✨｜sdxl

[small tree]

thats neat ig

but not worth doubling generation time for it lol

[worthy orbit]

yes I feel liberated of not putting nothing of that in the negs, really usually run it blank and just use to lead the IA, not for hoping it will get better things

[glad grove]

smallest sd negative

[native knot]

lul

[small tree]

lmao what is this dog shit

[native knot]

God...that neg is filled with a bunch of shit that doesn't actually make sense. There are some good words in there, though.

[wet raven]

i turned off --no-half-vae but it still isn't working.
I use
set COMMANDLINE_ARGS=--api --disable-nan-check --xformers --opt-split-attention --medvram

[small tree]

i just run (worst quality, low quuality:1.4), bad anatomy, bad proportions on everything and its fine

[heady vale]

what do you get if you put all that in the positive? 😮

[worthy orbit]

I'm running it with --medvram --opt-sub-quad-attention --opt-sdp-attention, it works but it is bad (not that bad) at changing general stuff, like it need the controlnet global harmonius thing, I'm hopping for an SD XL inpainting model soon 🙏 or Controlnet

[upbeat summit]

that is dangerous!

[worthy orbit]

I have AMD on Ubuntu

[wet raven]

it's worth trying with your settings

[heady vale]

yeah unlikely to be posted here

[native knot]

Results of my workflow with the following basic settings on that prompt, (Euler a normal, no refiner, baked VAE on the SDXL 1.0 checkpoint, 20 steps, 8 CFG
@ 1024^2):

Oh...and seed was 0. 😉

[glad grove]

this

[heady vale]

lol

[native knot]

I've I bump the settings to what I prefer, (DPM++ 3M SDE GPU exponential, no refiner, baked VAE on the JuggernautXL 3.0 checkpoint, 60 steps, 10 CFG @ 1024^2):

So whatever that neg is...it basically is removing a grayscale frog from your image. 🤣

[wet raven]

no luck yet, i'm playing around with mine and your settings to see

modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[glad grove]

you are using auto1111 thats where problem comes from

[wet raven]

i'm just trying to use SDXL inpainting on a1111

everything else works fine

if i can't get it to work, i'll try it on comfyui

with the workflow file

i may need to reinstall 1.6 anyway, because while it was doing something i canceled and now i always get a bunch of these:
fatal: not a git repository (or any of the parent directories): .git
fatal: not a git repository (or any of the parent directories): .git

UPDATE: As mentioned, I needed to run some git commands in the webui-user.bat one time
#🤝｜tech-support message
which are:

git init
git remote add origin https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
git fetch --all
git reset --hard origin/master
git branch --set-upstream-to=origin/master master
git pull

[glad grove]

Seagal mythos

[wet raven]

does it require a YAML or config.json any other file or just that 1 inpainting model file?

[supple knot]

in Comfyui where do I find " Evaluate Integers
Nodes" ?

[steel mist]

Are you getting an error/red box around a node in a workflow?

[supple knot]

I download this work flow from Civitai ,But it has missing error https://civitai.com/models/123048/sdxl-image2image-automated-workflow

[steel mist]

Ok, you're gonna want to grab the comfyUI manager: https://github.com/ltdrdata/ComfyUI-Manager

[native knot]

[steel mist]

Then once that gets loaded, when you load comfy you'll get this button

If you press manager when the workflow is loaded, you'll get an option to "install missing custom nodes" and it should be listed in there

(you'll need to restart comfy each time you add a custom_node, as well)

[supple knot]

I got"the Manager" but missing this node,not sure what package its from

[steel mist]

hrm nothing listed under missing nodes?

[supple knot]

it did 2 but not this one

[steel mist]

so it probaby grabbed two of these, see if you can find the third

WAS, Comfyroll, Efficiency Nodes

my guess is its WAS

[supple knot]

ok thanks

[west breach]

wikihow to not care about anything 😆

[native knot]

Looks like Evaluate Integers is in Efficiency Nodes.

https://github.com/LucianoCirino/efficiency-nodes-comfyui

[slow marsh]

Hi guys, I just deployed SDXL refiner model on Google cloud vertex ai, I know it is a text-2-image model but can I pass in “image” as input with prompt (with the KWARGS) in order to achieve good results like what I got from dream studio?

Looking forward to any possible help here🙏🏻

[glad grove]

these bull frogs are getting wilder

[worthy orbit]

yes I got message like that before the last update, ah it was --no-half

But it was really broken as later it will crash, really buggy.

[worthy orbit]

oh no, I'm inpainting with an SD XL model

[worthy orbit]

yes I was going to tell you to try a reinstall

[wet raven]

i wish they could just have something to check for missing or corrupt/incomplete files (hash sums) because my internet is slow

[worthy orbit]

reinstalling is quite fast, just make sure to put a model on the folder before launching so it doesn't download it

[worthy orbit]

well I hate inpainting so much that I'm just img2img the img at different denoising so I see if it corrects things here and there and then I photobashed in GIMP, don't have to mask unmask and lot of operations and maybe it gives better details in other parts

PS: yes it is kind what inpaint does but it takes so much operations that it seems easier to just use the same prompt and just move the denoising slider

[nimble heart]

it's cause the government keeps putting chemicals in the water

[native knot]

Using the following monstrosity, I've created what I now call...The Randomizer. It's taking a random string from the text file from sandlot here: #🌠｜show-and-tell message
and concatenating my positive prompt with a random terminology modifier applied from QoL Suite V2, and finally concatenating a random prompt from the WAS node suite.

It's given me a couple of beauties already as seen here.

[hardy cipher]

that's awesome

cursed

[native knot]

Well...I've kicked off a batch of 500. Let's see how this all goes when I get up in the morning.

[hardy cipher]

nice

[native knot]

2 more it made

It's still working on queuing up...lol

It's counted up just over 300 so far.

[hardy cipher]

I'm really trying to figure out exactly where things are going awry as far as conditioning and model data are concerned with the ipadapter. is it passing bad data to the sampler or omitting things? seems like it could be demystified with a bit of detective work

well I realize not everyone is having the issues I'm having, but I don't think anything is corrupted. seems like it just gets overloaded, or underloaded, and then the sampler outputs empty latent data

it's interesting to mess with knowing the paper on it came out less than a month ago. crazy how quickly this stuff gets implemented

[native knot]

Ok, this came out amazing...

The combined prompt from all the randomization was:
Part 1: 4 professional asian industrial designers are meeting to develop products, conference table, air purifier model, scratch paper, asian staff, computer, conference, desk, white model, design drawings,
+
Part 2: texture, intricately detailed, theatrical lighting, amazing, highly detailed, sharp focus, trending, 8k, 35mm,
+
Part 3:american romanticism, ,natural arches covered in bushes and moss, sunny day, long shot, matte painting, digital art, by greg rutkowski, in the style of studio ghibli, vivid colors, highly detailed, 8 k, establishing shot, smooth, trending on artstation, illustration, realistic

Checking all this out when I get up is going to be fun.

