✨｜sdxl

[cinder aspen]

and also programming ofc

[elfin flare]

And not a workstation GPU

[cinder aspen]

yes

workstation is overkill for now

[slender coral]

Those don't need high end workstations.

Simulating relies more on a CPU, than a GPU.

[cinder aspen]

Im not intending to build one for those

Since im gaming as well

[slender coral]

A gaming computer should cost max $3.5k usd.

[cinder aspen]

Impossible mission

A RTX 4090 alone is 2.000€

Or 1.800 lets say for some

[slender coral]

You don't need a 4090 for gaming...

[cinder aspen]

1.800/3.000 just with a GPU alone

[slender coral]

Again, a 4090 is overkill for gaming.

But it's your money.

[cinder aspen]

But kinda "need" it if i want to play 4K with raytracing at most extreme level possible

[steady grove]

at 1440p, my 4080 could do a little better at rdr2

[cinder aspen]

And still have over 100 fps

[steady grove]

i dip to about 50 when that wide open world is maxed up

[slender coral]

One day you will realize you will start a family and want to retire, but you're free to be a consumer.

[steady grove]

weird flex

[slender coral]

Granpa why are you living with us

Because I wanted those dank fps on Max settings.

[cinder aspen]

Others spend that money on other stuff

Cigarettes, alcohol, etc.

[slender coral]

Young men have this idea that because others do it I should too.

You

re free to do what you want though.

[cinder aspen]

I didnt need to spend thousands on my art workflow and workspace, but i did because i wanted it

Its about what you want

Expectations

Maybe 4090 will drop in price next year

[slender coral]

There's going to be a market crash, but you could keep warm with your new computer possibly 😄

[cinder aspen]

0_0

[slender coral]

I own a company and make quite a bit a year, if I can give you any advice, don't buy useless things you don't need. Unless it can give you a return. Invest in your skills and other sectors to grow wealth.

I have 2 4090's and they make me money, but other than generating for AI I wouldn't purchase the cards. But you're free to do what you want.

[cinder aspen]

@slender coral respectable. And you are actually right. Better dont find out that im still a beginner in art and co. and still spent all the money from the beginning

Hardware and software

Actually my idea is essentially to earn secondary income besides of my current full time job that i might eventually replace if i get good enough with programing and jump to the IT sector.

[slender coral]

Best decision you can take, I'm a Software engineer.

Spend that money on school.

[cinder aspen]

What do i want with that secondary income? I want to invest in even more stuff related to secondary income and art like buying a good 3D printer at some point and make a "business" out of it by making and selling 3d figures as well. Also i want to educate myself

Pay courses

Some of which are expensive af

Let alone online degrees

[slender coral]

The only thing I've used my 3D printer was for printing parts and upgrades for the 3D printer. Having tools and toys doesn't make money. Knowing people does, that's how you get contracts and projects.

[cinder aspen]

Yes. Networking and building necessary skills

[cinder aspen]

Having a product wont make me money automatically

Dont get me wrong, im not interested in this area just because i want money

My interest and fascination for those are on the first place, money earning with those is just a ideal case but not a necessity

[slender coral]

Programming gets boring.

Money allows you do do what you want.

[cinder aspen]

Money gives freedom, but i look for other factors as well

Free time, atmosphere in the company and the team are important for me as well

[spring fulcrum]

Does anyone know if LYCORIS works with ComfyUI

[vital ermine]

iA3 does not, nor DyLoRA.

[slender coral]

Sure. But those factors change. The whole idea of company culture in tech is bullshit HR talk to keep you not asking for raises.

I've worked on nfts the last few years, I hate nfts but it allows me to support my own projects and buying 2 4090s.

But yes, go to school, it will be the best investment you can do.

Learning programming online I wouldn't recommend. It takes about 2 years of acedemics to actually learn and understand it. Anyone trying to sell you a bootcamp or online cert is just scamming people.

[crisp owl]

Valid, and yes when I do my own changes I have a custom .py file (SolesbeesSaloon) I make all my additions or changes to.
I made that mistake once 😅

[spring fulcrum]

I'm going on the hunt for things that I can add to my workflow, Can you give me a list of all the things that DO currently work in ComfyUI with SDXL models? Please and Thank you...

[vital ermine]

I can't because I don't use it now since it refuses to work with the lora types I train, and use.

[cinder aspen]

NFTs? Damn i dislike those lol.

For the education, i definitelly want to do something but going to those in real life has a high price tag and not only financially. For studying i dont even need to talk the route for now at least.

Some of my friends got into the industry without a formal education, they were at some online courses but more importantly...built a portfolio while attending those and after as well

[high skiff]

anybody happen to know how to stop python from forcing your GPU to boost clock when at idle?

cause its using a lot of power, and making a lot of heat

I cant figure out how to stop it from sucking power when paused/idle

[slender coral]

use .Sleep not a pause

[high skiff]

My bad, I'm not running my own code

I have this problem when automatic, comfy UI, and oobabooga are idle

[slender coral]

The reason it's expensive is the same reason why you make 100k+ a year.

Take out a student loan, it will be the best investment for you.

[slender coral]

Unless there's some running threads (and I don't think there are) I don't see idle being an issue, the reserved memory should be released. GPU is different.

Check your cpu / gpu at idle and see your percentages.

[high skiff]

They're both below 2%, that's not the issue I'm having

The issue was that Python.exe is forcing my GPU to stay at boost clocks, even when it is idle

[slender coral]

Pyhton doesn't do that, it's whatever process you have running.

Check your PID processes and see what's consuming the most.

[high skiff]

So then it's command line?

[slender coral]

ps -ef | grep python

[craggy ibex]

With Pixai

https://pixai.art/submit/gen

This is probably the best Stable Diffusion provider with anime style

[vale eagle]

peace sign hand lora alpha

[steady grove]

afterburner could probably create a python specific profile for it. maybe even geforce experience

[craggy ibex]

It is like he's cheating for sure

[steady grove]

wait afterburner doesn't have per app settings. mb

[craggy ibex]

Hold on let me try again

[craggy ibex]

I cannot.

That's not even a fucking blue

[steady grove]

pytorch is doing it likely, since it has the cuda library talking to the drivers. but windows sees that as python.exe

[craggy ibex]

It's a FUCKING green

[steady grove]

green is not a creative color

[craggy ibex]

Now this IS BLUE

[slender coral]

Python just runs Python there's no processes it runs. So Python is all you see.

If you want to see what's running you have to list the Python modules.

[steel mist]

I'm using the comfy node that gives you a feed of images at the bottom—is there a way to load one of those images into the workflow?

[heady vale]

drag the image from your output folder into comfyui

[steel mist]

Yeah I just do like 50 gens at a time and it's just a touch of a hassle to find the right when when I have this gorgeous feed right there

[heady vale]

or go to history (theres a little button) and load

[steel mist]

Yeah there's just no clean mapping between that feed. Its not a huge deal would just be a nice lil QOL thing

[strong field]

@indigo carbon finally figured out what screwed up my speeds a while back, i nuked my CUDNN drivers somehow and that was the culprit

back to higher it/s on your workflow

[heady vale]

what was the change in speed?

[strong field]

8s difference in generation on Tdg workflow

using default workflow, it was down from 10s for 1024 (vanilla) to 7.99s (AIT+CUDNN)

20% change ish

[hardy cipher]

https://tenor.com/view/not-bad-ryan-reynolds-gif-23839048

[strong field]

wait shit

[steady grove]

thats what i mean. that's why they see python.exe using the gpu, because as far as the kernel is concerned, it is

[strong field]

thats without AIT

lol

with AIT it went down to 5.79s on default workflow 1024x1024

[hardy cipher]

https://tenor.com/view/much-better-guy-ryan-reynolds-free-guy-way-better-gif-22851674

[heady vale]

that sounds a lot better

[strong field]

https://tenor.com/view/woof-macaulay-culkin-home-alone-gif-5332907

[strong field]

it took FOREVER to figure out

[heady vale]

which version did you install?

[strong field]

cudnn_8.8.0.121_windows.exe

lol, the exe

[heady vale]

yep thats good

[strong field]

even tried to compile ait for 3090 but im stuck on compiling

aitemplate module is giving me issues in python

have ot rebuild the whl

[indigo carbon]

Huh, did you actually manage to set up the compilation script? I tried for AGES to do that to have my own modules, never worked

[strong field]

nope, im stuck at compile.py

tells me aitemplate isnt a module

and then i do a pip ls, and its sitting right there

🤷‍♂️

[indigo carbon]

Did you manage to run the build script? That's what I'm always stuck on

[strong field]

hahahha, that damn script is stupid on windows cause you dont have write permissions on the folder without a bunch of jumping through hops

so i just renamed the folders in setup.py

[worthy orbit]

Hi, for some reason Searge's workflow inpainting don't work for me (img2img works). Inpainting just delivers the drawing of the girl (img2img delivers results)

[indigo carbon]

Pog?

Did that fix it?

[strong field]

i got the whl to build but there is some incompatibility with one fo the dependancies

[indigo carbon]

Is it the kernals? That's what I get when python setup.py bdist_wheel

[strong field]

this is what i changed, "aitemplate2"

ah shit your error must be different, i was getting write errors because the file existed

[indigo carbon]

Man, idk how the fuck Hlky does this shit

[queen adder]

is low learning rate (0.0001) and high epoch count what trains the model towards detail? or are there other parameters too?

