✨｜sdxl

[sterile yoke]

is that a desktop or a laptop though

[glad grove]

desktop

[sterile yoke]

yeah

figured

[native knot]

Laptops are generally going to be shit compared to desktops on graphics. It's gotten a lot better over the years, but same "model" isn't really the same model in a lot of cases. Mainly because graphics cards have high power requirements which kinda plays against what laptops really need.

[sterile yoke]

well i'm only looking at laptops

this is just really disheartening

[native knot]

It's just a reality. You can do it on a laptop, it's just gonna take longer and/or gonna be more expensive.

[sterile yoke]

cause i feel like the only way i can get access to this cool new technology is to jump through a million hoops with google colab and it'll never work well anyways

[native knot]

To be fair, it's still very early in the tech's maturity. Give it some time and it'll continue to get better, be faster, more accessible, cheaper, etc.

[sterile yoke]

i guess

[glad grove]

because u only want a laptop thats where the extra expensive price comes from

[sterile yoke]

i know that

it's just frustrating

and i've had very little luck with sdxl on google colab and no one seems to be able to help me cause i always run into issues no one has ever seen before

[native knot]

When it comes to tech, saving up generally has a double-impact. Not only will you have more money to drop on whatever you're gonna buy, but the tech tends to get better and/or cheaper during the time you're saving up. So my advice would be to put some cash away for a bit and target a later date.

[sterile yoke]

well the problem is i kind of need a new computer

[native knot]

So just get something that's minimally viable for the moment.

[sterile yoke]

and i have way more important things to save up for (house)

eh
i'll try again to figure out colab

i tried stable swarm but it kept not working

same with the camenduru colab

[native knot]

Ah, yes...priorities.

Best of luck.

[sterile yoke]

it's whatever

[glad grove]

this probably the cheapest https://www.bhphotovideo.com/c/product/1758502-REG/hp_6u715ua_aba_17_3_omen_gaming_laptop.html/specs

[sterile yoke]

lemme see if there are any used ones

[steel mist]

This is friggin wild—thanks for sharing!

[sterile yoke]

https://www.hp.com/us-en/shop/pdp/omen-gaming-laptop-16t-wf000-161-76w28av-1
i assume this one is not the same and will not do as well?

[glad grove]

all 4060's have 128 bit bus

[sterile yoke]

ah

[heady vale]

https://www.reddit.com/r/LocalLLaMA/comments/14sti62/4060_ti_vs_3070_vs_3060_ti_stable_diffusion/?rdt=64654

[vital ermine]

Been following GAN since 2018 and I think that is where the future is not with diffusion based algorithms.

[sterile yoke]

well thank you guys for your help

[heady vale]

It does look cool, yet my gens never look like the left side. So im curious what generates that image

[steel mist]

I'm assuming the low res photos they used were downsamples of high res shots so they could compare the upscaler w/ ground truth

I'd be really interested to see if a downsampled SD generated image would upscale as well

[native knot]

Was thinking exactly that. I would imagine that you could do some really interesting things with that as a method combined with some conditioning.

[steel mist]

guess that's why they didn't release the model 🙃

[heady vale]

for my bike image at 1024sq I need to zoom in 1600% to see that same pixel level

[fathom hound]

Nvm, I'm dumb

[heady vale]

the zoomed screenshot is the top of the gas tank

unless youre making masssssive images it doesnt seem worth it

[glad grove]

what model is this

[steel mist]

It's from a research paper

[native knot]

No LORAs.

[hardy cipher]

creating beauty

oh, it's this guy again. I made him earlier

doesn't look like he won the genetic lottery

[steel mist]

do you say that just because he's got a row of teeth floating detached in his mouth?

[hardy cipher]

oh, that would point to superior genetics

but his gaze disturbs me a bit

[native knot]

A good checkpoint + good prompt work + fiddling with all the bits + a decent seed value = wonderful results

I still have a few small tweaks to make, but that is almost perfect.

[hardy cipher]

just like this beauty

[glad grove]

my waifu

[hardy cipher]

you'll have to get past this fierce beast first

hmm. I think I should see what it'd look like if they had babies and the babies grew up and we took a picture of them

anyone know what this error is about? defniitely not moving models or dispatching them on multiple devices

WARNING:accelerate.big_modeling:You shouldn't move a model when it is dispatched on multiple devices.
WARNING:accelerate.big_modeling:You shouldn't move a model when it is dispatched on multiple devices.

[native knot]

Good enough to post up now.

[upbeat summit]

so much to explore in these 🙂 nice

[hardy cipher]

I like this one

could maybe be a bit more refined, but like the aesthetic

now this guy is definitely refined

who knew these two would make him?

[heady vale]

wow...without uninstalling any games I reclaimed 50gb of storage on my C drive by removing crap that was just sitting there

[hardy cipher]

I often hit numbers like that just clearing cached BS. so many programs be like, "oh, ya got some space? how about we fill it?"

[heady vale]

exactly it was mostly cache

[hardy cipher]

now I will acknowledge that my primary drive is only 500 gb which is undersized imo. but I store pretty much all media and things on other drives. I'll clear 120 gb of data and the next day it'll be down to 30 somehow

[heady vale]

Im considering playing Starfield so that needs 125gb, then CP77 Phantom Liberty in less than a month probably another 40-50gb there

I should just get a 2gb M2 SSD and store all games on that

[strong field]

Plus two 16TB drives for the wealth of gens

[heady vale]

surprisingly all my saved gens are under 20gb

I delete a lot

oops realised I typed 2gb, I meant 2Tb drive😆

[hardy cipher]

16tb, lol. I mean, I'm not against it. but I've indiscriminately dumped things onto a 5 tb drive for the last year and it still has plenty of space

[heady vale]

sadly it doesnt know Beaker or Swedish Chef from the muppets

[glad grove]

hello ladies

[crisp owl]

close enough

[glad grove]

real

[crisp owl]

most noodie they allow here

[hardy cipher]

@hardy cipher

[short marsh]

this looks insane 🤯 https://magic-edit.github.io/

[vital ermine]

Don't let your SD blind you to what he was doing or saying. He was showing just how 128px became 1024pix etc...

[heady vale]

even 8x upscale, I rarely ever do. I would prefer see a 1k image at 4x and see what it adds then. Im sure its so small that current upscalers are just fine

[late marsh]

Couple randoms for you ❤️

[crisp owl]

Someone brought a gorilla suit to a gun fight

[glad grove]

and hes pointin the gun down there

[crisp owl]

It seems to be working

[crisp owl]

https://tenor.com/view/the-office-mexican-standoff-finger-guns-michael-scott-dwight-schrute-gif-15805326

reminded me of this scene

[crisp owl]

I've gotta say, I admire your tenacity in sticking to Steven Seagal images lol

[glad grove]

well i gotta kill time while i generate hundreds of imgs on 1.5 😂

[shy kelp]

whats the different between this two?

