✨｜sdxl

[uneven dove]

mm, that doesn't mean the files are gone but it certainly sounds unstable

have you adjusted any firmware params?

[high skiff]

nope

[upbeat summit]

either something happened to your OS installation or your hardware settings are not correct - check ram speed. cpu should be automatic if you don't got into overclocking

[uneven dove]

can you launch some kind of thermal monitor?

[high skiff]

the drive is showing up as 0 bytes

its nothing to do with my thermals

the drive is fucked

outside of that drive, everything else works

[uneven dove]

go into Device manager and see if it shows up as an exclamation

[visual glade]

unplug it, boot linux

[uneven dove]

i suggested it but i don't think he's comfortable with linux

[high skiff]

theres about a dozen exclamation devices

[visual glade]

then do a ddrescue

[high skiff]

I have 0 interest in going back to Linux

[uneven dove]

we're just saying to diagnose the issue

[visual glade]

I have no idea how to do recovery on windows but it's pretty straightforward on linux

[uneven dove]

it's a LOT easier to see hardware issues there

yeah, what comfy says

i've diagnosed hardware failures on linux and even used kernel methods to get damaged devices working again

if the device is damaged but the data is accessible, i would assume that's via Linux

SSDs are tricky because the ddrescue thing assumes the controller is still operational and able to speak to the nand

[west breach]

windows has chkdsk

[uneven dove]

the problem with running chkdsk in windows is that it thinks the device is 0 bytes

if you're insistent on diagnosing it in windows, go to the "event viewer"

red circles in there = bad stuff. i don't know what the heck to look for in there, but it'd probably look like HardwareDeviceError

[visual glade]

on linux it's just: dmesg -w
plug the thing in

[serene sedge]

Sounds more like a usb fault if it's an external drive

[visual glade]

and if it works try to dd it

[high skiff]

its full of hundreds of errors and warnings

[quasi remnant]

hot take probably but i'd advise against using external drives; their constant need to be connected means they're prone to corruption/failure under lots of conditions

[uneven dove]

i assumed it was a nvme thing

[high skiff]

178 errors

[uneven dove]

@quasi remnant read the room lol

[west breach]

did you do a fresh windows install with your new computer?

[high skiff]

yes

that drive is fine

something nuked my second drive

[visual glade]

external drives are fine you can always shuck them

[high skiff]

if that drive is gone, all of my SD everything is gone

[serene sedge]

Do you have another pc/laptop you can try it in?

[quasi remnant]

more for future reference. since if the usb randomly disconnects due to windows shenanigans while stuff is being written or any number of other things issues can crop up

[high skiff]

no

[visual glade]

what kind of external drive is it?

[uneven dove]

i mean, fair, but i've got a 5-disk USB-C enclosure and it works great, i assume because it has a robust power supply of its own

[high skiff]

m.2

these errors seem... weird

[uneven dove]

oh

if it's m.2 and it's external. its in an enclosure, right?

[serene sedge]

It's plugged into usb3 right?

[uneven dove]

maybe that thing got pooched

[visual glade]

I would shuck it and plug it in a computer directly

[high skiff]

USB 4

[high skiff]

I don't have a way to do that

[uneven dove]

a 12900k mb doesn't have two m.2 slots? wtaf

[high skiff]

not without disassembling this entire PC

I do have my moms laptop

let me try it on that

[uneven dove]

i'm assuming you tried a different port

[visual glade]

oh you have one of those hard to open cases?

[high skiff]

no, I need to remove my GPU and likely my CPU cooler to access my additional M.2 drive bays

[uneven dove]

i'm only ever using Antec cases from now on lol

as of my last build

cpu cooler?!

[visual glade]

I have a huge case because I wanted space when doing stuff in the computer

[uneven dove]

i have a Thermaltake X9 Core, Comfy. look that one up

[high skiff]

The IO operation at logical block address 0x5f1f08 for Disk 1 (PDO name: \Device\000000a6) was retried.

[visual glade]

I have a 7000D

[uneven dove]

i got that one because it looked like a modern DEC Alpha server

[high skiff]

OH

everything just popped up-

[uneven dove]

da fuk

[high skiff]

what in the hell

[visual glade]

dead usb port?

my old computer has a few of those

[uneven dove]

probably just loose connector

[high skiff]

and now its working fine

[uneven dove]

or dirty

[west breach]

windows and usb, match made in heaven

[high skiff]

no, i tried across 3 USB ports

I am not sure what happened

but all my files are here

[serene sedge]

Get some cloud storage and back it up 😁

[uneven dove]

sytan installing windows today

[high skiff]

and now everything is back to not reading

[uneven dove]

sounds like bad enclosure/cable

[high skiff]

and my file system is crashing

[serene sedge]

What brand disk is it?

[high skiff]

Solidigm

[uneven dove]

whut the hell is that

[high skiff]

Its a reputable brand

its by SK hynix, basically the best NVMe Drive manufacturer

[vale eagle]

An idea of the comfyui, could we load mutiple workflow in one workflow, e.g. use custom workflow as a component.

[west breach]

I'm going to sound like the most annoying microsoft support forum MVP representative but... "bios updated? drivers updated? windows updated?" 😄

[uneven dove]

sounds bootleg lmao

[high skiff]

SOLIDIGM is Intel

[visual glade]

or wait hard drives are 3.5

[high skiff]

SK Hynix bought out all of Intels storage facilities and rebranded them as SOLIDIGM

[visual glade]

I should probably go to sleep

[high skiff]

ok, the image loaded now

I am not sure what is up here

[quasi remnant]

back up what you can incase it's failing

[high skiff]

I have nothing to back it up to

my other PC is OOC, and my moms laptop only has 256GB storage

[serene sedge]

Cloud storage, OneDrive is 1tb.. not free though

[high skiff]

wait a sec

what the hell

[uneven dove]

i would leave unplugged then until you have a different solution because it could just disconnect/reconnect and zero its partition table

[high skiff]

my PC says that 50% of my total drive speed is being used at 0.3MB/s

[visual glade]

I would put it in a real m2 slot

then do a smartctl -a /dev/whatever

[high skiff]

what the hell is a COM Surrogate

[uneven dove]

it's for the computer to have babies

[quasi remnant]

normal thing with long explanation

[visual glade]

COM is a part of windows that you never want to deal with

[serene sedge]

Unless you can try the drive in another device it sounds like it's dieing

[high skiff]

Ok, shutting PC down, trying different port and cable, cause the files ARE there, it's just reading everything at like 0.3MB/s

[uneven dove]

what i said at first 😄

i overvolted my SSD once to overclock it and of course lost all my data. why? because power corrupts

[high skiff]

Just don't know how restarting my computer would have damaged the USB port or the cable

[visual glade]

stop playing around with it and put it in a real m2 slot

[high skiff]

Alright, fine then I'll do that next week or something

[uneven dove]

idk i think you should unplug and replug it in another 3,000 times

[high skiff]

I have only unplugged it as the PC was off

[visual glade]

then once it's plugged into a real m2 slot the first thing you want to do is check the SMART

[uneven dove]

USB has staggered sense and ground pins to prevent issues plugging in at runtime

it says that in the windows 95 thing i showed earlier actually

ahaha

[visual glade]

if the SMART is fine it was probably just the enclosure so you can probably continue using it as usual

[visual glade]

without that enclosure of course

[high skiff]

PC just updated

drive works fine now lmao

[uneven dove]

i hate your damn rollercoaster stories you tell us

[high skiff]

I'm sorryyyy lmao

stupid shit happens to me sometimes lol

its fine now

[uneven dove]

Narrator: And it was fine all along

[high skiff]

right lol

[uneven dove]

still throw that enclosure out imo

just cuz i hate it now

who even like, gimps nvme with USB

[high skiff]

The enclosure has been fine for a long time, i can only assume it was the windows update

[high skiff]

somebody who didn't have enough slots, thats who

[serene sedge]

Wtf.. nvme enclosure? That's what we're talking about?