[strong field]

DUDE, i feel like i know the guy, been reading so many of his threads

hes BUSY

[strong field]

What is your denoise value set to

[indigo carbon]

@strong field mind taking this to DM? I am also always trying to compile myself

[strong field]

lets go

[worthy orbit]

0.74, but it doesn't seem to matter (and I think I tried some other number) as it never produce any result, just paste the drawing of the girl

[strong field]

can you send me a screenshot

[steady grove]

https://www.reddit.com/user/hlky 👀 hope not this guy

i got curious cause y'all were excited bout hlky threads

[worthy orbit]

sure, here

I messed up the heigh and width but it show it doesn't take the image

[heady vale]

Dont think it knows Linus's face

[worthy orbit]

oh I figured it out, I need to put the image both in the source image and in the inpainting mask 😒

[heady vale]

@indigo carbon do you know if there was anything new with AIT for Lora and VAE decode?

[wet nacelle]

https://tenor.com/view/things-that-never-happended-things-never-happended-invent-fake-fake-news-gif-18019799

[heady vale]

https://tenor.com/view/my-computer-trash-bin-erase-delete-elderly-gif-17041770

[wet nacelle]

https://tenor.com/view/shocked-shock-shocking-omg-omg-meme-gif-26225226

[spice skiff]

This image was created with base SDXL + refiner.

[wet nacelle]

That's really kool Patt

[glad grove]

my new routine at the gym

[wet nacelle]

"I have no eyes but I must see"

[glad grove]

dont need no eyes when u look this beautiful

[wet nacelle]

https://tenor.com/view/you-fine-damn-beautiful-baby-pyt-heart-love-gif-14878673

[hardy cipher]

such detail

it's the little subtle things that make images like this seem real

[high skiff]

@visual glade sorry to bother you over this again, but I have talked to some other friends/colleagues, and they are also having this abnormally high idle VRAM draw from comfy, and other AI programs.

I am not sure why, as the GPU's are set to idle state, but they max out their clocks and just sit there doing nothing

[high skiff]

Would be curious to see the results of people in this server in terms of idle GPU power consumption and clock speeds when not generating in comfy

I have it in RVC, auto, comfy, ooba, kohya, and Derrian

I am assuming its just some underlying issue with python that causes the GPU to be clocked up for no reason wasting power

all I know is I really wish it wouldn't, cause its running up my cooling, and thus my electric bill, while already using unnecessary power

[heady vale]

mine clocks down, but I dont have gpu-only. Maybe that is the cause?

[high skiff]

I have this issue in all AI suites, and it's been on all of my GPU's as well

[high skiff]

When you just have comfy open and not generating, how much power does your GPU use?

[hardy cipher]

it could be nonsense. but I read today that 720 hours of a 4090 running hard will cost about 32 dollars on average in the US

something along those lines

[heady vale]

Oh ok. I havent seen that before also with A1111. When not generating in A111 or comfy open its like 40W

[visual glade]

it doesn't happen here

[high skiff]

I'm not sure why it works for you then...

For me, my GPU permanently stays clocked at boost clocks and just sits with 0% utilization while sucking down 140 Watts to do nothing

And I have talked to others with the same issue

[visual glade]

33W / 450W | 8617MiB / 24576MiB

[high skiff]

I don't understand

My GPU uses 140 Watts idling in comfy

[heady vale]

do you have hardware accel gpu scheduling turned on in Win?

[high skiff]

No, it's off

[crisp owl]

COmfy, 24 idle, ~135 generating

[high skiff]

I always make sure that's off

[heady vale]

hmmm Im not sure why

[high skiff]

Same here

Like I said, does it with oobabooga, auto, comfy, kohya, Derrian, and RVC

Just cranks up the clocks for no reason while doing literally nothing

[visual glade]

wattage calculation for GPUs is never really exact by the way

it's an estimation

[high skiff]

I know, but nearly 5x normal idle is definitely not off by just a little guess haha

[visual glade]

are the temps fine?

[high skiff]

And my GPU runs much hotter roo

*too

The temps are 14C higher

When idling with comfy open

32C normal idle vs 46c idle in comfy

[crisp owl]

I keep my pc running 24/7, work on it all day, so generally have something batched more or less, or working on something for it.

For 1 month use I'm at 141 kWh, costing me about 14 bucks.

[high skiff]

Normal idle is only a few 100 mhz, but "idle" in comfy is 1800mhz core and 9501 mhz VRAM

[high skiff]

God, what I wouldn't give for such cheap power lmao

[crisp owl]

valid, it's cheap here

[high skiff]

It's fucking expensive here

[visual glade]

I worry a lot more about the heat than the power usage

especially in summer

[high skiff]

And I also live in an extremely hot area, so hours of dumping 110 watts of unnecessary heat into a room also hurts my electric even more on cooling

[crisp owl]

I work in my basement, so I enjoy the extra heat, especially with my pc at my feet 😆

[steady grove]

yeah me too. but then i turn the AC on and worry about power usage

[high skiff]

My mom and I were just hit by a fucking absurd power bill. And now I'm worrying about every little power draw

[visual glade]

| 0% 37C P8 25W / 450W | 8608MiB / 24576MiB
and now the fans stopped with comfyui still open

so it must be an issue with your driver maybe?

[steady grove]

my gpu isn't idling high

[high skiff]

Maybe, but it's happened across 3 GPUs on 2 completely different PC's for over half a year now

[heady vale]

could it something related with remote PC access?

[high skiff]

I don't access my PC remotely

[heady vale]

werent you generating from your phone for a while?

[high skiff]

It's python.exe forcing my PC into boost even while reporting it at idle, it only happens when python.exe is open

[high skiff]

Yeah, but I haven't in well over a month

[steady grove]

even if you're just running the python terminal?

[high skiff]

That I haven't tested

[heady vale]

nothing special set in nvidia control panel 3D settings?

[high skiff]

I'll have to test tomorrow, I am off my PC now

[high skiff]

Nope, checked power profiles and preferred performance

First is optimized, and second is normal power scheduling

[heady vale]

out of ideas, sorry. Hope you can find the problem as I dont think thats normal for idle

[high skiff]

No worries, and no, it's definitely not. My GPU normally idles at 30 watts

I suppose I can switch my GPU to stock VBIOS and see if that helps, but I happened on my 3080 and 3060ti too

[heady vale]

only last thing would be to do another DDU wipe of drivers

[azure oxide]

did a quick test

wattage doesnt seem to be elevated when idle

im still currently at 22/24gb for vram though but that might be an AIT node thing

[heady vale]

mine clocks right down

[azure oxide]

ya nope nvm thats not an ait node thing, im assuming comfy just keeps the model cached in vram (super high usage with sdxl) but that might be by design or with --gpu-only

on the bright side, clocks goes right back down as expected

[glad grove]

novelai lookin good

👍🏿

[dense chasm]

this is pure python script of kohya with canny edge detection and lite controlnet generation,not novalAI

[soft zealot]

are you sure its not something else taking advantage of the fact that python is idle to use your GPU as I don't see that behaviour here.

[uncut fiber]

anybody fancy of a precious stone?

[candid walrus]

how are you monitoring your watts?

[upbeat summit]

with programs like gpu-z or hwinfo on windows for example

[uncut fiber]

How reliable is afterburner in watts measuring?

[soft zealot]

@high skiff
To the LHS of the green tick is no Python running at all
Green Tick I started ComfyUI
Next Spike is when I ran an Image Generation

@visual glade I am running with smart-memory disabled (orange tick) however the VRAM is no longer releasing at the end of the Image Generation as it used to do in earlier builds (although Noted sperately if I start a new generation you will see it briefly try to clear)

[upbeat summit]

you can also open a command prompt on windows (should work on linux too) and use the Nvidia SMI tool to monitor your GPU. It's part of the driver
nvidia-smi.exe -q -d utilization -l

[primal vault]

And that's why I included a readme file 😉

[upbeat summit]

it reads the sensors from the gpu - so it should be accurate

[uncut fiber]

@upbeat summit thank you!!!

[upbeat summit]

you're welcome. compare it with gpu-z or the nvidia smi tool if you are unsure. but they should be the same

[candid walrus]

i used that command for gb checks but didnt realize that it was also giving info no watts nice

[upbeat summit]

you can monitor it in real time using these parameters: nvidia-smi -q -d power -l

but only on supported GPUs

[soft zealot]

or just use HWINFO (free edition) and get little graphs 🙂

[uncut fiber]

i got report i am exceeding 119 power percent. 🙂

[soft zealot]

Im guessing you have set the Power limit in After Burner to 120% then

[uncut fiber]

is it Lovecraft, or rather Giger?
@soft zealot no, 100% and 83C
I think it is power of gpu+cpu probably?

[soft zealot]

depends where the warning came from,

[uncut fiber]

no warnings, just reading max values

[soft zealot]

from?