[hardy cipher]

if you don't know the difference I'd suggest method 2

it's more plug and play

option 1 requires more command line stuff

[shy kelp]

okay thankyou

[hardy cipher]

with option 2 it should create a .bat file you can use to open comfy

[shy kelp]

how to download this

[hardy cipher]

I'm assuming you're using windows?

[shy kelp]

yep

[hardy cipher]

click the little search button on the bottom of your screen

type in "cmd"

[shy kelp]

i can type in here right?

[hardy cipher]

that's where it should be installed

[shy kelp]

okay then

[hardy cipher]

woo, my custom nodes got moved up the list. I wonder how that works?

[shy kelp]

next how

[shy kelp]

just paste this?

[hardy cipher]

yeah, if it gives you an error then go from there

[glad grove]

🙇‍♂️

[shy kelp]

it shows this

[hardy cipher]

meh, not ideal but shouldn't cause anything to stop working

[shy kelp]

ohh okay

[hardy cipher]

there are so many finicky little things with these libraries

[glad grove]

witness him

[shy kelp]

but the file .bat doesn't show up

[hardy cipher]

well where are you looking for it?

[hardy cipher]

big chungus seagal

the size of that lad

[uncut fiber]

@toxic moss i made some IMG2IMG workflow just not sure if it works properly. If you want test it?

[peak dove]

Some morning SDXL/ComfyUI surrealism

[peak dove]

You can forget this. It's simply a note to the developer to change something later!

[soft zealot]

SO the image above was the ReVisioned output of my workflow

This is the IpAdaptor output of the same input images & seed & weights ;o)

I think I have some work to do on the IpAdaptor side of things 🙂

I shall call this one "Mr Hangover Cure"

[hardy cipher]

nice

so how does the ipadaptor really differ? I still haven't gotten around to trying it. I probably should

[soft zealot]

dunno yet, from what I've seen elsewhere it does more of a merge (potentiallY) but it really depends how its configured. Only just started playing with it

This is slightly better actually using a Prompt to "guide" Ipadaptor . Not saying thsi is the right way but....

I can see how its taken my prompt & the info in the 2 images to produce the final output

happy with the way my ReVision setup is working though 🙂

[hardy cipher]

yeah, I've been messing with clip vision a lot. not sure if that' revision, or unclip or what it's called these days. but interesting results sometimes

very strange how the exact same input can be so different depending on the prompt

like, not even remotely the same

[soft zealot]

dont forget that "ipadaptor" is different to "revision" even though they share some commonality

[hardy cipher]

I think I'll download the model while it's in my mind

[peak dove]

"Is iPad Aptor different to iPad Pro?!" 🥳

[hardy cipher]

yes

[soft zealot]

Yes. It’s onion encrypted

[floral island]

that is a weird puppy XD (i missed my A1111 randomizer)

[floral island]

implications that dairy is good for bones?

[ionic dragon]

lol

[ionic dragon]

you trained johnny depp's lora?

[floral island]

probably can be done with prompting only on sdxl models

[vital ermine]

SDXL know depp well

[west breach]

does it do depp when prompting for pirate?

[floral island]

that's propmting jack sparrow

[vital ermine]

No, pirate could be anyone unless you specify

johnny depp as a pirate

[shy kelp]

anyone know why is this happening?

[vital ermine]

Not I

[shy kelp]

it only render half way and pause

[floral island]

randomizer prompt going haywire ❤️

i just press generate, stuff comes out 😄

sdxl models are the best thing that happened to my prompting style

[floral island]

technically, she isn't showing any skin

QUICK, TO THE TURTLE-MOBILE

[gusty pebble]

I am currently playing around with SDXL 1.0, coming from 1.5. It seems like all my pictures are soft / way to much bokeh in SDXL. Is there any way to avoid it?

[floral island]

enough steps/proper workflow or up to date A1111?

[astral path]

never tried it but I've seen this here on reddit: https://www.reddit.com/r/StableDiffusion/comments/15wgbwr/new_soap_photo_enhancement_lora_bye_bye_blurry/

[indigo carbon]

dog+birthday cake=..?

[vale eagle]

@indigo carbon when would you release your workflow with image blend? want to try it out and see what different between revision

[indigo carbon]

I won't release it until I'm as confident with it as I was with the normal AIT one when I released it

[indigo carbon]

1 difference right from the start is revision doesn't use AIT, so it's automatically better

[delicate grotto]

how do you clip skip at comfy?

[native knot]

If you're on SDXL, you don't really need to.

But I believe you'd use CLIPSetLastLayer.

[delicate grotto]

which node does it?

[native knot]

That is the node.

[indigo carbon]

it will lose the ability to do text

[heady vale]

agree. I tried it and found it made worse images

[indigo carbon]

when clip skipping, coherency decreases

[native knot]

XL...ahem, excels at not needing it.

[delicate grotto]

nvm ill find myself

[native knot]

I literally told you the name of the node.

CLIPSetLastLayer

[delicate grotto]

i know, thank you

[indigo carbon]

CLIPSetLastLayer is a somewhat useless node.. it makes SDXL lose its high coherency

I can see why it would be useful for 1,5 models though

[west breach]

https://civitai.com/models/138024

Just sharing a lora I trained on frames from documentaries filmed in the 1960s

[delicate grotto]

seems alright even with clip skip, somewhat more coherent in some areas

[indigo carbon]

text is lost when doing that.. try it

[delicate grotto]

nothing lost, everything present
doesn't seem like a problem

[delicate grotto]

hand/leg/machinery seems less coherent

[indigo carbon]

text also

it can't do full sentences like it can without clip skip

[delicate grotto]

tbh i was never interested in text in sd

id rather use img2img for that

[indigo carbon]

dog+cake= dog on burger..?

[heady vale]

try a cake without layers

[wet nacelle]

goo goo gaa gaa. Where is the chocolate milk?

https://tenor.com/view/rock-one-eyebrow-raised-rock-staring-the-rock-gif-22113367

[wet nacelle]

THAT CAKE FLAT AS A FEMALE DOG!

I have recently found out that the word "bitch" is derogatory.

[indigo carbon]

relax man

i'm just combining images, not offending anyone

[wet nacelle]

No. I poorly conveyed my joke.
dang it.

Now I'm upset that I made you feel a way that is unintended.

[native knot]

Huge Jacked Man

[waxen cliff]

I think I've made perfect cover art for a game

[meager canopy]

That's just plugging a primitive into the clip_skip

clip skip -1

-2

[peak dove]

Recommend me some good SDXL LoRAs? Please 🙂

[uncut fiber]

it seems performance of A1111 is way better than it used to be, but still 1024x1024 almost 30 seconds. Still too much. Also getting back i felt bit confused 🙂

[west breach]

I just released a 1960s lora

[peak dove]

Cool - if it's at Civit - I will d/l

[west breach]

https://civitai.com/models/138024

[peak dove]

Thank You, me got it - just introduce this new LoRA into Comfy here ...

