🤝｜tech-support

[vocal burrow]

And about the conflict there s probably more info in the console but they re probably both rewriting the same section of code.... so there s not much that can be done. Choose one or the other.

[fair oxide]

download these from text_encoder folder And for checkpoint sd3_medium.safetensors is good enough

[dapper briar]

Hello guys, have a problem in the openOutpaint, any ways to solve? I have --api in the bat

[vocal burrow]

dont skip a line

set COMMANDLINE_ARGS=--api

[dapper briar]

still get this error for some reason. Should I get api key from somewhere or something?

[vocal burrow]

show your webui-user.bat

[dapper briar]

here

[vocal burrow]

and you run sd by doubleclicking this file, dont you?

[dapper briar]

hmm my shortcut setted on webui.bat, sorry know a little about how it works, I thought this to bat files connected

[dapper briar]

it working now, thanks.

[vocal burrow]

you can see in the file that webui-user.bat calls webui.bat. it does not go the other way

by running webui.bat directly, nothing that is set in webui-user.bat takes effect.

[dapper briar]

I see, thanks for explanation

[tame yoke]

when i generate image, it takes gigabytes of my space from drive C:, is there a way to change that drive

dont exactly got enough on there

[ornate elk]

Open up a cmd and type
Pip cache purge

That will give you back the space

[tame yoke]

but stop the generation too i assume?

perhaps a way to change pip cache location?

[ornate elk]

Nope

[ornate elk]

The installation of an SD webui stores temp files on C

With pip cache purge it will remove them (5-10gb)

They will only be redownloaded when needed

But not for an image generation

[tame yoke]

can that be changed

[ornate elk]

It can, but not recommended to mess with caching stuff

[tame yoke]

Why?

[ornate elk]

Better tot just delete stuff from C or move things away

Also better to have the webui on an other drive than C

[tame yoke]

I have it on E:

[ornate elk]

Because non techy users can break a lot by changing system paths

[tame yoke]

that would be me

thanks then

[ornate elk]

No problem 🙂

[ornate elk]

Also a good tool is WizTree to check what's taking up space on a drive

Spotify for example was 10gb on PC...
It downloaded songs for offline play lol xD

And yea run the pip cache purge command in a cmd.

[fierce hollow]

I have 2 questions.

1. Is there an analogue of Hires. fix in the ComfyUI plugin
2. In the screenshot there is a function from ComfyUI called "scheduler" how can I find it in the usual SD automatic111 without this plugin?
   https://cdn.discordapp.com/attachments/1002602742667280404/1251061503382130708/image.png?ex=666d3552&is=666be3d2&hm=88b785312db4cfc33dea40b6448ff2505a189342680df6467afbc3ed20a31e43&

[solemn moon]

lol idk

[fierce hollow]

I'm sorry, I'm communicating through a translator, I don't quite understand you...

[solemn moon]

idk = I, Dont, Know 🤷‍♂️

[fierce hollow]

I got it)

[wind folio]

1. there is no direct analog to high res fix in comfy UI, but there is a workflow that does work the same : https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/
2. Scheduler and Sampler_name in comfyUI are merged into a single parameter : "Sampling method" in A1111

[solemn moon]

yea just realized "karass" is threre

[wind folio]

for simplicity's sake, A1111 merged both of those into a single dropdown yeah

it's comfusing for a lot of people seing both parameters in comfyUI

[fierce hollow]

thank you very much, in fact, if I manage to solve the first problem, then there is no need for the second question)

[wind folio]

comfyUI's workflow explains a lot of what happens in that "high res fix". High res fix is mostly 2 steps :
1/ generating a smaller picture
2/ doing img2img to upscale that picture and add details.
be very careful with the upscaling factor and the denoising parameter if you tinker with this workflow, it can bring abherations

[solemn moon]

"automatic" is making more sense now o0

[fierce hollow]

(

[wind folio]

Automatic is the username of the developper ^^ but yeah, it does fit here. It's good for new commers, makes using SD easier, but it is also quite interesting to use Comfy and go further into that pipeline to get even better, in my opinion

[solemn moon]

defo's but for non-code bro's Comfy is daunting

less "patreon templates" would help 😅

[wind folio]

here is a workflow for that upscaling. I just Drag/droped the picture in the linked tutorial I gave you, and it made the full workflow for me

[fierce hollow]

THANK YOU VERY MUCH!!!