[uncut fiber]

From Afterburner

[soft zealot]

GPU power only

[uncut fiber]

i will post so you can check. But yes seems gpu only

[soft zealot]

no need too,

[uncut fiber]

weird it has limit to 100% probably so short, i noticed it is at start of generating, some shock to system 🙂

[soft zealot]

some cards spike , is it the 4080 or 4090 especially ?? Cant rememeber

[uncut fiber]

i have only 3070

[soft zealot]

or maybe it was the 30 series

https://www.google.com/search?q=3070+transient+power+spike&oq=3070+transient+power+spike&aqs=chrome..69i57.6863j0j7&sourceid=chrome&ie=UTF-8

[worthy orbit]

oh I'm reading right now, couldn't get to that part

[soft zealot]

youu would likely get more attention paid to it if you called it a "dontreadme" file or "ignoreme" ;o)

[upbeat elbow]

Anyone have an SDXL installation tutorial for beginners who have only used Easy Diffusion UI? I'm sure it's been asked many times. I'm looking through YouTube right now.

[worthy orbit]

but I think for inpainting for fixing details I'll use old Automatic, as it takes a lot to generated and results are I would say evasive and it is kind of exhausting to set it all. For great changes sort of img2img yes def.

[uncut fiber]

@upbeat elbow i would go with default workflow in ComfUI. It is easy to install and SDXL perform well. What is your GPU btw?

[worthy orbit]

Also don't know how to prompt for inpainting, I guess it is more like a general img2img, or just the part you want to change?

[upbeat elbow]

Gtx 2060, and thank you

[uncut fiber]

not sure how well it will works, but for sure better than A1111

[worthy orbit]

Anyway if anybody knows a good workflow for upscalling in SD XL, I apreciatted. I mean I tried some but gave me out of VRAM

[upbeat elbow]

How long would you all say you've been working with AI image generators anyway?

[uncut fiber]

not sure, suggesting resolution 1024x768 or other way round. For sure it takes more than previous models.

[uncut fiber]

is this pure Giger?
Trying to combine him with Lovecraft therefore asking 🙂

[rain pebble]

Anyone had issues with running Base SDXL 1.0 with https://huggingface.co/diffusers/controlnet-depth-sdxl-1.0 in webui?

Seems like results are not taking controlnet into the account Not sure I need to update something

[dense chasm]

you can run controlnet thr diffusers python code

btw the current canny edge detection model can't generate proper images with complicated contour

[rain pebble]

Okey code it is, thanks

[hardy cipher]

new virtual environment, everyone. very cool

[dense chasm]

like this

[hardy cipher]

looks a bit disheveled in there

[rain pebble]

yes 😄

[dense chasm]

this is the original image, i just detected the canny edge,it looks like have to adjust and modify with params

experiment failed with diffusers canny

[hardy cipher]

lol. nice. I get those wrong so often. from that canny image it looked like some sort of abandoned room that is being overtaken by vegetation. at least that's what my mind saw

[dense chasm]

it run well on this tornado on grass contour,it's strange maybe prompt should be more task-specific

[patent badger]

hey guys, I'm using auto1111 - and i have a veryyyy (at least very annoying) big problem. I have a RTX 3070 (non lhr) and when I try to generate images it takes about 11 seconds for the generation, and additional ~30 seconds just for the VAE decode / idk what else it's doing (I'm dumb).
anyways, anyone knows what could be the reason for this? it wasn't like this until a few days ago and I really can't think of anything I could have done to make it this slow.
-- already tried updating drivers
-- already tried removing and re-enabling xformers

[uncut fiber]

what model? SDXL?

[patent badger]

dreamshaper xl

[fresh path]

does the vram max out?

[uncut fiber]

oh, it is based on sdxl supposing and performance of A1111 and sdxl models is realy bad

[patent badger]

I mean even if it is, it wasn't that slow a few days ago - was like a few seconds at max

also just tried using a different SDXL model and same thing...

[uncut fiber]

i have as well 3070 just supposing LHR and it is pain, so more ppl leaning on A1111 therefore switched to different UI until this is solved.

[patent badger]

but it does seem to be a problem with sdxl - when I'm trying a 1.5 based model it's super fast

[patent badger]

you experiencing same shit as well?

[uncut fiber]

yes it is problem of A1111 and SDXL

[patent badger]

but I could swear it wasn't like that a few days ago

[uncut fiber]

yes, i am now using comfyUI therefore

[patent badger]

btw - set COMMANDLINE_ARGS=--no-half-vae --medvram --xformers --api

could the medvram affect the vae decoding time?

[fresh path]

:(

[uncut fiber]

not sure but for SDXL it is somehow Must

[fresh path]

what if you try lowvram

[patent badger]

damnnn I'm not complaining about slow speeds ever again

[patent badger]

lemme play with it a bit

[fresh path]

yeah spend 25 minutes just to have the vae not work

bc my 16gb vram maxed out (on 1.5)

[patent badger]

@uncut fiber you using medvram or lowvram?

[uncut fiber]

lowvram i think will not help you as it is another performance hit.
medvram

[hardy cipher]

16gb vram?

what sort of things are you doing there?

[uncut fiber]

But not now, in ComfUI
I am doing almost full HD images super quickly.

[patent badger]

didn't they say there's a special UI developed by the SD team?

[hardy cipher]

is the vram even getting used? that's pretty nutty

[uncut fiber]

@patent badger there is feature, making generating even faster for 3xxx and 4xxx

[patent badger]

yeah I used comfy for a while but A1111 is just so much better experience for me

[uncut fiber]

They did Stableswarm-UI i believe is it

[hardy cipher]

https://github.com/Stability-AI/StableSwarmUI

[patent badger]

anyone used / using it?

[hardy cipher]

uses comfy as the backend

I've used it

best of both worlds really

[patent badger]

k I'll test it out now

[hardy cipher]

front end is similar to a1111. not as many extensions. but also not as many headaches. and you can switch tabs to go full noodle with comfy if you ever get the urge

[patent badger]

lmao

[uncut fiber]

Just what you will choose. In ComfUI, i can create 1920x1024 in 25secs. Time matter.

[rustic garnet]

I don't know...

its MUCH worse than a111

[hardy cipher]

well I mean, it works

[rustic garnet]

If you want a proper UI, use auto111. if you want a fast software with many flexibility, use ComfyUI

for SwarmUI I would wait a few months until it is usable

[soft zealot]

rowlocks

[hardy cipher]

a1111 was all good until it crapped out.

[potent flame]

SwarmUI?

[soft zealot]

SwarmUI is prinmarily a front end to control multiple instance of A1111 of ComyUI from one pleace.

It's a UI to other UIs nt a magic bullet ;o)

[hardy cipher]

true enough

I just think it's a good thing for some people as it gives them the node option without requiring them

[rustic garnet]

yeah, sure, but the UI is horrible

I recommended a friend SwarmUI when he asked me which would be a user friendly SD UI

[hardy cipher]

not sure why some people are so intimidated by them. but I also understand the casual users that don't want to bother with it

[rustic garnet]

he was like "WTF is that?"

so I had to recommend him auto111 ^^°

[soft zealot]

If you 're new and simply want to be able to generate images locally without paying fro an Online service then consider Foocus

https://github.com/lllyasviel/Fooocus

If you want to tiunker, stay away from it

[fresh path]

how do I let multiple UIs like auto111, comfy, use the same folders for all the models I have downloaded?

[rustic garnet]

even I'm annoyed that I reinvent all my workflows all the time. There is not even a simple drowndown list for predefined workflows X_x

[soft zealot]

I use Symlinks

[potent flame]

yeah

Foocus is really basic

[hardy cipher]

I sometimes forget how deep I've gone into this stuff. and most of you guys know a lot more than me, lol

[potent flame]

I have used it for SDXL

[hardy cipher]

I knock my down like sandcastles all the time

[rustic garnet]

yes, me too

but its not very efficient

[soft zealot]

you save your workflows with a disntinct name in a folder called Workflows and open from there, not rocket sceince

[rustic garnet]

but I have the feeling loading an old workflow via drag&drop and modify it is also not efficient

[hardy cipher]

on the one hand I'd agree, but I've also learned a lot of new things using the nodes. I learned quite a bit with a1111. but didn't go nearly as deep

[rustic garnet]

all in all productivity is definitely higher with a1111

[rustic garnet]

I used diffusers for that xD

[hardy cipher]

a lot of the stuff I was doing I understood only so far as how it impacted the images

[north sluice]

Hi guys, hope this is the correct channel for this question..
I've downloaded SDXL 1.0 but it won't generate when I try, could it be my GPU is too old/weak? (RTX 2080)

[uncut fiber]

no i think no. What UI? @north sluice

[north sluice]

Using Automate1111

[uncut fiber]

it has A1111 issues with SDXL

[rustic garnet]

however, one thing I like on ComfyUI: In diffusers I so often had memory leaks, cause its a hell to do proper memory management in a jupyter notebook environment. In ComyUI it works really well how only models are loaded I currently need and how everything is cleaned up in memory

[north sluice]

Aha !

[soft zealot]

well im using an even older 1080ti without any issues

[north sluice]

Glad to hear, what UI you guys using ?

[hardy cipher]

I like how quickly it loads. well how quickly it will load when I finish reinstalling the virtual environment

[soft zealot]

comfyui , SDXL1 support in A1111 is still I believe hit 'n'miss.
ComfyUI is a steeper learning curve but glad I switched

[north sluice]

Hmm I'll look into it, thanks !

[hardy cipher]

I think I saw another UI too

on reddit or something

[cinder aspen]

Photoshop, no joke

[hardy cipher]

lots of people trying it out

[soft zealot]

but as I only posted a few minutes ago

#✨｜sdxl message

[hardy cipher]

I have the photoshop plugin or whatever, never actually used it though

[cinder aspen]

There are several of those

[uncut fiber]

@north sluice Just try stay with default UI for some time, there are more complicated and bit harder to understand them.