[uneven dove]

plug it right into the cpu or something

disappointed in intel for not having two slots

[high skiff]

I could get one of those shitty 4060's with 2x m.2 slots cause "The GPU only uses half the lanes" lmao

[high skiff]

it has 3, I already said that

my old mobo only had 2

[uneven dove]

o

[high skiff]

on this one, if I wanna access the other 2, I have to remove everything

[uneven dove]

but that sounds worthwhile lol

[west breach]

my mobo has 4, yeah 3 are under some metal plate things you have to unscrew

[uneven dove]

how many places does this nvme device go, is it like Flat Stanley? seeing the world?

[west breach]

since the gpu is so huge you have to remove that too

[high skiff]

Yeah, that's how it is in me

[high skiff]

No, but this enclosure is intended to bridge the workflow gap when I am working between laptop and PC

Again, like I said, I only have it cause my previous PC didn't have enough slots

And I could take get an expansion card

[uneven dove]

you could like build a 4U rackmount NAS server that has a backpack on it

every day is leg day

[visual glade]

that's why I put my SSD in the m2 slot under the GPU so the others were easy to access if I need them

[west breach]

saitama?

[high skiff]

You know what, I wonder if it was not the windows update, but it was the USB ports

If that's the case, that suggests that both of my USB 4.0 ports are toast

[sharp robin]

yes sir

[high skiff]

In which case, fuck

[dense chasm]

it's weird,Jordan's face feature extraction is totally damned

[uneven dove]

i am concerned for your love with MJ 😛

[dense chasm]

We will never be slaves，Imao

[quasi remnant]

i think most celebrities & very famous things had their data messed with a bit. for instance if you ever try to do mona lisa it's always incredibly overly-bright, same with trying to get the face of certain celebrities. chuck norris works fine though but he's chuck norris

[west breach]

is it because it tries to give everyone high cheekbones and a square jaw?

[quasi remnant]

¯_(ツ)_/¯

[uneven dove]

oh

set your aesthetic score to 2/3

that'll fix that

[quasi remnant]

the positive aesthetics score?

[uneven dove]

bam

fixed

[quasi remnant]

...can't tell if fixed or not

xD

[uneven dove]

it's a different person

[high skiff]

hmmmm

that does not seem right lol

[uneven dove]

lol is that a boombox?

[autumn forum]

that is an,, interesting expresion lol

[dense chasm]

i wanna to make a AI generation dialogue between MJ and Kobe,i'm stuck now

[uneven dove]

ohhh

[autumn forum]

what looks off?

[high skiff]

woah...

[dense chasm]

the neural network of deep learning algorithm changed always with features extraction,damn

[high skiff]

My generations are way faster now

and my GPU is HELLA boosting

way past what it should be lmao

its rated max boost is like 1870mhz

its going to 1970mhz

[high skiff]

Got too carried away overclocking lmfaoooooo

[uneven dove]

sytan's gpu right now

[autumn forum]

😦

[high skiff]

wow, this PC is so fast

[high skiff]

in that amount of time, I am already restarted, launched comfy, and am generating again lmao

Bryan Renault's

[winter raptor]

Meanwhile I’m over hear on my gaming laptop from when I was in college 💀

[high skiff]

this PC is so much faster OMG

even loading SDXL and SDXL refiner only takes like 8 seconds now

[autumn forum]

America by killing joke begins to play

[trim orbit]

modern pcie4 ssds are fun

[west breach]

dunno how to boost my gpu, has a '1 button oc' whatever that is

[uneven dove]

there's a bunch of people with 1660's in here so, ya know, don't rub it in too much 😛

[sharp robin]

i need to save up

buy a 3090 or something

[sharp robin]

cost like $650 around my area used

[dense chasm]

i doubt the model has some ethnology bias on specific figures

[autumn forum]

yall makin me think i got a bad deal. payed 800 for my 3090

[uneven dove]

very impressive shoulders

[high skiff]

I don't mean in that remark, that has an improved much, but the CPU

[uneven dove]

frank what the hell lol

[high skiff]

Like the processing time in order to load the models, and then especially the processing time between the first sampler and the second

[sharp robin]

[dense chasm]

air Jordan looks like an idiot and dumbass right now .Imao

[uneven dove]

is that what that is?

i got him

[delicate grotto]

Did it drown?

[dense chasm]

this is the black jesus genuinly

[uneven dove]

smol

[autumn forum]

when you click on an image and it gets smaller lol

[uneven dove]

i love img2img lmao

[sharp robin]

what u mean drown?

[dense chasm]

can u img2img with the real photo of mine?

[delicate grotto]

650 is extremely low

[dense chasm]

dude,you have a amputational Jordan,Imao

[sharp robin]

Not around my area

[delicate grotto]

Oh wow

[winter raptor]

It almost got the text right

[delicate grotto]

Holy shit wtf
Best i can find around will prob be ~1500

[dense chasm]

okay,the celeb of sdxl features extraction is totally sucks

[sharp robin]

Idk what to tellnyou prices coming down

Maybe crypto crashed or something

[uneven dove]

@dense chasm i asked for shrek version

[sharp robin]

I should try and lowball em all $500

[delicate grotto]

Finally then, cuz up to ~ 3months ago it would still cost 2k usd or up here

[high skiff]

I got my 3090 for $600, so its not that unreasonable

just make sure to research the models you are looking at

the one I bought had bad VRAM thermals (a killer for AI), but the guy also did AI and already repadded all of the VRAM chiplets

[sharp robin]

is 3090 what u got rn?

[uneven dove]

lmfao it's not good at reproducing MJ at 80% strength

[high skiff]

I have a 3080 right now, 3090 is in the mail

[sharp robin]

so u bought 2 gpus?

[high skiff]

I came from a 3060ti, got a 3080 for $320 basically brand new, and about a month later, I just decided to buy a 3090 (gonna sell the 3060ti and 3080 likely to pay it off and then some)

[west breach]

only a few days to go until 1.0 release ✨

[high skiff]

My GPU gets here the day AFTER 1.0 lmao

[uneven dove]

yep 1.0 is coming with a refiner

emad said earlier

[sharp robin]

did they train it further?

[uneven dove]

yup

[high skiff]

Nice, glad to hear they realize that 1.0 is not good enough without it lmao

[uneven dove]

well, we have been telling them that 😛

[high skiff]

oh you know what

I wonder if the faster model loading is cause I have 64GB now

cause it would smack into my 32GB cap before

I am sure its a mix of several times faster CPU, and also more RAM headroom

[uneven dove]

lol wtf why

[high skiff]

why what?