[hardy cipher]

quality

[native knot]

I really should go to sleep, but I can't keep my eyes off of what it's generating...

[worthy orbit]

what did you use to make in 3 parts?

[native knot]

#✨｜sdxl message

It's just hammering away. I need to go to sleep, so I'm gonna walk away now and let it do its thing. Hoping for a nice little cache of images to look at when I get up. 🙂

[worthy orbit]

Ah but it was all one big prompt? but it differentiated and didn't mix

[obtuse valley]

Suddenly using sdxl models in stable diffusion has started giving me colorful grainy low quality images. the images start off looking correct, then when theyre done loading, they turn 2 a pixelated mess... can anyone help me with this? it used to work fine....

[crisp owl]

VAE

[native knot]

It's all concatenated together, but as you'll see from some of the examples I've posted, it pulled the concepts together pretty well. It doesn't fare well in all cases, but many are surprisingly awesome.

[obtuse valley]

thanks soul can you show me how to add that? i never needed it before thats what i cant understand

sole*

[crisp owl]

These are the models for the VAE if you need it
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/vae

If in A1111, you can choose what VAE it uses in settings

[obtuse valley]

yes im in a1111, but why did it suddenly happen?

[crisp owl]

Some checkpoints have it baked in

[obtuse valley]

i started with the stable diffusion xl base and it used to work fine

[native knot]

You probably either changed VAEs at some point or switched to a checkpoint that doesn't have one baked in

[obtuse valley]

wow their is someone actually named ''soul'' i type that as a mistake when i was speaking 2 'sole''

[native knot]

You probably had the base with baked in VAE

[crisp owl]

haha yeah lol

[native knot]

We were both active when you typed, too

[obtuse valley]

lol interesting

ive been using the same base model of xl since it came out

[native knot]

soul and sole...we are omnipresent

[crisp owl]

mostly lol

[obtuse valley]

ive added new xl models like juggernaught and it was workin fine

[native knot]

some models have it baked, some don't

[crisp owl]

I forget how A1111 does it, but you should be able to specify to always use your manually selected VAE from settings

Should completely mitigate any issues of such in the future

just remember you would have to chnage that again if using a 1.5 model

[obtuse valley]

any links on how to do that sole?

[native knot]

It's a drop-down on your A1111 page

[crisp owl]

just go to settings, click show all settings, then do a page search for VAE and you should find where to input it

[obtuse valley]

i think this happened after it automatically updated to 1.6

does anyone use a1111 here? can someone upload a picture?

[crisp owl]

lemme boot it.
Might take a bit, I'm sure it'll update some things

[obtuse valley]

ok pal

[uncut fiber]

who i am

[glad grove]

obama

[crisp owl]

weird al

[uncut fiber]

Brian May! 🙂

[glad grove]

who are we

[uncut fiber]

seagals 🙂

[crisp owl]

Choose the sdxl vae, not the 1.5 as I have here

[obtuse valley]

thanks sole, i was there looking in that section but i dont see one for sdxl

[peak dove]

Dr Who/Bryan May

[crisp owl]

I linked it above, hugging face

place in here
stable-diffusion-webui\models\VAE

restart

[obtuse valley]

ok thanks im going to try that now

what do i download here?

[crisp owl]

Try the standard one first (bottom one). Click the button right of "LFS"

[crisp owl]

THose are the actual checkpoints with baked in vae

[obtuse valley]

is it not the second to last safetensor?

[nimble heart]

not quite what I expected but I kinda like it

[crisp owl]

VAE and Checkpoint are two completely different things

but the VAE can be baked into a Checkpoint, as these are in your image

[obtuse valley]

so can i put the sdlx vae safetensor in stable diffsion/models/vae?

[hardy cipher]

https://tenor.com/view/hmm-dot-dot-dot-stick-figure-intresting-gif-23376515

[obtuse valley]

sorry this is confusing

[crisp owl]

doens't hurt to try what you have

I personally load in the Diffusion_pytorch_model vae I linked to you.

[heady vale]

I see A1111 now switches from base to refiner with a new section

[obtuse valley]

ok thats answer i was looking 4. so ill put the diffusion pytorch inside of the directory you posted

[peak dove]

Moon Baseball

[worthy orbit]

yes that is pretty impresive.

[peak dove]

If I update A1111 and Vlad A1111, does that new Refiner Section update as well? Or is it an AddOn?

[obtuse valley]

so sole, i dont have to manually switch anything after i do the diffusion pytorch deal?

[crisp owl]

Just in the settings choosing it

and hitting the box to manually use that one

[obtuse valley]

so if i go back to using any of the 1.5 models ill have to go back in settings and change it back?

[crisp owl]

yup

[obtuse valley]

wow i never had to do that before

like i said this suddenly started happening

[crisp owl]

You were probably using all models that had it baked in already

But I know they just did a big update, so perhaps they changed the way the vae gets loaded

[obtuse valley]

the only changes i made was i deleted sdxl demo from stable diffusion and it has been updated to 1.6

ok im going 2 try to generate now...brb

same results....

[crisp owl]

Are you using some extension/model/anything that is meant for 1.5 in conjunction with sdxl?

[obtuse valley]

not that im aware of

[crisp owl]

Can you test with another checkpoint as well?

[hardy cipher]

I'd suggest disabling extensions and see if it's one of those. you can disable and then re-enable in the extensions tab

unless you guys already did that

[crisp owl]

no but I just asked that above haha

[obtuse valley]

sure ill try another checkpoint

[hardy cipher]

he has a 1.5 lora loaded or something

[crisp owl]

yah

[hardy cipher]

don't think a vae would do that

but maybe

[obtuse valley]

will this lora load automatically or something?

[hardy cipher]

who knows. I don't know what all your extensions do

[crisp owl]

someone earlier today was dealing with the same issue and ended up being their vae

[hardy cipher]

ahh

well either way, those extensions can do all sorts of things

and I'm not familiar with a few of them

[uncut fiber]

i think thats exactly what vae do.
Try copy it in model folder, or download model with baked VAE

[crisp owl]

for reference, this guy
#✨｜sdxl message

[obtuse valley]

i was so excited to use control net with the sdxl models but i cant get it to work after the update

[hardy cipher]

buddo, people are trying to help you here

[hardy cipher]

well now I know

https://tenor.com/view/the-more-you-know-gif-19118236

[hasty mortar]

This one I liked so much I've put the comfyui flow online

[obtuse valley]

SDXL FOOOCUS IS WORKING GREAT AND IS MUCH FASTER, i wish that had control net.

