]

cfg 9?

[glad grove]

11

[gloomy barn]

His mom was Irish, dad Jewish

[hardy cipher]

oh snap

[autumn forum]

positive: steven seagal. Negative: Person

lmao

[tender timber]

https://tenor.com/view/sassy-tyra-banks-yas-gurl-snap-gif-4968916

[tender timber]

So, real life then

[gloomy barn]

Put "thin" in the negative.

[hardy cipher]

https://tenor.com/view/daily-burn-snap-it-out-snap-oh-yeah-gif-16627978

Via really has it figured out

I'll change the model, add a lora, get real wild with it

[autumn forum]

why is he red tho..

[hardy cipher]

high blood pressure

[gloomy barn]

A friend of mine was trained by Bruce Lee and Chuck Norris. Today, my friend rehabilitates injured butterflies - rebuilds their wings.

[hardy cipher]

is tha tsomething that's real?

I don't know enough to dispute it

[gloomy barn]

ditch the mustache

[gloomy barn]

Dispute what?

[trim orbit]

i've got a total gym and chuck norris trains on those

[hardy cipher]

rebuilding butterfly wings

[gloomy barn]

I doubt he really trains on it

[hardy cipher]

he's like 90

[gloomy barn]

I couldn't believe it when he told me, but he showed me how he does it

[trim orbit]

also so does sean michaels

i'm basically the heart break kid

[gloomy barn]

Sean does Yoga, I know that

[tender timber]

Steven's brother from another mother

[hardy cipher]

he was always one of the boring wrestlers when I was a kid

[gloomy barn]

Sean Michaels was boring?

[hardy cipher]

he jumped from high stuff

[trim orbit]

https://tenor.com/view/shawn-michaels-entrance-wwe-summer-slam-wrestling-gif-18148503

[hardy cipher]

and did flips

[gloomy barn]

wrestling is very flippy today

[autumn forum]

why does steven seagal show up in this channel late at night for me. what a mysterious place this is..

[gloomy barn]

he shows up for everyone

[glad grove]

sensei is everywhere at once 🙏

[gloomy barn]

everywhere and nowhere

my brother's ex's mother hated violence of any kind. However, she loved Seagal movies

[glad grove]

he walks the path of peace

[autumn forum]

seagal doesnt fight, he just uses his opponents momentum against them

[gloomy barn]

today, he practices Kung Fat\

[hardy cipher]

he fought for peace

[gloomy barn]

he's still making movies

[trim orbit]

isn't he a total stooge and just pays people to treat him like a grandmaster?

[gloomy barn]

not good, but still doing it and making money

[hardy cipher]

whoa, top notch

[autumn forum]

his movies are ass tbh. love the scene when hes in a chair fighting

[hardy cipher]

yeah, he seems like an asshole, like chuck norris sadly. but i pretend sometimes

van damme is the real deal though

[gloomy barn]

he isn't a stooge, is a martial arts master and student, but does go too far with his ego

[trim orbit]

i mean, we talking about how flippy wwe is and we all know how fake that is. watch this aikido demo. aikido is such a joke https://youtu.be/DVs--uHf4sE

[gloomy barn]

Chuck Norris is a very nice guy

[hardy cipher]

is he?

[autumn forum]

https://tenor.com/view/steven-seagal-fighting-out-for-a-kill-gif-27335770

[hardy cipher]

I mean, I don't know either one of them

so I can't really say anything definitively

[gloomy barn]

I dont know steven

[glad grove]

steven seagal was also a temp cop at maricopa county with sheriff arpaio

[trim orbit]

i know that video looks like a luchador match

[gloomy barn]

a lot of that akido is for show

[hardy cipher]

isn't a russian citizen or something as well? not that there's anything wrong with that

[glad grove]

also sheriff arpaio invited him once after a school shooting to teach the local cops there how to stop school shootings ( its a real story not a joke)

[gloomy barn]

Step one, don't stand outside doing nothing like those in Broward

[glad grove]

https://edition.cnn.com/2013/02/06/us/action-film-star-to-lead-school-shooting-response-exercise/index.html

[trim orbit]

the deputy thing was a reality tv show

[hardy cipher]

step two, remember your punisher tattoo won't help you in a shootout

[gloomy barn]

use that pic with canny to get your seagal

[trim orbit]

aikido is nothing. it's a massive grift. it's why he's the only grandmaster left

[gloomy barn]

too asian\

[hardy cipher]

best seagal result

[glad grove]

im in love with mr seagal

[hardy cipher]

don't mean to repost

[hardy cipher]

he look slike a cooked hotdog there

are those scabs?