[wind folio]

comfy is all but comfortable to use, when you are not a coding bro, or at least very "fluent" in Stable diffusion, for sure. I do not recommend at first at all. But it does open quite some doors once you get it. You can have nodes for almost anything. I was using LLMs inside my workflow, or 3 different models at different steps of my picture making for example. it makes it possible to do lots of things, but yeah, I discourage any newcommers to start with it ^^

[solemn moon]

i did tuts and all, but i like buttons, also A1 seems more forgiving, like if something shouldnt work it gives you a messed image, but comfy givews you a red report.. idk.. "comfy is all but comfortable" 😂

[wind folio]

I like complicated things, what can I say x)

[burnt totem]

hello .

i got this error RuntimeError: Couldn't checkout commit cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf for Stable Diffusion. Command: "git" -C "D:\sd\forge webui\stable-diffusion-webui-forge\repositories\stable-diffusion-stability-ai" checkout cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf Error code: 1

while iinstalling forge with ipex how do i fix that arc a750 COMMANDLINE_ARGS=--use-ipex --skip-torch-cuda-test

anyone know how i solve this? i did delated k-diffusion folder tho

[jagged glade]

my pony xl is not working any suggestion?

[ornate elk]

Need more information

[jagged glade]

if i all same setting as in civit it will come pink screen \

if i edit some settings its giiving me low quality results

[trail arch]

Can anyone guide about the workflow to efficiently change outfit of models ?

[ornate elk]

Can you share an example or your settings

[burnt totem]

hello .please can you help me ?

[ornate elk]

What's the issue?

[burnt totem]

this

[burnt totem]

RuntimeError: Couldn't checkout commit cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf for Stable Diffusion. Command: "git" -C "D:\sd\forge webui\stable-diffusion-webui-forge\repositories\stable-diffusion-stability-ai" checkout cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf Error code: 1

[vocal burrow]

1/ you've got modifications in your webui-user.bat preventing your git pull from working => revert its modifications => create your own webui-user-forge.bat and put stuff in there and then git checkout --theirs webui-user.bat
2/ forge hasn't been updated in three months or so at this point, I would recommend staying on A1111 which is probably gonna be updated soon ™️

it's telling you everything right there

[burnt totem]

i did changed COMMANDLINE_ARGS=--use-ipex --skip-torch-cuda-test this only this causing the eroor

[vocal burrow]

yes I know

like I said put everything in its own file and revert webui-user.bat

[burnt totem]

got it

[burnt totem]

ive to put what? in my own forge.bat?

[vocal burrow]

ok .

rename your webui-user.bat to webui-user-forge.bat

run git checkout --theirs webui-user.bat

voilà

[thin nexus]

best option?

[vocal burrow]

it depends what you want to do with it and how willing you are to dive in the nitty-gritty

[ornate elk]

I would recommend starting with Automatic1111

[vocal burrow]

• fooocus => I just want to write text and get "good" results. But I'll get few control on what's actually going on.
• Automatic1111 => biggest community, also very easy to understand but give you access to more technical stuff and has a lot of custom extensions to dive into
• ComfyUI => more complex as it's "node based" so you won't get a simple "put text here, press button, here's your result". Also a pretty big community and a lot of extensions

(Forge-webui => forget about it was an experiment built upon A1111 and was never really made to replace it. It has not been updated in three months. It has made its points and most of its improvement are coming to A1111 anyways)

[burnt totem]

i heard this was 6times faster then auto111 was it false?

[robust vessel]

More like 20-30%

[vocal burrow]

ah ah ah yes definitely

it does depend of your machine and usage.

you can get maybe 2x faster "at best" but that's usually if you're using the worst hardware.

It did use less vram tho, also depends of your hardware and usage. Probably 10 to 25% if I had to make random numbers on the spot.

[white tide]

How much VRAM does my GPU need to have to run SD3 Medium with clips included?

[robust vessel]

In Comfy ~8 gb

[vocal burrow]

runs fine on my 8gb gpu

[robust vessel]

Peak at 7.6 gb

[white tide]

huh, odd. Mine always complains about out of memory errors in ComfyUI. Maybe I've got the setup wrong? 😅

[vocal burrow]

oh yeah, and if you want to try SD3, for now it's only supported by ComfyUI

[white tide]

Does this look correct? I think it's a workflow from ComfyUI for SD3.