[hasty mortar]

Let me know if you want the link, I don't wanna spam here

[obtuse valley]

i switched to juggernaut and still the same deal

[uncut fiber]

try copy that vae into model, stable diffusion folder. Copy, not move

[uncut fiber]

oh, you are using sdxl, try not using refiner, just for test

[obtuse valley]

ive tried that bernix but ill try it again, just a second

to disable the refiner do you just close the tab? there is no on or off tick

[uncut fiber]

you can also try have on top where clip skip is have to choose VAE.
There must be way skip refiner, i am using comfy but used to use A1111

[glad grove]

something on my tooth

[uncut fiber]

there is some off in screen

[crisp owl]

Humor me here and please select these options. (scroll through the list to choose them).
Then show me your top menu after restarting your UI

[obtuse valley]

bernix do you remember which option in settings allows you to put the vae where clip skip is?

[crisp owl]

my comment above

[peak dove]

^..^<

[obtuse valley]

ok sole

[glad grove]

🙌

[crisp owl]

Finally putting th emultiple fingers generation to good use

I'm heading to bed, but I'm still relatively convinced it's a VAE issue you're having. Those steps should resolve that if the correct VAE is loaded in. I'll check through messages in the morning to see what resolution ends up as

[obtuse valley]

thanks so much for your time sole. hopefully see you tomorrow

[obtuse valley]

wow now i cant get regular 1.5 to work anynmore

im using rtx 4090 128 gb ram

[zinc cargo]

why not go to sdxl with your 4090?

[uncut fiber]

it has VAE inside or is it VAE model? What size it has?

[obtuse valley]

sdxl is no longer working on stable diffusion 4 me klinter

bernix this particular model says realistic vision with vae

[uncut fiber]

check its size

[obtuse valley]

its 1.98 gb

[floral island]

restart webui

[uncut fiber]

so it is o.k. but restart UI and browser

yes

[floral island]

it's bugged if you get that message

[obtuse valley]

ill restart the cmd box

[floral island]

lel

[uncut fiber]

in medieval ages classic cure of anything! 🙂

[glad grove]

good as new

[uncut fiber]

Latent space is emerging real life world. Photo-realistic. Colorful.

This prompt should gave something between abstract and real thing?

i like some results

[glad grove]

lookin neat

[uncut fiber]

no colorful it is just white, like ambient oclussion? i like to play with abstract random things

[glad grove]

or try with saturated color,deep color,high contrast

[peak dove]

Cool!

[floral island]

too low resolution

sdxl needs to be 768x768 at least

1024x1024 even better

and change your vae -> it's a 1.5 vae

[obtuse valley]

i forgot about that but i doubt that will fix it

its so weird i never had to worry about 1.5 vae before. can you post the link to download it eface?

[floral island]

"oh shit"

[floral island]

just set vae to automatic

basically anything 1.5 and sdxl are not compatible with each other

[uncut fiber]

Oceans collision?

[floral island]

whole song text with some additions

running through ALL my old prompts

and... that's quite a few

[obtuse valley]

everytime i try to change the vae it goes back 2 the 1.5

[uncut fiber]

probably you have no other vaes there

@obtuse valley have you preview on? Just to see if it generate o.k. and get spoiled in last part

[obtuse valley]

yes it starts out fine and gets pixelated in the last part

[floral island]

that's a vae issue, what happens if you put it on automatic (for reference, it should stay there)

[uncut fiber]

i noticed there is off button, it is for disabling refine?

[obtuse valley]

guys it worked this time!!

i changed it 2 automatic

[floral island]

for sdxl_base you should use the refine

for some of the finetuned models it's not needed anymore

i'm using protovision, and refiner is just making it worse, so it's not needed in that case

[obtuse valley]

im going to try juggernaut xl

juggernaut xl does not have the vae in the model tho

[uncut fiber]

@floral island o.k. iam using nightvision or that something like copax and i think i dont need it as well

[floral island]

yeah, nightvision is also by socalguitarist, he made amazing models

[uncut fiber]

only top architects will survive 😄

[peak dove]

Red Haired a.m. Beauty (I'm back using A1111 as it has integrated a Refiner Stage!)

[obtuse valley]

yes it works

i love you guys thanks so much!!!

[floral island]

i'm looking forward to your generations ❤️

[floral island]

i pity the engineer that has to make this

[obtuse valley]

i notice my 4090 used 2 always run the fans when generating images...now its not

how can i ensure that a1111 is using my videocard?

[floral island]

also, make sure to install fancontrol and set up a custom fan scheme

[uncut fiber]

what is a. m.

[floral island]

all my case fans are activated by gpu activity

[obtuse valley]

this seems new for the 4090

i dose not always run the fans

only when its being used

[peak dove]

a.m. = morning p.m. = afternoon (It's morning here in London UK)

[uncut fiber]

i know this meaning, just weird in that prompt 😄

supposed it has different meaning

[mossy canopy]

why am i getting black images with ip adapter?!

[obtuse valley]

does anyone know anyway to make sure a1111 is using my gpu?

i hope microcenter didnt sell me a garbage 4090

[uncut fiber]

dl for example Afterburner, and when it generate image it should show you high usage of GPU

[floral island]

is it extremely slow?

[obtuse valley]

yes

[floral island]

reinstall venv perhaps?

[obtuse valley]

thats another thing, a1111 has gotten slower

[uncut fiber]

show your webui-user.bat

[obtuse valley]

ok

[floral island]

earthbender fashion

[obtuse valley]

show what part?

[uncut fiber]

all, should be short.

[obtuse valley]

do you mean right click the user bat.?

[uncut fiber]

no, you should start it by that file, webui-user.bat

right click and edit

[obtuse valley]

right

just a moment

[mossy canopy]

@floral island try running this..

[floral island]

you're using comfy?

[mossy canopy]

yes. finally got it working..

[floral island]

shit, i'm on a1111 😦

[obtuse valley]

im not using comfy

i was turned off by the user interface

and all those lines going 2 blocks

[peak dove]

--cors-allow-origins?!?!?! 🙂

[uncut fiber]

what is that link? You need it? I think you need only --xformers

[floral island]

the user interface isn't the part that's the problem: i want my dynamic prompts / wildcards / prompt editing

[obtuse valley]

i dont know what it is

ive been following youtube videos and adding things and forgetting what they do

[uncut fiber]

i think dynamic prompt is as well in Comfui

[floral island]

yeah, there is 🙂 but not the prompt editing afaik

[obtuse valley]

but its only recently been going slower

[floral island]

which can be replicated with creative node usage

[uncut fiber]

this {|} is included in comfui without extension

[obtuse valley]

i thought a 4090 would solve all of that

[floral island]

the [i want to : do other stuff:0.5]

[mossy canopy]

you have custom nodes for that

[floral island]

haven't been able to find one that satisfied me 😦

fire, water, stone, wind 😄

[uncut fiber]

@obtuse valley delete all commands in set commandline arg and keep there only --xformers

[obtuse valley]

bernix are you sure? im kind of afraid 2 do that

[floral island]

you can copy the line and put REM in front

rem make it so the bat file doesn't execute it

my bat file

did a lot of testing, so all the stuff i don't use anymore i put rem in front, so i can keep it for future reference

[obtuse valley]

i cant see the whole thing

[floral island]

for mine, not needed 🙂 just to show where to put rem if you copy your line

so copy and paste your current command to a new line, put REM in front

and then make the other line with --xformers only

so you can keep your old line, if there's something in there that's needed

[obtuse valley]

show me?