1960's LoRA and Underwater LoRA together ...

The 1960's LoRA introduces a small amount of realism into my current w/flow - its nice

[uneven gazelle]

Hey guys, i tried to install TemporalKit with HuggingFace, but it corrupted my whole Stable Diffusion.. Do you know where i can find a clean version that work of the tqdm.py file ? From \venv\Lib\site-packages\huggingface_hub\utils ?

[peak dove]

Oh man!

2 x 1960's LoRA - cool!

[I posted in HumbleMikey's Civitai Gallery]

[floral island]

this is a weird pizza

[indigo carbon]

dog-trooper

dog+storm-trooper=dog-trooper

[heady vale]

I like it. Nice work

[west breach]

@peak dove you might also like my melting candles lora, since it's more 'arty' ;D https://civitai.com/models/127822/melting-candles

[peak dove]

2 x 1960's LoRA in Sytan's W/Flow (no UpScale)

[floral island]

aight, imma do some groceries and leave pc on generate forever 😄

full random max mode is enabled, i'm sure some interesting stuff will come rolling out

[peak dove]

I d/loaded - thank you

[delicate grotto]

thanks mate

[peak dove]

Cool retro 60's look!

Some Anime jumped out

[west breach]

looks good!

[peak dove]

It has definition - very well defined.

My arty stuff is normally a lot looser ...

[west breach]

is that the candles lora?

[peak dove]

No, I'm trying Candles now, but mixed with 60's - 60's seems to be winning - no melts yet.

I will put both LoRAs to Candles ...

[strong field]

i got curious

so i did some research @soft zealot

[strong field]

@indigo carbon you might be interested aswell

[peak dove]

I think this was an attempt at a very crude melt

Not found its metier yet ...

[west breach]

did you put mlt in the prompt?

[peak dove]

No. I thought just the LoRA? OK, I will add mlt

[west breach]

yeah trigger word makes the effect much stronger, the trigger word for the 1960s lora is 'flm'

[soft zealot]

and the answer was?

[strong field]

hahha, well it depends what you are looking for. Clip vision interprets, with IPA interpreting more inconsistently, and IPA give you the best IMG2IMG bang for your buck

here's another example all at default weights

Just a reroll

Notice how #3 doesnt change

and how different #2 was

[peak dove]

It seems to be melting the whole picture - which sometimes looks incredible!!!

It actually brings candy-colored supersaturation - also looks incredible!!!

Nothing outright melty ... yet!

Good hands limbs fingers etc'

[strong field]

added one more, Revision + IPA at the top RHS

[peak dove]

My w/flow is not conducive to that melted candle wax look ... mebbe if I changed my prompt?

I am getting balloons - nothing in the prompt suggests balloons - but mlt seems to like them

mlt also "arraigns" the subject - bolt upright, to camera, centred, dark hair.

I will work on some prompts to see if I can encourage mlt to diverse and differentiated looks

[west breach]

The training images were melting objects on a table

[peak dove]

Perhaps the balloons are actually blobs of wax - mebbe my prompts are too much work for the LoRA to easily deal with!

But if I ever need candy-colored saturation - mlt will be there!

[west breach]

Could be yeah. But I also like the effect it has on colour and lighting

[peak dove]

Sharp, bright, saturated, outdoor - yes, its a cool look

[west breach]

If you do a basic prompt with only a few words you will probably get something closer to the training images

[meager canopy]

I've just posted the workflow that includes the clip skip on Civitai, if you want it.

https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[peak dove]

Yes, some LoRAs respond better to a simpler prompt. I will try that!

L8r - and thaks for the LoRAs!

[uncut fiber]

What is this? 🙂

[dusty jewel]

looks like you've visualized gravity 🤔 while making it look like food...

[strong field]

i would like to hire you to write my prompts

[dusty jewel]

lol

[strong field]

i uh, need to lie down

[floral island]

full random gave me some cool stuff ^^

interesting architecture

[strong field]

i dont know how i got here

[upbeat summit]

so you've build a latent space kaleidoscope

[strong field]

Im getting lost

[upbeat summit]

very nice 🙂

[strong field]

these are generated by recycling clip space

[native knot]

Those look great

[uncut fiber]

This is random as well

[strong field]

I wasnt trying to build this, but now that its built, i cant stop

[upbeat summit]

not sure yet where I ended up here, but let's explore it 😉

[dusty jewel]

is it possible to keep the exact same angle with img2img?

[shy kelp]

can nmkd use sdxl models?/

[strong field]

do you mean, can you use nmkd upscalers on sdxl?

Yes mostly

[shy kelp]

nmkd sd gui

[thorny frost]

Pure Lucian Freud. While others upload fake loras, i completely recommend fr3u xl loras. They work great.

[dusty jewel]

lol nice ufo

ever thought of making those

[waxen cliff]

I mean at least it got proper 5 fingers

[dusty jewel]

let me make the ones that we have seen and that come from the most advanced civilizations

[waxen cliff]

not sure if 3 black blob are even hands

[dusty jewel]

there are ofcourse less smart civilizations visiting us as well, like the ones that keep crashing

between you and me, i dont think they crashed. i think other aliens shot them.

[mossy canopy]

ufo time

[dusty jewel]

oh its on

hehe

[thorny frost]

Lucian Freud manga

[waxen cliff]

technically if you threw a rock really fast to someone it become UFO for a bit

[waxen cliff]

cuz UFO aren't necessity space craft

they're all unidentified flying object

technically underwear can be UFO if they didn't find out

[upbeat summit]

awesome images 🙂

[shy kelp]

nmkd sd gui have you tried it?

[strong field]

No, sorry not familiar

[mossy canopy]

once long time ago..

[upbeat summit]

if it's in the sky and I can't identify it it's an UFO or UAP. if it's in the water it's an USO

[shy kelp]

i just wanna know does it work well with sdxl

i haven't install it yet

[waxen cliff]

what's USO? United Service Organizations?

👀

prob not right

[thorny frost]

Coffee spot Dali style

[upbeat summit]

Unidentified submerged object

[waxen cliff]

also here's my attempt at saucer

[strong field]

interesting interpretation

[mossy canopy]

i am going for as real looking as possible

[waxen cliff]

if you want realistic make ||chinese spy balloon||

🤣

[native knot]

Yay for causal racism, amirite guys?

