✨｜sdxl

[eternal fog]

    from .AITemplate.AITemplate import AITemplateLoader, AITemplateControlNetLoader, AITemplateVAEDecode, AITemplateVAEEncode, VAEEncodeForInpaint
  File "D:\Code\Stable-Diffusion\ComfyUI\custom_nodes\AIT\AITemplate\AITemplate.py", line 334, in <module>
    from comfy.sd import ControlBase
ImportError: cannot import name 'ControlBase' from 'comfy.sd' (D:\Code\Stable-Diffusion\ComfyUI\comfy\sd.py)

Cannot import D:\Code\Stable-Diffusion\ComfyUI\custom_nodes\AIT module for custom nodes: cannot import name 'ControlBase' from 'comfy.sd' (D:\Code\Stable-Diffusion\ComfyUI\comfy\sd.py)```

[visual glade]

if you update the node is it fixed?

[eternal fog]

If you are talking to me, the AIT Node is fully up to date, according to git pull anyway

Apparently IPAdapter is broken too

Input type (torch.FloatTensor) and weight type (torch.HalfTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor```

That's not been updated for 2 days though, so I'm guessing the change you've made has broken how they are doing it

[cedar tusk]

I reverted back to commit 1300a1bb4c87246b846c6209ab65ef0553004290 and I'm not seeing the error with IPAdapter. So I think the changes to the clip vision logic did stop something from working due to changed syntax maybe

[eternal fog]

Yeah it's almost certainly clip vision, as he's changed a few things with it.

[soft bone]

anyone know?

[cedar tusk]

Does hitting 'refresh' not do it? @soft bone ?

[eternal fog]

Could you remind me how to do this please?

I thought it was git checkout and then the hash, but it complains about that

[cedar tusk]

Hmm, that's what worked for me

I followed it with a git reset --hard

And then redid my symlinks for models

[eternal fog]

It didn't like me using the full commit id

I had to do the short one

[cedar tusk]

weird

[soft bone]

hmm not sure how to do that from api. but seems like it should work. will attempt

[cedar tusk]

Oh, sorry didn't see you were asking about the API

[soft bone]

forgot to mention it. still i think that could be the answer

[eternal fog]

I wonder why the AIT module isn't working as well, I've seen people using it, maybe they've just not updated Comfy

Hang on, I think it wasn't actually updating via the Manager, just done it manually and it's not throwing errors anymore.

[cedar tusk]

With the Comfy revert?

[eternal fog]

Nah it wasn't working anyway

Guess I must have fucked it up somehow

[soft bone]

manager is very inconsistent for me

[eternal fog]

Yeah I'm not sure I trust it now

However AIT seems to push me over the VRAM limit when using it with the IPAdapter 😦

[cedar tusk]

What is AIT again?

[eternal fog]

AI Template, Optimisation that speeds up interference.

[wet nacelle]

⠀⠀⠀⠀⢀⣤⣴⠶⠶⠷⠶⠶⣦⣤⡀⠀⠀⠀⠀
⠀⠀⢠⡶⠟⠉⠀⠀⠀⠀⠀⠀⠀⠉⠻⢶⡄⠀⠀
⠀⣴⡟⣥⣤⣤⣤⣤⣤⣤⣤⣤⡀⠀⠀⠈⢻⣦⠀
⢰⡟⠀⠻⠿⠿⠟⠁⣼⣿⣿⣿⣷⡀⠀⠀⠀⢻⡆
⣿⠃⠀⠀⠀⠀⠀⣼⣿⡏⠙⣿⣿⣇⠀⠀⠀⠘⣿
⣿⡄⠀⠀⠀⠀⣴⣿⠟⠃⠀⠙⣿⣿⡆⠀⠀⢠⣿
⠹⣧⠀⠀⠀⢸⡟⠁⠀⠀⠀⠀⠘⢿⣷⡀⠀⣼⠏
⠀⠻⣦⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠻⠃⣴⠟⠀
⠀⠀⠙⢷⣤⣀⠀⠀⠀⠀⠀⠀⠀⣀⣤⡾⠋⠀⠀
⠀⠀⠀⠀⠈⠛⠻⠶⠶⠶⠶⠶⠟⠛⠁⠀⠀⠀⠀

[eternal fog]

Yeah AI Template increases my VRAM usage by over 1.5GB if used with IPAdapter

[wet nacelle]

@cedar tusk Aphex?

[cedar tusk]

Oh yeah

But why though?

[wet nacelle]

https://tenor.com/view/aphex-twin-come-to-daddy-screaming-creepy-scary-gif-13712953

[cedar tusk]

... doesn't really answer the question

Are you saying she looks like the big mouth from Come to Daddy?

[wet nacelle]

ni gga

no

[glad grove]

steven seagal x belle delphine photo when

[wet nacelle]

now???

I'm going to work on it

[glad grove]

[slender coral]

Hey lads, anyone know of a workflow or a way to have a list of prompts to que up and do?

Just want to automate the process.

[terse pewter]

how do I see all my generations in the bot? I can't find the generation I made like a week ago XD

[wet nacelle]

@glad grove

[eternal fog]

lmao

[glad grove]

sensei finally ascended 🫡

[glad grove]

perfect all american family

[terse pewter]

how do I see all my generations in the bot? I can't find the generation I made like a week ago ?

[wet nacelle]

It's like her mother is taking a photo of her and her dad.

[glad grove]

heard she calls steven "daddy"

[high skiff]

finally, a break from the cursed shit in this chat

[terse pewter]

lol i asked it to create a modern kitchen in northpole, it created this

[cedar tusk]

Ha, does what it says on the tin

[high skiff]

nailed it haha

[eternal fog]

He's? Coming

Here's a cursed one

[cedar tusk]

now you've got me plugging in game screenshots

[eternal fog]

Yeah those work very well

[cedar tusk]

100% de_dust

[eternal fog]

Yup

And I used an image of a castle with Revision

So that's why the background is like that too

[cedar tusk]

Try CS plus something cel-shaded like borderlands or XIII

I wonder how well the style will transfer

[eternal fog]

i can give it a go

[cedar tusk]

there we go, haha

[eternal fog]

Someones hit him really hard in the head lmao

[cedar tusk]

Man I really need more VRAM

[eternal fog]

yup lol

[cedar tusk]

You using base SDXL or one of the community models?

[eternal fog]

I'm using NightvisionXL

[cedar tusk]

I've tried dreamshaperXL and juggernautXL but didnt' see too much improvement over base so far

[eternal fog]

This just seems to work more consistently, without ht eneed for the refiner.

[hardy cipher]

I'm way out of my element here. wallish, what do you know about the input and output of these nodes? all the sudden mine seem to be labeled wrong, output is "detected_map" which doesn't even seem to be anywhere in the project code. I think I'm missing something that should be obvious and have no idea what it is

[cedar tusk]

which nodes?

[hardy cipher]

it's one I'm making. it worked before today. but current version didn't work, and none of the previous versions work now

last night it worked, updated, doesn't work now

[cedar tusk]

I had to roll back my Comfy to a previous commit

[hardy cipher]

it doesn't even run the debugging within the node

just says it's invalid

so can't see what it's actually doing. most perplexing thing is that it's calling the output "detected_map" I cannot figure out where that would have come from

[cedar tusk]

weird question, but are you sure that it's your node that's failing?

like does the console spit out lines from your code?

[hardy cipher]

no

[cedar tusk]

hmmm

[hardy cipher]

Failed to validate prompt for output 3:

• AdaptiveCanny 2:
  • Required input is missing: image
    Output will be ignored
    invalid prompt: {'type': 'prompt_outputs_failed_validation', 'message': 'Prompt outputs failed validation', 'details': '', 'extra_info': {}}

[hardy cipher]

I'm just not sure what it's not doing

other nodes work though, so if I swap it out for another node that inputs and outputs images it works fine

[cedar tusk]

Is there a place I can see the node code? at least the header part

[hardy cipher]

one moment. it's a bit butchered since I've been messing with it a lot

[hardy cipher]

you know, I just noticed something I want to check really quick.

[cedar tusk]

ha, that's how you fix problems

last minute 'hey wait a second...'

[hardy cipher]

well I was like, "what will someone else tell me is wrong" and noticed a missing comma, lol

[glad grove]

she lost her shoulder 😔

[slender coral]

Hey lads, anyone know of a workflow or a way to have a list of prompts to que up and do?

Just want to automate the process.

[hardy cipher]

you still there?

[wooden pagoda]

How do you stop getting multiplle limbs and limbs in places they really shouldn't be?

[slender coral]

what resolution you prompting at?

[wooden pagoda]

I keep changing to see if it will help - it seems nothing's helping. .