[obtuse valley]

in command liine arg, all i need is xformers?

[uncut fiber]

i am telling you keep just --xformers

[floral island]

added colored gemstones for enhanced coloring flavour

[uncut fiber]

for 4090 yes

[floral island]

yeah,that should be enough

[obtuse valley]

oh ic

great let me try

[floral island]

and i think --no-half is causing your slowdown

that prevents it from working on fp16 and only does fp32 (which is 50% of the speed)

[uncut fiber]

now played with latent space and interior architecture

[floral island]

looks a bit blurry tho, are you getting enough steps?

[uncut fiber]

no 😄 can make more. i have only 3070 🙂 going to try 40

[obtuse valley]

like this

[uncut fiber]

also git pull isnt necesery

[obtuse valley]

bernix doess your gpu fans only come on @ a certain time?

[floral island]

30 should be enough, depending on sampler

[uncut fiber]

yes sometime i hear fan

[obtuse valley]

bernix the git pull just updates stable diffusion automatically without me having to do it manuually

thats what i was told

[uncut fiber]

o.k. will try 32 then what sampler you suggest me @floral island

[mossy canopy]

a1111 has ip adapter for sdxl?

[floral island]

no, just prompting 😛

but yeah, there is ip adapter for sdxl

under controlnet

[obtuse valley]

bernix did you review my last post of the edit?

[peak dove]

(--cors-allow-origins) What does it do?

[obtuse valley]

i dont konow torcello

[floral island]

i'm loving the restart sampler, it's pretty heavy and slow, but the output is way better than the other sampler imho

[obtuse valley]

i forgot, i hope its not anything i needed

[mossy canopy]

i haven't been using a1111 ever since sdxl came out... so the lack of knowledge

[floral island]

the 1.6 release added the refiner option, and then models came out that don't need the refiner lol 😛

[obtuse valley]

lol eface

[uncut fiber]

i think those plants can looks weird, but i think it is because latent thing. Will see.

I think refiner option was earlier.

[obtuse valley]

i guess this is what open source entails

[floral island]

@uncut fiber yeah there were extensions, but with 1.6 it was made native

[peak dove]

I am using SDXL on A1111 - and there is NO fan-noise whatsoever!!! ComfyUI was well noisy ...

[floral island]

check your temps with msi-afterburner

[obtuse valley]

i was used to seeing my 4090 fans occasionally coming on when generating images... now theyre not moving

[floral island]

sd(xl) makes your memory run VERY hot, and they are usually without a sensor

[peak dove]

Its a lot better and easier now that the Refiner Stage is included - and has way more extensions than Comfy!

[floral island]

won't be fun if you fry your gpu memory

[obtuse valley]

i check task manager and its using 17gigs of vram

[uncut fiber]

@obtuse valley you can try make more images at once

[obtuse valley]

yes, b4 i was generating 4 images @ a time but since it started going so slow, i switched 2 one image

but bernix i didnt c your reply to my post

is this right?

[mossy canopy]

comfy i get to choose how every step of the way would work.. and i get to use ait which gives me 200% speed boost.. with a1111 i dont get that

[floral island]

seems about right

[peak dove]

I've yet to try ComfyUI+AIT ... mebbe I'll be convinced 🙂

[obtuse valley]

should there b a space after 'args'?

or does it matter?

[floral island]

added light and dark too 😄

@obtuse valley doesn't matter afaik

[uncut fiber]

better?

@obtuse valley it doesnt matter

[floral island]

way cleaner ❤️

i see you are using the vae 1.0

[uncut fiber]

me?

[floral island]

yeah

[obtuse valley]

you guys need 2 try SDXL FOOOCUS

[uncut fiber]

i should use vae in model... Going to switch to nightvision

[floral island]

you're going to be surprised ❤️

[mossy canopy]

fooocus is way too limited for me...

[uncut fiber]

How is it in A1111 with AITemplate? Does it exist for it? Should

[obtuse valley]

all it needs is control net

[mossy canopy]

you can easily replicate foocus incomfyui with comfybox frontend..

[worthy orbit]

I was checking a video but yes also it looked limited, but in resources speed, it makes some difference, i.e. it let you resize more than Auto, or just for generation, etc?

[supple knot]

Try this "faceswap" tutorial for ComfyUI getting this error anyone know what I can do to fix this

[mossy canopy]

what do you mean...

[worthy orbit]

I wondered if this focus ismore optimized than Automatic1111 in some way, I guess it is in a way for 4 GB VRAM but for more VRAM it does a difference in performance?

or let you do bigger images

[mossy canopy]

you can do all this in comfy . with low vram too..

[obtuse valley]

yea my 4090 fans are not coming on, im generating 4 images using sdxl model. i just hear the cpu fans gettin louder tho

[worthy orbit]

I had try Comfy but I found after the last update Automatic1111 peformed better. So not some magic performance thing.

[glad grove]

not better than comfy

[uncut fiber]

This is nightvision. Leaving arch for now 🙂 but what.
@obtuse valley post your cli what is there. Can be so light work for 4090

[mossy canopy]

still a1111 requires more ram just to operate..

[obtuse valley]

cli?

i dont think im gettin the most out of my 4090

yea the fans didnt come on once

[obtuse valley]

cli?

[supple knot]

Did you buy the 4090 used?

[obtuse valley]

hell no

[uncut fiber]

that black window command line interface

[obtuse valley]

i paid 1300

@ least

[worthy orbit]

in my case, I don't know if I had something bad in the settings but I got bumps when generating with Comfy. Automatic was worse at that time as it couldn't handle XL without crashes. But now after the update it is very stable.

Also it was fun at the beginning in Comfy but really too much, even using a workflow.

[obtuse valley]

brand new from MicroCenter

[supple knot]

is it running hot? use a app to check it . like "msi afterburner"

[obtuse valley]

i heard people whispering 'thats a 4090'' as i left the store...

[uncut fiber]

check your apps that comes with gpu, to modify fan curves, @supple knot yes AB is best

[worthy orbit]

well, maybe the bumps was Comfy optimizing the generation but freezing the video, I don't know. I hear stuff like that about Comfy but automatic works so, better in much ways.