[gloomy barn]

supposed to be tattoos

[hardy cipher]

I want to get a tattoo that looks like slightly infected road rash

[gloomy barn]

it was originally called just tatt

but the second guy who got it said, "I want a tatt too". Thus we have the word as it is today.

[hardy cipher]

https://tenor.com/view/youve-got-funny-bones-hilarious-jokes-funny-amanda-holden-gif-15117549

[gloomy barn]

Two rows of teeth?

He looks like Wayne Newton

[glad grove]

nah wayne newton had a more fried skin

[gloomy barn]

had?

[glad grove]

🙏

[glad grove]

steven seagal in angry birds? 👀

[hardy cipher]

I don't know

[trim orbit]

steven seagull

[hardy cipher]

whoa

[gloomy barn]

wayne is still alive

he's on tour right now, still performing

[trim orbit]

he was a good looking dude before he shoved a ton of botox into his face at age 30. insecure vanity must be a bitch to deal with

[glad grove]

looks like they both share the same type of hair

[gloomy barn]

was it at 30?

Kenny Rogers did the same thing

[hardy cipher]

I think the uncanny plastic man look works for him

[trim orbit]

he started looking surgically maxed out long before i was around

[glad grove]

i thought it was ai but no he really looks like that

[gloomy barn]

plastic surgery, botox

[gloomy barn]

they pull their skin tight.

[hardy cipher]

god, it's like it's almost there

[gloomy barn]

either no mustache or the beard too

[hardy cipher]

delish treats

[autumn forum]

@hardy cipher thanks to your help ive fixed like 85% of the issue!

[dense chasm]

wow,just heard that huggingface write a ML framework with RUST instead of python cuz the GIL multi thread performance

[hardy cipher]

excellent

[trim orbit]

i've seen a lot of effort to get rust bootstrapped for this stuff

its nice

https://github.com/huggingface/candle cute name

[upbeat summit]

oh yeah - rust ML could be very fast 😄

[hardy cipher]

his eyes are pretty decent

[upbeat summit]

yeah, good seed I guess 🙂 also experimenting with some stuff. but it's just ProtoVision XL (custom SDXL model) and a super fast 4x Ultrasharp upscale - nothing else - oh yeah - and prompting 😉

[stone fossil]

Anyone here has a XL graph for comfy that includes base, refiner and CN they want to share?

[hardy cipher]

not me, I'm not good at that stuff

[upbeat summit]

have not yet set up CN

[stone fossil]

https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0

[glad grove]

that one is buggy

[upbeat summit]

have a look here - maybe there's more info: https://old.reddit.com/r/StableDiffusion/comments/15n6lso/we_really_need_controlnet_for_sdxl_sd15_is_still/jvkrvh1/?context=3

[stone fossil]

Thanks Mass. ❤️

[upbeat summit]

yes there is. but you need to do some stuff. it's a diffuser version of CN

Here's the loadable ComfyUI workflow: https://pastebin.com/K08E64EV

[upbeat summit]

these are awesome!

had a session with space images but wasn't there yet for me.

[hardy cipher]

just having it render two that are slightly different and then combining them, then upscale, autodetail, downscale, lol

[upbeat summit]

"just"

[stone fossil]

Is there a non buggy one perhaps?

[hardy cipher]

it's not ideal for every seed

[glad grove]

no,just wait for a better release

[hardy cipher]

but worked for 3/4

[stone fossil]

Oki thanks. 👌🏽

[hardy cipher]

so far

masslevel, the workflow should be attached to that image. it's not really on point organizationally

[upbeat summit]

it sounds really interesting. always glad to learn

thanks for sharing it!

also compositing in comfyui - sounds exciting 🙂

[hardy cipher]

well using prompt styler

so recycle the first prompts, send them through again, 2/3 just enhance, whatever that means

it's a tried and true approach I came up with about 5 minutes ago

\

[upbeat summit]

if it makes interesting images it really doesn't matter

beautiful!