[wet nacelle]

This is super fucking odd. I'm trying to update Comfy UI and this is happening.

[wooden pagoda]

I'm not using ComfyUI

I'm using the webui

Rename your files,.. long names can scupper odds of opening or running files IIRC.

[wet nacelle]

@visual glade Do you have any clue of how I could fix this?

[wooden pagoda]

Esp the use of under-scores

[visual glade]

remove that file it's complaining about

[wet nacelle]

ah

I'll try that

[hardy cipher]

you're probably using the wrong resolution. if you're using sdxl needs to be 1 megapixel

[wooden pagoda]

1024 x 1024,. yeah, it's happening in that res'

[wet nacelle]

@visual gladeIt's doing it now with another image. I think it may do it for all of the other images too actually.

I'l move the images out of the folder to see if that works.

[eternal fog]

You need to be careful with that Windows has file path size limits, those filenames are way too long

[hardy cipher]

you know, it's doing something funky with my images too. seems like it's listing them as a prompt or something

like the entire list of cached images in load image

[wet nacelle]

That's wicked odd. I never would've thought about that.

[hardy cipher]

['0_Close shot of Steven Seagal eating a raw potato in_esrgan-v1-x2plus.png', 'snoop_dogg_photo_by_estevan_oriol_archive_photos_getty_455616412.jpg', 'Untitled-2 (1).png', 'Untitled-2.png', 'view.png', 'Zdar (1).png', 'Zdar (2).png', 'Zdar (3).png', 'Zdar (4).png', 'Zdar.png']

it was a lot longer than that, but you get the idea

[wooden pagoda]

So, no ideas how to stop multiple limbs in stupid places?

[hardy cipher]

well, I don't know what your setup is, so hard to say

I mean software

what you're using

[wooden pagoda]

Webui,.. SDXL 1... erm, what else do you need to know? If so, you might need to tell me how to find out

[wet nacelle]

@hardy cipherStay calm. I swear they mean no harm.

[hardy cipher]

and you're sure you're rendering the image in 1 megapixel?

[wooden pagoda]

If 1024 x 1024 is indeed 1Mp, yes

[hardy cipher]

well, what model? are you using anything else?

[wooden pagoda]

Ok, wait a sec, will find out,.. I've been using various models to try to see which sorts this mess out but so far none seem to

[hardy cipher]

because extra limbs and things happen regardless. but shouldn't be super common unless something funky is going on

[wooden pagoda]

Kinda annoying that Checkpoints and Loras show up in the pop-up menu when you only want a checkpoint file,.. makes keepoing track of ONLY the checkpoint files almost impossible.

[wet nacelle]

https://tenor.com/view/ghostface-scream-dbd-deadbydaylight-gif-22756066

[hardy cipher]

do you have them in the same folder?

[wooden pagoda]

I have saved the latest 4 or 5 Lora fiels into the Lora folder,. but there are Safetensor and checkpoint files showing up in the models folder, but no Lora files,.. DESPITE the fact that there are Lora files in there when I open the list in WebUI

[hardy cipher]

oh yeah, it's a mess. too many things going on

[wooden pagoda]

Why are Lora files and Checkpoint files suffixed with .safetensor ? That's kinda stupid.

[wooden pagoda]

Surely Lora files should be e.g. .lor or something. Sorry for calling you Shirley. 🙂

[hardy cipher]

well you see, it's all very new stuff. I do believe in time these sorts of protocols will be implemented

[wooden pagoda]

I'm stunned they haven't already.

[hardy cipher]

as for the extra limbs, hard to say. I'd say it just happens, but really doesn't all that much with sdxl. as you can see from this image it's not always perfect with everything. but not a bunch of extra limbs there

[wooden pagoda]

There are way too many Samplers to know what to do with, the checkpoint files and Lora files need to be easily differentiated and the list of SD Checkpoints must only ever show Checkpoint files.

ATM, it's a royal mess of epic proportions.

[hardy cipher]

well therein lies the other problem. there are lots of different things going on there

[wooden pagoda]

It's no wonder it's so complicated.

[hardy cipher]

you are using models based off of stable diffusion xl

but, these loras and things aren't theirs

[wooden pagoda]

I THINK so, yes

[hardy cipher]

nor is the ui you are using

and it's all open source stuff, so anyone can make whatever they want really

[wooden pagoda]

Oh,. eh? Pff, sorry,. it's v. late here and my brain's shutting down

[hardy cipher]

so you're downloading loras from the wonderful beautiful civitai. and they're named absolutely nonsensically

then you have different extensions expecting your loras and things to be in different locations, because that's just what they want to do I guess

[wooden pagoda]

THAT I noticed. Nor are they organised.

[hardy cipher]

at least you can finally type the names into the searchbar and op up the models

[slender coral]

Hey lads, anyone know of a node or a workflow that would allow me to generate prompts with string concatination / automate prompting?

[hardy cipher]

there are all sorts of add on nodes that can do that sort of thing. not sure precisely what you're looking for, but look in the manager list and you'll see a few node packs with that stuff

[wooden pagoda]

^^^ Clueless.

[hardy cipher]

the more I learn the more clueless I realize I am

[slender coral]

I basically want to read from a file something like large {TypeOfAnimal} sitting on a cloud

[wooden pagoda]

https://tenor.com/view/calculating-puzzled-math-confused-confused-look-gif-14677181

[slender coral]

Read from a file or a list of animals.

[hardy cipher]

https://tenor.com/view/the-universe-tim-and-eric-mind-blown-mind-blown-meme-mind-explosion-mind-explosion-meme-gif-18002878

[wooden pagoda]

That one's if you've learned so much you now know you know you don't know as much as you thought you knew.

[slender coral]

but look in the manager list and you'll see a few node packs with that stuff

Could I ask what this is?

[steel mist]

Do you have the comfyi UI manager custom node installed?

if so you'll have a side bar that looks like this

[vital ermine]

AIT the first two and without AIT on the second two (base then refiner) on BS4

[hardy cipher]

yes, exactly

[slender coral]

I don't

[steel mist]

Ah, well that's what they're referring to

[wooden pagoda]

Are you sure? 😄

[hardy cipher]

my bad, bruseph. you need that

[slender coral]

https://github.com/ltdrdata/ComfyUI-Manager
Correct?

[steel mist]

yup

[slender coral]

Thanks

[wet nacelle]

Comfy really is fun to use. I don't make workflows. I spend hours getting some workflows optomized for my Vram. I also add an upscaler each and every time too.

I love it.

[steel mist]

Get that installed and it makes getting all the other nodes and models 10x easier

[hardy cipher]

and the I'd suggest creating a quick node to string together random items from this list. https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/sd1_tokenizer/vocab.json

it really grinds my gears that this node of mine is no longer working. it was like 95 percent there. but now it's being spat out like trash before it even starts to run, lol

[slender coral]

Stupid question but I hate visual scripting / nodes, can I do python for these nodes?

[hardy cipher]

how do you mean?

if you want to create your own https://github.com/comfyanonymous/ComfyUI/blob/master/custom_nodes/example_node.py.example

[steel mist]

No, you don't need a gui

You can definitely write a python app that would do it if you know how pytorch works

[slender coral]

In terms of what I want it's literally:

for item in OpenFileAndReadLies():
  prompt("something {item} on a cloud")

[hardy cipher]

I'd say just find a node that does something similar and modify it to your liking

[steel mist]

I mean you could probably write a node that would allow you to execute arbitrary py in it

[slender coral]

Okay, can you reccomend a place to search for something like that?

Got the manager working, so I'll just google around. Unless you know of a node that could open a text file, read all lines and just concatinate and prompt each line.

[steel mist]

https://github.com/search?q=comfyui custom node&type=repositories github 😉

[slender coral]

Perfect, this is what I was after Thank you.

[hardy cipher]

yeah, I'm not sure off the top of my head

[steel mist]

But if you go into the manager the descriptions are pretty ok

[hardy cipher]

although they're listed in the manager

[steel mist]

and you can search, so "text" or "string" to start

[wet nacelle]

yeah.

[steel mist]

nightmare fuel

[glad grove]

looks like a new season of jersey shore on mtv

[wet nacelle]

Sure does

[slender coral]

Ah okay so it does have wildcards, but I would like to read from file, thanks lads. https://github.com/comfyanonymous/ComfyUI/discussions/862

[hardy cipher]

wow. perfection finally achieved

he's promoting good health

[glad grove]

sensei always trying to protect us 🙇‍♂️

[vital ermine]

Is there a batch count node for comfy?

[vital ermine]

I'll take the silence as a no

[hardy cipher]

in what way? seems like it keeps track internally

[vital ermine]

I need it as I had it in auto as clicking 160 times for 4 images at a time sucks

[hardy cipher]

you mean like batching batches?