[obtuse valley]

what is that fuzzy donut program that they use 2 stress test?

i 4get the name, its on the tip of my tongue

[floral island]

powermax

[shy kelp]

how to completely uninstall/remove comfy ui?

[floral island]

just remove it

XD

[supple knot]

its all in one folder

[shy kelp]

does it have relation on phyton or something?

[uncut fiber]

donut is as well in afterburner

[obtuse valley]

furmark

thats the name

i dont know how to use afterburnerrr

furmark is much simpler

[uncut fiber]

o.k. just afterburner can help you with many things, performing stress test, modify fan curve, simply it is swis knive for GPUs

[supple knot]

theres also GPU Z

[uncut fiber]

you just download kombustor for tests

[obtuse valley]

ok bernix

these fans havent moved 1 time

im gettin worried

[supple knot]

what brand is the 4090?

[mossy canopy]

it creates venv. so no.

[uncut fiber]

it may not reach enough power, or it can be so silent

[obtuse valley]

zotac gamingg

[obtuse valley]

bernix but it came on b4 under much less sdxl generations

[uncut fiber]

do you know what is your psu?
b4 is 4 at once?

[obtuse valley]

agt 1000

yes bernix

4 @ once

[uncut fiber]

should be o.k.
If i do b4 i would be worrying about my gpu

[supple knot]

looks like Zotac hass there own software call firestorm

https://www.zotac.com/us/page/zotac-gaming-firestorm

[obtuse valley]

wow thanks make b

is a ryzen 7 bottlenecking my 4090??

[supple knot]

I think for AI Image Creation it all about the vram on the gpu.

But but if your having issues with the best card my guess would be the heat

[mossy canopy]

its about cuda cores

[obtuse valley]

my god the fan came on now!! after that stress test

[glass snow]

you guys know how to take an image of my own art (photography)and change it with stable diffusion but keep it mostly the same like same person just different places in background. is it possible to change the faces to change expressions?

[icy brook]

Nice!

[glass snow]

those already run hot

[obtuse valley]

they havent turned off tho

its runnin full blast

[glass snow]

im running it with a 1070ti lol but i have 11gb of ram so works

vram

[obtuse valley]

wow a 1070??

[glass snow]

yea man

not firestorm

[floral island]

ait on comfy crashes for me 😦

[obtuse valley]

these fans havent slowed down yet

[obtuse valley]

i turned off furmark

[glass snow]

i would watercool if i was you

[obtuse valley]

i am watercooling

[glass snow]

don't wanna burn up a 2k card

great

[floral island]

fancontrol

that's the tool you need

[icy brook]

You got it working training LoRA’s 🙂

[obtuse valley]

no, my water cooler is on my cpu

[floral island]

mine is too

[vital ermine]

No, I can't get lora, or locon to work I train DB and extract

[floral island]

but with fan control, i can enable that cooler so it starts venting the whole case when the gpu becomes active

[glass snow]

you guys know how to use sdxl to change a face to frown in a photo?

[obtuse valley]

i dont have a case

[vital ermine]

That is my 2.1 attempt come to life for XL I will release today or tomorrow.

[obtuse valley]

its one of those termaltake builds

thermaltake

everything is out n the open

[floral island]

well, if it's cool ^^

[glass snow]

how many gpu's you using to train?

just one?

[obtuse valley]

can a 4090 train by itself?

[glass snow]

24gb can probably train a smaller model but it may take like 15 days at least for like 3d ai gen it took my buddy 21 days to train the models they used to make 3d shoes for metaverse

[glass snow]

he did it with a 3090

[obtuse valley]

growth can you show me how?

i make shoes on secondlife and sansar

[mossy canopy]

gpu?

[floral island]

3090

[obtuse valley]

man if i can make 3d objects with this thing

[glass snow]

bro it was my buddy, i don't have that kind of hardware, we're mainly focused on chatbots not art

[obtuse valley]

i wasted so much time learning zbrush

[uncut fiber]

you have to go back @floral island + disable keep in memory

[floral island]

keep in memory was not the fix

[glass snow]

but yea it is possible to make a 3d gen that focuses on one object

[floral island]

driver perhaps?

[mossy canopy]

workflow?

[glass snow]

and you can productize that

[vital ermine]

I see even Maya is getting AI now

[uncut fiber]

@floral island git reset f92074b --hard

[uncut fiber]

not keep in memory. I have same issue and somebody suggested this and it is working @floral island

[floral island]

the git reset fixed it

[uncut fiber]

not wasted day 😄

[obtuse valley]

my gpu says 30 degrees celsius but its still blowing full power from when i used furmark

how do i make the fans turn back off

[icy brook]

Ok! On which trainer?

[mossy canopy]

whats your it/s using ait?

[vital ermine]

Kohya

[floral island]

~4.9it/s

not totally fair, because i ran at 1280x1024 before, lemme check

[floral island]

damn, having the l prompt back helps tho

[mossy canopy]

tf?!

[floral island]

i mean, that's pretty good, isn't it?

[mossy canopy]

what was before that?

before ait

[uncut fiber]

i without it 2it/s with it 3it/s but only 3070

[floral island]

on a1111 using restart(heavy) sampler: 1it/s

[vital ermine]

non ait using comfyUI

loading new
100%|##################################################################################| 30/30 [00:04<00:00, 7.22it/s]
100%|####################################################################################| 5/5 [00:00<00:00, 7.15it/s]

[floral island]

i'm at 1280x1024 single shot

i mean, if i drop down to 768x768 it should be pretty fast

[vital ermine]

1280x1024 no ait

[floral island]

dang 😮

what gpu u have @vital ermine

[vital ermine]

I am also undervolting limiting power to 80% as it helps

[mossy canopy]

4090?

[vital ermine]

4090 Zotac

[floral island]

yeah, i heard undervolting actually makes it faster?

(more headroom)

[vital ermine]

I will go faster next year when I upgrade to the mobo/ram/cpu

it did, ironically.

[icy brook]

So you got around buying one? 🙂

[vital ermine]

1 month ago as I almost bought the AMD but last second switched. paid off the purchase on wed

that hurt as all that money could have been spent for a new system and have money left over but gpus are far too expensive these days

[floral island]

@vital ermine undervolt using which tool? i wanna try that too

[vital ermine]

I use msi afterburner and tried all that nonsense with curves, etc... and found the simple way the best for me

[floral island]

so it's editing that curve?

do i move it up or down?

[vital ermine]

no way that made it worse

[uncut fiber]

in msi AB it is easy as pie, on other hand, i thought it has not any performance impact

down

[vital ermine]

Best for my card after a month of trying all vids, sites, own experimentation.