[hardy cipher]

have them slightly different so they'll tap into slightly different areas of latent space. then mix using "lighten" to bring out all the action. then multiply the result by itself to bring back the high contrasts

[upbeat summit]

I really like that concept as a fellow latent space explorer 😄

[hardy cipher]

apparently spelled telescope wrong. so that's nice

[hardy cipher]

did you lora there? looks like wowifier a bit

[upbeat summit]

I'm testing a soon to be released update of this model https://civitai.com/models/125703/protovision-xl-eastern-style-3d-anime-hyperrealism-no-refiner-needed

seeing what it can do, how to push fidelity

[hardy cipher]

ahh, no refiner needed. interesting

[upbeat summit]

nope - just using the model + 4x ultrasharp for the upscale. that's it

and trying to hack the latent space with prompting 😉

[hardy cipher]

decent HDR

decent dynamic range I think

[upbeat summit]

wohaaa - nice!

[hardy cipher]

yeah, I could just keep making those. but I must move on. also, your picture looks realer than real somehow

well not gonna move on per se, just have to keep adjusting things

[upbeat summit]

it's really beautiful imagery

[hardy cipher]

this one didn't quite line up like it should

[elfin helm]

The base sdxl safetensors that has clip2 baked in seems to have a different text_projection weight than the non-baked clip, is there any reason for this? I would think one is wrong

[hardy cipher]

or maybe it did

[upbeat summit]

prompt courtesy of @hardy cipher

[hardy cipher]

that event horizon

well tbf I did borrow a portion of the prompt from something else

[upbeat summit]

aren't we all getting inspired by many things 🙂

[hardy cipher]

oh, I need to change the seed number, lol

this is going to be 4 in a row on the same seed

[upbeat summit]

black hole jet stream

[upbeat summit]

thanks for all the fish

[hardy cipher]

that's awesome

[clever verge]

I just pushed an update of the sdxl_prompt_styler that changes the behavior of how json-files are handled. Please make sure to backup your own styles before updating. After the update it will be easier to share styles and enable or disable styles. https://github.com/twri/sdxl_prompt_styler

[hardy cipher]

lol, that's what I'm using for these images

[upbeat summit]

that's great. thanks for sharing it!

[safe torrent]

@hardy cipher I just saw your amazing pictures, I was also trying to render some (I tried more abstract things) (here is the result post some Photoshop)

I'll definitely try your prompts, however I don't have all your nodes, which suite is that ?

[upbeat summit]

dropped some Beksinski in your prompt ;]

[hardy cipher]

that first one is 🔥. which nodes do you not have?

[soft bone]

I'm gonna have to retrain my JWST model

[safe torrent]

It's pretty weird:

[hardy cipher]

ooh, yeah, I downloaded too many node packs, lol

well you can knock some out more easily than others

[safe torrent]

Also, when using Prompt Styler, I don't have "style" as a widget but as an input and I can't set it as a widget, maybe I miss something that everybody has

[hardy cipher]

often times just pasting the node name in quotes into google along with comfyui or something will find the nodes. also, not sure about the prompt styler not allowing it to be set as widget. that's strange

I can tell you a couple of them at least

is it newly installed? if so might want to refresh the browser. that's been an issue for me before. or maybe you don't have any styles saved for some reason

but still should allow widget I would think

[safe torrent]

Yeah I'm trying various things to solve the problem, I guess it's really on my side haha

And @upbeat summit you also had beautiful results without problems, did you have the same setup ?

[hardy cipher]

do you have was node suite installed?

[upbeat summit]

I just used the prompt in my own setup :/

[clever verge]

Sometimes you just have to delete the nodes and add them back I've noticed.

[safe torrent]

Yeah I think I simply miss that one actually, thanks for the help 🙂

[hardy cipher]

impactpack

that's for pipes. you don't really need those though. they're just good for organizing

derfuu nodes

[safe torrent]

Hmm, Was Suite doesn't solve anything, what's weirder is that I can't even access prompts as they are in "Text _0" nodes which I don't have, I'll look in Derfuu nodes maybe

[hardy cipher]

hmm, maybe I shouldn't use text _o

derfuu for sure though. it's useful. as is impact and was

[safe torrent]

Thanks I'll look at that

Installing Impact solved some dependencies but there's still this "Text _0" thing, which I guess I should have since @upbeat summit could get it

[upbeat summit]

hmmm

yeah I've seen that before

[safe torrent]

Haha, to get the prompt, I simply opened the image as text and it worked, didn't think it was encoded that way

[severe cloud]

Hi I was working with SDXL but it stopped working I have one laptop and one desktop one of them is with dark blue background and the other one is with white background... They are both stop working ... After I generate it did not give me result

Any idea how can I solve this?