[vital ermine]

batch size is a max of 4 so I need a batch count. Batch count of 20*4 it would loop over those 4 20 times

[hardy cipher]

ahh. it's 4? did not realize that

[vital ermine]

yep, AIT has a max of 4

[hardy cipher]

oh that. it slowed everything down when I used it. I'm sure I'd be told I set it up incorrectly, but whatever

[vital ermine]

plus AIT doesn't allow PMS, UniPC or DDIM it errors out but that is another issue

as I showed above 50% increase, where most people are saying the same but you gotta have a newer card for it to do anything.

my point is, even without AIT, I need a batch count so it can loop over long runs

not sure why the ksampler doesn't have that

rather the empty latent image node

[indigo carbon]

same, I asked Comfy why and he said he doesn't want features like A1111 had

[hardy cipher]

for the love of god

got prompt
Failed to validate prompt for output 3:
* AdaptiveCanny 2:
  - Required input is missing: images
Output will be ignored
invalid prompt: {'type': 'prompt_outputs_failed_validation', 'message': 'Prompt outputs failed validation', 'details': '', 'extra_info': {}}

nothing I do changes this, lol

wtf am I doing wrong all the sudden

it worked yesterday, worked 3 days ago. now no version works

[indigo carbon]

idk what to tell you. when I asked that here they said the solution is to have the folder with the outputs in the background for batch count. but no preview of the batch count like A1111 had.

[hardy cipher]

the old version I had 5 days ago that worked then doesn't work now. current version from yesterday that worked then doesn't work now either. nothing in between works

[vital ermine]

I just want a way to gen as many images I can unattended. The lack of that for such a stupid reason given I dunno what to say.

[indigo carbon]

ask Comfy, there is no node that can do that. it's a limitation of ComfyUI itself.

[wet nacelle]

Check it! https://open.spotify.com/track/0rr3y6RTzTxhR3UHSSYHX4?si=d3598dc6500e42db

wicked sound

[vital ermine]

ComfyUI frontend or the backend?

[glad grove]

i think fooocus has that you can try to use it

[hardy cipher]

really seems like it wouldn't be that hard to implement a node like that

[vital ermine]

depends if the backend can do it

if not then forget it

[hardy cipher]

yeah. I've noticed a couple things that have been a bit tricky to do

[vital ermine]

yep

[wet nacelle]

I can't eat this you know? I'm not some animal!

[indigo carbon]

the ComfyUI with the nodes can't. I already specified why.

[hardy cipher]

have you ever tried to set the widget slider to only hit odd numbers?

[vital ermine]

For this task I would go back to auto to get it done but his has some serious noise issues with XL

[glad grove]

looks like hes about to smash my head with that melon

[vital ermine]

for the AIT does the keep loaded really do anything?

I use the gpu flag so might be why it seems to do nothing

[indigo carbon]

no. it's meant for workflows that only use 1 checkpoint for sampling, with SDXL+refiner it should be off.

[vital ermine]

I hope this works

[indigo carbon]

it doesn't, it will just loop an X amount of times and will show you in the save image node only the last image

[vital ermine]

It worked perfectly

[hardy cipher]

https://tenor.com/view/oh-snap-ohsnap-gif-18271233

[vital ermine]

Showed me my BS4 then did the next pass

showed it and continues on showing all 4

Might be an AIT thing but as long as it works. I do miss DDIM for this

[glad grove]

this is awesome now ima steal it

[indigo carbon]

yes, but what about batch count higher than 1? if you batch count two with batch size 4, does it show you 8 images?

[glad grove]

Seagalvinci 🙏

[vital ermine]

Since I am limited to a BS4 that is what I am using so no idea. It shows the 4 then goes another round. If you check the box it does the batch count and will start all over doing another X batch count forever

[glad grove]

this from the Seagalvinci movie where he fights against king arthur

[vital ermine]

is that even using my lora or no?

[indigo carbon]

yeah, that's what I meant. I just thought it would be nice to not have to go to file explorer to view the entire batch count if you leave it generating unmonitored

[hardy cipher]

you got a lora, bro?

[vital ermine]

yes, I made it for the memes n here

[hardy cipher]

I did not know

[vital ermine]

https://civitai.com/models/135544

[vital ermine]

I am having mem issues though but it still works

Error occurred when executing KSampler:

Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated : 12.81 GiB
Requested : 40.50 MiB
Device limit : 23.99 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB

seems a pytorch issue

pics are fine though

[vital ermine]

even BS1 I got it so I restarted comfy

I know why I got that because I rolled my drivers back to 531.79 the last one before the mem stuff

[indigo carbon]

the 531.79 drivers are too aggressive with VRAM, I also got that issue, but changing to latest drivers fixed it

[spring fulcrum]

So when I am doing anything that involves skin I am getting this scratchy result... Can anyone suggest a way to reduce this effect... Here is a sample of what I am talking about... I think it has to do with the upscalers adding details to make skin look less smooth or perfect... something like that. I will also show some settings I think might be effecting it.

Any suggestions would be appreciated. @indigo carbon @soft zealot @high skiff

also be advised I cannot show the rest of that image

[vital ermine]

WHich has the OOM issue being a slowdown for ADA cards as it is aggressive going the other way thanks to gamers

As jps said as that is a known issue even Nvidia has acknowledged.

[vital ermine]

I figure they will get around it via some setting in the control panel, but I hope they figure out how to fix ADA cards being too slow for the power they wield. Even ADA 6000 pro cards are having AI slowness issues.

[indigo carbon]

that's just another case of NVIDIA screwing with us

[vital ermine]

I am a conspiracy person and after seeing Jensen I wouldn't put it pass him to have it a LHR type deal of some sort to protect his Hopper sales. Might not be it but something is happening as either malevolence or ignorance. They are not ignorant.

gamers none of the slow downs touch only AI stuff

[visual glade]

yeah the new nvidia drivers are a bit of a pain but hopefully my bf16 vae commits improve things a bit there

thankfully it's only on windows, linux doesn't have that problem

[soft zealot]

when peopel say slow downs on newer cards are we talking 3-10 seconds slower per generation or like 10-20- times slower ?

[glad grove]

looks like a mix between guy pearce and steven seagal 🤣

[hardy cipher]

Ginger Seagal

sort of got my canny node working. figured out why I'm an idiot and it wasn't working

[vital ermine]

All about the it/s or s/it. A 4090, doing precisely the same tasks, on the same machine using the same paramters, etc... should be 35-57% faster than a 3090 for the 4090. Well, it is a FAR FAR cry from that. Matter of a fact a lot of cases it is slower. People with ADA cards are more than a bit pissed.

Take for instance Kohya sd-scripts. all equal only I was slower than a T4 using windows. Switched to Ubuntu and about the same speed as a T4. I switch back to pre XL sd-scripts I am 3 times faster, BUT I am always slower than a 3090 being all equal otherwise.

It is a known issue for the AI tasks but at the hardware level all AI tasks should be 57% faster than a 3090.

[soft zealot]

oh dear, how sad, never mind ;o)

downgrade the drivers 🙂

[vital ermine]

Done

[soft zealot]

sother eis no issue then?
Move along
next 🙂

[high skiff]

@vital ermineyou got a 4090?

[vital ermine]

For a month now

[high skiff]

never thought you'd actually do it, you talked yourself in and out of it for over half a year lol

nice, welcome to 24GB VRAM gang

[vital ermine]

I almost didn't as the 7900xtx was about to be pressed to buy and I said no, I saw too much and read too much I grabbed the 4090

[high skiff]

I grabbed a used 3090. Much better deal, but I know you are against used components

[vital ermine]

Only from strangers, yeah

[high skiff]

Currently trying to solve some extremely annoying power issues on my GPU. I had it on my others too

[soft zealot]

a stranger is a friend you havent met yet

[vital ermine]

1500 miles away they will remain a stranger

[high skiff]

when idle not running AI but with AI loaded, my GPU never drops from boost clocks, so it just stays clocked up wasting power and making head for literally no reason

[pure crystal]

GPU power issues across cards, same motherboard/PSU?

[high skiff]

I was in a long distance relationship with a man for 5 years, and he was over 6k miles away, and we are still best friends lmao

[high skiff]

different mobo and PSU

everybody I have talked to has said they haven't had this issue

but I always have, and its a problem where I live

[vital ermine]

On Sept 6, I have to pay for this beast and I am not happy. More than I ever imagined so when it isn't even performing as it should for all ADA users sort of demoralizing. I figure it eventually will though.