[uneven dove]

that's SO much sys mem use

[high skiff]

yeah, thats consistent with basically everybody that uses it

I have talked to at least 2 dozen people saying 32 is like the minimum to run it

hell, when I was training that glacier speed LoRA for SDXL, it was using 31.8GB RAM

[west breach]

mroe ram def helps load and swap models quickly

[high skiff]

I have 64GB of high speed DDR5 now, and it seems to help a lot, I think haha

[uneven dove]

mine uses 19GB of system memory

[west breach]

made a huge diff for 1.5 when i went from 16 to 64 on my laptop

[high skiff]

before it was probably a solid 30 seconds for me to load them

[high skiff]

likely a difference between comfy and diffusers

but for comfy, it uses 32GB for sure

[uneven dove]

it's probably because i move everything to the GPU

[shy kelp]

so XL needs 64 gig to run smooth? damn

[high skiff]

32 works

but its probably the minimum

24 GB RAM would probably struggle hard

[shy kelp]

ive got 32

[uneven dove]

it's really like 16G but that's dependent on comfyUI settings being correct. i doubt A1111 has a chance, count on 64G there

[high skiff]

for me and several other 32 GB boys, it would smack into the cieling and then purge other things in RAM

its the whole reason I went 64GB in my new PC

that, and its shockingly affordable

[sharp robin]

tf, im running it fine on 16gb ram lol

wth

[winter raptor]

meanwhile im on 6gb LMAO

[high skiff]

honestly, I am not sure

[autumn forum]

i may end up buying 2 more sticks to get 64gb of ram if training loras actually maxes it out

[high skiff]

we are talking system RAM

not VRAM

6GB VRAM is fine

[winter raptor]

ohhhh, 16 then

[high skiff]

weird, maybe its another optimization thing

I know for me when I was on 32GB, it would load both models at the same time

likely for you guys, it loads one, then the other

and on 64, it definitely does both at the same time

[trim orbit]

i caught 64 gb of the lowest timings and fastest bandwidth ddr4 i could find for my alderlake. it's a zippy one now. i'll upgrade to a ddr5 board when there's more point and less price gouging

[high skiff]

DDR5 is shockingly affordable now

[trim orbit]

i migrated my old system's ddr4 over and man, bottlenecked me by about 30%

[high skiff]

I got 64GB fast 6000mhz CL32-40-40-80 for just $150

not even like a special deal, thats just what it always is

[trim orbit]

nice. doing 3600 with 16-16-16-18 here

paid bout that

[high skiff]

thats what I had in my previous PC!

3600 16-16-16-18 as well

[trim orbit]

by more point, my use case wouldn't flex it i mean

[high skiff]

I paid a pretty penny for that

[trim orbit]

loading models i think i'd shave a few seconds off. i don't do much else

[high skiff]

32GB DDR4 at that speed cost me like $270

[trim orbit]

hah yeah at one time that was HIGH end

[winter raptor]

this is horrifying, the people in the background

[high skiff]

but yeah, i was gonna get 32GB more DDR4, but I realized that DDR5 was so much more affordable than when I was first looking lol

[molten breach]

I think stable diffusion sometimes combines Michael Jordan and Michael B Jordan into an unrecognizable dude

[high skiff]

ok, so, new install with the faster CPU and more RAM and a non power limited GPU, and I can now do 4 images in 35 seconds, or ~8.75 seconds an image

thats so much faster than my old PC

just trying to justify this purchase to myself lol

hopefully the 3090 will see another huge reduction in time

WHY ALWAYS THAT ONE

[uneven dove]

moist johnny knoxville

"riding the slide at the waterpark" jesus really

[west breach]

found the oc scanner for my gpu, running now. should also look into the amd expo thing

[high skiff]

pretty happy with some of my MJ results

[west breach]

bullbs

[high skiff]

bulbs

😩

:)

[uneven dove]

well i have moar buttons for my bot now 😄

resubmit with new seed, copy settings, take seed, resubmit prompt alone, and then variants and self-serve delete

[dense chasm]

🤣

[ionic dragon]

What's that?

[arctic bloom]

https://github.com/oobabooga/text-generation-webui/

[winter raptor]

Nice!

[delicate grotto]

oi, careful with au1111, loading models might BSOD you, no perma crash though

latest update correct to yesterday

[ionic dragon]

we are working on only sd prompts

not an all round model

[grizzled warren]

I'd chicken out! ;D

[spring fulcrum]

Hi everyone. I just got sdxl going with comfyui. I can get the image to render but when it hits the refiner it errors out. I have a 4090 I am not sure why it is saying its out of memory.

Any help would be appreciated:

[sharp robin]

Seem u clicked run cpu?

Instead of gpu?

[spring fulcrum]

It would seem that way but I promise I didn't

[sharp robin]

Oh wait it renders then hits oom at refiner.

[spring fulcrum]

yessir

[sharp robin]

Can u screenshot ur workflow?

[spring fulcrum]

maybe it could be settings I just left everything default except the prompt

[spring fulcrum]

I have 64GB DDR5. I just set it to system managed. It was already set to 9055MB. Hopefully this works... I do a lot with docker containers and docker eats up a lot of RAM but I don't have Docker running ATM.

[sharp robin]

Hopefully that fixes it but also try and keep the same sampler and scheduler.

[spring fulcrum]

I set it to 16MB min 32768 MB max and its working perfectly.... Thank you.

Now that I got my default go to cat image I am on to much more interesting things with SDXL 🙂

[dense chasm]

does anyone knows comfy can use the LyCORIS (LoCon/LoHA) model checkpoint now?

[ionic dragon]

for sdxl, no

[dense chasm]

this model is perfect,how can i use it in comfy?

[west breach]

you need a 1.5 model

[ionic dragon]

use this node

is inpainting supported in sdxl?

[dense chasm]

put the download safetensor file in the directory of ComfyUI/models
/loras/?

[ionic dragon]

yep

[dense chasm]

sure,1.5 model is my default setting

i just found D-ID has an authorized limitation now ,i can't uplaod the celebrity image,any alternative img2video application recommend yet?

[dense chasm]

how to chain the lora with other nodes,u have any workflow examples?

[ionic dragon]

https://comfyanonymous.github.io/ComfyUI_examples/lora/

[dense chasm]

thx buddy

[delicate grotto]

holy fuck
comfyui just did a 2048x2048 using a freaking 6gb vram gpu

is it possible to train embeddings/use them on comfy?

[dense chasm]

https://www.youtube.com/watch?v=oVcWj7PcdkM anyone knows how to make a video like this?as far as i know,the dialogue script from chatgpt4,avatar from midjourney,voice from some rvc training maybe,but still can't figure out how to generate the video

[arctic bloom]

Within comfyui, you mean?

[grim otter]

d-id

[dense chasm]

but D-ID is limited for celebrity image uploading,violates the authority rules, i can't upload some celebs images already,so it confuse me

[dense chasm]

are u a starcraft game player?

[dense chasm]

Sauron's army is prevalance huh

[sweet bane]

against bots yes lol

[dense chasm]

for the humanity,never be the slavery,whatever machine or human beings Imao

[west breach]

been working on a new node for comy to add style to your prompt, also simplify things a little bit

[lusty prairie]

cool! @west breach thx!

[west breach]

just need to improve some styles before I publish 😄

[upbeat summit]

that's awesome!

[patent badger]

can anyone please link a guide to do lora's on SDXL? I'm not super technical so anything would help!

[hollow halo]

Not sure there are any yet, but I have had good success so far if you want to dm me

[shy kelp]

is it tomorrow?

[clever verge]

Trying to download the xl base model from hugging face with StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, variant="fp16", use_safetensors=True) but complains about "text_encoder_2\model.fp16.safetensors not found". I seem to be able to get other non gated models. I'm logged on with token and it's in the .cache/token. Anyone faced something similar?

[frank sleet]

how to use

[halcyon tusk]

SDXL? try any of the bots!

[lilac wren]

Jim Carrey, Mafia boss in LS

[upbeat summit]

gta style works pretty well! cool

[hollow halo]

lots of toes

[lilac wren]

🙂

[upbeat summit]

those are some great hybrids!

[hollow halo]

Normal toes (I think, never seen a kangaroo irl) lol

[hollow halo]

Thanks! Trained a Locon on 40 images of humanoid animals

[upbeat summit]

ohh... that's why they are so good 🙂 nice

[lilac wren]

Little challenge guys, you take a legendary actor, and make him an NPC / Protagonist of GTA painting.
And without stealing other people's prompts huh 🙂

[sweet bane]

easy

[lilac wren]

A wacky situation is welcome

[sweet bane]

joking about easy, not sure what i can do

[hollow halo]

What kind of dog is that?