[waxen cliff]

wasn't rly being racist yet

just talking thing that happened

[strong field]

hmm, im almost out of fruit, might need to go get some more

[strong field]

if it makes me scared then its a great concept

[mossy canopy]

sure it's an ufo

[waxen cliff]

nice barrier it got there

[thorny frost]

most senseless thing ever

[strong field]

im scaring myself

because whatever that thing is i dont want to see it again

[strong field]

i um....yikes

[native knot]

supertornadoes

[strong field]

heres a bigger version, looks pretty dark

[native knot]

Legit those can exist.

[strong field]

Ive always wanted to see one, and then i got some sense, and now never want to see one

[strong field]

^

[thorny frost]

A curious facts about UFOs is now cameras are much better than in the 60s but UFO photographs are equally crappy

[dusty jewel]

mines coming up

[thorny frost]

i would like some HD UFO pics

[thorny frost]

how have you done it, with revision?

[dusty jewel]

have you seen that @EyesOnCinema channel that just popped up on youtube?

it's got all these reports, interviews and witness testimonies, as if someone from the u.s. government is behind it.

it's like one of the most remarkable channels on youtube.

let me emphasize credible. this ain't no history channel BS.

[strong field]

Revision, IPA Clip Vision Only, IPA + Clip Vision, IPA, Revision + IPA

[thorny frost]

Will investigate a bit. The only IPA i know is the beer

[strong field]

Pick your favorite of those pictures and focus on that, this workflow is just to demonstrate the possibilities

[mossy canopy]

the end..

[uncut fiber]

b&w ufo reminds me plan9 🙂

[mossy canopy]

whats that?

[uncut fiber]

Movie, classic made by one weird moviemaker. Also he was bit naive 😄

https://www.imdb.com/title/tt0052077/

[mossy canopy]

lmao.. that trailer cracked me up ..

[mossy canopy]

true classic indeed

[thorny frost]

The man has turned into a legend

check Tim Burton movie about him: https://www.imdb.com/title/tt0109707/

[mossy canopy]

UFOs are UAP now!! https://www.cbsnews.com/news/what-are-uaps-unexplained-aerial-phenomenon-ufos-new-name/

[dusty jewel]

that's nothing yo

check out the @EyesOnCinema im not kidding

aka. @EyesOnUFOs

[peak dove]

Some more using the flm 1960's LoRA (inter alia)!

[dusty jewel]

man, all these cases

like the ariel school encounter in zimbabwe, some 30-50 children saw it

[uncut fiber]

colors spared by generating of AI images

[dusty jewel]

it landed, and people came out

every child had the exact same story, its impossble they could have all lied

majority white children, few blacks

[dusty jewel]

i'm trying to describe that ufo that was seen outside the coast of san diego i think?

[dusty jewel]

like everybody thinks that the more advanced a civilization is, the more details are on their craft

it's by our observation the exact opposite

"a single transparent bubble sphere enclosing one solid dark grey geometric cube with its edges touching the inside of the sphere, in the sky outside california, high angle, few clouds, ufo, photograph, ultra realistic, otherworldly, void of details"

[dusty jewel]

almost! nice bg btw

[crisp owl]

just noticing some checkpoints I've been using don't have the fixed vae baked in, so they have that vae issue still in the images

[crisp owl]

open eye image messed up vae
closed eyes fixed vae

Another example

[nimble heart]

use the 0.9 vae

[crisp owl]

Yeah, I'd just have expected finetunes to know that and have that baked in

but htey clearly don't

[nimble heart]

I gave up on using baked in vaes a long time ago

always explicitly specify

even on sd 1.5 there's only like two people use

[crisp owl]

Yeah gonna for suure do that now.
Just thought I'd bring it up in here in case others don't know

[crisp owl]

What checkpoint?
I haven't gone through all mine yet, but checked Zavychroma and sdvn7RealArtxl

Protovision looks ok

[mossy canopy]

nightvisionxl

[native knot]

Nothing going on here...

[crisp owl]

ah, seems they used the right vae then, or you have the right one manually haha

[mossy canopy]

i am using external vae..

[dusty jewel]

agnes kittelsen 😁

[uncut fiber]

nightvision got baked vae i think, depends probably on version

[crisp owl]

Gotcha, then yeah that'll mitigate the issue.
Again, I just had expected the finetuners to know to not use the broken vae

[native knot]

Just a normal group of people sitting around having conversations in a bedroom.

[uncut fiber]

dim light

[mossy canopy]

whats up with the faces?!

[native knot]

Why, whatever do you mean? Those are totally just hu-man faces.
👽

[mossy canopy]

tf

[uncut fiber]

cute animals on end 😄

[mossy canopy]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/sdxl_vae.safetensors use this

[native knot]

I don't know about the rest of you, but I'm totally into hanging out, fully clothed, just standing around in someone's bedroom in awkward groupings.

[mossy canopy]

sure... me too

[crisp owl]

I have it for sure, just didn't think I'd have to specify using it after the finetunes started coming out

[mossy canopy]

sdxl baked vaes never gave me any issues.. may be u r doing something wrong

[crisp owl]

There's nothing to do wrong.
use baked in vae, get vae issue. use manually selected vae, get no vae issue.
haven't checked many other checkpoints, but again, Zavy has the issue, sdvn7Realart has the issue.
Not sure beyond those two yet

[mossy canopy]

well.. comfyui?

[crisp owl]

yah comfy

[mossy canopy]

image res?

[crisp owl]

832x1216
1344x768

[mossy canopy]

412x 609

[crisp owl]

crop

[mossy canopy]

i always do 2k or 4k . so i dont usually notice those small artifacts

[crisp owl]

Yeah I upscale 2x usually.
I hadn't noticed the issue until just now because I was messing with face detailer and actually zooming in on my images to check how it stitched. Then started to notice the pattern with the base images.

These are the two images for these eyes

so at a glance, not gonna notice at all

but the issue is there. Most won't care because it's hardly noticeable from general viewing of the images. But when it comes to being able to fix it vs not fixing it, I'll always opt to fix it

[mossy canopy]

nice

[upbeat summit]

this also looks like the same artifacts when using the broken VAE of SDXL 1.0 and not the working 0.9 one

[crisp owl]

Yup that's exactly it

[mossy canopy]

sdxl original vae

[crisp owl]

Original 1.0 vae

The generally recognized vae issue they released the 0.9 vae to resolve

0.9 fixed vae

hmm
anyone know if there's a node similar to primitive which would loop back through the list once the incremental counting hits the end of the list?
Primitive counts up, then stays at the end of the list

[hardy cipher]

decided to try out adobe firefly just to remove something from an image, and it plastered it with a huge bright red adobe logo. pretty top notch service there

[crisp owl]

May look at the primitive node code and see if I can add that loop in somehow

[upbeat summit]

there's a WAS node where you can restrict a range. not sure though if it loops

[native knot]

"Oh, you want this removed? No problem, here you go!"