MY GOD the heat from it when it trains

I need new case fans for it

[high skiff]

Thats why I went with a 3090. Same capability, not much lower perf, less power, less than half the cost

[vital ermine]

know what my system idles at with this 4090?

[high skiff]

I am super happy with my purchase, even if the first one was a scam

[high skiff]

whole system?

or just the 4090?

what model 4090?

[vital ermine]

whole system is 60-92w at idle. The card uses 10w more at idle than my 1060. That floored me

[high skiff]

my 3090 only draws like 25 watts, yeah. They are extremely good at being efficient when needed

you can severely power limit them and lose minimal perf too

I run my 3090 at 50% power limit most times, and I get anywhere from 70-90% my perf

I can gen at the same speed as my previous 3080 with half the power on this 3090

[vital ermine]

ADA cards changed all that, unfortuantely. They didn't like us doing that for some odd reason.

[high skiff]

I just need to figure out how to fix this damn power issue

with comfy open, my GPU idles at the same power it takes to run maxed out cyberpunk 2077

[vital ermine]

Surprising to me coming from Pascal that 3k and 4k cards limit the min fan to 30%. 0 or 30%+

I had that on my 1060

it was a setting

[high skiff]

it depends on the card model

my previous 3090 could off all of its fans, my new one can off 2/3 of its fans

one of them is always at 30%

[vital ermine]

mine is always off but it ramps up then fans come on falls back to 35c from 45c and off

[high skiff]

this power waste is horrible. I need it to stop

[vital ermine]

@high skiff My 1060 had a low power setting too but I never used it while this 4090 only has this. I'd check this.

[soft zealot]

turn the pc off when not generating, that will save power ;o)

[high skiff]

that has been removed for like a year now. It was changed in drivers

and mine is set to normal

[vital ermine]

My 1060 had 536.xx and was there for it

[high skiff]

weird, no idea then

I used to have it on my 3060ti, and then it was removed

[vital ermine]

Yeah, I ran to it to make sure that was not on by default and noticed it was gone. I was just in that setting a week earlier on the same driver too so I suspect gone for later models

I really wish they would allow fans to always be on at 20% as I used that on my 1060.

20% to 100%

[high skiff]

most fans can't go below 30% ever

I have never heard of a GPU being able to do below 30%

[vital ermine]

all of my fans can on my gpu and I had to replace them 3 times thanks to AI stuff

pascal, and I believe turing still have no minimum fan speed

that is enforced

I even had mine spin at 15% but its cooling sucked

[soft zealot]

Must be card specific

Im on the 537.13 drivers with my 1080ti and I still have that option

[high skiff]

I guess so, weird

[vital ermine]

See, Pascal still has it

Not sure why that is removed for Ampere, and Ada cards.

Dunno about turing

I know this I want to waterblock this investment

about 170 USD plus tax for one block though, sheesh

[soft zealot]

thats quite cheap IMHO

well sort of cheap

about what I would have expoecetded ball park

[vital ermine]

The blocks used to be around 95

I only have an AIO but I do want to go open loop

Now open loop/custom loop is really expensive now and even Jayz2cents was shocked

[vital ermine]

he has been water cooling for decades

[soft zealot]

thats more the range of prices I was expecting you to come back with

[glad grove]

sounds like a waste of money

[vital ermine]

not for me it isn't as I love quiet

In USA we don't have any stinking GST only a local sales tax

[glad grove]

still delta server fans outperform all of those

[soft zealot]

plus noisier 🙂

[vital ermine]

did you miss the part about quietness?

[sand pier]

Out of curiosity, does sdxl use any meaningful CPU or ram outside the GPU? I've forgotten to actually monitor things.

[high skiff]

it can use huge system RAM when loading models

[glad grove]

yes right now it eats all the ram

[soft zealot]

In the USA your shelf prices dont include the tax (whatever you decide to call it) rest of the civilised world its included automatically ;o)

[sand pier]

Alrighty ty

[high skiff]

its enough that my 32GB system RAM was choking

I have 64GB now, which is no problem

[sand pier]

Just considering changing my system a bit,and may go for 64 gigs for next build

[soft zealot]

yay join the club 🙂

[sand pier]

I use 5900x, 32 GB currently. Could be nice to swap to a 7800x3d/64gb

[high skiff]

that doesn't seem like much of an upgrade on the CPU

[tepid surge]

32 giggies

the llm's i could run

[vital ermine]

I consider all taxes to be theft so I will take not being civilized.

[high skiff]

giggle bytes

[tepid surge]

cx

[high skiff]

giggi biggies

[tepid surge]

giggi's big biggie

cx

[sand pier]

Yes and no, it removes some stuttery stuff. As well as lets me upgrade my server from a 4 core to a 12 core xD

[high skiff]

ah yeah, if you have another use for the 5900x, that makes sense then

[tepid surge]

||i'm gonna se if i can run stable lm, bye for now x.x <3||

[high skiff]

best of luck

I need to figure out why the hell my GPU is wasting power

it has something to do with AI

cause it does it for ooba, kohya, auto, comfy, and RVC

[sand pier]

Wasting power? Using more then expected while doing ai tasks?

[high skiff]

no, using a shit ton when doing nothing

literally no model loaded, completely idle, nothing happening, GPU stays pinned at boost clocks, wasting 130 watts of power an hour doing nothing

[sand pier]

Odd, which GPU?

[high skiff]

it happens on 3 different GPU's

3060ti FE, 3080 MIS gaming Z, and FTW 3 ultra 3090

I have looked all over google, and I can't find any info on why its happening

[vital ermine]

My psu keeps track of my power usage and reports to me (hooked up to my USB) and it says I have used 2kw in a bit over 800 hours so about 22c of power.

[sand pier]

I need to check how much my 4090 draws when I get home, I usually have a1111 and comfy running in the background

[high skiff]

god, i would kill for an electric rate that low lmao

maybe my mom and I wouldn't be getting fucked over for electricity

[vital ermine]

Our rate went up during the 1st pandemic as it was 10c per kwh

[high skiff]

still a fraction of ours

[vital ermine]

so about 22-23c

[high skiff]

jealousy for sure

[vital ermine]

elec is cheap the rest they really fuck us with

oh, about 8 weeks ago they changed all our strret lights to LED and they suck but on top of them are camera like things

[high skiff]

bro your rate is low, and you're saying you're being fucked 😅

[vital ermine]

yeah, because we even have to pay for street lights

[sand pier]

Luckily my power usage is still dumb low

[high skiff]

we use about 1300KWH/M in these searing hot months here in this new house

[vital ermine]

where I moved from that was the city here we all pay. garbage is 40 per month and they strike it goes up another 10

they constantly strike

[sand pier]

Summer here and we really don't go over 25c, and winter I have external heating

[vital ermine]

well, come this winter I have my 4090 to warm my room.

Not joking

[sand pier]

Lol technically same

[high skiff]

yeah, I have my 3090 to do the same, tho, my room will still be hot in the winter

[vital ermine]

This fucker gets so hot the glass side panel is almost too hot

[high skiff]

and our electric rate is fixed year round, so I save nothing in winter

[sand pier]

If I do a couple of hours AI art my temp goes up 2 degrees

[vital ermine]

fixed here as well

[high skiff]

bro, if I do a single hour, my room raises several degrees

oh, you mean C

[sand pier]

Yeah

[vital ermine]

I need new fans but first comes to pay this off and wait for my 140s to come back in stock as they are out everywhere

these 120s that came with my case I can blow harder than they can.

[sand pier]

At least fans can be really cheap with certain brands

[vital ermine]

45-49 for a box of 3

[high skiff]

I have 200mm fans, and they are dope

[vital ermine]

I need 4 :/.

[high skiff]

love my case, no cooling issues here

[vital ermine]

200mm, yeah

[high skiff]

200mm, 2000rpm. They move stupid amounts of air lol

I have no other fans on my case

[vital ermine]

140s will correct this as I look in the front and the 120 can't push the air that is being ejected on the side of my gpu. 140 it will.

[high skiff]

and it moves enough air to blow paper off the top

[sand pier]

Arctic is like 35 for a pack of 140mm

[high skiff]

I got them big ol honkin phans

[vital ermine]

these stupid ass fans is a max of 1200rpm (never makes it)

Kaku, that is what these are the pressure ones but I want the rgb models as they do 0-1900 and have more cfm/h0 than the non RGB 200-2k rpm version

[high skiff]

oh my bad, I lied lmao

I have no idea where I got 2k RPM

they are 900RPM

[soft zealot]

Yes but that via local council taxes or whatever you call it .

Unless you really do have a streetlight hooked up to your house and you have to pay fo ir on your electricity bill ;o)

[vital ermine]

costs about 5-6 bucks more but OOS everywhere. I hope they come back in Oct.