[west breach]

pomeranian

[dense chasm]

your style prefer to darkness,it's interesting

[hollow halo]

He has no nipples so it should be okay lol, let me know if not

[lilac wren]

Ewan Mc Gregor, and his hot dog business

[upbeat summit]

that's some pop art 🙂

[hollow halo]

@west breach I'm dying right now 😂

[west breach]

supposed to be 'comic-book' 😄

[upbeat summit]

I dialed in the gta style, but the coherence is horrendous 😄

[fresh path]

1.0 release when

[lilac wren]

18 july

[fresh path]

confirmed?

cant wait 🙂

[lilac wren]

No, i choose a random date

[fresh path]

good enough

[sharp robin]

[lilac wren]

Will Ferrell gang biker

[upbeat summit]

nicolas cage in a GTA 5 loading screen, concept art

[lilac wren]

you should type "GTA V painting", and type in negative photography.
But great idea lol

[upbeat summit]

I like this style too 😄

[lilac wren]

Refined x1,5

[urban fjord]

Hey, does anyone recognize this movie and can help me find it?

[upbeat summit]

bruce willis at a burger joint in a GTA 5 loading screen, concept art, GTA V painting

[west breach]

okay pushed the prompt with style node to my github: https://github.com/bash-j/mikey_nodes

[upbeat summit]

git pull in progress 😁

[west breach]

try the prompt_with_styles.json workflow

[hollow halo]

Pulled it immediately lol

[lilac wren]

Great

[upbeat summit]

thank you so much for your work. I will check it shortly

[west breach]

the styles are kept in the styles.json file, so should be fairly easy to add your own styles. you don't need to reload comfy if you make changes to the styles.json file

any feedback more than welcome 🙂

[lilac wren]

Will smith, gangster of grove street

[hollow halo]

Generating a test image with your setup right now, will let you know what I think of the setup soon 🙂

[hollow halo]

Photograph style vs Crayon style, same seed and prompt for the inputs, besides what the style adds

works with Lora's and locons too!

[upbeat summit]

those are great!

[west breach]

awesome! thanks for trying it out ❤️

[hollow halo]

If this pixel art one works I'm going to scream with joy for the possibilities you opened

I'm contradicting it a bit with my prompt but didn't want to change any variables, this is good though ❤️

[west breach]

that was the hardest style to get working

[hollow halo]

You one upped ClipDrop with the style choices too, so bravo

[west breach]

haha, i tried 😄

[hollow halo]

The naming structure of the images with that other new node is nice too

[west breach]

yeah, I didn't like the built in style, and I like to sort the files by filename descending to get the most recent, instead of using the sort by date which is really slow in windows when you have over 1000 images

[void skiff]

I still see so many that use 8K, 16K, UHD, HDR, 64 Megapixel and RAW photo, do those tags do anything when the AI often just do 1024x1024? It sound almost impossible to imagine how an AI would be able to do 64 megapixels into 1024x1024.

[autumn forum]

I did a test yesterday to see if those tags really worked. It seemed to only make the photo possibly a little more realistic in the base prompt. But it’s hit or miss

[high skiff]

honestly, SDXL's non noise enhanced images for realism like the one I just posted look kinda bad

[hollow halo]

those tags would have been on high quality images, so it makes sense to see it improve things occasionally

[high skiff]

they look way more realistic when using some form of method to add in some more noise

[urban fjord]

It is not needed now, and I don't know if it was needed before. But it is not about doing 64 megapixels, but what images are captioned as 64 megapixels.

[lilac wren]

Share this workflow?

[high skiff]

its not done yet

[upbeat summit]

Johnny Cash sitting on a couch excited to see you in a GTA 5 loading screen, concept art, GTA V painting

[lilac wren]

amazing

[high skiff]

also, its not consistent enough to share as the normal workflow for my github

[upbeat summit]

well - some coherence errors - but well straight from the generator 😉

[lilac wren]

https://www.youtube.com/watch?v=uaGF6WPuSpM

[void skiff]

I use bot-1 hare so I do not know how to test, but so many use 4k, RAW photo, UHD and so on, I often say that folk after a year use things they know is working, but sometime I think AI is 60 % magical words that folk just use for other use those words.

Had been fun if AI could be used to analyze prompts and result to see what are unessesary words,

[hollow halo]

Album art, Neon-punk, origami, and low-poly

[west breach]

oh yeah, forgot about the album art style 😄

[uneven dove]

it... depends.. if its a fine tune especially

[high skiff]

considerable textural difference

[upbeat summit]

between?

[high skiff]

between what?

[upbeat summit]

what is different? 🙂

[uneven dove]

adding noise you now say is a good thing?

[high skiff]

its my fractional offset

its been good since the day I found it 😅

just not reliable enough to make the main workflow

[hollow halo]

Texture it appears

[uneven dove]

that aint adding noise

[upbeat summit]

How is your computer?

[high skiff]

its doing good now

[upbeat summit]

great!

[high skiff]

seems all that scare I had was either a broken USB or a bad windows update

its working nice now

[high skiff]

true, its just not consistently denoising it

[lilac wren]

Tom Hanks LSPD

[uneven dove]

you are giving it one too few steps right?

[high skiff]

I am giving it less than one too little step

[uneven dove]

the incoming latents have like 6 steps of noise left and you did 5

[high skiff]

a full step too little destroys the full image

[uneven dove]

its a full step lol fractional steps arent a thing

they cant be

[high skiff]

don't know what to say, if you make it a full step offset, it destroys everything

[uneven dove]

offsetting where they begin will do that, but offsetting where they end will not

they have to overlap the stage 1 end and stage 2 begin

[high skiff]

I am offsetting where the refiner starts compared to where the base ends, and its not a full step offset

[uneven dove]

the stage 2 can have its length reduced by a step

which seems to introduce more

[high skiff]

have a numbers example of what you mean?

[uneven dove]

like stage one does 10 of 21 steps and refiner does 10-20

[autumn forum]

That would be cool if someone did do that. Turn a single step into multiple steps that way you have control over a fraction of steps.

[uneven dove]

i did that with the john cusack jester and it added a lot of coherent detail but you can absolutely push it too far

[high skiff]

my example ios 0-13/21 and 13-20/20

so there is a misalignment between 13/21 and 13/20, which is that "fractional" gap in latent that adds considerably more detail

if I make it 13/20 and 14/20, there is way too much noise and it destroys the image. That would be a full step offset

[uneven dove]

thats pndmscheduler that starts out at 50 steps and has a lot of flexibility but takes a while to generate

yeah you have to look at the actual sigmas that theyre starting and ending ons

the sigmas have to be the same when the refiner picks up where the base left off

[upbeat summit]

just had tom hanks going, but yours is way better 🙂 nice one

[uneven dove]

and on top of that i added a line of code that would randomly select a step to add more noise on and that fixed some coherence of the incorrect starting step but made the picture so wild looking, overall breaks stuff lmao

[high skiff]

I have a potential fix for the issues introduced by my fractional misalignment

it seems to make things a little less... crusty

[uneven dove]

i know how to make them melty instead

[dense chasm]

this is the truly Mj,fuxxing awesome

sdxl be short of lots of additional models and extension nodes

[upbeat summit]

it sure will have lots of additional content, but it's not even out yet

[dense chasm]

sure, hard work should be appreciated and to be continue..1.0 will released soon?