[north sluice]

default Comfy you mean ?

[cinder aspen]

Alkaid.art/ImageCreator, Alpaca (is new), Gyre AI which i heard about from a user days ago and then there is Firefly, generative fill which is basically based on FF and DALL E plugin etc

[dusk mica]

how is fooocus? in terms of results

[hardy cipher]

it does the work for you

not bad results as I understand it

but also not as much control

[uncut fiber]

@north sluice default comfyUI workflow

[north sluice]

Thanks will keep in mind

[hardy cipher]

maybe start with that, but one thing I've learned is everyone has their own version of the truth when it comes to comfy setups

one thing most people would agree on is don't start installing all the node packs all at once. look them over, decide what interests you, install as necessary. but don't install a whole stack of them all at once unless you're feeling very ambitious

[cinder aspen]

True

I installed someone elses workflow before deinstalling this and it was too much lol

[uncut fiber]

This should be happy picture.....

[hardy cipher]

I'm weird so I actually liked noodling right off the bat

but I guess that's not the case for a lot of people

[cinder aspen]

People dont have the same workflows and one should end up having one based on his needs

Not what others have

[hardy cipher]

yes. it's fine to emulate and use what other people do. I don't think there's a right or wrong way. but I'd suggest learning how the stuff works on your own as well. so many people asking for workflows, lol

I like to make really impractical things just to see what happens and then abandon them to do something else equally impractical

but that's just me

[soft zealot]

and then theres me, I leave my workflow in my images but if you do decide to blow your mind make sure you have full read and inwardlydigested this message from the sponsors before asking any bloody questions!!!

[cinder aspen]

Pretty similar to art workflows

With all the softwares

[hardy cipher]

I always leave my workflow in images. but I'm fairly certain people look at them and move on, lol

[uncut fiber]

Is there way to update nodes? If some new version i mean, by manager?

[hardy cipher]

yes

or git pull if you're feeling fancy

[uncut fiber]

i got portable only version

[soft zealot]

thisd all makes sense to me lol

[hardy cipher]

pretty sure that wouldn't change the git pull. but do what works for you

[uncut fiber]

going to try it

[hardy cipher]

I only use it as a last resort option though. sometimes for whatever reason the manager goofs and doesn't install something quite right

[cinder aspen]

Been there

Had to go the Git way

[potent flame]

Generating the femcel ( thought is is not SDXL )

I think i gonna try SDXL again and see if it will still continue my 30 minutes bicycle ride

[hardy cipher]

top notch

[potent flame]

you guy can't put SD1.5 LORA to SDXL right?

[uncut fiber]

cats version 🙂

[hardy cipher]

well you can. but you probably shouldn't

figured out a weird approach with the canny 1.5 model and sdxl

used it to make weird noise nonsense with an xl model. then used that to render on

[dense chasm]

prompt:a woman, close up, detailed, beautiful, street photography, photorealistic, detailed, Kodak ektar 100, natural, candid shot.the diffusers/controlnet-canny-sdxl-1.0-small runs faster than the previous unet full attension fine-tuning model

[quartz sequoia]

is there a on/off switch node for comfy?

[heady vale]

comfyroll custom nodes has logic switch nodes

[hardy cipher]

you can right click and deactivate

but yeah, also on/off nodes that take boolean input. and then you can customize nodes with on off switches if you feel ambitious.

[uncut fiber]

or bypass them ctrl+b i think

[slim wren]

nope can't do

[patent badger]

@uncut fiber do you have the refiner extension on auto1111 by any chance?

[uncut fiber]

i have but not activated.

[patent badger]

it's weird, when messing with it a bit I cut my VAE decoding time by like 60%

actually nvm I'm stupid

[uncut fiber]

many of workflow creators claims they next version will be without it.
Yes it is 5GB model is another 6GB it is large

you have 3070? then in Comfy, there is big boost thing, making it working even faster.
I can send you it in image, just using very basic layout. @patent badger

[hardy cipher]

well it beats up what you're doing, and no one really knows how it works, lol

ask someone how to set up the refiner prompt

literally no answer out there

[uncut fiber]

you can only set number of steps in A1111 UI for refiner extension. Not any special prompt. Just 10 steps default.

[hardy cipher]

ahhh, a1111

[rustic garnet]

you would usually use same prompt as for base model

refiner is basically an image2image step

[vale eagle]

did anyone try different prompt for refiner?

[uncut fiber]

supposing it is that language model and supported is refiner, isnt it that way?

[thorny frost]

I just found the sd ultinate upscaler node has an interesting 'dirty' effect when used with a value of 1 upscaling (so not upscaling at all) don't know if it is a bug or a feature

[rustic garnet]

the refiner is only using CLIP-G and not CLIP-L. So its text encoding input is slightly different than for base

[rustic garnet]

don't know what you mean

[uncut fiber]

i will make screenshot. It is in Sytan 1st workflow

[hardy cipher]

that's what I've been doing. but then what about loras and other things. it's just ambiguous to me. I play around with it and make educated guesses

[rustic garnet]

ah, in Sytans workflow there are different prompts for CLIP-L and CLIP-G

[hardy cipher]

well what else does the refiner use besides clip and L?

I mean base

[rustic garnet]

you can mess around a lot in SD, trying many different things, using different prompts, different models, switching models or prompts during generation and all that stuff

[hardy cipher]

I thought refiner didn't even use 2 clip models

[rustic garnet]

yes, refiner is using only clip-g

[vital wraith]

just got into SDXL today...Loving how far you can push the resoluion.

[rustic garnet]

which is called "linguistic prompt" in Sytans workflow

[hardy cipher]

I've just been doing things like separating L and G for base, then combining them for refiner

[rustic garnet]

the thing is: the "correct" way of using it is using ONE prompt for clip-l and clip-g and the same for refiner

[hardy cipher]

well sometimes just 2 positives, but send the same negative to all 3

[rustic garnet]

where "correct" means the way SDXL was trained and developed.

[rustic garnet]

yes. "correct" way is using the same text in "linguistic" as in "supporting"

but as said, you can mess around in SD and do things differently. Just play around with it if you want

[hardy cipher]

also, I think a lot of people are looking for the magic formula when there really isn't one necessarily

some things obviously work out better under normal circumstances

[uncut fiber]

I think it is reason why workflow creators claimed next time without it 🙂

[hardy cipher]

who are these workflow creators?

I create workflows, lol

[uncut fiber]

and you drop refiner?

[rustic garnet]

the refiner is always just optional

[uncut fiber]

Sytan and another one i know but dont know name

[rustic garnet]

it is not necessarily improving your image

[hardy cipher]

I don't think there's a united team of workflow creators all with the same mindset

[rustic garnet]

and dropping it makes things easier for sure (in particular if you use different checkpoint models or loras)

[uncut fiber]

i mean workflows creators, like many individuals, not as groups

[hardy cipher]

I need to work on getting images to come together completely without it

[patent badger]

the only reason I'm using auto1111 is the easier img2img and inpainting UI - it's my main usage anyways, so there is a huge benefit by using auto1111, anyways, I downloaded this "tiled vae" extension, and the vae decoding is basically immediately now. like 0-2 seconds at max, with almost identical results (not better not worse, just a tiny tiny bit different)

[rustic garnet]

I would just always create two images: base and refined, and then decide for the better one

[hardy cipher]

I save way too many steps of my workflow

[patent badger]

@uncut fiber one of them is without the tiled vae, one of them is with it. guess which one is which😂

[soft zealot]

This is how my Prompt Createion gets split out and wrapped insid e Styles and wheat evenetually ends up where

[hardy cipher]

currently trying to figure out these stupid custom nodes and broke something, lol

get a little something wrong with the node mapping and you won't have a fun time

[uncut fiber]

i hardly spot difference, not sure which is which honestly

[hardy cipher]

tired of nodes I want not existing so gotta make my own

and spend five hours making what it'd take some people in here 15 minutes to do

[patent badger]

also tried zooming in and try to see if there is any difference in quality, tbh after testing like 3 different prompts, there is literally 0 difference

without using the VAE tile

with the VAE tile

[uncut fiber]

what is your card? 2080 or 3070. I forgot.

[slim wren]

hmmmm

DIdn't you switch the screenshots ?

Event TiledVAE is saying that there is no point tiling with the size you're asking for

[hardy cipher]

lol

[patent badger]

3070

[patent badger]

tbh I tried testing some different sizes in it just auto-disabled so I counted it as turned off

I can try again with the option turned off😂 one sec

@slim wren first generation without the extension, second one with it

~14s for latent for both (rtx 3070)
+~25 seconds for vae
+~2 seconds for vae with tiling

[dry pollen]

Is it possible to run SDXL locally on AMD 6600XT (8GB)?

[fresh path]

yes, but it still be extremely slow, like 20 minutes per gen

because amd needs 2x vram compared to nvidia so you need 16gb

from sdxl release: Linux users are also able to use a compatible AMD card with 16GB VRAM.

[dry pollen]

damn, part of me was hoping that the latest iterations will be also aiming to reduce the system requirments and optimizing for larger variety of consumer tech. What do you suggest as best alternative of usage away from my own PC, but still safe in terms of privacy?