[vital ermine]

Nope, here it is on each utility bill and one company owned by the city. It covers all utilities, waste water, and refuse collection, admin fees min of 40 USD, and street lights.

[high skiff]

jesus, these fans are 98CFM EACH

movin air

[soft zealot]

so you pay for domestic electric to the city ???

No choice of supplier?

[vital ermine]

No choice it is owned by the city and no net metering allowed if you even had solar.

[soft zealot]

sucks ;o)

[sand pier]

I use 2 200mm Lian li fans front, 3 120mm ml120s top and 1 ml140 back

[high skiff]

we have 2 provider choices here, both are expensive out the asssss

[high skiff]

Lian Li Lan Cool 215?

[vital ermine]

The best electric I ever had was a co-op

[sand pier]

216 i think it is

[high skiff]

our current provider is $0.43/kwh all year

And the alternative is:

Summer: off ($0.36) on ($0.61)
Winter: off ($0.31) on ($0.53)

[sand pier]

Ouchy that's expensive

[high skiff]

oh no shit, I had no idea they made a 216

[high skiff]

our $500 eletric bills agree

I have the 215

love this case

[soft zealot]

@vital ermine here in the UK for Electric & Gas we have a completly free choice of supplier .

Water & Sewerage is which ever company covers your area

Things like refuse collections, street lighting , policing costs, fire service coverage etc gat paid to your local council based on the value of your home (roughly)

[vital ermine]

police and fire is from the property taxes

[high skiff]

its a handsome little case

[sand pier]

I limit effectively out on 0.15 bucks, mostly thanks to government support to power

[high skiff]

pft, must be nice

[sand pier]

Currently it's like 0.015 though

[vital ermine]

My dream is to live out in the wilderness with solar and LEOS for my inet.

[soft zealot]

@vital ermine Current energy Tariffs for me are

[high skiff]

if mom and I can figure out how to get some form of second hand solar system, we can jump to their extreme on off plan, which is:

off ($0.21) on ($0.79)

[sand pier]

They made some improvements on the 216, I forgot which

[vital ermine]

DAMN, that is cheap on gas and hideous on elec

[sand pier]

It's such a dumb cheap case for what it is lol

[vital ermine]

@high skiff Are you on something connected to PG?

[soft zealot]

I really need to hook my aquarium up to the gas lol

[high skiff]

PG?

[sand pier]

I paid like 95 bucks plus tax

[high skiff]

I got mine for $60!

[vital ermine]

P&G the elec cause they suck ass and over charge and infrastructure is shit

[high skiff]

I love this case lmao

[sand pier]

That's a killer deal

[high skiff]

oh no, we have their way worse older brother, SCE

[vital ermine]

Oh shit

[high skiff]

electric infrastructure nearing 130 years old

Almost every year they have to file for bankruptcy lmao

just a few years back they got his with a $12B lawsuit that put them under like 3x over

turns out when your power systems are linked to over 1M acres of wild fire that happened in a single fire season, people get mad

There is some good however

They do have a really good rebates for solar energy, and we also have local metering.

And we also have net metering as well, and it has 12 month rollover too which is dope If you have solar

[soft zealot]

although it should be noted we use more Gas than Electric

[vital ermine]

How? I thought gas lighting went out with the Victorian era in England?

[high skiff]

See, our electric use last month was 1289 KWH

And I think our electric was like 160 KWH

[soft zealot]

heating

[vital ermine]

heating? In July too?

[high skiff]

Some parts of the world are in their cold months right now lol

[vital ermine]

Not the UK

[soft zealot]

well no although it is used for hotwater & cooking

Heating is the big user come the cooler months

[vital ermine]

Aye

Now the big push, globally, is to remove gas

I like gas it cooks, heats, etc... so damn well

[soft zealot]

yup for cooking its a million times nicer than electric to use

[crisp owl]

I converted my oven/stove to gas two years ago. Electric sucks

[high skiff]

Gas stoves are actually monumentally less efficient and more dangerous for your health compared to radiation stoves or induction stoves

[vital ermine]

yeah. Electric has hot spots and loves to burn

[soft zealot]

thats as may be but as a UI they are much nicer 🙂

[high skiff]

Oh, that's first generation electric, that's like 30 plus years old

Ceramic radiation cooktops boil water faster with less energy than gas stoves do now, and induction cooktops are even faster and more efficient

[crisp owl]

and I use cast iron/carbon steel strictly

[vital ermine]

Tell that to a chef.

Oven not burners

[high skiff]

In that case, it depends on which model you get

If you have a double-decker of an option, a double decker electric oven heats faster than a standard gas oven does

I'm hoping to get my mother and I to switch over to a double decker electric stove. Gas stoves are not good for your health, especially for people with breathing issues

[crisp owl]

hood vents are installed and used for a reason

[high skiff]

I'm referencing like an actual scientific process of comparison, not just some form of anecdotal evidence.

There have been various tests conducted against gas and electric stoves, and on average, electric stoves can boil water considerably faster, while wasting monumentally less energy

[high skiff]

Not all locations have hood vents, and they are only effective for back burners. That has also been scientifically tested as well. And additionally to that, gas stoves decrease air quality even when not in use

[soft zealot]

yes but gas is so much nicer to use ;o)

[high skiff]

But why do you feel that way?

If it has to do with thermal inertia, then that's definitely a personal preference

[crisp owl]

old gas stoves that require a pilot to be lit 24/7 perhaps

[soft zealot]

oooh that is old

[high skiff]

But generally speaking, radiative cooktops do not burn you when you reach over to stir boiling water or dishes, radiative stoves do not introduce any form of toxic emissions into the air, radiative stoves use less energy by comparison to kilowatt output of gas, which can be cheaper in some regions

[vital ermine]

Mine it clicks

[high skiff]

No, this was on modern stoves. Even with the gas lines off, they still leak

I'm not saying gas is bad, I have a gas stove still, but I'm saying that a lot of the reasons that people think electric are bad have actually been disproven

[soft zealot]

I dont disagree with any of those points but cooking on a gas hob is just more satisfying 🙂

[high skiff]

For whatever reason a lot of people say that electric stoves take way longer to heat up water, when by comparison, on a similarly rated burner, they can be up to 45% faster to reach temperature

[glad grove]

still prefer normal electric stoves and not induction cooktops because of the EMF radiation they emit,they say the radiation is minimal but i prefer to be safe than sorry

[crisp owl]

I guess everything has downsides. Electric still requires power to be produced somewhere

[high skiff]

The three biggest benefits of radiative electric cooktops is one, extremely easy to clean, two, zero gas emissions or toxic VOCs, and three, no burning yourself when you're cooking by stirring things

[soft zealot]

well they do if you live ina country that only has 120v by default instead of 240v lol (and yes I'm fully aware of the fact you have split phases and that 240v is availble to you for things such as air conn etc)

[crisp owl]

How else am I supposed to keep my lower arm hairs shorter if I don't singe them over the flames when cooking?

[high skiff]

This was also tested in the United States, on 120 volt electric.

A fully rated electric radiative cooktop needs to be plugged into a 30 amp circuit, as some of them can draw in excess of 8000 Watts when used concurrent

[soft zealot]

Man use fire
Grunt

[glad grove]

u can burn your arm instead by touching the pot on top of the induction cooktop

[high skiff]

I never realized before just how severe the CO2 emissions are off of gas stoves, and I will definitely be using our hood vent as much as possible, though they don't really help much

[soft zealot]

dont induction hobs need specific materials to be used with them?

Can you use cast iron with them?

[high skiff]

That's induction, I'm not talking about induction, induction is extremely expensive and very finicky

I'm talking about radiative

Where it's a glowing red wire underneath the ceramic cook surface, which absorbs the heat reflected from the wire into a ceramic surface, then transfers that directly into your cookware

[soft zealot]

can you sat fre to your alchol based sauces on demand using an ekectric hob?

[high skiff]

Induction is better for some things, but it does require specific pans, and you can't use metal utensils around them

[soft zealot]

god thats ancient shit

oh sorry you dont mean ....lol sorry

[high skiff]

Yeah, but it's still considered the most efficient and safe type of stove. They have been updated since then

My grandmother had one in her house for probably two plus decades

[crisp owl]

we had a nicer glasstop electric stove. Hated it to all hell

[high skiff]

But why though?

Like I'm genuinely curious to know why

[crisp owl]

sucked at cooking

I love cooking

[high skiff]

In what way?