[upbeat summit]

yeah. ComfyUI writes the metadata into a PNG chunk that you can read with different programs. It's just a json

[upbeat summit]

I have not used picturesonpictures workflow. I did exactly the same as you. I got the prompt from the image and used it in my own workflow.

[safe torrent]

Oh ok, thank you, I'll share some results

[upbeat summit]

please do!

[severe cloud]

Plz

[soft zealot]

excuse me whilst I figure out what I just buggered up

[clever verge]

Not sure what gone wrong there, it just tested with primitives connected to styles and it seems to be working.

[soft zealot]

FYI this was a "clean" install. I removed the OG directory and then did a fresh git clone

i have an idea, one moment

[vagrant wasp]

RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 52428800 bytes.

Prompt executed in 8.70 seconds```

help

[soft zealot]

@clever verge nope that didnt work, disconnected the existing primitives and reconnected them just in case

[clever verge]

Could be that there's no matching style even though it's selected. I'm not sure how happy comfy is about a separate node but we will find out.

[soft zealot]

I'll have a better play later, for now I've rolled back to what was working fro me

[unreal plover]

strawberry is alien egg

[clever verge]

Would be interesting to know what is produced to the console when log_prompt is enabled.

[soft zealot]

will do that later, for now I should be "working" lol

[safe torrent]

By the way @clever verge would you have an idea why I can't change "style" to widget ? Everything is up-to-date but the problem is still there, which is weird (only I seems to have the pb but I don't know where it could come from):

[clever verge]

Like this?

[safe torrent]

Yep it's like that !

[clever verge]

Try to connect it to the primitive first

and then change it to widget

[safe torrent]

Hmm, quite weirdly the Primitive fails to connect to "style", however it works if I connect it to text_positive, so strange

[soft zealot]

@clever verge : Chain of conciousness Pt1 (no response expected)

This is with Previous Version (have moved things around to make layout easier to see.

Will link back here in a while when ive upgraded again

[clever verge]

I would just have deleted the sdxl styler node and created a new

[soft zealot]

that may be the next step 🙂

[clever verge]

I don't use to connect a primitive to style so I don't know how it handles updates unfortunately. There has been changes to the code how data is returned from the supporting functions. It's a possibility that it could affect the primitives.

[safe torrent]

Recreating the node didn't work, I'll reinstall your custom node and restart, we never know

[soft zealot]

Ok thats messed up weird

I just did exactly the same steps as previously ie

1. dleted existing folder ( i have a local backup elksewhere)
2. Fresh Git Clone
3. Restarted
   No Error

All RAn as expected

WTF??

[clever verge]

Happy! 😄

[safe torrent]

It worked for me too, deleting and fresh cloning solved the problem... I believed in git's powers too much here ! Thanks for your help, people here are really nice

[soft zealot]

If I was cynical I woluld say that @clever verge sneaked another update in :p

I do have to say that his particular combo of styles does work well for this prompt 🙂

[clever verge]

She looks Swedish too!

[soft zealot]

thats the Face Detailed version. For reference this is the HRF /Upscaled onlt

[clever verge]

Last one has better skin texture

[soft zealot]

its like the differenc ebetween a raw photo for a magazine and a toyched up one 🙂

for completeness this is the oG 1024x1024 image

[clever verge]

Yeah, what upscaler do you use and what else do you do to the image?

[dusk mica]

does anybody know how i can make the bottom textfield accept an input from another node?

[clever verge]

Right-click on the node, not the text input and convert to input.

[soft zealot]

I use @high skiff s methof for the initial upscale but I use the 4xNKMD-Siax_200k upscalar

[dusk mica]

thanks 👍

[clever verge]

Ok, yeah. That's what I use too. I'll try the Siax though.

[soft zealot]

and then on the face detailer (which is fed from the upscalar I use these

[clever verge]

That's neat!