[autumn forum]

We were told not to train on 0.9 unless we are just testing. 1.0 is coming out on the 18th

[dense chasm]

i just hope celebs render well,it will save a lot necessary efforts

[autumn forum]

Some do some don’t. I’m sure someone will come out with a celeb Lora or a model. You just have to wait and see

[dense chasm]

yep,life is like a box of chocolates. you never know what you're gonna get,i taste different flavor since 1980's last century

[high skiff]

ok, I might have found a potential fix to the destruction my fractional offset could cause

[dense chasm]

excuse me ,why u run these models locally?it seemd u devote a lot of resources and efforts

[high skiff]

what do you mean why do I do it?

I do it to try and find the best ways to use these models

[paper phoenix]

its much like "why have a car when you can take the bus?" It allows more freedom to play around with things, do things your own way in your own time.

[dense chasm]

i admire you excellent work，just for curiosity,keep doing bro,your workflow is so cool

[paper phoenix]

plus there are constant extensions for text to image UIs which i couldnt honestly expect a third party to install or keep abreast of on my whims. that's why i run locally.

[high skiff]

I don't think it was meant to be rude, they seem to have a language barrier is all

[paper phoenix]

i didnt take it as rude 🙂

its a fair question. what are the pros and cons of local AI models?

[dense chasm]

yes,still a lot of work and efforts to give it a short,i just found automatic and midjourney is ahead,anyway we can can try some new ideas and stuff together

[eternal fog]

When you run it through things like bots on this discord you are limited to how it's been configured.

If it's on your own machine then you can do what you want to it.

[chilly night]

kinda offtopic but can someone help me with #🤝｜tech-support

[high skiff]

I am pretty proud to say I can consistently generate better looking realism images locally compared to the 1.0 candidate models

HES BACK

[dense chasm]

i made some adjustments of Stan's workflow,AI is prevalance across different aspects of our lives,AIGC is ultimate topic eventually

[high skiff]

Stan

[arctic bloom]

https://tenor.com/view/stan-writing-angry-enime-gif-20035572

[dense chasm]

it's good to know,let's embrace the whole new AI world

[dense chasm]

sorry bro,Sytan,mispelling, english is my second language,excuse me

[high skiff]

all good haha

its crazy how one missed later takes me from Satan loving chaos boy to "I just filed for my pension"

[uneven dove]

accidentally recreated monkey christ

[sudden cliff]

For me, even if for nothing else, it's nice to just automatically save every generated image to my hard drive. I can generate many images a minute, which would be a pain to manually open and review before saving. I can queue up hundreds of images and check them out at my leasure without any UI challenges.

[uneven dove]

you can do that with remote APIs too

[dense chasm]

both good and evil Paradox or mercurial is human nature,Imao

[uneven dove]

one myst later and we're all moist

[high skiff]

[hollow halo]

Anyone have any decent negative prompts I can copy paste real quick? Accidentally cleared mine and am trying to quickly test something

[boreal bough]

they definitely have an impact, just not what most people would expect. 4k & 8k introduce at lot more background blur & bokeh - if you stack enough of these words, you literally turn the whole image blurry XD

[hollow halo]

for photo realism I should mention

[delicate grotto]

someone give me a prompt with a seed for sdxl + scheduler

[azure oxide]

The last message I read before going to bed last night was that you lost everything @high skiff lmao glad you're good now

[uneven dove]

img2img be like

[azure oxide]

The people need their comfy sdxl leader

[sudden cliff]

Well I guess the final nub would be, there aren't too many cloud resources I can use for free like I can use my local resources

[high skiff]

yeaahhh, turns out either my USB port on my brand new MOBO is dead, or the windows update I did fixed it

blue tiger :D

[azure oxide]

I always just shrug those mysteries off and move on lmaoo, if it works then it works

[eternal fog]

Not what I had in mind, but interesting.

Also a 2nd img2img pass after an upscale actually worked well on this one, as it looked like this before

[winter raptor]

@west breach for your custom node is the top text input positive, bottom negative?

[autumn forum]

8k, uhd, ultra high resolution, 4K, insaneres, absurdres, QHD,OLED, 69MEGAPIXELS,9001QHDPRORESMACBOOKPROSUPERSAMPLEQUADULTRA. = the whole image is blurred lmao😂

[high skiff]

great, I leaked one of my prompts again? >:C

[boreal bough]

negative be like "blurry"
person asking? why sdxl so blurry? nothing works

[eternal fog]

That GTA V style is pretty strong isn't it lol

[boreal bough]

btw, you willing to share your upscale workflow yet? ❤️

[winter raptor]

This is fantastic! Would love to see the workflow you used to achieve this type of look

[autumn forum]

Oops:p

[high skiff]

perhaps soon

[dense chasm]

hey, medieval or ancient mysterious walker,your render is awesome too

[high skiff]

I just uhh

[upbeat summit]

LoL Arcane is also present - lots of new stuff

[eternal fog]

I can share mine if you want, it works reasonably well. I'm using a custom aspect ratio node but you can change that.

[high skiff]

need to get back into my github on my new PC lmao

[boreal bough]

weeeeeeee ❤️

[winter raptor]

yes please lol

[upbeat summit]

Will you release you fine-tuning once you have a SDXL 1.0 version? it looks terrific!

[uneven dove]

i think it's just a lora?

[eternal fog]

@winter raptor @boreal bough You can take it from this image

[hollow halo]

Absolutely, I'm training so much right now to figure out how this stuff works early. But yeah it's just a LoRA, was going to attempt finetuning once 1.0 is out, takes too much time compared to LoRA's

[eternal fog]

Actually I've just saved it @winter raptor @boreal bough

Couldn't remember which one it was in my list XD

[winter raptor]

Thanks!

[eternal fog]

It still likes to smooth photos sometimes, but you can mess with the settings to get it to do what you want.

[dense chasm]

lora or LyCORIS is out of control for small training datasets

[delicate grotto]

what prompts are you using? all im getting garbage from sdxl

[boreal bough]

'out of control'?

[hollow halo]

Humanoidimals, An Anthropomorphic, Humanoid white tiger, wearing light, leather armor, red hood, standing in a village at night.

Humanoidimals was the token I kept when training the Locon

[delicate grotto]

so its a lora?

[hollow halo]

And the supporting prompt is this

Highly detailed, dramatic contrast, magical door, rtx, vfx, glowing, ethereal lighting, bloom, global illumination, thick mist and fog

[hollow halo]

yep

[boreal bough]

negatives are important.
not in the sense of having a lot, but rather having as little as possible - while achieving your goal

[dense chasm]

META released new models and deep learning algorithms frequently,no mention Elon setup the new xAI group and purchased 10 thousand GPU in advance

[raw hearth]

hi

[delicate grotto]

mmm i tried lots of negs, even without
still getting worse than 1.4 quality

[eternal fog]

Not sure how you are managing that

[delicate grotto]

me too

[eternal fog]

Even without negatives it makes good photo stuff

[uneven dove]

someone explain to me please why tim curry looks like this in sdxl

[boreal bough]

can you post a reference picture of what you're trying to make? anything really

just so I get a handle on the style

[eternal fog]

Looks like he's had a bad day

[hollow halo]

Anyone got an animal/human hybrid I can attempt?