[patent badger]

guyssss I need sdxl controlnet tilessss. images are just so low quality (when using subject loras) and I can't keep upscaling them in sd1.5

[fresh path]

I would try it regardless and see the speeds, I am using runpod right now but its kind of expensive

gpu market right now is gonna get even worse as AI demand increases

and the average consumer wont be able to upgrade for a few years

I put $25 on runpod and I get get like 54 hours with a 3090

[dry pollen]

yeaah, that is bit too much for me, I am currently using Mage Free account that uses Web GUI and runs SDXL, but was hoping my specs are enough for having under one minute creations on latest SDXL or free cloud option with some anonymity enabled.

[soft zealot]

I run a 1080Ti that cost me <£200 a few months ago (with fitted waterblock plus the OEM cooler) and that can generate standard sized 1024 x 1024 images using SDXL models in Comfyui workflows in 50-60 seconds.

[dry pollen]

yeah, that's the Nvidia bias 🙂

[soft zealot]

ah you're an AMD person . My acommiserations ;o)

[dry pollen]

haha, yeah poor AMD again, constantly beaten in the game 🙂

[shy kelp]

how many steps for 60 seconds?

[soft zealot]

keep the AMD for gaming and buy a 1080Ti for platyng with Stable Diffusion. Sure its not the latest gen but it meets your <60s image generation criteria and theyre relatively inexpensive on the pre loved market

[soft zealot]

I srick at 25 steps for everything by default

[shy kelp]

20/25 steps does for 30 seconds in amd last year gpu

was doing 25/30 at 36s or something

[soft zealot]

well thats nice isnt it

[shy kelp]

yea it's cool

it cost more cuz it was the whole covid thing at the time but they're cheap now

[woeful patio]

anyone know if easy diffusion supports controlnet?

[soft zealot]

wonders which of us is being the most sarcastic ;o)

[shy kelp]

wym

[soft zealot]

sigh Never mind, moving on out ;o)

[shy kelp]

im not sure if it good or not lol

drawing football is wonky

[woeful patio]

Trying to reinstall a1111 and keep getting this error:
ERROR: Could not find a version that satisfies the requirement tb-nightly (from gfpgan) (from versions: none)
ERROR: No matching distribution found for tb-nightly

anyone have any idea?

[uncut fiber]

How you reinstall? Clean install?

[shy kelp]

hands is hard

[soft zealot]

oh I have no compalinst about the hands, they look to have the right number of fingers erc.

They just dont appear to match the rest of the body lol

[uncut fiber]

Mickey mouse gloves 😄

[shy kelp]

how long does it take you to generate that image tho

[soft zealot]

deosnt he only have 3 fingers?

[uncut fiber]

4

but his gloves hide everything 🙂

[spring fulcrum]

12 years

[soft zealot]

upscaled to that size? around 120-130 seconds

Less time than it takes to make a cup of coffee or take a dump ;o)

[uncut fiber]

2048x1152 it is divideable by 16 and as well 16/9

[shy kelp]

cool

[woeful patio]

for future people coming across this problem.
this is how to solve it:
Focus on the error "No matching distribution found for tb-nightly". This error means that there is no corresponding "tb-nightly" dependency package in our pip source.

View the configuration of pip: pip config list, see the configuration as follows:

global.index-url='https://pypi.tuna.tsinghua.edu.cn/simple'

This shows that the tb-nightly package is missing from the pip source of Tsinghua University. Then let's change the source of Ali and use the following command:

pip config set global.index-url https://mirrors.aliyun.com/pypi/simple

Check it out again: pip config list, the configuration has become Ali's address:

global.index-url='https://mirrors.aliyun.com/pypi/simple'。

Re-execute the installation: ./webui.sh

[spring fulcrum]

It is pretty fantastic... AI art has come a long way in the last year for sure.

[soft zealot]

why the fascination over how long it takes?

Speed isnt everything.

A Mini will get you from A to B as well as a Ferrari .

and yes I am fully cogniscent of the fact I'm not running the worlds fastest GPU . So what ?

Tortoise & Hare 🙂

(apologiers the whole speed is good is a bugbear of mine lol)

[uncut fiber]

longer it takes one appreciates it more

[spring fulcrum]

That's actually pretty awesome considering the tech you are using. That just proves this process is improving a lot.

[shy kelp]

sorry mate

[soft zealot]

rest of the system is built around 5600X and 64Gb RAM.

I've never been a gamer per se. Originally I was running an old 980Ti in here , the 1080ti was an upgarde to get 5Gb more VRAM

[spring fulcrum]

Mine all take between 60 and 90 seconds upscaled to roughly 4k ballpark depending an chosen resolution... I prefer a greater level of control over speed and I know you do to Mr. Woof... great workflow

[soft zealot]

oh and with the water coolong GPU temps never go above 38c when generating inages lol

[fair stag]

I literally see this in every thread. People obsessed with it/s

[soft zealot]

I would love it/s lol

Meanwhile back in the real world is s/it ;o)

[shy kelp]

whats a good upscaler?

[fair stag]

Haha

[spring fulcrum]

wow Mine usually get upto around 65c when generating.... But hey cooling an i9 and a 4090 isn't easy and there is no AC down here... Its an oven... Especially with my wife's pc with similar specs and my dual xeon server running 24/7

[shy kelp]

dont think anyone is obsessed i was just asking if it was normal cuz someone said amd doesnt work

[soft zealot]

Im not using AMD either lol

[spring fulcrum]

amd CPU should be fine... Its AMD GPUs that seem to be an issue

if you aren't using CPU to generate that is

[shy kelp]

yea im running amd gpu

[spring fulcrum]

there's the rub

[fair stag]

I'm using AMD

[shy kelp]

can you train on it?

[fair stag]

Yup

[shy kelp]

it was failing on me with kohya train_network

[soft zealot]

I would rather poke my own eyes out with hot needles than run on CPU only lol

[fair stag]

Lora, haven't tried dreambooth full model

[spring fulcrum]

I've been getting pretty good results with my setup and workflow.... check this out

[fair stag]

Which might need more VRAM

[shy kelp]

how do you do that?

[spring fulcrum]

ya the only way I could see that working is well never.... unless nvidia starts making their own consumer CPUs and get rid of the need for GPUs... that would be great

[soft zealot]

wonder how a ThreadRipper would do ?

[spring fulcrum]

lots of practice

[spring fulcrum]

that might do alright but still slow compared to GPU... structure isn't made for it.

[fair stag]

Which card do you have, and what was the error... Might just need some different settings

[shy kelp]

6800, was giving size mismatch after loading the model and saying Missing key(s) in state_dict: "down_blocks.0.attentions.0.norm.weight",

[soft zealot]

@spring fulcrum dio you klike my idle temps 🙂

[spring fulcrum]

That is awesome... My idle temps are sitting right at 44c right now.... when absolutely idle they are about 33c

[fair stag]

I have the 6800xt.... You should be able to get it working. I used train_unet_only flag in extra options, disable xformers, enable the other memory one, I think flash memory attention or something, I don't have it open ATM. Use adafactor optimizer, with constant learning, and LR of 2

adafactor is better on VRAM

[shy kelp]

thanks

[fair stag]

I did bf16 mixed precision, but fp16 should also work

[soft zealot]

I may not have the fastest GPU in the world but it may possibly be the least stressed GPU in the worl lol

This si about about 5 or 6 minutes into a batch run (whilst foing other stuff such as "work" lol

[spring fulcrum]

I wish I could keep my card that cool it is really way too hot in this room for high temps to be coming from my setup

[fair stag]

And if you're training sdxl, you'll want the res at 1024x1024, default is 512

[uncut fiber]

sdxl seems almost perfect for watch dial Roman or Classical numbers.

@fair stag somebody yesterday posted resolution on which it was trained. Not just and only 1024x1024

[spring fulcrum]

Perfect eyes on the close up image.... Great hands on the zoomed out image

[fair stag]

I haven't played around with other res, but I think training at 512 wouldn't be good

Simply because the base model isn't good at producing it

[uncut fiber]

there is 576 even 512x2048 maybe, cant find it unfortunately

going to try 512x1024

[fair stag]

That might be a good thing to try, if 1024x1024 gives oom

[soft zealot]

also helps my ambients arent too high at aroun 22c

[uncut fiber]

it is working 512x1024

here i have almost 8 degrees more 🙂

[spring fulcrum]

my ambient temp is about 30c

[uncut fiber]

apple-punk, i like that anthene 🙂

[full idol]

Do I need to duplicate my SD checkpoints from auto1111 folders to comfyUI folders?

cant comfyUI just take the models from the auto1111 checkpoint folder

[uncut fiber]

i think it is not NEEDED :), and it is described on github comfUI page

[soft zealot]

nope way to warm for me lol

[uneven verge]

crazyy good, is there anywhere i can learn to make images as good as yours?

[spring fulcrum]

If you use comfyui you can just drag and drop the image right into your browser and it should load up the exact workflow

[polar epoch]

What node/node pack can scale by height? It has one for width, just not height. As i'm trying to regenerate it, but wider

[soft zealot]

Take a look at this from Deerfu

weill upsacale or downscale to the lntghth given based on an input images longest side

[high skiff]

I am certain

[soft zealot]

'nuff said

[high skiff]

GPU z, hardware info 64, and afterburner. I am able to see the power draw from all three of my AX pin connectors, as well as my PCIe Lane

[high skiff]

I have tested it extensively by killing every process on my computer, and the only point in which my GPU clocks down to base idle clocks is once I close out of command line as a whole

As I've said, it's been a problem for about half a year now

Or at least, I've known of it for about a half a year

[soft zealot]

most ofdd, has anyone else said theyre seeing thesame pattern of behaviour?