ADOBE™️

[crisp owl]

hmm, would it stick to the value it's connected to? Like in this instance connecting the primitive to a boolean_number, I just simply want it to go from 0,1,0,1,0,1,0,1 each time I hit queue prompt

[hardy cipher]

I find it somewhat ironic that i'll now be using adobe photoshop to remove their watermark

[native knot]

I've used SD 1.5 inpainting to remove watermarks on images before.

Worked great.

[upbeat summit]

there're over 100 nodes in the WAS package - so maybe, but you could definitely do it with the EVAL nodes of Efficiency Nodes (https://github.com/LucianoCirino/efficiency-nodes-comfyui) because you can just use SimpleEval python expressions.

so this could work as a workaround for what you want to do

[native knot]

Inpainting, depending upon what you're trying to remove, can be really good at getting rid of watermarks. I had a tricky gradient area that I had to clean up and it was able to navigate that very well.

[hardy cipher]

yeah, I will probably do that in the future. but I haven't inpainted much with comfy. not saying it's complex, but also wanted to check out firefly

[crisp owl]

🤔 Hmm didn't think about that, I'll fiddle with this, thanks for the extra brain haha

[hardy cipher]

tbf it did do a good job of removing what I wanted, lol

but didn't realize they'd add something else

[native knot]

I didn't end up using comfy for that job...just was a fast job in A1^4

[hardy cipher]

I mean, it'll be super easy to fix. but just wanted to remove the logo from the gray scale image of the shirt so I could put something else on it

[upbeat summit]

or check the Number Operation or Number Counter (which has a reset_bool pin) node in the WAS package

[hardy cipher]

they basically just decided to give me a couple minutes of extra busy work

[mossy canopy]

just use photoshop beta

[native knot]

I'd have been tempted to pad a ton of white space around the image so that their watermark wouldn't overlap the shirt

[hardy cipher]

yes, that's what I'm doing. but does photoshop beta have the same tool in it? if so I haven't been able to find it

[crisp owl]

ooh that may work also. Gives me a few things to test out here
Appreciate it

[hardy cipher]

well I have the original image so I'll paste the new one and then erase the logo. it's just dumb and pointless

I just joined their discord to raise a ruckus, lol. well not that much, but I don't really think it's a reasonable approach. put a watermark on if you want, but could be a bit more subtle

[mossy canopy]

its there.. no way to miss it.. just select area type prompt or leave it blank thats it..

[hardy cipher]

huh? I've never seen an option to type a prompt into anything but then I've never really tried I guess

okay, crisis averted

[mossy canopy]

how come ideogram.ai is so good with text?!

[native knot]

Is that good?

[mossy canopy]

with text

[native knot]

Except it's not adjusting the letters very well for the way the fabric is sitting.

[hardy cipher]

so what I'm doing is removing one logo, then adding another one. and using that image I posted as a displacement map

[vale eagle]

ideogram.ai is exceptionally good at text

[hardy cipher]

told my friend I'd help him out with his silly little t-shirt business dream

[strange mist]

phenomenal with text 😄
struggles with hand

[primal vault]

The Searge-SDXL v4.0 workflow and custom nodes for ComfyUI are now released on CivitAI and the projects Github page. There is now an install script that should make it easier than before to update to the latest version and get everything set up correctly.

[shy kelp]

does anyone know whats wrong with this?

[native knot]

There's nothing "wrong" with it.

[hardy cipher]

need to add that folder to paths in system environment settings

[native knot]

It's literally telling you what to do.

[hardy cipher]

type "env" into search

you'll see system properties or something. click that, go to advanced

[native knot]

The path will be something like:
C:\Users\userrname\AppData\Local\Programs\Python\Python311\Scripts
But whatever version you're running.

[hardy cipher]

and there will be an "environment variables" option

[native knot]

Actually...your screenshot shows your path to add.

[hardy cipher]

indeed

ugh

maybe need to update or something

[uncut fiber]

i got those errors as well, as well i was puzzled.

[hardy cipher]

https://tenor.com/view/police-dog-dog-doggo-cute-adorable-gif-17942786

[upbeat summit]

1B-parameter GigaGAN

two minute papers video with some examples: https://www.youtube.com/watch?v=UyoXmHS-KGc
more examples:
https://mingukkang.github.io/GigaGAN/
implementation:
https://github.com/lucidrains/gigagan-pytorch

[mortal pollen]

Hi, if I was wanting to do an animation using sdxl what would be the best way?
I’ve used an animation script with automation 1111 before, and messed with deforum, but that’s not available for sdxl I think

[vital ermine]

Yep, that is the future not this diffusion based stuff I firmly believe. Speed alone will win out.

[tender timber]

I didn’t see it here, so here you go: InvokeAI has been spaghettified (aka now has working nodes)

https://github.com/invoke-ai/InvokeAI/releases/tag/v3.1.0

[sand pier]

Just toying a bit around with ipadapter for 1.5 and sdxl, and the sdxl version seems to want to significantly want to copy the style of the drawing, and way to prevent that?

[pure crystal]

90s comedy sitcom cover

[strong field]

Reduce the strength to below .3

You can also use the clip vision > unclip conditioning to add weight to your prompt

[thorny frost]

do you have a workflow with the ip adapter nodes?

[sand pier]

Headed off to bed now, I'll try it tomorrow thanks

I fixed it partly with sdxl text styles

At least it looked better

[strong field]

Not at desk but I can send something when I get back to it, what are you looking for. I have allot of workflows 😉

[thorny frost]

ty. One general wotkflow with ipa nodes to learn to use it

rubensian beauty

[strong field]

SDXL?

[strong field]

Have Fun

Play with the weights a bunch to get a feel for it

[ionic gulch]

Major update of my workflow with many improvements (new layout of the menu, use image size or size from settings, wildcard prompts, style options, resizing of images to recommended SDXL resolutions, own nodes to reduce amount of nodes needed for some calculations, etc.)
https://user-images.githubusercontent.com/142158778/264802323-99d60e13-d150-4b04-af0f-27c0b80189fa.png
https://user-images.githubusercontent.com/142158778/264802453-19d867a4-6fb9-423b-915f-121a2484733c.png
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[strong field]

Looking forward to checking this one out

[steel mist]

There’s no model here though, right? Like—this is great if you’re interested in training your own 1B parameter model over the course of a month on cloud A100s but it’s practically useless to non researchers?

[thorny frost]

Yep

[ionic gulch]

@strong field the styles feature is great. allows to create much better images in many different styles with simple prompt by using SAI styles and some others.

[thorny frost]

Ty! Not in my PC anymore. Will try it tomorrow!