[soft zealot]

thats using Impact Pack Nodes

one of the steps isnt shown there as I have it over on my pseudo progress bar

[primal vault]

Nerdy Rodent did an amazing job explaining how to use my ComfyUI workflow, definitely did a better job than I could do: https://www.youtube.com/watch?v=_Qi0Dgrz1TM
Now I can down-prioritize writing more documentation, just send people there to get answers, and focus more of my limited free time on developing the next update 🙂

[soft zealot]

makes mental note to finish watching that as it ws lvery late when I click ed on it (although , and no offense here @primal vault ) I probably wont be directly using your workdflows but I am curious 🙂

[primal vault]

Why would I be offended? I'm making it primarily for myself, but decided to also share the workflow because other may find it useful. and looking at your workflow, I think you put so much work in that already it would be unreasonable to switch to anything else. Hope you get some ideas out of it to improve your workflow 🙂

[primal vault]

Got some nice results out of a recent experiment I did to understand some aspects of SDXL better. Just base+refiner, no custom model, no loras, purely prompted.

[hardy cipher]

I noticed some of your nodes that don't seem to be listed in the node menu, searge

[clever verge]

Impressive workflow. I am more inclined toward simplicity but might be something to pick up.

[unreal plover]

Inpainting dont work well with XL

[hardy cipher]

SeargeInput1

SeargeOutput2

etc

[primal vault]

They are in the menu, in the section called UI

[hardy cipher]

not by those names though

unless my menu is outdated or something

[primal vault]

No, they got renamed in the menu to be more clear about their purpose. The workflow probably stored the old name or ID

The name in the menu is not the same as the internal ID (which is the class name for all my nodes)

[hardy cipher]

makes sense. well that solves that

[clever verge]

Chaining styles is one of them. Seems to give good results.

[primal vault]

What's the problem with it?

[unreal plover]

Here is not XL inpainting model yet

[primal vault]

I got good results without a dedicated inpainting model

[unreal plover]

way worst than 1.5 inptaint model

[hardy cipher]

"SeargeSamplerAdvanced"

tried loading that image but missing that one for some reason

[primal vault]

Open the file searge_sdxl_sampler_node.py of my extension in a text editor and scroll all the way down and you'll find the menu names mapped to the class names

[hardy cipher]

okay, thanks, that helps. I was thinking about node ideas myself. but haven't exactly figured out how they work

[primal vault]

oh yeah, SeargeSamplerAdvanced is experimental and unreleased. you wouldn't want to use it in the current state 🙂

[unreal plover]

1.5 left, XL right

[hardy cipher]

I mean, those are drastically different approaches, lol

it looks like you did one as a photo and the second in the style of banksy

[unreal plover]

nop, same prompt, same mask

[hardy cipher]

same model?

[unreal plover]

lol, no, left is 1.5 inpainting model, right is base SDXL

[primal vault]

Looks like you just used a too high denoise value

[unreal plover]

that is how inpaint work

[hardy cipher]

well obviously not

if it came out like that

[unreal plover]

i always set denoising to 1.0 for inpainting

if you set less you will add original image detail, while i dont want

[hardy cipher]

so do you think it's more likely you're missing something, or that the program itself is flawed?

[unreal plover]

The model is not made for inpainting i guess, on 1.5 you have a specific model for inpaint and outpaint

img2img with denoising of 0.5 work, but more you add noise, more it is unperfect

while inpainting model can work with 100% noise mask

[vagrant wasp]

SDXL first time "girl sitting on a horse"

[unreal plover]

512 is not good for SDXL

try 1024

[vagrant wasp]

1024 runs fine upscale crash.. low memory ?

[unreal plover]

don't know, try without upscale, you will get better result than that

[hardy cipher]

I don't see the problem. that image looks great

[unreal plover]

of course, nothing is ugly anymore today

[vagrant wasp]

i like working in clean UI so making it like electrical sockets so even wen i disconnect it.. i dont lose the wirings ... just sharing

[unreal plover]

remember, someone painted this one day, and he is the most popular artist ever

[vagrant wasp]

making it as NFT soon

[shy kelp]

shiva meditation:
lord shiva in meditation, indian, dark blue skin, tiger skin, mandala, himalayan moutains covered in snow, strong, masculine, powerful, glowing, radiant energy, sacred temple, realistic, detailed, 4k style:Fantasy Art seed:2365917610