[high skiff]

I talked to a few other people who said that they were having this behavior in my research group, but it seems to be a minuscular amount of people

Figures, it only happens to the person who lives in a region that is extremely hot and has exceptionally high electricity rates lol

[steady grove]

do you have anything like nvidia instant replay or whatever its called running?

[soft zealot]

fortunatelky I dont live in a hot part of the world (as you may have seen my Ambient & GPU temps are quite cool lol #✨｜sdxl message)

Electric is expensive everywhere.

For reference Im paying
Standard unit rate 7.513 p per kWh (approx 0.101USD /kWh)
Standing charge 29.106p per day (approx 0.369 USD /Day)

I also think it probably costs more to keep my 300l (80 US Gallon) aquarium at a constant 26.5c/79.7f in ambents of 18-22c /65-72f than it costs to run my GPU although maybe I should actually calculate that one day and scare myself lol

[vital wraith]

I live in the senoran desert. My GPU is a little heater. And as for the comments on resolution earlier, you can adjust from 1024x1024 if you maintain specific aspect a ratio such as 1152x896.

Exact pixel count is pretty important I’ve seen.

[polar epoch]

Aye. But i don't want it's side, i want it to scale based on height to retain height, but "outpaint it" to be wider by regenerating the entire image since we don't have controlnet outpaint for SDXL yet.

[soft zealot]

Use the Gimp and reimport

[soft zealot]

or use this which I believe is core COmfyUI

[high skiff]

Unfortunately, our electric prices are considerably higher than that

[soft zealot]

really ?? And yet I've seen seevral NA Youtubers commenting on that in their opionion prices over this side of the puddle were higher

[high skiff]

Our current plan is a 24-hour fixed rate, and it is $0.43/kwh

Or, we could switch over to a day night cycle plan, where our night will be $0.26, an hour day will be $0.68

[ionic dragon]

@high skiff whats the current status of your refund?

[high skiff]

I escalated it to a supervisor, called her out on her bullshit, proved her wrong through her own company's emails, and forced her to comply because somebody above her told me improper information in my favor. I have a scheduled call with her later today to handle a PayPal personal payout rather than an actual refund

[ionic dragon]

oh ok

i hope everything goes smooth and you get your money back

[soft zealot]

theres an interesting tangential thought given not all US customary Units are Identical to UK Imperial measurements (eg a US Gallon is smaller than a UK Gallon) . areis a us kW/h the same as a UK kW/h

Hmmm

[steady grove]

we got 2 tier charging here. 10c /kwh for the basic usage, than anything over the threshold is 15c/kwh. all my appliances eat the basic rate so i count gpu at 15c

[high skiff]

It's a kilowatt, it's 1,000 w

[soft zealot]

I guess this lies then
https://www.energysage.com/local-data/electricity-cost/

[high skiff]

Yeah, that's average across the US

I live in a very hot and expensive part of the US

[steady grove]

the US uses metric in some cases, just like the UK uses imperial in some cases. a metric gallon is more of a french gallon aint it?

[fresh path]

what would imperial version of kwh be

[soft zealot]

theyre also both called gallons but yours is only 3.78 litres whereas ours oare 4.54 lites (meaning our pints are bigger as well lol)

[high skiff]

this is our rate if we switched to hourly, its actually higher than I remember

[fresh path]

true

[soft zealot]

oof

[fresh path]

thats so much

[high skiff]

yeah, my mom and I are getting taken for all we have

[fresh path]

ever thought about solar? with those prices it would pay off in a few years

[high skiff]

we aren't even living comfortable, and our electric bill is almost $500 in a decently small house

[high skiff]

yes, but thats somebody my mom can't afford at the moment, so I will be looking into a custom reclaimed solar array when I get a job

[ionic dragon]

wait, a gallon is different across countries?
isnt a unit meant to be constant?
it breaks the fundamental of calling it a unit

[high skiff]

no, they are different units with the same name, its hard to explain

I think ton and tonne are different as well IIRC

[ionic dragon]

500?
thats too costly if considered in my country, like really costly.
idk abt the electricity bills in the us

[soft zealot]

Im a knob, those figures were my Gas Rates not my Electric rates!!!

so that equivalent to appros US$0.37 per kWH with a approx US$0.66 standing charge per day

[high skiff]

ok, so not too far off from us

Our flat rate through the whole day is $0.42, or we can do the 28/65

[soft zealot]

ones a US gallon, the others a UK gallon / Both are defined against the litre

[ionic dragon]

dude, even $0.37/hr is too costly for electricity, isnt it $0.07/kwh if you choose to pay for consumption?

[high skiff]

no lmao

[soft zealot]

well I didnt know that

--
Way back - at the time we all lived together - we had used the Queen Anne's gallon of 3.785l to measure wine. We also had different volumes and names to measure both beer and grain. When we realised that this was silly, in 1824, the UK chose a single imperial term and lumped together all three measurements, picking a roughly-average volume and calling it a gallon. American colonists, though, stuck to tradition and kept just the Queen Anne's volume for their own gallon definition.

---

https://www.thecalculatorsite.com/articles/units/us-gallons-imperial-gallons.php

[high skiff]

I wish we were still in 1990 with rates like $0.07kw/h

instead, my mom and I are being shook down like nerds by a school bully for lunch money

[ionic dragon]

in my country, its less than $0.001/KwH

wait

its $0.084/KwH

[steady grove]

i got an idea. new technology energy companies could use.

https://tenor.com/view/fat-gogeta-lmao-lmfao-lol-gif-13380086

[soft zealot]

Mind you wehn comparing cost of living its not as straightforward as compationg costs directly.

Average Wages etc also need to be taken into account.

Strictly speaking for a true comparison it should be the cost of an item as a percentage of avearge wage (or some such metric)

[high skiff]

yeah

[ionic dragon]

yeah

the avg income in my country is pretty low

[steady grove]

average wages being accounted for i think drives costs up. its what i am seeing here in vancouver region. minimum wage goes up and increases the average wage. landlords and realtors all account for that and drive the cost up calculated from what they think a full time worker would make

[high skiff]

thats exactly how economics works

raising the minimum wage never helps, just closes the middle class gap to the poverty floor

[soft zealot]

I liked Vancouver when I vistied there back in 85

[ionic dragon]

are the sdxl CN models officially released by SAI

[steady grove]

its a pretty obvious theory. BC has the highest minimum wage in the continent and the highest rent prices. https://twitter.com/j_mcelroy/status/1691663491135099095

Justin McElroy (@j_mcelroy)

hahahahaha we’re doomed

Likes

6618

Retweets

561

▶ Play video

Twitter · Aug 16, 2023, 4:09 AM

[soft zealot]

travvelled across the contry bu train to Toronto stopping off en route.

Local breweries were on strike so could only drink imported beers and whilst I was in Toronto Air India Flight 182 got blown up mid atlantic

[high skiff]

our minimum wage here is basically $16.50

[ionic dragon]

oh damn, CA way back in 85, what was the population like?

[high skiff]

120 people

:p

[steady grove]

lol to prevent scraping i guess twitter has disabled video embeds now. content delivery is getting so shitty

[high skiff]

Twitter is dying anyways

[soft zealot]

sont you men "X" ;o)

[high skiff]

I hate to say it, but threads is gonna take over, one they fix a couple issues and default to dark mode

[steady grove]

vancouver was huge in 85. i think 86 is when the world's EXPO was there

literally the world's fair

[high skiff]

Elon is too much of a fucking joke for me to call his little shit baby site by its code name lmao

[soft zealot]

yes that was 86, they were building stuff ready for it when I was there

[steady grove]

x.com was first. it predates paypal and is one of the original dotcom boom successes

[soft zealot]

notice even he cant get x.com lol

[high skiff]

yeah

[steady grove]

hhe's got it. it redirects to twitter

[high skiff]

I think people forget that Elon has never been smart, he has done like nothing

[soft zealot]

bugger

[steady grove]

i think they're using the old domain mainly for legacy support

[ionic dragon]

i dont think so, as the controversy kings wouldnt move there instantly and even famous celebs posting on X wouldn't change

[steady grove]

landing rockets is huge but that wasn't him. engineering teams under his employement

[high skiff]

well yeah, X will stay around for the radical biggots until they all die off and X fizzles out in a $50B whimper

[high skiff]

They can't even manage to do that consistently for like more than 3 times in a row lol

their newest test lasted for what, 1 minute in air?

[dense chasm]

Japanese is the second largest active group on X,lol..

[steady grove]

what do you mean? they've been landing rockets for years

[high skiff]

and also caused like $75M damages to their take off site, while also damaging infrastructure for miles

[steady grove]

the last hover test was for starship booster iirc. that thing is bigger than any first stage thats ever been built

lol yeah that was epic

[high skiff]

yeah, and it was broken on launch

they made a big booster using a bunch of impromperly made small boosters

[upbeat summit]

Some analog photography images made with Aether Real - a SDXL model by @icy brook

[dense chasm]

global market and supply chain,don't forget manufactory asks for more parts

[high skiff]

where is Aether real out?