[uncut fiber]

just coreclocks move to left, save as preset load preset

[uncut fiber]

what if it reach those power limit? It stops slowdown or turn down PC? @vital ermine

[vital ermine]

never goes above

[uncut fiber]

get kombustor 🙂

[vital ermine]

I hate that one

For me, and I feel this btw, I have two issues. 1 pci gen 3x16 and this card is gen 4 x16. second is my monitor is 60hz

[uncut fiber]

@floral island there is tool to overclock your gpu safely, not sure in 3000 and 4000 in worst case it should told you it is not posible.

[vital ermine]

don't o/c a 4x card

it is worthless as all the overclockers will tell you

[uncut fiber]

@vital ermine i think it is just fraction of performance pcie4

he got 3090 i think and me as well have not OC it.

[vital ermine]

no, x16 gen 4 would be like gen 3 x32 it has twice the data transfer and it does show it too. my card is bottlenecked by the cpu, and mobo, and probably the 60hz in that order.

I am cool with it knowing I will have this card for a very very long time so I can grow into it

I do want to water cool it but the block is almost 200usd and while people in europe may think that is cheap I don't

[floral island]

i'd have to create a custom water loop to get my gpu on water too

would be hella cool tho ❤️

[vital ermine]

I just went aio after sitting on the fence for 20 years. I am hooked

[floral island]

aio for my cpu too

but having the option to have it support my gpu is even better ❤️

[vital ermine]

My case has never been so uncluttered in my life

[floral island]

(gpu temp up -> cpu and case fans start exhausting hot air)

[vital ermine]

yep, this was a test to see if I like it and I do. Next is the open loop/custom loop but omg the price now

[floral island]

next step is putting my pc exhaust air on a direct line outside lol

goddamn is that a lot of heat entering my house

[vital ermine]

that would rock

[floral island]

my house A+ insulated, i haven't had the heating on since februari

[uncut fiber]

@vital ermine dream keeper. Reminds me person from episode Scooby doo Mysteries ltd.

[floral island]

voltage tuning give me about 5% extra it/s

[floral island]

not worth imho

[uncut fiber]

@floral island is undervolting bad for gpu?

[floral island]

nah, gpu's have a boost clock available, so if it notices it isn't thermal throttling, it'll boost higher

or power throttling

[uncut fiber]

o.k.

[floral island]

so if you go too low, it makse your gpu unstable

just reset in that case

[nimble heart]

basically just the difference between having a youtube video on or not lol

I'd still undervolt non-aggressively just to reduce the heat in whatever room the computer is in.

I've got my ryzen @ -30 points for the PBO undervolter and it can turbo one core up to 5.9 GHZ without the fans ramping

[vital ermine]

I leave the one core stuff alone but I do have mine +25 (had +50) and the pbo optimizer at -30

[nimble heart]

my max clock is +200 mhz and the curve optimizer is -30 for all cores

[vital ermine]

on Noctua HSF best I could get for the same temps was 300mhz less than this AIO. I was sold instantly.

[nimble heart]

it just doesn't give the full 5.9 after a few cores cause I capped the thermals to 85

[vital ermine]

I no longer chase the dragon's tail so I am fine at 4475mhz max all cores and 57c temps on a 5600

[nimble heart]

I would try setting the max clock higher but my bios doesn't let me

[floral island]

cool, restart sampler is available on comfy too

^^

[uncut fiber]

going to test it 🙂

what those segments mean? Going switch to not simple

non simple seems better performance bit

[floral island]

that's when it wants to restart and how many extra steps it should move back and add

i think the format is [steps to move back, steps to add, minimum percentage it happens, maximum percentage it happens]

[uncut fiber]

US i didnt spot it

[floral island]

restart vs no restart (less artifacts/better composition imho)

[vital ermine]

lol

[uncut fiber]

vinyl figurine?

[thorny frost]

has somebody managed to install kandisky extension for A1111?

[thorny frost]

or at leat tried

[uncut fiber]

1st is restart 2nd normal?

[floral island]

yeah

a lot less "sd artifacts", imho

things that are in strange places

or just generally weird

[uncut fiber]

you can try send those images in that software saying AI or not, what a difference 🙂

[mossy canopy]

normal

restart

[floral island]

combined with ait, that's a hefty speedup 😮

[floral island]

less artifacts going on

[mossy canopy]

how many steps minimum with restarts

[floral island]

i'm running 40

[mossy canopy]

minimum?

dpmpp 2m karras

[floral island]

not sure which sampler to use with it

what gives best results

[uncut fiber]

there are two schedulers why?

[mossy canopy]

i always go with the fastest sampler

[floral island]

one for the denoising, i think one for the re-introduction of noise

i'm not sure

[uncut fiber]

o.k. now risking PC health 😄

enjoying abstract interiors 🙂

[peak dove]

Scaahlett O'Hara (Gone With The Wind) "Quite frankly my dear, I don't give a d***!" 😄

[peak dove]

Coochy coo!

[glad grove]

bro watched the censored version

[peak dove]

There was a censored version?!?!?!? 😄

[glad grove]

yea they remove the word damn and also remove all scenes where ladies show their ankles

[peak dove]

Some redhead daahlings!

[uncut fiber]

try tizian's red

i want photo 😄

[peak dove]

You got text to work!!!

[uncut fiber]

yes 🙂 just only one

[peak dove]

Some nice blank background people!

[uncut fiber]

have you tried cut them out of background?

[peak dove]

Not yet ... but will use Generative Fill @ Photoshop

[uncut fiber]

oh. i did few there is extension to do it in a1111

[peak dove]

They look quite saleable with those blank backgrounds!

OK, I will search for new A1111 Extensions ...

Not sure what --cors-allow-origins means at all?!

[uncut fiber]

i dont know. To remove background i used. Wait a moment

rembg

[quaint cosmos]

cors is a security feature, basically servers can refuse requests from domain that they don't know

https://en.wikipedia.org/wiki/Cross-origin_resource_sharing

[peak dove]

Good to know - not certain what you can do with it? 🙂

[static lodge]

anyone know if its possible to reprompt an image that only changes position. everything else exactly the same. like raising hand up, move head right/left, close eyes e.g

[uncut fiber]

inpainting probably only chance, but lot of work for sure.

[static lodge]

the problem with inpainting is, lets say i inpaint the current arm, then it will generate a new arm there, but i would need a feature that has like "Old destination" "new destination"

[uncut fiber]

@peak dove Titian hair. More brown than red

[static lodge]

"Remove arm here" , "put arm there"

[peak dove]

Gawjus

[static lodge]

or could i mess arround with contolNet 🤔

[uncut fiber]

not simple solution for sure. For eyes inpaint is best for example

or pix2pix

[static lodge]

whats that

only know img2img

[mossy canopy]

try pfp noise for latent get even more details..

[floral island]

pfp noise?