[hollow halo]

bonus points for giving me an outfit

[boreal bough]

in case you mean generally, or for everything, follow this
#✨｜sdxl message

[upbeat summit]

he was too intense when the dataset was normalized

[dense chasm]

base model is almost in priority,sufficient train datasets,but the overfit is a big problem,it's a trade-off strategy for some small optimized model like lora as well

[hollow halo]

Doesn't seem overfit to me, it trains really well too, even with lora

[dense chasm]

celebrity is a good example as far as i know

[hollow halo]

Celebrities were most likely pruned for legal reasons

icons are still there but minor celebs are gone

[eternal fog]

Nah there's loads of Celebrities in it

But they still can't get Rowan Atkinson right

[hollow halo]

Rowan Atkinson can't get Rowan Atkinson right

[delicate grotto]

pretty much anything i make looks bad idk why
a street lamp was the only masterpiece so far

[eternal fog]

I just imagine the dataset is full of Mr Bean pulling silly faces

[dense chasm]

can u show me how u generate michael jordan in realistic as himself in real life thr sd models, i wanna to learn and grasp

[boreal bough]

I'm asking because I finished all my lora tests, and got everything working. (when I say working, i mean 4/5 seeds return images I'm 'happy' with, that are considered good enough to post)
• Concept of specific Person
• Clothing
• Style
• Composition
• 5k image dataset, with 100+ concepts, all successfully trained into a 43mb lora
• faces <- while working, doesn't count, since I'm overfitting (on purpose) to solve certain issues with refiner

[uneven dove]

frank i tried to make a supervillain vs normal version of you, is there any part i can improve on it

[hollow halo]

She looks familiar, is that from something?

[eternal fog]

It's a "French Woman" With a Final Fantasy FMV type style

[hollow halo]

I can't do that, but here's him as a giraffe/human hybrid

[hollow halo]

Definitely has final fantasy hair

[eternal fog]

yup, it's a really strong tag it seem

s

[dense chasm]

u should pay attention to my icon,i am more elegant and charm than this dummy guy by AI generation

[eternal fog]

Again not what I intended

[hollow halo]

New pixar movie when?

[uneven dove]

well this is useful feedback

@dense chasm more elegant/charming?

you can only blame me so much for this, i put your name into the prompt 😄

[dense chasm]

don't be some stereotype with eastern dumb ass impression dude,the new AI model has some preferences either in my opinion,you can search for some infos on google about different job occupation and class division ethnology

[uneven dove]

i typed frank chieng, so

[boreal bough]

and people said it doesnt do celebrities

[uneven dove]

well, give me your real portrait, i will use img2img

[dense chasm]

are u a LGBT?why not talk about youself and we can know more about u?

[boreal bough]

give us 20 images. I'll make you a lora that can even be turned into a lithograf

[uneven dove]

yes, i "am LGBT"

[vale eagle]

The training set of sdxl has some Asian biase which doesn't represent the real Asian group.

[high skiff]

the energy of this statement is both threatening and welcoming lmao

[dense chasm]

michael jordan is a hard issue for all of us,but midjourney is a piece of cake, fuxx it

[high skiff]

what in the world lmfao

[boreal bough]

smirks enchantingly at pseudo 😏
I may be straight... but I'll make an exception for you

[high skiff]

right lmao

the energy in the studio is very uncomfortable today lmao

[uneven dove]

my thoughts exactly lmfao

[dense chasm]

idk, black lives matters?what a fuxx

[uneven dove]

you animal

[uneven dove]

frank frank frank

what the hell frank

[eternal fog]

I'd suggest stopping with this weird line of conversation

[uneven dove]

[boreal bough]

oh the biases are real for all stereotypes XD
but you can get rid of them with enough negatives, or by adding styles, or just writing the scene this person finds themself in, in more detail

[dense chasm]

solve the existing problem,bro

[vale eagle]

calm down

[tender timber]

Apologies for the late reply, real life gets in the way sometimes 🙂 Thank you for the link. When I loaded it it had two "Load checkpoints", I'm assuming the top left one is the refiner based on the box to its right?

[high skiff]

Call Frank a sliding glass door, cause bro is unhinged

[boreal bough]

yep!

[vale eagle]

That for sure

[boreal bough]

also, you can use notes to store your favorite prompts

just an fyi, so you dont end up like me with random notepad notes XD
much easier to just permanently add them to the ui

[tender timber]

Oh cool, I'll need to get a 43" monitor now 😉

[dense chasm]

we can make more progress together, it's a joke leftover anyway

[eternal fog]

SDXL makes some very good space stuff

On the older ones I always found it did really strange stuff to things like planets and stars

They'd never be actual Spheres for a start

[viscid girder]

How do I get access to sdxl 1.0 Inpainting?

[viscid girder]

How do I get access to sdxl 1.0 Inpainting?

[winter raptor]

this is sick

[vale eagle]

I am planning to make some basic templates like image2image, inpaint and I will share them on civitai. It should help people to step in comfyui and get A1111 similar function to work on.

[high skiff]

if you would like to collaborate, I would love to see what you come up with

We could combine your workflow with my mixxed diffusion for faster and higher quality results

[viscid girder]

Help How do I get access to sdxl 1.0 Inpainting?

[winter raptor]

Wym by mixed diffusion?

[delicate grotto]

what prompt did you use?

[high skiff]

thats the term @uneven dove came up for it lmao

it was a proposed way of using the base into the refiner that has them work half way through the image each

It results in much better results for realism, and proved to be good enough that Diffusers now uses it as their base method of Img2Img

[eternal fog]

Text_G: "Final Fantasy Remake A lone woman in a spacesuit walking on an alien planet, dust, fog, sunset"

Text_L: "Final Fantasy Remake dramatic close-up face shot" (It ignored this one a bit)

Negative: "melted, compressed, artifacts, earrings, deformed, amputee"

[boreal bough]

everything we know, comes from this channel
#📣｜announcements message

[delicate grotto]

what is text_g text_L?

[eternal fog]

The two text encoders for SDXL

[vale eagle]

I just know today, comfyui could load multiple templates into one workflow. It means you could get connected group of nodes as a component like.

[high skiff]

SDXL has 2 text encoders lol

[delicate grotto]

this might explain my problem ahahaha

[eternal fog]

I'm guessing you arent using the CLIPTextEncodeSDXL Node

[high skiff]

my workflow leverages both of them

[winter raptor]

What's the difference between them?

[delicate grotto]

and i've no idea what that is

[eternal fog]

One of the nodes in ComfyUI

[high skiff]

from my research, one seems to be good at linguistic prompting, where you talk to it like a person, and the otehr is much betetr for tagged prompting, where you can strengthen certain ideas of the linguistic prompt

excuse my wonky typing lol

[winter raptor]

Which is which? And no problem, that's super useful info. Do you have an example? Thanks!

[high skiff]

I have my github page where I have the 0.5 beta release of my workflow for SDXL

[boreal bough]

I hear you, and I've suffered attempting it

[high skiff]

it explains the differences and also suggests how to prompt it

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[boreal bough]

but without better reroute node support, I'd recommend against it for now

[winter raptor]

Awesome, thank you!

[urban fjord]

What is wrong with the reroute node atm?

[high skiff]

getting ready to launch the experimental branch, where I add in a section for my fractional offset, and also my upscale beta workflow

[high skiff]

it can't reroute primatives

Comfy said he wanted to fix it

[boreal bough]

positive/style prompt, if you loaded the one I linked to you

[eternal fog]

With this do you always do half and half. I've found personally it seems to do better with only a couple of steps at the end. However this is in a workscale to tidy up upscaled images

[high skiff]

so f or my normal workflow, the base does a little more, about 2/3 of the work

For my newly proposed fractional offset, its more like 3/4

not that I have seen

I have found no way to get a reroute to work into a primitive

[delicate grotto]

so that's what it was

[eternal fog]

Is that doing the whole 21/20 thing? I did that by accident and found it looked better lol

[high skiff]

yes, thats the goal

[vale eagle]

I would give it a try later. If template doesn't work. I would try to implement custom_node to get similar function

[high skiff]

it does look better, but it can cause some issues

so I am reworking my splits and total sample counts and such

[boreal bough]

its renamed, so it makes more sense on how to use it

[eternal fog]

I have my pass after pixel upscale like this at the moment

[uneven dove]

it's in Diffusers main branch now

[boreal bough]

use as such

[uneven dove]

search "SDXL Huggingface Diffusers inpainting example" on Google or Bing lmao

[high skiff]

the 1.0 inpainting actually uses @uneven dove and I's mixed diffusion to get better results haha

They found it helped solve some problems they had

[eternal fog]

I've found as well sometimes really low step counts give interesting images

[delicate grotto]

the G is the positive, the L is the style?