[vocal burrow]

On my end I ended up using the included sd3_medium_example_workflow_basic.json at https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/comfy_example_workflows

[white tide]

It's working, just very slow on my 6800 XT

[vocal burrow]

oh, it might act differently on AMD, because.... AMD

[ornate elk]

Are you using Zluda or directml?

[vocal burrow]

(was about to invoke you cs)

[burnt totem]

so ya i did messed up and deleted forge ui but yesterday i installed stability matrix and comfy in it which built in its own inference ui, , its run on cuda rocm direct ml , but i wanted to use ipex because i got intel gpu so there https://github.com/Disty0/ipex_to_cuda, i dont know coding like clue less asf and where to put stuff, isnt there way to modify and use ipex?

[white tide]

I've never heard about Zluda

[ornate elk]

Oh its the saviour for us AMD users.
Its much much faster than directml and uses less vram at the same time.

[white tide]

So Directl ml though. I used python main.py --directml to start ComfyUI

[white tide]

Appears I'll have to check pinned messages.

[ornate elk]

Yea in the first link You'll find my AMD Comfyui with Zluda guide

Also a zluda guide for Auto1111

[white tide]

Actually idk, reinstalling my Python version might mess up a lot of things. 😅

[ornate elk]

What's your python version?

[white tide]

Python 3.12.2 (tags/v3.12.2:6abddd9, Feb 6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] on win32

[vocal burrow]

those are the instructions to run A1111 on intel stuff https://github.com/openvinotoolkit/stable-diffusion-webui/wiki/Installation-on-Intel-Silicon or you can try sd.next https://github.com/vladmandic/automatic/discussions/2023

[ornate elk]

Thats the wrong version for comfyui and auto1111

And are you on windows 10 or 7 ? xD

[white tide]

Windows 11

[ornate elk]

Okay

[ornate elk]

Do you need python 3.12 for anything? If not you can just uninstall it and install 3.10.11 64bit

[burnt totem]

so comfy isnt too comfy but is fastest, want easy ui built in like stability matrix not possible?

[white tide]

I'm not sure, but I don't want to break anything either.

Not to say that my ComfyUI installation is working, because sadly, it isn't. Ofc dependency hell when I go to update it from not touching it in months.

[ornate elk]

Yea then start fresh with my guide. Its worth it

[vocal burrow]

stability matrix is "just" an installer for other clients with an additional UI making in my opinion everything more complex to understand for beginners

[white tide]

I don't know what "starting fresh" looks like with Python when I have other projects that utilize it too. It's hard for me to sit down and do any of that honestly.

[ornate elk]

You can have multiple python versions installed at the same time.
But you can also try my guide with your 3.12 version. It may can have some bugs or it works perfectly fine

[burnt totem]

seems easy to me tho but the ipex isnt supporting makes it streessing af

[white tide]

I thought DirectML would work fine

It did for every other model I've tried before when I've used ComfyUI in the past. Just not with SD3. Edit: The update broke it. Could I just delete my conda environment for ComfyUI and restart that?

[ornate elk]

Directml is slow and uses the whole vram even when using 1.5 models.
Its not optimised at all.
In February, Zluda got releases which emulates cuda for AMD ROCm.
With that you have more than double the performance while using a lot less vram at the same time.
In Auto1111 I was getting 6it/s for 1.5 models with directml.
Now I get 16it/s with zluda

[white tide]

So what version of ComfyUI should I be running? The normal installation or the portable version?

[ornate elk]

The installation from my guide

Its not the portable one

Its an fork of comfyui which integrates changes for zluda to work

[white tide]

Seems rather complex just to get image generation working again when it was working fine before. I'll have to pass up on this honestly.

[ornate elk]

No problem its up to you. It's really a game changer and even people with a 6600xt got it working and never look back to directml

[white tide]

I fear I'd break something honestly with any other environments I've setup before. I'm not good with that kinda stuff. 😦

[ornate elk]

Nah you can't break anything

The comfyui version sets up its own venv

So it won't break other stuff

Its like auto1111

Also everything can be reverted

Nothing to lose

[white tide]

So you're saying I could keep this, but also have Python3.10.11 (64bit) along side it?