[uncut fiber]

you basicaly say let her close eyes and it should close her eyes

[floral island]

which node does it come from??

ah never mind, found it

[uncut fiber]

this way?

Thank you!

[vale eagle]

I have an idea about the workflow. Separate Subject and background sampling and combine resampling again

[west breach]

new McDonald's burger, the McSpider

[native knot]

The Randomizer really put together some great stuff overnight.

[indigo carbon]

it's like a jimmy neutron villain

[west breach]

with extra chicken nuggets

[floral island]

flaming hot

[west breach]

I'd like to see Helen outrun this

[floral island]

terminator waifu tried to emulate hair with fire

[idle zenith]

sdxl vae makes image cleaner but also introduces extra rgb pixels or noises.. how do you deal with that?

[floral island]

sdxl1.0 error

get the 0.9 vae

[idle zenith]

0.9 vae with the sdxl1.0?

[floral island]

yeah

that should fix it

or get any of the finetuned models, they usually ship with correct vae, and no need for refiner

[peak dove]

Some curly-haired types

[idle zenith]

i believe the model i used come finetuned [FormulaXL] but i'm not quite sure
using sdxl vae made the image cleaner

[floral island]

i just know the 1.0 vae shipped i with an error

lol

[idle zenith]

can you give me the link for 0.9 vae?

[floral island]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

get the safetensors, i think this one should fix it

[idle zenith]

i only need the vae here or anything else?

[floral island]

only the safetensors file is enough, if the vae is the only thing that needs replacing

[idle zenith]

alright, thanks

[floral island]

police violence in USA escalates even further

omg i fed the team america song again... this shit is awesome!

(the cringe is f'n real)

[floral island]

THUNDERMOUSE

[uncut fiber]

would be nice make collection of texts 🙂

curious if proper stop because it is traffic sign

text is what i want with exception of 3rd almost word, not intentional.

[peak dove]

S H T T

[uncut fiber]

realy unintentional, Shut is source i believe

this looks like AIrt

[peak dove]

Angry Young Men grrrrrrr!

[zinc cargo]

this style is so cool!

[cyan crown]

thanks. It'a a feshly new Lora I've just trained

[peak dove]

Libromancer? 🙂

[uncut fiber]

is this proper English? Or should be deal ends?

[uncut fiber]

remind me some game

fractals it reminds me

[peak dove]

Deal End could be an English Village name! 😄

Or "End by Fire OrDeal!"

[floral island]

we are in the wrong timeline boys

[uncut fiber]

almost close, will collect and post at once one day 🙂

[uncut fiber]

Mass Effect

[peak dove]

That's a tad more lively!

[uncut fiber]

bit of feeling of shawn the sheep or original which i forgot its name 🙂

[cyan crown]

I call this Lora ShaunSDXL

[uncut fiber]

😄

[peak dove]

This is like Wallace and Gromit - Aardvark Animation Studio in the UK - it is way cool! Have you placed the LoRA at Civit at all?

[uncut fiber]

yes original is wallace and gromit. Park

Nick Park

[peak dove]

My bro knows the guy

[cyan crown]

Not placed because I've problem with Lora dimensions (1.7 Gb).

[uncut fiber]

it is almost whole model

[peak dove]

OK, but when that is sorted, I'd really like to d/load!

[cyan crown]

To obtain extremely good results I use 256 Network dim

[floral island]

lets go on a trip bois!

[uncut fiber]

this is for me unknown territory. I never tried train lora or so 🙂

[peak dove]

She looks a tad zany here!

[cyan crown]

it works also with real people

[uncut fiber]

or celebrity deathmatch, if anybody remember

[peak dove]

Less golden ...

[cyan crown]

and obviously with sheeps

[uncut fiber]

how about that brown dog and pigs 😄

[cyan crown]

pigs comes out this way

[peak dove]

Can you do Kayleigh Cuoco in that Shaun LoRA? 😄

[cyan crown]

ok I try

this is the first try

[peak dove]

Or Barbie?! Or Margot Robbie? 😄

[peak dove]

Good try!

[cyan crown]

I can upscale if you need

[uncut fiber]

how do you upscale?

@cyan crown you are using comfyui?

[cyan crown]

comfyui workflow

[uncut fiber]

o.k. i made my own to try upscale video, just for test.

[cyan crown]

Margot

[peak dove]

If you can afford GigaPixel - it is much much faster as an upscaler. I got a free copy as I Beta Test @Topazlabs (lucky me!!!)

[zinc cargo]

will you upload it to civit?

[uncut fiber]

there is good page with models, some instead of resizing restoring image...

[cyan crown]

Yes. I' use it for photography and also for upscale AI images, but I prefer a first 2x upscale with COmfyUI because it produces better results

[uncut fiber]

i am not upscaling and my eyes are not good enough to spot differencies.

https://openmodeldb.info/
working with comfy most i thing

[cyan crown]

It's 1.7 GB so It's too big. But smaller loras doen't work with my way of training (don't know well why)

[zinc cargo]

https://tenor.com/view/huge-bitch-deuce-bigalow-rob-schneider-gif-12618957

[zinc cargo]

they do share 6gb of models, so i think they can take...

[cyan crown]

this is a Lora, not a model

[uncut fiber]

i think he mean they can make exception

[zinc cargo]

i understand. i've wrote it in the sense that the site can take it.

[uncut fiber]

more such loras for SDXL and they will have to 🙂

[peak dove]

Margot Robbie as Barbie!!! 🥳

[cyan crown]

this is another Lora. But all of them are 1.7gb since network dim

[indigo carbon]

it's LOAB, she's back!

every time we get rid of her she returns LOL

[peak dove]

I love this LoRA!!!

Put it in a shareable Google Drive Folder ... ?

🙂

[peak dove]

Margot Robbie Aardman Animations stylee

[cyan crown]

That is why i prefer training Lora. Because SDXL is good but not so good with styles. Before training that Lora i tried Aadrman animation style prompt. Then I decided to go with a Lora

[hardy cipher]

[peak dove]

Yes, Aardman prompt is OK, but the LoRA is much richer in detail and character

[cyan crown]

Another Margot

[cyan crown]

Yes I can

but please not share

[peak dove]

OK

[cyan crown]

pm a shared drive and I can upload

[crisp owl]

the chest hair is a nice touch

[cyan crown]

And last two

more or less these are my trained loras

[peak dove]

Gawjus

[peak dove]

OK

[half ivy]

Why does this speak to me

[ionic dragon]

can we use it with sdxl?
https://civitai.com/models/24969/raiden-shogun-boss-genshin-impact

[crisp owl]

no

[upbeat summit]

looks like it's not made for SDXL

[ionic dragon]

oh

any raiden shogun loras for sdxl

[hardy cipher]

please, I need this for sdxl https://civitai.com/models/61868/socks-and-sandals

[ionic dragon]

how can we know if a lora is made for sdxl or 1.5/2.1?