[ionic dragon]

oh, never knew that

[polar epoch]

Neat! So it will regenerate OG, but add more to left and right? As currently now, the wider just squishes top and bottom by the looks of it

[upbeat summit]

you have to ask joachim about that

[uncut fiber]

do you know style?

[steady grove]

no that's not how the engineering on this beast works. it uses all new types of engines

hate elon sure, but spacex is awesome and doing it

[soft zealot]

random aside:

I rmeber as a 5 year old child being allowed to stay up late at my grandparents and watch the Eagle land on the moon

[high skiff]

SpaceX is nearly as big of a loss of cash flow as X, and has just as fake and misleading promises as Tesla

[soft zealot]

no idea, not used it. just found it

[steady grove]

elon's an idiot and you can be assured that he's just doing busy work over there, if anything

spacex is designed to be a cash eater. the investors there want elysium

[high skiff]

I am not even talking about Elon, SpaceX is also another failing company

[steady grove]

i dont see spacex failing. literally going to the moon

their strategy is outside of profits. it's a big big deal

[high skiff]

they lose like $6B a year, and have made nearly no successful trips other than putting some satelites in the atmosphere (and sometimes still exploding), in like 2 decades

[uncut fiber]

some sattelites? It is almost ring

[high skiff]

also, their numbers they stated for the projected growth are impossible, which is interesting

[steady grove]

$6b a year is a drop in the bucket compared to the profits heavy lift systems will bring them

[high skiff]

thats all speculation

[steady grove]

it's a frontier

the final frontier

[uncut fiber]

wait 6b like 9 or 12 zeroes 😄

[steady grove]

these are the voyages of the starship spacex

[high skiff]

You are huffing too much of their promises without looking at the numbers lol

[polar epoch]

Nice, thanks:P It did something, but not outpaint sadly

[ionic dragon]

i wouldn't consider $6B loss a big loss considering its valuation

[uncut fiber]

OpenOutpaint?

[high skiff]

Elon is the same guy that said we would already have a colony of humans on mars by today on multiple occasions lol

Yet they can't even get a rocket out of low orbit

[steady grove]

pennies for them really

[high skiff]

where do you think they are getting money?

[uncut fiber]

european billions are 1000 x more

[polar epoch]

I will try it

[ionic dragon]

yeah exactly, the new age startups lose billions, this isnt very big for spacex

[steady grove]

the falcon can insert into high orbit. maybe you're huffing to much anti elon sauce

where do you think the ISS is stationed?

[high skiff]

oooo, high earth orbit, that is practically on the surface of mars a year ago when Elon musk predicted

[dense chasm]

India on rise

[ionic dragon]

wait,

[uncut fiber]

@polar epoch i tried it only with A1111

[polar epoch]

Or not. As comfy is way way lighter on vram than automatic's, and openoutpaint is an extension of automatic, i can't use it as i plan to make it 1080p/1440p ultrawide.

[steady grove]

lets hope their lander touches down softer this time

[high skiff]

I remember that lmao

[steady grove]

i'm impressed they've already got a second one launched

moving fast

[soft zealot]

How are the CHinese & Indians doing these days anyways?

[high skiff]

I am just wondering how long it will take for Elon musk to go away for life for the sheer dozens of times he has lied to investors and consumers

Only thing I can think of that hes made properly is Tesla's, and even that is debatable

[soft zealot]

everyone wonders what Musk is up too but there are otheres.............

[high skiff]

yeah, there are smart people

[soft zealot]

that started by converting a Lotus lol

[high skiff]

then there is Elon lol

[high skiff]

exactly

[steady grove]

chinese are getting a space station up there bigger than the iss

[high skiff]

by a CEO before Elon

[ionic dragon]

the govt is shit, they are pro-religious political party

[steady grove]

thye've got a massive heavy lift campaign going on and are launching huge modules

[soft zealot]

and he bought in to it rather than createdit

[high skiff]

I think people forget that Elon got rich off of being fired from PayPal

[steady grove]

paypal bought x.com so they had no competition. they named elon a co-founder. hardly "fired"

[dense chasm]

tax payers lol..

[steady grove]

that's a huge win

[ionic dragon]

thats not what i would be concerned about, if they get their hands on GPUs, then i am worried

[steady grove]

you think they aren't producing models? they are

[ionic dragon]

and the tax is also is too high

[ionic dragon]

they are, but i am talking about semiconductors

China had hard time cracking semiconductors

[dense chasm]

it's more like geopolitical

[soft zealot]

wonder how effificent a GPU running in a vacuum at sub zero temps is??

[ionic dragon]

you need to be precise, so i can answer

[ionic dragon]

wait, are you talking abt superconductors

[uncut fiber]

@polar epoch yes i forgot i am in SDXL. But for A1111 it is miracle. You can outpaint in it very large image, not limited by sizes even it is A1111 1.5 or similar models

[steady grove]

i often wonder about supercondictive transistors. isn't resistence required to prevent the current from passing throuhg the circuit?

all bits are 1 with no resistence right?

i dont know

[polar epoch]

That's why i use stable diffusion infinity. Largest canvas i made there was 5000x5000 something, and it barely touches my ram and video memory :P

[high skiff]

I look forward to BYD coming over to America more and pushes Tesla out of the US market

[ionic dragon]

does BYD have auto-driving?

[high skiff]

not that I know of, and neither does Tesla-

[ionic dragon]

Tesla atleast has some kinda auto-driving

but thats not the main thing, but Tesla is very hyped

[high skiff]

it has an "auto" pilot of which has caused several crashes

They are far behind their competitors by a lot in self driving, its kinda sad

[ionic dragon]

so i dont think its gonna fall down anytime soom

[high skiff]

Teslas FSD is a joke, and a sham

[steady grove]

i think it's ready and good to go. just a bunch of red tape in the way

[high skiff]

Tesla is only at Level 2 Autonomy

[steady grove]

safer than any human on the road imo

[high skiff]

bruh

[steady grove]

seirous. people suck at driving. fucking maniacs out there

[ionic dragon]

Elon is way better than many CEOs

he knows his shit

[high skiff]

oh my god, the amount of huffing in here is insane

[steady grove]

i dont care if its tesla self driving, apple, toyota, whoevers self driving. human drivers are just dumb

lets get them off the roads

congestion? it's all because of dumbass drivers. roads can support all the traffic we throw at them until the human element is in play

[dense chasm]

rich ppl usually do not be a driver,it's a very complicated situation of electric car development and the gasoline car

[high skiff]

fair enough, but Tesla is nowhere near good enough to do that while they are still causing car accidents that are avoidable by humans with basic perception skills

[steady grove]

red tape is the only thing preventing deployement of these systems really

[high skiff]

Thats funny, cause Waymo is already at Level 4 autonomy and they need 0 human monitoring for their steering wheel and have been running for years 😅

[soft zealot]

no I was just thinking out loud

[dense chasm]

it's a niche market,food, close,resident,the last is automobile,climate change is another big issue

[high skiff]

granted, Waymo is still far off from what we need, but they are leagues past Tesla FSD

[steady grove]

only in what? 3 or 4 major cities? red tape

does google sell waymos to everyoen?

[high skiff]

The problem with waymo is in order to be way better than tesla, which they are, they need a lot more data than tesla

but I mean, teslas are still getting into car accidents when being summoned, and slamming on breaks for no reason in FSD mode, which has caused several accidents

[steady grove]

i remember when waymo first hit the scene and i said to myself "that design won't last long"

one of those "wtf where they thinking" moments

their new model is the "Waymo one!" trying to pretend that thing didn't happen

[elfin flare]

The golden melon graces us with its presence.

[dense chasm]

have u ever heard of Faraday Future

[steady grove]

yeah i think thye're ugoh too

[steady grove]

bit of a joke imo

their marketing hype. it's an all ability car with hyper+

all-hyper hyper+ systems

they're going to make people take electric cars less serious. it's like if homer simpson made a car

[fossil thorn]

Self driving without human intervention is still very much an unsolved problem. It might be barely capable of some highways or in highly documented urban areas, but no one has been able to demonstrate a system that is effective given arbitrary roads. Especially poorly maintained roads which are unfortunately way too common.

The whole thing with SD suddenly adding extra legs or fingers? Self driving cars have their own version of that problem, except they see false road signs or miss signs that are actually there.

It is getting better, but the technology isn't there yet. At best, we might soon get cars that let a driver take a nap while on roads that the car is capable of like well maintained highways, then wake the driver up when entering secondary roads.

Which would frankly revolutionize things like long-haul trucking, but is a long way from the hype 7-8 years ago that human controlled cars would be gone in the next five years.

[steady grove]

just an engineering problem really

[graceful wadi]

hello! i'm getting an error from my own trained sdxl model, can anyone help out here?

[autumn forum]

Hey mass is that aether real model public yet? Looks amazing!

[ionic gulch]

bad idea to use ai for things that allow zero mistakes. much better to use it for things that allow 10% error rate as long as ai saves 15% time.

[upbeat summit]

Thank you! Glad you like the images. I'm currently working on and testing @icy brook's model. It's not publicly available yet.

[autumn forum]

Nice, I’m looking forward to playin with it

[thorny frost]

is there a node to allow you using local LLMs to generate prompts in ComyUI?

[wet nacelle]

At this point in time is it reliant on a high positive and negative to get a good output in regards to texture?