[ornate elk]

Yea but like I said you can also try it with 3.12 first

[white tide]

I suppose it doesn't harm giving things a try, but would you be my assistance if I need guidance? <_<

[ornate elk]

Yea sure. I helped a lot of people getting SD running here 🙂

[white tide]

Just Install now?

[ornate elk]

Yep

[white tide]

On this part: git clone https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu && cd stable-diffusion-webui-amdgpu && git submodule init && git submodule update

Should I create a new virtual environement with Conda first in the SD-Zluda folder?

[ornate elk]

You have conda installed to?

[white tide]

Yes

[ornate elk]

Ah okay

[white tide]

So I could do a Conda python3.10.11 env?

[ornate elk]

Do that after you git cloned

Inside the comfyui folder

But my launch script will also create a venv folder so I don't think it would be necessary

[white tide]

okay

[ornate elk]

Maybe first try the guide way

If needed you can create a conda env later

[white tide]

Optional: Move one 1.5 based model from your other existing directml webui install into the models/stable-diffusion folder. Or skip this if you dont have one.

I don't believe I have a SD 1.5 based model

[ornate elk]

Okay then take any other model

But not SD3 for now

[white tide]

Will any of these do that isn't SD3?

[ornate elk]

Juggernaut

[ornate elk]

And thanks for pointing out I need to change it to models/checkpoints

[white tide]

Ohh, got it, will change that myself too

Do I create the checkpoints folder manually?

[ornate elk]

There should be one already in the comfyui folder

Models/

[white tide]

There isn't one in the forked version I cloned from your tutorial

No worries, I created it and put the juggernaught model checkpoint in there.

[ornate elk]

Wait

Your following the Auto1111 guide right now

I thought you wanted to go for Comfyui

[white tide]

ohhhh wtf I am 🤦‍♂️

I jumped in too quickly

Can I stop this?

[ornate elk]

Sure

[white tide]

<_<

Rookie move

[ornate elk]

Close it and then you can delete it

No worries xD

[white tide]

Yeah no worries 😅

Just might be some junk left over in a virtual environment somewhere.

Ehh, so about ComfyUI... this works with the Zluda thing too?

[ornate elk]

Exactly

You can use Zluda with Auto1111, Comfyui and Fooocus

I'm gonna go eating dinner real quick

[idle bolt]

What are the best settings to train an SDXL character model with kohya at the moment? All the videos on YouTube are pretty old. For example, how many steps are necessary? What other settings should be made? Thank you ...

[white tide]

I found the proper tutorial now. 😅

My apoligies for being a dimwit and jumping straight into what I thought were the correct instructions initially. I didn't even know what I was thinking! 😅

Ironic + more work for me. I already have 5.7.0 installed. Hopefully it's a simple upgrade and not uninstall it all manually and install it all again.

uh oh... the Start-Comfuyi.bat got done installing all dependencies as far as I am aware of, and then when it went to start the webUI, the .bat file exited

It's complaining about no NVIDIA driver on your system. Of course though. I am using AMD? <_<

The 5.7.1 installer shows 5.7.0 ROCm applications for literally all of them? <_<

[fossil dew]

Sorry to interupt, where can I find upscalers? I had someone help me before, but I can't find where Id found them before~

[ornate elk]

All but not the AMD Pro driver

Make sure it's not selected

[white tide]

It's showing 5.7.0, which I already had installed. Will it upgrade them all to 5.7.1, or am I already good for this step?

[ornate elk]

I think it will upgrade them

[ornate elk]

Here:
https://openmodeldb.info/

[fossil dew]

ty

[white tide]

It's removing 5.7.0 ROCM, will lyk what it decides to install

Still 5.7.0 lol

[white tide]

I installed from the HIP SDK download link of the ROCm version I highlighted - 5.7.1 🤷‍♂️

[fossil dew]

would you mind reminding me what file I drop the downloaded file into? 🥲 Sorry I know you're busy~

its my final question ❤️

[white tide]

5.7.1 just sounds like a fluke now?

https://rocm.docs.amd.com/en/docs-5.7.1/deploy/windows/quick_start.html

[ornate elk]

Okay then let it be on 5.7.0

Or maybe restart the PC after installing hip sdk

Thats needed anyway

[white tide]

I just did that

I restarted my PC and uninstalled all the HIP SDKs and then restarted my PC once more.

The same setup wants to install 5.7.0 again.