[crisp owl]

Did some more testing of various checkpoints with the checkpoints included vae (I of course don't have all the checkpoints). Noted which ones contain the 1.0 vae issue. Mitigate by using a vae loader for the 0.9 vae instead manually.

altxl_V50 GOOD
bluepencilxl_V021 GOOD
cherrypickerxl_v20 GOOD
copaxvividxl_v10 GOOD
copaxvividxlv2 GOOD
counterfeitxl BAD
crystalclearxl_ccxl Strangely just a little, but considered bad
dreamshaperxl10_alpha2xl10 GOOD
duchaitenaiartsdxl_v09 GOOD
dynavisionxlallinonestyalized_beta0411bakedvae GOOD
juggernautxl_version1 GOOD
nightvisionxlphotorealisticportrait_beta0702bakedvae GOOD
protovisionxlhighfidelity3d_beta0520bakedvae GOOD
pyrosnsfwsdxl_v012e6 GOOD
sdvn7realartxl_beta2 BAD
zavychromaxl_b2 BAD

[strong field]

Excellent info here

[native knot]

Just in case anyone here was waiting until they got one...

[hardy cipher]

you know a bit about the IPAdaptor stuff. I was wondering about a couple things

I've had a few outputs that were just black images. but seems like restricting image size might fix that? or am I imagining things?

also haven't really figured out what the mask input is for. I guess I coudl send a mask to it and see

[hardy cipher]

Steven Z'Dar Seagal

alright, the IPAdaptor is pretty cool. should've got this earlier

Taylor Z'Dar Swift

[native knot]

The sexy nobody knew they needed.

🤣

[pure crystal]

I can't stop staring

[fresh path]

malfoy

[glad grove]

hes fightin cancer

[hardy cipher]

he fights for our well being

whoa

[pure crystal]

one day

[ionic gulch]

]]a]]

[keen nimbus]

https://www.youtube.com/watch?v=DQJPDW6domk

full game

[glad grove]

real picture

[shy kelp]

what is this actually, i dont even move my models

[hardy cipher]

I keep getting that. I think it's an error error. Are you running more than one instance of comfy?

[shy kelp]

nope

[vale eagle]

How could I save an image in custom path outside comfyui folder? Just enter the full path?

[heady vale]

edit the bat file and add this, with your folder
--output-directory D:\SD_Output

[hardy cipher]

Could you add it to this? https://github.com/comfyanonymous/ComfyUI/blob/master/extra_model_paths.yaml.example

[vale eagle]

Thanks for your help. But I want to change the dir in process

[heady vale]

Im not sure you can

[vast galleon]

I've spent some time upgrading the color transfer method even further. It now runs on GPU for batch processing (animation ready) and works by extracting, mapping, and swapping color palettes. AND it can be used with masks. The code is tested and working nicely. Now I just need to get it into ComfyUI.

Hoping to have it out this weekend.

[indigo carbon]

=]
https://civitai.com/models/138381?modelVersionId=152897

published, as requested.

[vale eagle]

appreciated

[hoary saddle]

getting this today on AIT

any thoughts?

updated

[indigo carbon]

batch 4 isn't on linux yet

[hoary saddle]

ah

[indigo carbon]

change to 1

@hoary saddle if you want to contribute; you can compile the modules for linux yourself, people will appreciate that

the reason I decided to upload it, is because I'm confident with it now; it can blend images very nicely

the craziest part about this is; I managed to make this workflow in a way that it still uses AIT while achieving img+img very nicely

[heady vale]

Awesome work. I am having better results not using the base XL checkpioint for some reason. Finetune models seem blend easier

[high skiff]

@visual glade Is there a way to change the output image folder by chance? I need to keep some images from different workflows separate

[indigo carbon]

I'm using Copax_colorful, but that also seems to work!

[vale eagle]

Just asked this question. LUL

[high skiff]

oh shit, really? lmao

[hardy cipher]

this is my workflow

[vale eagle]

#✨｜sdxl message

[high skiff]

no shit lmao

thanks lol

[crisp owl]

test/ComfyUI in the save image node will create a new folder named "test"

[vale eagle]

You could set output-directory to a root folder. The save image function able to create subfolder

[high skiff]

ok, so I need a second bat file, gotcha

perfect

[high skiff]

dope!

I remember doing that once, but I need these images on a different drive to save space

[crisp owl]

ah

[vale eagle]

You could use symbolic link (mklink command) to link different driver to a folder

[high skiff]

I need to windirstat my PC and delete some old big files off my drives

[indigo carbon]

you can also weight them accordingly to output if you want it to be more imageA than imageB

[high skiff]

need a lot of E in D

[heady vale]

not sure the weights were the problem as the blended image was totally different to both input images

[indigo carbon]

on base right?

[heady vale]

yeah

its fine I just wont use base for that workflow

[indigo carbon]

well, the finetunings are definately better than base anyways so..

[high skiff]

lotta files to scrub

[vital ermine]

getting that with 3 and 4 on Windows

[high skiff]

I need to get a big HDD to offload a lot of my cache files

[heady vale]

I did that yesterday and reclaimed 50gb

[hardy cipher]

lol. I feel your pain

[heady vale]

so much cache garbage sitting around

[indigo carbon]

you don't have the latest version of the node then. it should do BS1,2,4 just fine if you're on windows

[hardy cipher]

google drive really takes a lot of liberties

[high skiff]

I have like 300GB of caches

[hardy cipher]

just deleted all of those

pip likes to cache hard as well

[high skiff]

this one is a 130GB cache of my moms old drive

[vital ermine]

Well, cool to know cause I am not chasing updates daily. Get the update tomorrow if there is one. How do I even know what version I am using?

[high skiff]

oh no, I mean I have caches for files I save

[hardy cipher]

ahh, I have an external drive for that stuff. need a bigger ssd though

[high skiff]

yeah, my old SD install is the issue

280GB

[indigo carbon]

nodes get updates constantly, I think there is a button on manager to update all nodes

[vital ermine]

Would be nice to know what version it is though

I got burned a lot chasing updates so I am once bitten twice shy with this stuff

[hardy cipher]

yeah, mine is 500. I'll clear up 150 gb then a few days later I'll have 100+ gb of trash cache garbage

[vital ermine]

I told it to fetch updates and all it said was there is a node that needs to be updated but didn't do it for me.