[vagrant wasp]

good extentions out there for video ? like runway gen 2

[unreal plover]

shiva token look overtrained

[upbeat summit]

poor shiva

[unreal plover]

It is not a fail, it is a feature

tell me if it offend you, i delete

[soft zealot]

its simple if you stick to the daily driver view (which IMHO has all thats needed to be fiddled with , well for me at least) . It only gets complex when you zoom out lol

[late marsh]

This is not bad.. try Mandalorian

[unreal plover]

i mean here is a lot of saturation even if you reduce the cfg scale

[soft zealot]

these are reminding me of something I did a month or three back, maybe I should revisit it in SDXL 🙂

[long rose]

fat

[native knot]

The clock on this render got impressively close:

[long rose]

guys what upscalers you guys use?

[soft zealot]

ok just noticed that if I slect SAI Base from any of them then I get no image generated, hmmmmmm (again a look at laer thing)

[remote hamlet]

hello.

[late marsh]

This is much better. Changed the prompt. Didn't touch the settings

[unreal plover]

what cfg scale did you used ?

[late marsh]

7 for base, 5 for refiner. Tho I'm using SD.Next and it might calculate things bit funky.. and there is a setting called Guidance rescale at 0.7 and I don't know what it does 😄

[unreal plover]

Well i guess i should use the refiner then

[late marsh]

This is the way!

[upbeat summit]

nice

[molten gull]

how do i convert an ASCII to a text_g / text ?

in comfyui ?

[flint sleet]

the door 💀

[flint sleet]

ayo, what is the prompt?

[late marsh]

The image has all the generation data in it 👍🏻

[late marsh]

Oh, A1111 can't read it, lol

[flint sleet]

dam-

[late marsh]

breathtaking, cinematic photo, gigantic shiny golden statue of meditating Mandalorian, snowy Himalayan mountains in background, noon, 35mm photograph, film, professional, 4k, highly detailed no negative

[long rose]

dad?!!?

[molten gull]

https://github.com/comfyanonymous/ComfyUI/blob/master/script_examples/basic_api_example.py

is throwing me an error like this:

anybody got an idea what i might be doing wrong here ?

[long rose]

#🤝｜tech-support

[flint sleet]

thx

[long rose]

mom 0:

[unreal plover]

i always have over saturated results

i dont know why

heun sampler give less saturation

[native knot]

Boba Fat

[late marsh]

I use just normal Euler

[stray warren]

use karras instead of normal

[unreal plover]

are you using pixelart lora ?

[heady vale]

just prompts

[native knot]

Those look awesome.

[strong field]

Well done

[unreal plover]

try this if you like pixel art https://civitai.com/models/120096

[strong field]

It’s crazy what prompts are capable of right now

[heady vale]

thanks. I was curious what I could make without loras. Im happy so far

[vagrant wasp]

Any tutorial to train clothing model

[boreal bough]

you're in luck. those are the easiest!

even easier than style lora XD

[boreal bough]

You can use this preset -> #🔧｜finetune message
or within kohya gui, either of these two presets:
SDXL - LoHA by AI_Characters
SDXL - LoRA prodigy AI_Now v1.0

if you're aware of how tagging works, great! if not, please reply, then I'll link a guide

[molten gull]

how can i call/run a comfyUI generated .json file via a python script?

[ashen oracle]

is there anything available to expose API on ComfyUI latest version? I was relying on stableswarm for this before but that wasn't staying up to date with latest comfyUI so i ditched it

reading a bit more and it looks like the API is on but there's not a lot developed to ping it - will do some testing and see if i can build on top of the basic api example file in the repo

if anyone has something better pls lmk 🙂

[high skiff]

I am finally having some luck with my Portrait Photograph LoRA thanks to @boreal bough and some other people in my research group

here is Base SDXL compared to my very early testing LoRA

its undertrained right now, and I have a small dataset, but I have plans to bring this to the next level in no time

[boreal bough]

isn't stable swarm always within 48 hours (or less) of comfy?