[boreal bough]

yes

[urban fjord]

I guess there is no inpainting controlnet then. Has there been any more news on controlnet support for 1.0? I feel they have been very vague here.

[eternal fog]

From what I've heard it's supposedly coming very soon after 1.0 release

But

¯_(ツ)_/¯

[boreal bough]

tbf, we're -2 days on 1.0 release XD feels a bit unfair to ask for all features pre-release

[high skiff]

@brave halo Are you up by chance?

[eternal fog]

I mean Emad has been putting it out there https://twitter.com/EMostaque/status/1648812237279166464?s=20

Emad (@EMostaque)

New Stable Diffusion update cooking nicely by the applied team, no longer 512x512

Getting loads of feedback data for the reinforcement learning step that comes after this update, wonder where we will end up.

Will be variants for controlnet and other bits too

#SDXL

Likes

412

Twitter · Apr 19, 2023, 10:14 PM

Thats from back in April, talking about ControlNet for SDXL

[high skiff]

Emad is far from a reliable source, TBF

[eternal fog]

I'd say don't hype your shit if you aren't going to deliver it.

And that's what other people think, which is why people are asking "Where Controlnet"

[urban fjord]

Yup which is why I asked if there has been any clarifications.

[boreal bough]

oh yeah, but I feel like we should take those with a bucket of salt.
Currently the announcements are as hyped as No Mans Sky promises.

Important is the post release support & communication

[high skiff]

Emad is a neverending hype train with non-concrete goals of delivering lol

this ^

[hazy basin]

So SDXL has an NSFW filter integrated, even self hosted ?

[high skiff]

not sefl hosted

tho it sucks at NSFW locally

[boreal bough]

'filter' is the wrong word. But also, this is a taboo topic here.

[hazy basin]

it wasn't trained on any NSFW content I guess

[hazy basin]

Is it ? I mean, internet is for porn

[eternal fog]

There's NSFW in it

You just have to prompt out things covering it up

[high skiff]

not to an extent of producing any particularly good NSFW images

[uneven dove]

i told them that two months ago and was suggested that i'm merely hypercritical

[hazy basin]

So the previous version is better at it ?

[eternal fog]

Hey man, if they want people calling them liars, let them go for it.

[winter raptor]

Is that lora public? Looks awesome!

[high skiff]

I think all people should be pretty disillusioned to SAI's over hyped and under delivering promises on average

[uneven dove]

it seems like they genuinely believe the stuff they say which is why they're confused when people get upset or are critical

[boreal bough]

but on the official Stability AI discord, hosted by the corporation Stability AI, where screenshots often get taken and posted on social media - things should remain on a sfw level, since many people ARE at work here

[dense chasm]

@uneven dove oh,man ,your icon reminds me of Jim Carrey as Gingy in Shrek, unique personality,it's good to know,Imao

[past cedar]

what does the resize button do ?

[uneven dove]

@dense chasm thank you, it was my friend's prompt, a delicious jar of shrek

and jim carrey played Gingy? whoa

[urban fjord]

CeoBeast: It gives you another aspect ratio I think

[past cedar]

oh ok

[hazy basin]

I feel like your answer contains criticism about SFAI, is there any drama happening ?

[uneven dove]

it was not jim carrey

[high skiff]

||No, he didn't lmao||

yeah lmfao

[uneven dove]

that blew my mind for a sec

[eternal fog]

People thought it was Jim Carrey?

[high skiff]

https://tenor.com/view/outslicked-evan-raynr-raynday-gaming-evdog805-outplayed-gif-25241669

[uneven dove]

look who figured out GIFs

[high skiff]

part of me feels depserate enough to try and train an SDXL LoRA on my new PC again lol

[boreal bough]

nah, I just work at a corporation myself and often interact with marketing, so I feel the pain

[high skiff]

they added embed perms lol

[uneven dove]

here's a GIFt for you

aw damn it, it didn't animate

[hazy basin]

Oh alright

[eternal fog]

This face doesn't match the hair lmao

[boreal bough]

basically the conversation about it is legit, this just isnt the place to have it

[autumn forum]

curious how you got the sampler to preview the image. how did you do that?

[uneven dove]

oh, it's not a gif

wait, what the hell

why did i think that'd be animated gif

[eternal fog]

It's a startup command

python main.py --disable-xformers --preview-method auto

The Preview-Method

[autumn forum]

thanks

[eternal fog]

It will give a low quality preview by default

Look on the github page and it gives you a file to download for better ones

[urban fjord]

Any reason you disable xformers?

[eternal fog]

I use Torch 2.0.1, so SDP is faster

[high skiff]

I could use some help semi soon setting up a more experimental branch of my github, if you would be so kind?

I have some ideas to set up both a fractional offset test workflow and an upscale test workflow

[high skiff]

hmmm, I should try this

[eternal fog]

Tiny bit faster

[visual glade]

if you want to use the pytorch attention function just do: --use-pytorch-cross-attention

don't need to disable xformers

[eternal fog]

I mean what difference does it make doing it that way? As it's going to use SDP both ways isn't it

[boreal bough]

so far, running nightly version of comfy with gpu only + bf16 vae has given me fastest gen time
(only relevant for 3080/3090,4080,4090)

[eternal fog]

Could GPU only cause issues though on a 10GB 3080?

[urban fjord]

comfy: Has there been any other changes to LoRA performance after your "Speed up lora loading a bit." commit on high VRAM systems?

[eternal fog]

And a BF16 Vae can cause accuracy issues can't it

[boreal bough]

good question
wouldn't know since not many people are on the nightly version with that setting enabled

[visual glade]

yes I pushed another commit to reduce the peak ram usage when using loras

[dense chasm]

yeah,sherk and The Mask major character is the same green color,so versatile, and u know a lot too,Imao ,similarity

[eternal fog]

@boreal bough How do you get the nightly version of Comfy? I only see 1 branch on github

[urban fjord]

Isn't that fix only for normal and low vram mode?

[visual glade]

yes

[boreal bough]

https://github.com/comfyanonymous/ComfyUI/releases
unstable nightly
, also you can just use the bf16 vae, and not use the gpu only

[high skiff]

weird, just tested, its slower for me

[eternal fog]

Oh I just pull straight from Git, so I'll have newer than that anyway

[visual glade]

it's not a github branch, it's the same code as the "stable" but with nightly pytorch instead

[high skiff]

let me do some benchmarking

[boreal bough]

it has different stuff inside, you then STILL need to run the update XD

[eternal fog]

I've got my own venv with all the stuff up to date

So I have nightly torch

[boreal bough]

in that case, try adding the bf16 vae option

--bf16-vae

[visual glade]

then you are good, that package is for people who don't want to bother with python/venv/etc..