[crisp owl]

Heat control

[high skiff]

Like these are just statements, that have been scientifically disproven in terms of all of the things that I mentioned

Okay yes, the one real negative of a ceramic cooktop is thermal inertia

[soft zealot]

drop nything heacy on them and they can shatter

[crisp owl]

heat control is the main thing when cooking lol

[high skiff]

That is something that you have to learn how to use properly, and can be very easy for some people, or very difficult for others

[glad grove]

yea those are the best the only reason ppl here dont use them is because electricity is too expensive ( mafia)

[hardy cipher]

heat distribution debates

[high skiff]

Yeah, totally valid, an electric cooktop would be outrageously expensive where I live

[hardy cipher]

very hot topic

[high skiff]

Though, after seeing the actual numbers as to how severely a gas cooktop can degrade air quality with just one burner, I don't know man lol

I don't remember the exact numbers, so don't quote me, but I think I remember it said something like 400 parts per million of CO2 emitted from a single 2400 BTU burner over the course of 30 minutes in a 600 ft² room or something like that

Which means within 30 minutes, that single burner can bring the air quality in a room above 1,000 PPM, which is considered unsafe

[soft zealot]

Oh yeah I had forgotten this about induction hobs

[high skiff]

And while a hood does help somewhat for the back burners, it's effectively useless for the front burners, which are the burners that most people use, myself included

[high skiff]

Yeah, I have absolutely no arguments on the issues of induction cooktop, they are messy

[soft zealot]

open a window and the back door whilst cooking

[high skiff]

I say radiative way over induction

[high skiff]

That's not an option for everybody lol

[soft zealot]

although they are the most efficient so.......................

[tepid surge]

i need a 4th monitor just for discord

[crisp owl]

I have an open stove on my island, hood comes down over it from the ceiling.
I have a CO detector installed damn near overhead it, and have never once seen the numbers raise remotely

[high skiff]

Yeah, induction is incredibly efficient, then radiative is also decently efficient, then you have gas, which is the least efficient because of the heated gas convection, which causes most of the heat from the flame to curve around the surface of the cookware, and dissipate outwards

[soft zealot]

or live ina draghty old house that was built in 1824 (ours scleberates irs 200th birthday next year) lol

[tepid surge]

that's cause you arent huffing the stove gases! cx

[soft zealot]

but gas is more controllable and so much more satisfying visually

[high skiff]

I was referring to the fact that it's way too hot where I live to open a window like that lol

[tepid surge]

https://tenor.com/view/hincha-de-nets-jmveintisiete-netslatam-brooklyn-nets-gif-23476908

[high skiff]

Totally understandable, can't dispute that

My only statements that I'm making are in effort to disprove false statements, and a matter of preference is not a false statement

But when people say that they are slow to heat, that's actually been disproven many times, that equally power rated electric stoves are considerably faster than gas versions

[glad grove]

he doesnt like gas stoves either

[high skiff]

Lol

I really do wish I lived in a climate where we could just open a window and blow in some air and not have to deal with the toxic emissions from natural gas combustion

[soft zealot]

however , radiant heaters also take longer to coold down when turned off. Turna flame off and the heat has gone 🙂

[high skiff]

Yeah, the heat from the flame is gone, but not the cast iron grates that the pot still touching

The craziest thing that always blows my mind when people bring up that argument is the fact that people forget you can just lift the pot, and which the heat from the electric stove instantly stops lol

Like I said, thermal inertia is a problem for some people, and some people can adapt to it easily

[soft zealot]

thats manual labour

[high skiff]

[tepid surge]

seagal my beloved 🥹

iu love the little circle in the corner cx

[high skiff]

https://tenor.com/view/screaming-seagull-meme-gif-18539479

[glad grove]

he do be fire tho

[tepid surge]

even the art in the background is cool

[glad grove]

yes he lives in texas

[tepid surge]

i love Texas. Texas, Cali, and New York = USA

[high skiff]

I need to bother my electrical engineer friend to help me source the parts for a reclaimed self-made solar array

[crisp owl]

I went to new york for a work trip once.
Never again will I visit the city there again.

[high skiff]

I'm going to San Francisco in October, both excited and scared for it lol

[tepid surge]

power inveters/conveters (no idea what it's called) is where my brain starts to go whaa with electrical stuff

[high skiff]

He has a killer off grid system right now, and he built it for a couple thousand dollars, so I'm going to ask him if he can give me a little more info about it

[tepid surge]

xD cx i'd prob be culture shocked too and i live in LA

[high skiff]

All we really need is about a 5 kilowatt solar panel array, and we should be pretty comfortable with our electric usage

[tepid surge]

https://tenor.com/view/star-trek-star-trek-tos-scotty-electrocute-electricity-gif-16908385

[high skiff]

As for my trip to San Francisco, I'm going on business, so I really have no idea if I'll actually be able to see the city, or what things are going to look like

All I know is that it's a fucking scary city, super super dangerous

[hardy cipher]

lol. that's so crazy

[high skiff]

What's crazy?

[hardy cipher]

how ridiculous san francisco became

[crisp owl]

I grew up in OC, lived in Inland Empire for several years, now live in Utah

[high skiff]

It's genuinely scary, like I'm nervous and excited going

[glad grove]

not just san francisco its almost everywhere ( except alaska or hawaii)

[rigid bobcat]

Hey

Would someone help me out a little bit? SDXL worked perfectly fine, however all of the sudden, it's not even loading up!

[high skiff]

Hawaii has severe crime rates too in the populated areas

But San Francisco is on another level

[glad grove]

😔

[crisp owl]

Personally heading to bed, but give people details on what you did.
Updates, restarts, etc. The more details the merrier

[glad grove]

new york too,u go to the subway and u either get mugged or theres a guy yelling hes jesus

[high skiff]

LA and San Francisco have a similar crime rate, however LA crime is a lot more gang related, whereas San Francisco crime is random victims

Never mind, I just googled it

[rigid bobcat]

Thanks 🙂
I have the mountains of cmd lines it gives me but I'm not even sure where I could post that

[high skiff]

San Francisco's crime rate is like two and a half times higher than LA LMFAO

You could screenshot it and we could take a look

[rigid bobcat]

Oh I have pic perms?

Cool!

[high skiff]

Yeah, everybody should

[rigid bobcat]

Thank you, hope I won't be bothering the ongoing conversation

[high skiff]

No worries, this is an SDXL focus chat, so SDXL issues take priority

We just get into long off topic conversations when things get slow lol

[hardy cipher]

I really don't mind helping anyone as long as they are attempting to be reasonably competent. I've went through all the BS, I know how frustrating it can be

[tepid surge]

ah oc, and first for me hearing Inland Empire but i see

and i keep hearing Utah, what pulled you there?

[hardy cipher]

this new canny node is quite fast, guys

[crisp owl]

Less people, beautiful mountains

[high skiff]

Same

I'm always willing to help somebody who's willing to help themselves

[hardy cipher]

only thing that bothers me is the occasional person that's demanding and acts like it's our fault they can't get it to work. I do wonder about those folks

[tepid surge]

Now i gotta Arazona v Utah cx

[rigid bobcat]

Well me and coding are worlds apart, I had to google some ridiculously basic stuff so far lol, hope that won't be a big annoyance

[crisp owl]

Arizona is waaaaay to hot for me lol
I'd move to Alaska if my wife allowed it 😂😂😂

[rigid bobcat]

But I get this, for like 3 more pages

[crisp owl]

Anyways, bed for me ✌️

[rigid bobcat]

See you later 🙂

[tepid surge]

i was thinking it'd prob be cooler up there cx rest well being, thx for the info✨

[hardy cipher]

ooh, some a1111 fun

[crisp owl]

Last comment. Is DarkSushi an sdxl model? You using them together? If so, non sdxl stuff will conflict with sdxl

I'm out ✌️

[rigid bobcat]

DarkSushi should be seperate

It just switches back to whatever I used last

[tepid surge]

brbs gonna food, beer and AC6 watch

[rigid bobcat]

I guess the last thing I did was maybe update controlnet?

[hardy cipher]

it worked before?

[rigid bobcat]

For making videos

[hardy cipher]

then went to that?

[rigid bobcat]

Yeah perfectly

It was legit fast too and everything

[hardy cipher]

tensor size mismatches

[rigid bobcat]

Is this useful info?

[hardy cipher]

so something funky going on

[glad grove]

about to drop some 🔥

[vital ermine]

I have 31 images, 31 captions then what is it caching latents of/for 93?

[high skiff]

Are you using flip augmentation?

[hardy cipher]

this is like the wish.com version of the original image

[vital ermine]

oh, thank you

that would be it

[high skiff]

Yeah, flip augmentation will introduce more latents

[vital ermine]

yep

[high skiff]

No worries, been there before lol

[vital ermine]

I always wondered

[high skiff]

I only thought of it because it was a multiple

[vital ermine]

how is my captions a latent

?