[ornate elk]

@white tide give me 10 minutes and I'll have time

Okay now I'm back

[white tide]

Hi

Has anyone else you've helped ever had issues with upgrading ROCm to 5.7.1 with a card that supports it? :c

Could it be my Windows version? I am using 24H2.

[ornate elk]

Never saw one that had issues with rocm

I can't recheck on my PC if the ROCm version is 5.7.1 or 5.7.0 after install

[white tide]

Okay darn it, I will reinstall 5.7.0 and restart my PC and continue the steps. Hopefully things just work, and works well. 😅

WAIT WUT NO?

I thought I had that unchecked!

That's misleading ASF I think. My Adrenaline drivers appear fine. AMD giving me a jump scare.

[ornate elk]

If your AMD Adrenalin software is Red colored everything is fine

If adrenaline is blue its pro

[white tide]

All is good. I freaked out a little. I had it disabled here too, which is what you want disabled.

[ornate elk]

Its a bit misleading that the whole installer is called pro

[ornate elk]

They go into models/ESRGAN folder

[fossil dew]

Am I safe to ask what folder I drop the upscalers into?

I just uninsalled stable diffusion and uploaded XL, having clearity issues- makes the image bright, but then it goes dull at the last few seconds before completion

[fossil dew]

awesome uhh OH FOUND IT

[white tide]

I stepped outside to let my dog out, but I’ll come back to this afterwards.

[fossil dew]

ok so putting it in that folder, and restarted the UI and it's not showing up

[ornate elk]

Alright let me know then

[ornate elk]

You dont See them when enabling hires fix ?

They should be in the upscaler dropdown

[fossil dew]

There it is in the file

let me restart UI again

[white tide]

Still 5.7.0 HIP SDK, but we'll just have to roll with it

[ornate elk]

Okay then it will work with that

[fossil dew]

so yeah no its not showing up 🥲

[ornate elk]

Can you try to download an other upscaler? Like Ultrasharp?

[fossil dew]

im doing anime pics so should I do the anime one?

[ornate elk]

Yea

[fossil dew]

downloading the zip file rn

[white tide]

Add the C:\ZLUDA folder and %HIP_PATH%bin to your PATH in the System Variables. Like shown here:
What is %HIP_PATH%bin? Am I literally setting that in my PATH system variables?

[fossil dew]

then ill beep bop it over to the file and unzip it

[ornate elk]

You only need the .pth file from it

[fossil dew]

uh it all came together i guess

[ornate elk]

Its an shortcut to C:\Programm Files\AMD\Rocm\bin

It will auto adjust the path to the rocm folder

[white tide]

So set it as shown then?

%HIP_PATH%bin

[ornate elk]

Yep and make sure your in the Path settings of system variables and not user variables

[fossil dew]

in the file, gonna reboot UI

annnnd nuffin'

[ornate elk]

And you closed the whole cmd and webui?

And relaunched

[fossil dew]

uhhh

ill try that

[ornate elk]

Perfect

[white tide]

Which step would I move onto from here 3, if step 4 doesn't apply to me; step 4.5 or 5? I have a RX 6800 XT GPU. My apologies for so many questions. 😅

[ornate elk]

With that card you can skip 4 and 4.5

[fossil dew]

anime sharp!

Thank you @ornate elk

[white tide]

...

It's the same error 😦

[fossil dew]

Now I just need to figure out why its going from Colorful during generation, and dull when its complete

[white tide]

gonna restart my PC. I know Windows environment variables are weird.

[ornate elk]

You need to use a VAE file

For color correction

[fossil dew]

OHHHH yep

thats what im forgetting! TY

[white tide]

It's the same result.

[ornate elk]

No problem

[ornate elk]

Did you renamed and copied the zluda files correctly?

[fossil dew]

WOOT WOOT!
Thanks again SC1o! 😄

Hope you fix your issue G30

I'll get out of your way now ❤️

[white tide]

I think I did it wrong, as I have ZLUDA/zluda

Simply an extra folder due to how I extracted it I guess.

[ornate elk]

Ahh well

[white tide]

I can fix if it should be C:\ZLUDA\fileshere

That did something..

[mystic basin]

is it normal that when i try to change the model i want to use, it just load for a few second then go back to the first model in the list. so albedo in my case.

[mystic basin]

my log

[white tide]

@ornate elk Does SD3 work on this version of ComfyUI?