[crisp owl]

upload date was a dead giveaway for that one.
Usually they specify where that one just says "other"

[crisp owl]

https://tenor.com/view/mondoshawan-fifth-element-alien-key-gif-16625457

[eternal fog]

If anyones missed it, new version of IPAdapter for SDXL - https://github.com/tencent-ailab/IP-Adapter#sdxl_10

Improvements in new version (2023.9.8):

Switch to CLIP-ViT-H: we trained the new IP-Adapter with OpenCLIP-ViT-H-14 instead of OpenCLIP-ViT-bigG-14. Although ViT-bigG is much larger than ViT-H, our experimental results did not find a significant difference, and the smaller model can reduce the memory usage in the inference phase.
A Faster and better training recipe: In our previous version, training directly at a resolution of 1024x1024 proved to be highly inefficient. However, in the new version, we have implemented a more effective two-stage training strategy. Firstly, we perform pre-training at a resolution of 512x512. Then, we employ a multi-scale strategy for fine-tuning. (Maybe this training strategy can also be used to speed up the training of controlnet).

Should use less memory now, hopefully I can run 2 IPAdapters without going into shared memory now

[crisp owl]

So is that on the main branch?

[eternal fog]

It's an updated model

[crisp owl]

ah

[eternal fog]

They seem to have done it as a safetensors this time as well

[crisp owl]

I'll have to update my flow, thanks for the heads up

[eternal fog]

They've confusingly put it in the models folder and not the SDXL models folder

https://huggingface.co/h94/IP-Adapter/blob/main/models/ip-adapter_sdxl_vit-h.bin

That's the new one

Will replace the IPAdapter.bin file in your custom_nodes\IPAdapter-ComfyUI\models folder, if you are using ComfyUI

[crisp owl]

I love seeing the file reduction of stuff overtime haha

[eternal fog]

It's because they are using vit-h now instead of vit-g, because they found it didn't make much difference.

I'll do a quick comparison

[hardy cipher]

excellent

I wonder if it'll help with all the nans

          [nan, nan, nan],
          [nan, nan, nan],
          [nan, nan, nan]

[eternal fog]

Ok I was wrong about the file size, they've fucked up their repo folders and it was confusing

[hardy cipher]

lolol

[eternal fog]

The file size is the same, but it says it will use less memory

[eternal fog]

Oh ok, it also uses the Clip_Vision model for SD1.5. Guess it makes sense as it's using VIT-H

[hardy cipher]

well let's see how this turns out

[eternal fog]

I need to test it, just downloading the SD1.5 Clip_Vision as they didn't explain this

Only found out when I saw an issue on the Comfy Custom Nodes Github page

[hardy cipher]

this seems to be for 1.5 or something?

[eternal fog]

No it's for SDXL, but it's using Clip Vit-h and not Vit-g

Vit-g is the Bigger Text Encoder in SDXL

So your current Clip Vision model won't work with it

[hardy cipher]

you know, it does say sdxl right in the name

[eternal fog]

https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors
https://huggingface.co/h94/IP-Adapter/blob/main/models/ip-adapter_sdxl_vit-h.bin

So you need both of these

[hardy cipher]

wait, but I'm not use the clip g clip vision model

[eternal fog]

Previous version used this
https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/image_encoder/model.safetensors

[eternal fog]

You will be, the one they supplied is the exact same as the Clip_Vision_G model

[hardy cipher]

well fair enough. I'll see what happens

[eternal fog]

Hmm it's still not working in ComfyUI :/

[hardy cipher]

I'd been using pytorch_model.bin before. that's what their github recommended

[eternal fog]

Oh it's because I had a Clip-G text encoder going through

[upbeat summit]

hmm - does anybody know how to enable the node "ids"? I've seen it in a couple of screenshots now

[eternal fog]

Yeah that's the same as the Clip_Vision_G model. I compared them and they were the exact same.

[hardy cipher]

wait, are you sure? why would they make a point of saying not to use clip g then?

[eternal fog]

So first pass I've just done with the newer version, it does seem to use less memory, but it seems worse so far.

[eternal fog]

No idea, but I loaded both in the Load CLIP Vision node on the old version and they did the exact same outputs

[upbeat summit]

where, how?

I don't have those options

[upbeat summit]

yeah don't have those. maybe it's from comfyui manager. I don't have that installed

[hardy cipher]

yeah, I do believe it is

[upbeat summit]

thank you 🙂

[hardy cipher]

or maybe not? I don't really know

lol

[upbeat summit]

I like to manage my nodes manually, so I have not installed it

[eternal fog]

Ok yeah this new IPAdapter just completely ruins my workflow

You can't use the SDXL Text Encoder, because it tries to feed in a Vit-G clip which it doesn't support

So I can't control the conditioning based on the IPAdapter

So I have to remove the clip conditioning section from it. Which then means I can't do a proper side by side test. And the new version looks awful compared to the old one

[hardy cipher]

I would say that perhaps you don't fully understand how it works or maybe they haven't released all the accompanying tools?

[eternal fog]

Nah I do and I know why it's not working, because I have all my conditioning setup to use both Text Encoders.

But I've removed that requirement now and it still looks bad

And the memory usage difference is negligable

Old IPAdapter:

New IPAdapter:

All with the same settings

[hardy cipher]

well do you expect it to respond precisely the same and have the same optimal settings?

[eternal fog]

No I don't expect it to be the same, I don't expect anything. I'm just saying it seems to look worse in the way I'm using it.

I'm just looking at what they've said. It didn't make a significant difference and that it reduces memory usage.

None of these seem to apply at the moment. At least for how I've been using it and getting good results.

It will likely vary depending on the models though

And how much the models have trained the text encoder most likely

[cyan crown]

Finally I decided to share on CivTAI : https://civitai.com/models/142352?modelVersionId=157839

[crisp owl]

I rarely ever use Lora's, but I will grab this one, love the look

[vital ermine]

Is there any way to stop this?

I can't find a way to disable it in comfyui

[crisp owl]

where is that even?

[vital ermine]

happens when I load comfy and I removed all modules/nodes/we and it was there so must be inside comfy.

[crisp owl]

hmm, unsure on that. Not sure I'm really tracking what you mean though

[vital ermine]

I mean that I don't like it, I don't use it, and I wish to have it completely gone so I don't have to X it out each time I load comfy

comes from this

that img icon.

[crisp owl]

That's on your settings menu?

[vital ermine]

yes, and I tried removing everything and it is still there

[crisp owl]

how odd...I'd say it's an extension since I don't have that, but unsure as it sounds like you tried removing them all

[vital ermine]

yep, I thought so too but dang if I know what did it to me. Maybe it installed something that never goes away :/