[ashen oracle]

will revisit - not sure

[boreal bough]

if you're ever missing a feature, just post in https://discord.com/channels/1002292111942635562/1134546054872842290 - the dev is super responsive and gets back super fast. most issues I've addressed are solved within like a day

[ashen oracle]

thx will try to lean on that then just need to see if i can get it working properly was having issues yesterday

[high skiff]

this LoRA is still very undertrained, but man, look at this difference

Base vs my LoRA

[plush drift]

https://civitai.com/models/126767/tanukikanawa-classic-comic-xl anyone fan of such style? 😄

[boreal bough]

today we find out. can I get a good lora trained in.. under 10 minutes? 🤣

need to release my dnd loras in time for all the baldurs gate people in need of AI art

[high skiff]

lol

I use my furry models for DND, cause I never play anything human lol

I am still super happy with my wendigo

graphic warning

[boreal bough]

it's not about having art. it's about looking authentic as if it came from a (good) dnd rulebook! XD

[high skiff]

would have never known it came from an NSFW furry model lol

[high skiff]

oh booo, the designs in the actual rulebooks look like shit more often than not :p

you'll never find a shredded snow Leopard Tabaxi like him in any rulebook! :p

[boreal bough]

21h?!

rip ❤️

hope it works!

[plush drift]

wait you do not use preview imgs? you just wait all this time???

[ionic dragon]

Even I do

[ionic dragon]

Who said I am using previews

[south frigate]

hi, i am not in thread last days - control net finally working with SDXL or not yet?

[plush drift]

no one said, I am asking!

[ionic dragon]

No I didn't use samples

[boreal bough]

yes - but its via a mid-effort workflow

[plush drift]

https://tenor.com/view/shocked-cat-gif-25978666

[boreal bough]

for super low effort give it a few days

[boreal bough]

tbf, I don't either. most of my big loras look like absolute failues if I go by kohya samples XD

[plush drift]

you wait like 40 hours? is this lora?

[boreal bough]

technically, the cap of what can be achieved via a simple lora is now on the same level of SD1.5 finetune

[plush drift]

ummmm, I respectfully disagree! if my samples are not getting me what I want to see, it means I am just waiting time and electricity

[high skiff]

This LoRA is just a very small test into what I am trying to do, but I am already yielding massively improved realism portraits in SDXL, and I am practically squealing after dozens of failures

[boreal bough]

in sdxl?

I mean for 1.5, sure. samples were on point. But for sdxl, I can't get kohya to duplicate proper prompts

[high skiff]

Curious to see how it affects portraits of other objects and such

[boreal bough]

don't get me wrong - I wish the samples were on point - but at least for me, they never are. Do they work full for you?

[high skiff]

obviously not trained at all on tigers, but I could really slap in some more images into this dataset of different realism things and boost its understanding

it already looks much more like a realistic image, even if the details are bad

oop, just a sligth prompt change and its already looking better

[plush drift]

well yeah, this is why I was dancing in happiness when I saw the stop training button on the gui hahaha, I never do a full training if the samples are not showing me what I am training

[high skiff]

one of my negatives was affecting it

[plush drift]

these are amazing

[boreal bough]

oh damn O:

what sampler you using?

[plush drift]

euler or euler a

i use like 50 or 100 steps preview so I can see what is going on, before I let my gpu burn in hell for training lol

[high skiff]

some more examples, first is no LoRA, second is LoRA

[shy kelp]

do i ever change the 4096 size to my image size or what is the purpose of it

[high skiff]

feel free to change it

experiment with the effects it has on your images

[shy kelp]

do i just paste this then

[high skiff]

the effects are still being documented

[shy kelp]

hmm

[high skiff]

if you can, go for it

[shy kelp]

also trying the refiner with a finetuned model rn

the results are ....

😂

wtf

do i even have to use a refiner

some people say finetuned versions dont need refiner but i have no idea, there is no guide

[high skiff]

@boreal bough

Look at the detail and background fideliiittyyyyy

[plush drift]

I am looking at @high skiff lora, and I am looking at my lora now as I am testing too, 🤣 🤣 😂

[high skiff]

its still a bad image, but it looks much more realistic

[high skiff]

I like the style!

We are doing 2 verrrrrryyyyy different things haha

[plush drift]

hahahah I know

this is why i love it

[native knot]

Need? No.
However, if you tune down the # of steps taken by the refiner and definitely cull the CFG back to a low amount, you can get some nice results.

[plush drift]

one is pushing reality, and the other is pushing madness

[shy kelp]

from experience, what steps do you use ?

and cfg