[high skiff]

I have 2x 1TB NVMe Drives, 1x2TB NVMe drive, and a 512GB SATA SSD

I wanna get like a 4TB HDD

[vital ermine]

didn't even tell me which one either, lol

[hardy cipher]

I have a 5 tb hdd. I dump any and all things on it and still have a couple tb of space left

[vital ermine]

yep, a lot of ones show update but funny how fetch updates doesn't actually fetch the updates

[high skiff]

holy shit, I found my culprit

my old SD install

just deleted 31GB of just old VENV's

[hardy cipher]

lol

I just need to get a larger ssd

[vital ermine]

Rookie league

I had 1 venv of about 15GB. Not sure why that thing was so huge but buh bye to it

[hardy cipher]

my oobabooga venv hit 95 gb somehow and didn't even work so I zapped that one

wait

that was wsl

lol

[vital ermine]

LOL, that one I did keep but it gets HUGE

[hardy cipher]

wrong letters

[vital ermine]

ooba does work though I just barely use it

[hardy cipher]

yeah, 95 gb would be a beast

I have it installed but haven't messed with it lately

gotta use deepspeed with it. not sure exactly how deepspeed works but sped mine up substantially

[vital ermine]

deepspeed I could not get to work

[indigo carbon]

it's only for LLMs and it's only for Linux

[hardy cipher]

were you running it in windows?

[vital ermine]

you would think Microsoft, and windows it would just work

[hardy cipher]

yeah, wouldn't work for me except with linux

[vital ermine]

I don't think I have tried it on Linux but if I did it didn't work either.

it was for kohya_ss

[hardy cipher]

had no issues with it in linux

[vital ermine]

come to find out for training a LoRA it can't be done. well, shit.

[hardy cipher]

lol

I want to know how to train llm loras. I haven't found a solid source that explains it

[vital ermine]

I know training llm is being done but not sure if lora or not

[hardy cipher]

they're a thing, but also really restrictive at the moment. you need to use them with the same exact model architecture they were trained on

[vital ermine]

Since I haven't seen anything with loading a lora llm yet not sure

[hardy cipher]

and there are a thousand different architecture. open source llms are like weed strains

[vital ermine]

yes

[hardy cipher]

they're named like them, there are countless crossbreeds

[vital ermine]

well, loras for SD is supposed to be like that and if it works on any other model great or it may do weird shit or not work at all. Probably the same in llm world

[high skiff]

YOOOOO

[hardy cipher]

wizard vicuna blue dolphin kush

[high skiff]

I JUST FOUND SOMETHING ON ACCIDENT

I AM SO HAPPY

[hardy cipher]

go on

[high skiff]

I found all of my 3D printer profiles backed up to a little folder I forgot I made!

I have been struggling so hard to replicate all of these profiles!

These took like 2 years to finetune

[vital ermine]

Nice find

[high skiff]

I am so happy

[hardy cipher]

excellent. put them on the cloud!

[vital ermine]

Once a 3d printer is finetuned those are gold

[high skiff]

will do! the file is only like 20kb

uploaded to drive now

[vital ermine]

SOB, updating broke AIT

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

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

Prompt executed in 7.50 seconds

Yep, I removed the ait nodes and it works

what was that about updating again?

[heady vale]

you have some funky things happening regularly on your PC. I update all the time and never have errors

might be worth clean wiping your windows/linux and starting fresh

[indigo carbon]

The newest ComfyUI commit is incompatible, merge with an older one.

[upbeat summit]

looks like General Awareness isn't the only one with AIT problems after the latest update

[indigo carbon]

Idk why Comfy keeps breaking AIT and requires us to change the code after every few commits

[vital ermine]

Wait, the update I did for comfy right before it broke it? Ugh. Thanks

[indigo carbon]

Yes. It almost feels intentional LOL. Idk why would Comfy make the UI incompatible with AIT after every few commits..

Might make a fork that is compatible

[vital ermine]

Sounds a bit like what Auto did with his program that pissed off all the extension devs.

[indigo carbon]

If it ain't broken, don't break it..

[vital ermine]

LOL

[indigo carbon]

That's a moto that people like Comfy should consider before making updates that break important nodes almost intentionally(no offense, this is just how it feels..)

I might make a fork of a compatible merged branch or something, because this is just ridiculous at this point unless Comfy will fix compatibility

[hollow halo]

Why is SDXL so bad? It trains okay but leans too heavily towards realism

[vital ermine]

Auto is bad about that too but they gotta break stuff

[hollow halo]

Even perfect trains 3/10 times turn to realism if its another style

thats 30 percent

[vital ermine]

I know with Auto the extension devs said it was because their is no back channel for them to test and stuff.

[hollow halo]

GA you got an answer? You have a great style that keeps turning to realism so I know you get what I'm saying

[vital ermine]

Mine did as well. Change seed and it pops into realism but I never trained on realistic art.

[indigo carbon]

Anyways @vital ermine the fix for that issue is to merge with a branch from a few days ago. I hope Comfy won't do this again

[hollow halo]

This channel loves to ignore questions they dont have an answer too 🤷‍♂️

[vital ermine]

Do you have a commit? do a hard reset to it?

[indigo carbon]

No, just merge with the earliest commit from August 28th, that's the one I use

[vital ermine]

True, very true. I wish I had an answer as it bit me in the ass but I suppose I just need to erase my drive and reinstall Windows to fix that.

[hollow halo]

Lol fuck reinstalling windows, thats not a solution. If you have to reinstall windows, someone messed up their python code

[vital ermine]

Well, that is what a venv is supposed to be for so you will never need to zap the os

[hollow halo]

If it leaks out of the venv someones code is bad

[vital ermine]

Yep, and there are a lot of bad devs out there.

[hollow halo]

Most of these devs don't even answer github issues so but Comfy wont even acknowledge features are needed, its all about the "back-end" and apparently this front end they made "to learn" doesn't matter anymore

[heady vale]

In this case it was comfy, but you have had so many things break it has to be more at issue.

[vital ermine]

Dude, for real? Okay, if you say so.

[hollow halo]

Bruh has brand new hardware, a fresh hard drive, and you're going to say its his fault when all he installs is SD related things?

half of them being made by comfy or Kohya et al

[vital ermine]

He has to be trolling or just pulling my leg for fun. No worries.

[heady vale]

I never blamed anyone for fault.

[visual glade]

https://github.com/FizzleDorf/AIT/pull/44

[willow bane]

can i use custom image at the end of deforum ?

under animation

[hollow halo]

Well you did say he should reinstall windows, which implies that he fucked up somewhere

[visual glade]

the reason this extension breaks is because it overrides a bunch of comfyui code

but that one was my fault

[heady vale]

if you see it that way. whatever

[hollow halo]

How did you mean it? I'd love to be wrong

[heady vale]

software isnt perfect. it doesnt mean he was at fault

[indigo carbon]

Ah I see. So for now we just stay on older commits?

[hollow halo]

No but somewhere, along the way, someone did something wrong

Software doesnt break on its own once working

[willow bane]

hey man plz help me a bit i have a question

can i use custom image at the end of deforum ?
under animation

[vital ermine]

Fact is venv's very existence is why what he said was full of it. It was made to protect the system so infection of bad code should not be able to escape into the main system, and besides if he could read code he would have seen that error was saying it isn't the end usersd fault. Got an int expected a tuple. simple.