[polar epoch]

Am i missing something? Wanna make a portrait image of a bunny holding a cheese, but the geneation makes a annoyingly tiny one instead, Using a 1080x1920 white image as the img to img

[steel mist]

Didn’t see it in scrollback… did you get an answer to this?

[crisp owl]

#✨｜sdxl message

[uncut fiber]

@polar epoch 1080 is bad idea. not divideable by 16, try 1088 or 1024

[vale eagle]

Why would you use llm in comfyui? You could just use it seperate

[vocal stream]

anyone have a decent easy to adjust Comfy workflow? There really are very few that get traction on civtai

[high skiff]

ping?

[high skiff]

oh, no

its not out

[wet nacelle]

https://tenor.com/view/2000-500-gif-7343783

[steel mist]

Bummer, thanks!

[autumn forum]

What do you mean easy to adjust? Like just a simple workflow with some knobs and a output?

[vocal stream]

Like something where I can just easily click to change models or use/dont use refiner/upscaler/img2img

but I'd take anything people like to test out really

[autumn forum]

I mean there’s plenty of workflows that do that on civit. Sarge, sdxl ultimate to name a couple

[vocal stream]

Searge doesnt even have a knob to disable refiner and multiple other things

and I oom hard with it

[steel mist]

The magic for me was learning that ctrl-b bypasses nodes

[vocal stream]

tho currently instead I get a way more bizare vram oom with sarge

Requested               : 3.75 MiB
Device limit            : 23.99 GiB
Free (according to CUDA): 14.62 GiB```

[steel mist]

So on those big complex ones you can just ctrl-b on the retailers and upscalers etc and then just drag out a single preview from the base or base+refiner

And it makes them a little easier to work with because you’re not trying to understand a giant 50 node workflow

[vocal stream]

getting pretty disappointed that stability decided that comfy is the future and that that it'll get enough momentum because of it to become the standard

[steel mist]

I don’t think that’s true?

[uncut fiber]

nothing is mandatory, and it running on background, SwarmUI will be official front end

[crisp owl]

Stability said nothing of the sort

[azure oxide]

comfy should be the standard, node based is absolutely based

[vocal stream]

how is it not true? Didn't the guy behind Comfy had access early while automatic didn't which made them ahead + is hired and chosen by stability thus actually gets funds to work on it

[steel mist]

I am honestly not sold nodes as the best workflow

[azure oxide]

while automatic didn't
did you read this from the reddit thread

this is how fast misinformation is spread lmfao

[vocal stream]

and people are already using Comfy because it's ahead due to stability picking it even when they dislike the workflow which means things will get built on top of it because it's head

[steel mist]

Well it’s also the most flexible, so it kinda is what you’d want to build on top of

[vocal stream]

no from here mostly but I am asking is it not true?

[azure oxide]

SAI have said that they worked with the big bois from the very start with sdxl: comfy, a1, kohya

[vocal stream]

is the Comfy guy not hired by Stability and paid to work on Comfy?

[azure oxide]

a1 didnt care to make an updated workflow as much as comfy did

[hardy cipher]

should I add more LoRAs? make it a 5 stack or a 10 stack?

[vocal stream]

the stability team definitely pushed Comfy a lot in the launches and said that's what they prefer to use and fund

[azure oxide]

yea is there something wrong with that lol

[steel mist]

Is comfy funded by SAI? I was actually looking for a tip jar cause I’d love to kick in for all the work

[hardy cipher]

I think they're allowed to have preferences

[crisp owl]

Well, when you have a preferred workflow, and one platform supports that, and another platform won't update to allow that....what would you do?

[vocal stream]

not wrong, but as I said I'm disappointed we are going to get stuck with it because of it since it's way less pleasant to work with

[crisp owl]

That's a developers problem, not a stability problem

[azure oxide]

i dont see how a1 is affected by comfy getting more popular

[crisp owl]

Get mad at A1111

[vocal stream]

they can obviously make decisions by their preferences, but I can be personally less pleased for myself when they meet my needs way worse

[hardy cipher]

when I first downloaded a1111 I pretty much assumed that it was the official UI for stable diffusion

so they had the ball in hand

[vocal stream]

I mean isn't a111 doing it for free basically why Comfy is getting funded by stability?

if that's not true then sure it's just worse vs better development

[hardy cipher]

I hope he is getting funded

people that make useful things that are used by a bunch of people should get compensated

[crisp owl]

I won't make any official statements as I don't know, but I highly doubt the A1111 dev is doing all that work for free

[vocal stream]

The lead developer of Comfy was hired to Stability
https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Motivations.md

[wet nacelle]

My first UI was InvokeAI. I then moved to Auto1111. I now use Comfy and Fooocus UI.

I still think that Invoke has the most user friendly UI yet.

[azure oxide]

everyone in the scene has always been doing it for free, thats kinda what foss is
just because someone gets hired on doesnt mean quality loses or something

[hardy cipher]

I think that a1111 just got bogged down by two many people or something

[crisp owl]

Stable Swarm is not Comfy. It uses Comfy

[vocal stream]

yes.. how does that change the quoted statement?

[hardy cipher]

tenoke, what's your issue?

[wet nacelle]

Rather than them reaching out (I believe since it is opensource) they are just using it.

[vocal stream]

again it's not an issue. Just personally I don't enjoy Comfy and personally it is worse for me that it is on its way to becoming the standard because stability chose it

[wet nacelle]

since snese? which should I use there?

[hardy cipher]

stability didn't change anything for you

the developers of a1111 aren't keeping up

[vocal stream]

if they pick one UI, and that UI is head then others work on top of them and make extensions/tooling for them before anything else

[azure oxide]

thats the thing, its fine to have preferences for one or the other but how does comfy getting more popular hurt you or a1

[wet nacelle]

DUCK!

https://tenor.com/view/wave-hello-duck-hi-howard-the-duck-gif-7246353

[steady grove]

Just using it?

They literally got comfydev on the payroll

[hardy cipher]

making extensions and implementing code for a1111 is no different now than it was before

[vocal stream]

A ton of extensions were only for a111 before since it was the most popular. Now they are going to be on Comfy and not a111 or something else I actually like working with

[crisp owl]

Yeah that's the entire underlying issue.
A1111 doesn't want to change to support, or at least they haven't, to fully support the workflow of SDXL.
That's entirely an A1111 issue, and zero to do with Stability

[wet nacelle]

Then I'm wrong and should've had better wording.

[steady grove]

Open source code is easy to port usually

[azure oxide]

ya its all open source so anyone is free to create new extensions

[vocal stream]

in theory sure. In practice were all extensions ported for the 3rd most popular UI before? not really

[crisp owl]

Again, developer problems

[hardy cipher]

unfortunately the world does not cater to any of us

[steady grove]

Anyone can port an extension.

This has only been around for barely a year. Confirmation bias is a bitch

[vocal stream]

like is there even stuff like deforum or roop on Comfy now for example?

[hardy cipher]

yes

[cobalt oracle]

Hi, has anyone here finetuned the SDXL model with Kohya_ss gui?

[wet nacelle]

Does it run well?

[uncut fiber]

DUCK Howard?

[hardy cipher]

anyway, rather than be upset with comfy or with stability, maybe focus your frustration on the a1111 developers for dropping the ball

because they had the lead by a huge margin

and goofed

[vocal stream]

as of 3 days ago the issue on their repo says no way to install deforum and no repo shows up https://github.com/comfyanonymous/ComfyUI/issues/1134

[steady grove]

Not deforum but tools like deforum yes. Roop yes

A1111 could've switched to better architecture long ago but voldy argued and knew better

[steady grove]

No one stopping anyone from using automatic. Judging from what's on civit, it's what most use still.

Not a lot of comfy meta data on civit

[vocal stream]

again, cutting edge stuff mostly get done for the most popular one and there's momentum where the next person builds on top of it. Yeh you can use a less popular UI but the less popular it is the less support it will have for newer stuff

[uncut fiber]

yes and lot of users went comfy just temporary until SDXL will work with A1111

[steady grove]

Comfy is cutting edge and wasn't popular at first

[steel mist]

SDXL doesn’t work with A1111? What? That’s what I started with like 2 weeks ago

[vocal stream]

it's cutting edge and ahead in part because that's the one stability is paying to be developed full time

[steel mist]

Was there a regression?

[compact hatch]

a1111 requires more vram than comfy to run it

[steady grove]

A ton of the bleeding cutting edge stuff comes out for a scripted interface most of the time

[compact hatch]

comfy is lean which helps when using big xl models

[steel mist]

This is like, how software eng works? The closer you are to “the metal” the faster you can release

[steady grove]

Then it gets ported by users of different projects

[vocal stream]

it kind of works with a111 but there's even more memory leaks, problems with stuff like refiner or more complex workflows for now

[crisp owl]

There is no regression, A1111 does not allow for splitting of clip, as is a large underlying thing behind SDXL

[hardy cipher]

you're so far off base, lol

[crisp owl]

That's an A1111 issue, not an sdxl issue

[vocal stream]

in what way?

[steady grove]

Most of comfy got made on comfydevs time

Only recently was he hired

[hardy cipher]

yes, and most of the nodes are made by people for literally no money, or maybe a nominal amount

same for workflows

people getting some donations from people that appreciate them

[steady grove]

A1111 has to gut his entire pipeline to introduce multiple clip support I think. Why it's taking him so long

[hardy cipher]

but beyond that not like they're on payroll