[rigid bobcat]

Do you think this is a memory issue? @hardy cipher

[hardy cipher]

when something like that happens I'd generally just try to backtrack or reinstall whatever it is you might have altered recently. someone might be able to pinpoint what's causing all those tensor size errors. but it's not something I can really figure out by looking at it

[high skiff]

Now that I have no idea

[vital ermine]

If I don't use captions latents will be less

[hardy cipher]

I don't know. wouldn't be my first guess. but I guess it's possible.

[high skiff]

Are you using cache latents to disk?

[vital ermine]

sometimes

though when I do only the images are .ngz

[high skiff]

If you are, then it's only caching the new images/captions

. Ngz is the cached latent

[vital ermine]

yes

[rigid bobcat]

I think the last thing I got was something called Temporalnet

[hardy cipher]

look at line 2041 in modules.py

start seeing if you can trace it back

[vital ermine]

which is why I wonder how captions raises the latent count kohya reports?

[hardy cipher]

all I can suggest. or try and reinstall

[rigid bobcat]

Oh, okay!

[hardy cipher]

whatever it is you installed. or uninstall

[sand pier]

Did the ait controlnet bug get resolved?

Caused some headaches for me last night lol

[rigid bobcat]

sry how do I check it?

I see the file

I think I found it, edit with IDLE right?

[hardy cipher]

find modules.py. you might want to get something like notepad++ if you don't have anything to look at python code with.

I don't know if you want to edit it exactly

but that's where the error is coming from

whoa

[sand pier]

If you don't know how to open python code, the file will probably not tell you anything

[rigid bobcat]

You are correct

[pure crystal]

alien nation

[rigid bobcat]

What might this mean?

[uncut fiber]

i need bit of help with some workflow i am playing with. Is it sometimes must to have two samplers in one branch?
To me it seems weird, but not having it this way, it rotate whole image...

[hardy cipher]

some kind of formatting, dictionary error. I don't know, man. uninstall whatever you've installed recently. something isn't doign what it's supposed to do

[sand pier]

You have to find what gets put in the error message variable

[hardy cipher]

yeah, that should be in the code if you scroll up

[sand pier]

But seriously don't waste your time debugging if you don't know the base of coding

[rigid bobcat]

I shall push forward

[sand pier]

There are so much stuff that can just break things in comfy and auto1111

[hardy cipher]

if you want to really want to figure it out start pasting snippets for bard to analyze

huge help, that bard

bard will make up libraries for you to use

they're not real, but you'll think they are. for a while

[sand pier]

Chatgpt can be very confidently wrong lol, but so impressive regardless

[hardy cipher]

they're helpful psychopathic liars. but they mean no harm

[hardy cipher]

count doggula

[rigid bobcat]

Hmm, thinking about a full-on reinstall

[vital ermine]

hahaha

[hardy cipher]

I tried to make images of him with willy nelson way back. ended up getting snoop willy dogg nelson

[vital ermine]

lol

[uncut fiber]

guess who 🙂

[glad grove]

meatloaf

[vital ermine]

that is who I was thinking

but no

[uncut fiber]

The Cure!

[hardy cipher]

are those pictures from the cure's latest tour?

[uncut fiber]

seems like 😄

[hardy cipher]

you should redo them and have guy fieri be their lead singer

and then it will be emo smashmouth

I'm telling you guys, this new canny node is kind of a banger, lol. under .1 seconds to run.

[vital ermine]

I don't do any CN since XL

[hardy cipher]

well the other one took a few seconds on my slow gpu, and it didn't give me the results I wanted. so I made my own. and I had no idea what I was doing so it took a little longer than it should have

[uncut steeple]

[vital ermine]

I have never even looked under the hood

[sand pier]

Controlnet still so nice with XL

[hardy cipher]

yes, now that it's not so resource intensive

[uncut fiber]

can anybody help me with my testing wf to rotate latent space? I think to have 2 inline same samplers is weird.

[hardy cipher]

this canny node though, like 50x faster than the other one for me. kind of a significant difference

[hardy cipher]

what's happening?

[uncut fiber]

in that cure image should be workflow. I bet i am horribly wrong, must be much easier. But when easier, it rotate image, not latent space

[hardy cipher]

why are you rotating it?

[uncut fiber]

to reuse seed 😄 Just trying to learn

[rigid bobcat]

@hardy cipher Hey, I'm thinking that I should just reinstall SD and get everything I had before the error. Is that a good start?

[tepid surge]

what would be the goal of rotating a latent/latent space?

[uncut fiber]

4 images from same seed.

saying just trying to learn

[hardy cipher]

yeah, just save the stuff you want to keep, get rid of the rest. or if you're installing in venv you can just get rid of the virtual environment maybe. but sometimes that won't even do it

[rigid bobcat]

By "save the stuff you want to keep" do you mean literally put them into a separate folder?

[hardy cipher]

tbh I haven't messed with rotating latents that much. looks like you're doing it right. just not sure what the result is in the end

[rigid bobcat]

Because that's the first thing that comes to my mind lol

[hardy cipher]

or delete around it. or just install over what's there

[uncut fiber]

both cure images are just rotated latent space

[hardy cipher]

but with a1111 I'd just nuke the whole install

[rigid bobcat]

Hmm okay, I highkey forgot how to install it though do I have to go through all those hours of looking up what codes to run?

ngl that was painful

[hardy cipher]

just look on github. it's really not that much. you just have to understand what's important

[rigid bobcat]

oh just this? https://github.com/AUTOMATIC1111/stable-diffusion-webui

[hardy cipher]

or, if you want to get even more complicated, come join us on the noodle train

yes

[rigid bobcat]

Thanks 😊 you can see I really have no clue what I'm doing LOL!

[uncut steeple]

Why use noodles if you can have straight lines

[hardy cipher]

please refer to my recent screenshot to see my preferred noodle shape

[tepid surge]

Latenspace encoding:
"The first neural network (the vae encoder) maps the input variable to a latent space that corresponds to the parameters of a variational distribution. In this way, the encoder can produce multiple different samples that all come from the same distribution. The decoder has the opposite function, which is to map from the latent space to the input space, in order to produce or generate data points. Both networks are typically trained together with the usage of the reparameterization trick, although the variance of the noise model can be learned separately..."

[uncut steeple]

Uh oh a direct line guy

I think straight looks cleaner

[hardy cipher]

straight looks cleaner until you try to use it. then it becomes a mess. at least for me

[tepid surge]

so a latent is just a representation of an image in the form of whatever a latent is, which my assumption is just a 1d list of all the pixels ie may be diff from the 8/16 color bit depth info images have, like it being remapped to 0 to 1 x y and z (rather then r g and b though the "same" in sense of it being mapped to it) or whatever the training worked with, which is why the input would need to be translated to that form

[hardy cipher]

well straight like you're driving on roads. I prefer straight like the crow flies

also helps to really line things up and get OCD with it all

[uncut fiber]

yes yes @tepid surge somehow i have to learn how modular things are working, i picked this. Not much use, but i am learning, now bit optimalized it

[tepid surge]

dont you yes yes me cx this started cause of your inqury cx and this be info

and saaame

[hardy cipher]

whoa

hottie

[uncut fiber]

i only inquire what is wrong on my side, seems too complicated 2 samplers for each branch, which i solved now. I haven't asked about if is it useful, or what is it. But thank you for info.
It is nice to play with it 🙂

[ionic gulch]

what is the latest status regarding the broken ip adapter nodes you were talking to comfy last night? could you fix it? seems to be still broken for me after installing all updates.

[hardy cipher]

hmm, I don't think that was me? I was just asking him about how to set the widgets to odd numbers only, and I don't think he responded, lol. but I worked around it. I should submit these to that manager soon

[ionic gulch]

so ip adapter works for you or don't you use it?

[hardy cipher]

should I use it?

this is something different than the ait stuff? because that actually slowed me down for whatever reason

[ionic gulch]

ip adapter is image promt in addition or instead of text promt. but if you don't use it don't worry. you just were in there with some posts in that discussion from last night, but i've checked again and your topic was something different.

will just go to work now and check again in 8 hours - maybe it's already fixed by then.

[hardy cipher]

ooh

[uncut fiber]

also rotating latent space it makes 2 different layouts if non square, portrait and landscape

[hardy cipher]

the clip vision thing

okay, so what's the advantage of using ip adaptor?

[tepid surge]

So you have the audacity to assume my ambient speaking is specifically about the constructs you're working through and then have a lil shindig about it? cx what a head spin cx

[hardy cipher]

over just using the clip vision?