[ornate elk]

Yep

[white tide]

Wait, so this will take a while to only generate the FIRST image, right?

[ornate elk]

Yes only the fist image will take so long

[mystic basin]

oh i get it. SD3 doesn't work with AUTOMATIC1111 i think

[ornate elk]

After thats its blazing fast

[ornate elk]

Correct for now

We have to wait for an update

[mystic basin]

thanks for the info

[white tide]

This is the only thing I've got to tell me that it's doing anything. Hopefully it's just working. 😄

[ornate elk]

Yea it won't show much there until its done xD

Give it 15-40 mins

[white tide]

Is this normal for my card?

[ornate elk]

Can't compare it rn with my comfyui but it should be normal

[white tide]

It loaded SD3 (or is beginning to) 😄

[fossil dew]

mmm one last question 😐
https://github.com/lllyasviel/ControlNet

I put this link in on extensions- install from URL - but I dont see ControlNet anywhere?

[ornate elk]

Okay, but sd3 can error out when not used with the right workflow

[white tide]

Applications using ZLUDA are slow to start.
On the first start ZLUDA needs to compile GPU code for the application. This is a one-time cost, compiled GPU code is cached in %LOCALAPPDATA% on Windows and in $XDG_CACHE_HOME or $HOME/.cache on Linux.
Some applications will gradually load the GPU code as it is used. If that is undesirable you can try setting environment variable CUDA_MODULE_LOADING=EAGER. It depends on how the application was programmed, but it might force to load (and compile) all the kernels on startup, no matter if they are used or not.

[white tide]

I'm on the VAE Decode part now, so we will see. 😅

[white tide]

Now is this a one time thing every time I open up ComfyUI from a fresh server start, or literally just this one time?

[ornate elk]

Thats the wrong extension for Controlnet.
You have to delete it again.
Then in the extensions tab click on "Available", then on "Load from"
In the list search for sd-webui-controlnet
Then install that

[fossil dew]

how do i delete it 😐

[ornate elk]

It can also take that long again after
Updating GPU drivers.
Upgrading torch versions
But its always only one time. And after that its fast

[ornate elk]

Go into the extensions folder

There delete Controlnet

[fossil dew]

OH

im stupid nvm

[white tide]

@ornate elk I thank you very much for your guide on this. You've rightfully earned a star on your Stable Diffusion docs repo, where the guide is to be found. ❤️ 🙂

[ornate elk]

No problem 🙂
Thanks for the Star!
If you have any other questions regarding SD and AMD feel free to ask me anytime

[fossil dew]

is there a faster way to look for controlnet in the list 😐

this is extensive heh

[white tide]

Not bad 🙂 (1024x1024 image)

[ornate elk]

Click in the search bar at the top

[fossil dew]

oh wow i didnt even see it dont mind me

🤔

[ornate elk]

Not that

Its called sd-webui-controlnet

[ornate elk]

Very good for an SD3

[ornate elk]

The 3

[fossil dew]

oki ty

[ornate elk]

Last one there

[fossil dew]

so its install then what 🥲

literally the last thing I need help with 🥲 promise 😂

[fossil dew]

oh its here

nvm n_n

ok so after install ill just reboot it all n_n

awesome

yaaaay CS1o fixed everything n_n

Thanks again!! 😄

[ornate elk]

No problem but you still need the Controlnet models

For 1.5 models you get them here:
https://civitai.com/models/38784/controlnet-11-models

[ornate elk]

Also just an information that not every node in comfyui will work with AMD beeing it zluda or directml

[white tide]

Yeah haha I know. I take things easy, I was just bummed out that things weren't working for me. :3

Hot

[ornate elk]

Nice

How fast is ollama for you?

It should work with Rocm

[white tide]

I get about 80 tok/s on average with 7-8B parameter models using Ollama. I am happy there. 🙂

16B gives me like 45ish tok/s

[ornate elk]

Ohhh nice

Will set it up too when I'm back home

[toxic plank]

can anyone help me setting up stable diffusion? i changed pcs and now i keep getting all kinds of errors, need some tips

[ornate elk]

Hey, make sure you followed the latest guide for your GPU here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

[toxic plank]

the problem is mostly about Pip and Torch, im not sure exactly what to do, tried a lot of tutorials, cant really seem to find it in this site either

[ornate elk]

Do you have an error?