[high skiff]

12.30
11.55
12.2
11.99
11.52
11.56
11.61
11.54
11.45
11.57

Those are the times with SDP. 10x the same image with an increasing seed

[eternal fog]

Looks like I can't use that

"upsample_nearest2d_out_frame" not implemented for 'BFloat16'```

[high skiff]

Average time is 11.792s per image with SDP

[eternal fog]

Maybe I don't have the torch nightly, I thought I did. Might need to try downloading that again

[visual glade]

yeah that error means you are not on nightly

[eternal fog]

ok I'll sort that

[eternal fog]

What CUDA Version would you recommend? 11.8 or 12.1

[boreal bough]

@visual glade
one of the custom nodes added this as an additional feature
so far this has been the biggest qol feature I've gotten, since it lets me insta-update my ui's & designs (since default save/load doesn't default to a directory, but instead to last used - wherever that was)

Could you add an equivalent of this to the main branch?
It didn't occur to me until I switched to nightly just how much frustration this avoids

[dense chasm]

SD has embedding layer with KNN nearest neighbor algorithm like mathmetics cosine similar with chatpgt between prompt and latent image

[high skiff]

woah

@visual glade I have a question

just noticed something considerable about gen speed in comfy

[high skiff]

if I have my mouse on any of the screens that command prompt isn't on, I lose a considerably amount of gen speed

btu if the mouse is in the same screen, it goes back up

and now its not doing it, what the hell

something feels really off here

[eternal fog]

Have you guys noticed the classic SD Same Face in SDXL?

When using simple prompts like Woman. It doesn't always happen, but with certain prompts it seems to lock into a single face.

[dense chasm]

negative prompt is especially important since sd 2.1,sdxl is indispensable

[high skiff]

I am going from 2.61it/s all the way up to 2.98it/s on random

I am not sure why its so inconsistent

[sage basin]

I've noticed on Firefox, at least, the CPU usage of the browser shoots up when moving the mouse around the interface

[high skiff]

oh right, I need to set firefox to run on my iGPU

I forgot about that

[boreal bough]

much words. very meme

[uneven dove]

sorry i was nonresponsive. i was busy working on that

[high skiff]

ok @eternal fog

i can confirm, non xformers is infact faster

average 12.3 seconds an image down to just 11.6

[uneven dove]

lul at you guys, half a second is a non-statistically significant deviation

[eternal fog]

@visual glade which version of xformers do I need with the nightly torch?

I get errors on startup with a version mismatch and if I try re-install xformers it tries to remove torch.

[high skiff]

I did 20 images with xformers and 20 images with SDP, and SDP was consistently that much faster

[uneven dove]

that's in line with random noise causing contention during tests

that doesn't make sense though Sytan, because SDP is using the Xformers code 😛

[dense chasm]

tbh, u are obsessed with shrek, who the hell are u, what's your realistic life lol..

[uneven dove]

they essentially pulled it from Xformers into Pytorch 2, according to @visual glade

[eternal fog]

@boreal bough Do you use the ComfyUI-Impact-Pack?

[high skiff]

IDK what to say, but disabling xformers in command line args makes it faster

[viscid girder]

Can u provide a link?

[uneven dove]

you'd need to do some kind of advanced profiling for determining why it's faster

run the generations inside perf and record a data chart of the timings

[high skiff]

no idea, but it is consistently faster

inside perf?

[uneven dove]

then you can generate a flame graph

https://www.brendangregg.com/flamegraphs.html

[high skiff]

not sure what you mean

[uneven dove]

https://www.brendangregg.com/FlameGraphs/cpu-mysql-updated.svg

[eternal fog]

I've always found SDP to be faster than xformers, I assumed it was because they'd optimised it directly in torch

[uneven dove]

this is a flame graph of CPU use of MySQL during a test

you can see the different ioctls and sys calls, stacked up

the longer the call, the more time it takes

[high skiff]

I don't really care much as to why it is faster, but it is

[uneven dove]

Nep 😛

[eternal fog]

ok --gpu-only doesn't work on 10GB GPU's. It gives me OOM as soon as it loads up the refiner.

[high skiff]

oop lol

wonder if that will work on my 3090

[eternal fog]

Think it will

[dense chasm]

@uneven dove can u generate this mimic so realistic mj profile not for the anime style🤝

[uneven dove]

GPU-only needs like 16gb vram so a 3090 will do fine

[eternal fog]

I might have to go back off the nightly torch though as it breaks the ComfyUI Impact Pack

[uneven dove]

damn, that is one crusty pic of MJ

[high skiff]

coolio

[eternal fog]

And it seems like it's slowed down generations

[visual glade]

try disabling the GPU hardware scheduling if you have not

basically SD uses the compute queue which is the lowest priority one

[uneven dove]

that sounds like Windows weirdness?

[high skiff]

it is off

[visual glade]

since it's the lowest priority anything that uses the GPU will make it go slower

[high skiff]

it seems to only have that huge fluctuation with xformers

[eternal fog]

I don't have those issues with Hardware GPU Scheduling turned on

[high skiff]

it was enough to be the difference between a sub 12 second gen and a nearly 14 second gen

[uneven dove]

unless you want to profile it to find out why i'd just accept that's how it is 😛

[high skiff]

but only with xformers on

[dense chasm]

xformers with the prefix is always awesome, maybe that's renson why elon create the new company xAI😜

[high skiff]

id be down to do it, but I can't be bothered to setup all of that additional stuff to figure it out, so for now I will just accept that that's how it is, and I will adopt not using xformers now

[eternal fog]

hmm, I've lost half an it/s generation time and I dont know why

[viscid girder]

I couldn’t find it

[uneven dove]

https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl#inpainting

[dense chasm]

btw,openai's competant not noly google claude ,meta's LLaMA,additionally Elon, cuz muck release Threads competes with Elon's twitter, Elon counterattack response

[uneven dove]

i completely forgot he existed until he floundered into the news this week, trying to remain relevant

[dense chasm]

anyway,if u are famous, u should be judged historically

[uneven dove]

yeah let's compare him to Genghis Khan so he has a fighting chance

[eternal fog]

Elon Musk could beat Genshis Khan in a UFC Fight but his Mum won't let him

[dense chasm]

Steve Job passed way with apple legacy,Bill Gates still alive,retired from Microsoft ,Gates is a good friend of Warran Buffet,neither with Elon Musk,it's obvious the relationship,lol..

[glad fulcrum]

yesterday sdxl demo on auto1111 extension worked perfectly. Now it seems to generate but the images are not loaded or saved.

[dense chasm]

that's only the technique talk,Bill Gate is so clever,builds good relationship with all kinds of different goverments and commercials, musk is a follower,but musk is a good challenger

[dense chasm]

open source products dilemma, so much to be concerned

[uneven dove]

look frank if you honestly believe that, generate it as an AI image and show us

[dense chasm]

independent thinking, just freedom talking,i live in China ,u know,Imao

[uneven dove]

i never assume

for all i know, you live in a helicopter, constantly in air

[ionic dragon]

get gpus

[glad fulcrum]

So I am not the only one?

well nevermind it worked now

[static prawn]

why u use an extension @glad fulcrum ? there is out of the box support

[eternal fog]

These look like they could be screenshots from some Monster Hunter game or something

[dense chasm]

probably,maybe that's the reason i live in a country i will mind blowing

[glad fulcrum]

I don't know how does it work without the extension? I got the sdxl branch but don't know how to run the refiner

[boreal bough]

photo of frank chieng, 30 years old, talking about open source products dilemma, so much to be concerned

its missing the inspirational quote though XD

[upbeat summit]

Hi Caith 🙂

[dense chasm]

i tested it already,even the celebrity of the model sometimes sucks

[static prawn]

oh no then its fine, i think the refiner is not supported yet

[boreal bough]

same prompt, but 1:1

[glad fulcrum]

there's a commit saying it is, but i don't know how the workflow is supposed to work

sdxl demo extension seems to break when doing more than 1 batch size

[upbeat summit]

nice coherence