[uncut fiber]

@tepid surge peace, my english is bad, i can easily misunderstand 🙂

[hardy cipher]

I was talking the unclip models up too, and people weren't listening

[tepid surge]

https://tenor.com/view/gun-anime-anime-gun-sakuga-sakuga-gif-23104054

careful there you might get me to undo a covering construct people like

https://tenor.com/view/deal-with-it-sunglasses-cool-happy-vtuber-gif-21940823

i am a they of science

cx

[soft zealot]

Pink Floyd: Set the Controls for the heart of the sun

[hasty mortar]

Guys

Any suggested model for cyberpunk cityscapes on SDXL so far?

https://civitai.com/models/131703/brixl-or-a-must-in-your-toolbox

I found only this one

[uncut fiber]

@hasty mortar not cherry picked. Something like this is o.k. ? Then NightvisionXL

[hasty mortar]

Looks cool

[hasty mortar]

May be good paired with a detailing lora

[uncut fiber]

yes this is just raw only

[hasty mortar]

Was trying to dig into a cityscape comfyui workflow for that it looked promising but it's just too advanced for me

I can share if you use comfyui

Thank you 🙂

[uncut fiber]

i think those i posted are only 768x768

so 1024x1024 is proper for XL and could be more detailed.

you can share, but i will look lately

[vital ermine]

See, he can smile

[glad grove]

steven seagal movie directed by tim burton

[soft zealot]

this was generated using Base SDXL & Refiner.
Don't write Base off and don't underestimate the power of the prompt 🙂

[hasty mortar]

Absolutely

It recognizes all my favourite artists

😉

[uncut gull]

Has anyone seen an SDXL Refiner model fine-tune on Civitai? Are there any out there? (There are tons of SDXL Base fine-tunes, but the Refiner would be harder to train.)

[soft zealot]

but would it really ?

Does calling it "The Refiner" add an air of mystique it doesn't deserve?

Isn't it just a.n.other model thats used in a second ksampler after the first pass?

[peak dove]

I'm doing excellent art but without the Refiner. Should I consider seeing a shrink?! 🥳

[idle zenith]

Is there any sdxl model than can deliver midjourny 5 like output?

Sdxl is great but it still looks like halfway through mj's

[soft zealot]

If you want MJ, pay for MJ 🙂

[peak dove]

MJ has a polished finish. It seems to use sharpening and glow, and some hsl in finishing each picture. A lot of its color output has a natural brio which in other ai generators would need to use Photoshop.

[ionic dragon]

whom does she look like?

[peak dove]

Try free bluewillow

[soft zealot]

oh she looks like , you know, errm , whats her name, the girl from all the Stable Diffusion images ;o)

[peak dove]

Daisy Diffusion

[soft zealot]

thats the one

[peak dove]

[soft zealot]

she shoul dbe SAIs official Mascot

[thorny frost]

Daisy is the girl with six fingers?

[peak dove]

That's Daisy Dozen-Fingers and not Daisy Diffusion 🤠

[ionic dragon]

doesnt she look like Selena Gomez?

[soft zealot]

who?

[uncut gull]

The refiner has twice as many parameters. I promise, it takes much longer and training it needs max-spec consumer hardware or renting a lambda clouds instance or the like.
(Actually, now I'm wondering if I'm remembering this wrong. Hang on.)

[soft zealot]

but do you really need to finetune it ?

Majority of fine tuned models seem to be designed not to use the refiner and seem to be setup for single sampler pass useage.

Personally use one model in the traditional base ksampler & eithe the same or different model (to chnage th effect) in the trafitional refiner ksampler

[ionic dragon]

Selena gomez 😅

[uncut gull]

Okay, after checking, I'm still not sure where I heard the refiner is bigger than the base.
To clarify: I am not asking about philosophy or best practices.

Does anyone know of a fine-tuned version of the Refiner posted on Civitai? I am looking for a link.
Using the Refiner does improve base model images, but tends to distort faces when used on a fine-tuned model.

[soft zealot]

Tennis player?

[soft zealot]

wouldnt a refiner for a fine tuned model need to be based on the fine tuned model?

ie wouldnt every FT moidel need its own matching FT refiner?

(or you could just do what I did lol)

[ionic dragon]

Singer

[soft zealot]

she's a sewing machine???

[uncut steeple]

My new workflow with ControlNet Support

[indigo carbon]

hey guys, I think SDXL is better without negative prompt..? (left is no negative, right is ordinary negative)

[strong field]

i use little to no negative with sdxl

[soft zealot]

yup that was sort of mentioned back at the eroginal launch that many people were over egging it

and that all of the magic feel good words werent needed

YMMV

This is Stable Diffusion, This is the Way

[thorny frost]

SDxl doesn't need refiner nor negative nor anything.

Only vram

[soft zealot]

but it makes me feel better ;o)

[soft zealot]

dont take away my comfort blanket please

[sand pier]

Why,us,this, the,case,let,me,just,add,a,word,for,every, possible, outcome,that,could,impact,my,image, without,realizing,the,actual, negative,impacts,it,may, have

[indigo carbon]

right? am I trippin' or is the left one better?

[soft zealot]

I 'm almost at the stage where a Negative is a Signature

[soft zealot]

there is no better, they ar eequally good.
Aethetically though my preference is the RH one

[wet nacelle]

Are we collectively joking right now? Because I can show examples of why a refiner should be needed.

[thorny frost]

Go ahead!

[strong field]

I prefer the left, winston is wrong, lets start a comment war, go

[soft zealot]

"should be needed" or "is needed"

The former implies its needed from a Process POV rather than from a quality POV

[soft zealot]

I like the middle one

[indigo carbon]

idk, if "eequally" is the term, the "aethetics" are same chrematistics; and yet is somewhat clearer IMO.

[soft zealot]

you dont have 59 year old eyesight

[wet nacelle]

Keep in mind that the base is insomnia and the refiner is launch SDXL 1.0 Refiner.

Also keep in mind that detail may be harder to see through discord due to compression.

Images are from base to refiner.

![ROACHES]!![ROACHES]!![ROACHES]!![ROACHES]!![ROACHES]!![ROACHES]!![ROACHES]!![ROACHES]!

[indigo carbon]

yeah, no one have tuned the refiner yet; idk why, might be because it isn't intended to do anything besides increase detail and remove anomalies

[wet nacelle]

That's really the only reson that I use it.

[indigo carbon]

same. it's best to get the latent right from the first K_sampler, then upscale- much more efficient and is the way the refiner is intended to be used

[wet nacelle]

I agree here too.

[wet nacelle]

I do think that the launch refiner still has some interesting aspects to look at. Doing a latent upscale can achieve those results but not always in the way that the refiner can.

[thorny frost]

Yep, the refined ones are more detailed. But I guees you can get more detail using any other model. The cockroaches made me sick btw

I have just lunch

[wet nacelle]

fixed

[strong field]

insects behind a spoiler for the squeemish, not pointing fingers...

[wet nacelle]

double fixed

[indigo carbon]

it's less efficient to do any upscaling before there is another K_sampler. SDXL refiner is way more than precise enough to not require something like hiresfix instead, upscaling should go after all the sampling is over

[strong field]

i havent tried this way, and never bothered to have refiner in my pipeline lol

[indigo carbon]

that's how the AIT workflow I made is built, did you tweak it to only have 1 stage or something?

[wet nacelle]

I used to do this. I stopped after workflows became too complicated for me to edit.

[strong field]

wait i just reread and realize i meant I havent tried what Mack recommened

[indigo carbon]

the efficient AIT workflow I made works that way, and it's very coherent

[strong field]

speaking of complicated workflows

that monstrosity is my daily driver

[soft zealot]

I use 4 Samplers in my flow

Sampler 0 is a Preconditioner that does 3 steps using Model2/Refiner

Sampler 1 is the trad Base Step using Model1

Sampler 2 is the trad Refiner Step using Model 2

Sampler 3 Is part of my Upscale and uses Model1

So on a typical (for me now ) 35 step run

03 steps on Sampler 0
19 Steps on Sampler 1
13 Steps on Sampler 2
Plus
5 Steps on Sampler 3(US Optional)

[strong field]

thanks for explaining what your stages meant, was wondering

[soft zealot]

I know whatt heyre based on however I prefer to think of it as a Mutli Sampler Based Workflow rather than being rigidly fixed to base & refiner etc

as I could use the same model for all sample steps

or I could mix them

[indigo carbon]

you shouldn't use the refiner at beginning of run, it's meant to refine, not create the shapes within the image. also ideally; you could add the 5 final steps to sampler 2, saves time and doesn't change stuff whatsoever. can't see your CLIP flow from the images, so I can't tell you if there are areas to improve there

[soft zealot]

oh "i shouldn;r " ?

really ?

Well dust me diown with a tickling feather and cal me Emily