[hollow halo]

Yeah int expected tuple is always a code issue

[heady vale]

well I spent 2-3 hours not long ago helping him sort out software conflicts.. so yeah

[azure oxide]

pretty sure comfy himself mentioned this a long time ago
#✨｜sdxl message

[hollow halo]

I help him daily, though he is very capable on his own, your point?

[vital ermine]

If iA3 breaks comfy code and is not implemented how did AIT code that breaks a good chunk of comfy code in this?

[hollow halo]

Why is AIT supported if iA3 isn't. Thats my question. AIT breaks more than adding an extra network type does

Not being Auto1111 is not an answer

[visual glade]

AIT isn't properly integrated yet

once it's integrated in master it won't break

[hollow halo]

But why did you even attempt to if iA3 isn't in the works?

AIT is more complicated than letting the unet adjust to the model instead of model adjust to unet

[visual glade]

AIT is way more important than iA3 support

[hollow halo]

But Why?

[visual glade]

because iA3 isn't an actual lora, it's a patch on the network

[hollow halo]

AIT, 50% speed increase is not better than something that trains perfectly with the same settings between styles, concepts, and subjects

iA3 isn't a lora but why does that matter

It's a new network

[vital ermine]

BINGO

[hollow halo]

Which should be supported

Why is LoRA the only supported network?

[vital ermine]

Hmmm, good question.

[visual glade]

lora just require modifying the weights

iA3 require modifying the network like hypernetworks

[hollow halo]

Kohya's official extension even had options for more networks, but never took off on that. I'm just genuinely curious why no one though outside of LoRA?

[vital ermine]

which HN rocked

[hollow halo]

What is SAI paying you?

[indigo carbon]

Are you referring to the ComfyUI repo, or the AIT repo?

[visual glade]

the ComfyUI repo

[hollow halo]

To not support anythign besides lora

[indigo carbon]

Huh, you implemented it yet?

[hollow halo]

I know SAI doesn't want anything besides LoRA because SDXL is broken, and has a lot of issues, but someone needs to admit that

[visual glade]

yes I just want to try reducing memory usage a bit before publishing my version

[hollow halo]

THe refiner is the main sayer of all that

[indigo carbon]

What about speed? Is it as fast as the node?

[visual glade]

yeah it's the same speed

I just want to try reducing memory usage a bit so that it can fit well on 8GB GPUs

[hollow halo]

Remember when LoRA was the start of "Additional Networks"? Guess that is gone

[vale eagle]

Nice to hear. Does it mean AIT coming soon with comfyui latest update when released?

[hollow halo]

We just have LoRA, DB, and FT, which is just a new model

Comfy, do you even know calculus?

[crisp owl]

cmon dude

[hollow halo]

You don't know this shit, and just learned the FastAI implementation of CNN's, admit it

[crisp owl]

why the salt

[willow bane]

hlo

help me plz

my deforum tab is not appearing

[hollow halo]

Because he won't acknowledge the ACTUAL questions being asked

[indigo carbon]

Will it be shipped with compile scripts for people that want to compile on their own?

[visual glade]

of course, there will be an easy compile script with instructions for windows

and it will let you compile every type of model that comfyui supports

[indigo carbon]

I see. What about CLIP_vision and ESRGAN?

[hollow halo]

but not iA3, the best LoRA equivalent currently available

[visual glade]

it already does but like I said I want to try to improve some things first

oh the first release is going to be for the unet only

[crisp owl]

I wouldn't either if someone was trying to talk to me that way 🤷🏻‍♂️

[visual glade]

it will use the pytorch implementations for the clip and VAE

[indigo carbon]

The custom node has CNET and CLIP vision, is it an issue with the compilation scripts or you just didn't compile the modules?

[hollow halo]

I'm asking in a factual manner, any REAL dev would aknowledge the questions but comfy wont because he is an SAI shill

[crisp owl]

lol blocks me

[visual glade]

I'm just focusing completely on the unet because that's the most important part

and the part in the custom node that breaks every time I change something

[hollow halo]

You;ve never provided anything of substance to this channel, I was the first one to share my sdxl training settings and no one was grateful, so yeah, lol

[crisp owl]

ok 👍🏻

[glad grove]

https://tenor.com/view/jetstream-sam-metal-gear-rising-gif-25742644

[indigo carbon]

Yeah, I know. But will it be usable with CLIP vision and CNET? Or will it break when used with stuff other than K_sampler

[hollow halo]

It will break with anything other than ksampler because comfy is copy pasting code form the huggingface diffusers section

[visual glade]

there's zero code from diffusers in my codebase

[indigo carbon]

Jesus, leave Comfy alone. He is doing a great job and you're being rude for no reason. I advise the mods to step in at this point.

[hollow halo]

He doesn;t answer actual questions about his software, just the same rehashed questions from reddit and github, so no

[visual glade]

yeah I'll make it work with controlnets/t2i

[willow bane]

hey anyone knows stable diffusion deforum

i want to ask something about that

[indigo carbon]

Will it be in the form of a node like the custom one is or something you select in K_sampler?

[willow bane]

i want to add custom image at end of the animation

[visual glade]

it's going to be completely transparent, that means you set it up and the regular sampler node will use it automatically

[hollow halo]

Yo sorry comfy but that was literally a test of the mods, thats sad that no one stepped in, I was being a dick

[glad grove]

https://tenor.com/view/tyler1-rage-gif-18980459

[hollow halo]

You're doing fine

[indigo carbon]

So to make the workflows I made compatible just remove the AIT node?

[visual glade]

yeah that's how it's going to work, just use the regular sampler node

[vital ermine]

Damn, I was afraid for the cat. :/

[visual glade]

if you pick settings that are not supported it will switch to the pytorch implementation

[hollow halo]

Any plans for support for none LoRA networks? Not iA3 but just something new, things are getting stale imo

[vital ermine]

which commit do I do a hard reset to so I can finish making 650 reg images?

[vale eagle]

Could we know when it was AIT on or off?

Other than the generating speed

[visual glade]

yeah I'll make it print something for debugging

[uncut fiber]

cant be AIT by default keep in memory = disabled, as it cause lot of crashes when one forgot?

[hollow halo]

ComfyUI on Ideogram.ai

[vital ermine]

Does clip skip really do anything in XL as 2.x it did nothing due to the different clip being used and XL uses the same clip I think.

I know in 2.1 I tried it and never saw any difference then finally read why that was. Made sense but in 1.5 it did make a difference.

[indigo carbon]

Is there a release date for when that's going to happen?

[peak dove]

Using HumbleMikey's 1960's LoRA

[visual glade]

I don't commit to release dates

[indigo carbon]

I'm asking because I'm unsure if I should provide modules for ESRGAN due to you likely releasing it before