You could share

[toxic plank]

send the whole thing here?

[terse plaza]

hi, anything specific you want to see?

the only uncommented line is export COMMANDLINE_ARGS=" --xformers --medvram --no-half"

sorry for the almost 1 day delay

[toxic plank]

'--xformers' is not recognized as an internal or external command,
operable program or batch file.
'--medvram' is not recognized as an internal or external command,
operable program or batch file.
'--no-half' is not recognized as an internal or external command,
operable program or batch file.
venv "D:\DaKimans\AI\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.9.4
Commit hash: feee37d75f1b168768014e4634dcb156ee649c05
Installing torch and torchvision
D:\DaKimans\AI\stable-diffusion-webui\venv\Scripts\python.exe: No module named pip
Traceback (most recent call last):
File "D:\DaKimans\AI\stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "D:\DaKimans\AI\stable-diffusion-webui\launch.py", line 39, in main
prepare_environment()
File "D:\DaKimans\AI\stable-diffusion-webui\modules\launch_utils.py", line 380, in prepare_environment
run(f'"{python}" -m {torch_command}', "Installing torch and torchvision", "Couldn't install torch", live=True)
File "D:\DaKimans\AI\stable-diffusion-webui\modules\launch_utils.py", line 115, in run
raise RuntimeError("\n".join(error_bits))
RuntimeError: Couldn't install torch.
Command: "D:\DaKimans\AI\stable-diffusion-webui\venv\Scripts\python.exe" -m pip install torch==2.1.2 torchvision==0.16.2 --extra-index-url https://download.pytorch.org/whl/cu121
Error code: 1

[ornate elk]

Okay, can you show your webui-user.bat?

[ornate elk]

Looks good.

[pastel kettle]

how do i fix this error for sd3 comfyui?

TripleCLIPLoader:
    - Value not in list: clip_name1: 'clip_g.safetensors' not in []
    - Value not in list: clip_name2: 'clip_l.safetensors' not in []
    - Value not in list: clip_name3: 't5xxl_fp8_e4m3fn.safetensors' not in []```

[ornate elk]

What's your GPU?
The launch args are in the wrong line

[toxic plank]

figured, 1s

[ornate elk]

Don't use --no-half if your GPU is not a gtx16

[ornate elk]

You need to download the Clip text encoders and put them into the clip folder

@terse plaza and you launched the webui.sh ?
Can you show the log when starting it?

[toxic plank]

intrel R uhd graphics, not sure where to find my gpu

4gb memory if its to any use

[ornate elk]

Taskmanager, performance tab, then on GPU 0 or GPU 1

[toxic plank]

NVIDIA GeForce GTX 1650

[ornate elk]

Ah okay then yea you need --xformers --medvram --no-half

At the line commandline_args=

[pastel kettle]

lowvram

1650 is 4g

[ornate elk]

Nah

Only for sdxl

[toxic plank]

like this ?

[ornate elk]

Behind the =

In one row

[toxic plank]

done

same error

exactly the same

[terse plaza]

there is an error

[ornate elk]

Yea now we need to fix the real error

Your python is broken

Because your antivirus Programm broke it

[toxic plank]

barely downloaded it ..

[toxic plank]

probably, kept getting scanned

its off now

[ornate elk]

Yea, you need to uninstall python.
Them disable the AV.
Then install python 3.10.11 64bit
Then enable the AV

Then delete the venv folder

And then relaunch the webui-user.bat

[toxic plank]

AV

aa whats that

[ornate elk]

Antivirus

Programm

Sry xD

[toxic plank]

it should be off, but dont i need 3.10.6

[ornate elk]

Nope better use 3.10.11

Will work the same but is newer, lesser bugs

[toxic plank]

alright torch is downloading, didnt happen before

[ornate elk]

Seems like something is missing for xformers

What you can try is to delete the venv folder and relaunch

[terse plaza]

ok, i m trying that

[toxic plank]

installing requirements for like 5mins now

think its supposed to do that

if i recall

[ornate elk]

Yea it can take a bit

[toxic plank]

ayy it worked, thank you

thanks a lot

[terse plaza]

same error

xformers was installed, i have xformers: 0.0.23.post1 at the bottom of the webUI

the only prompt was the word 'couple' with thte defoult settings

Arguments: ('task(2fndn2laf8c4f9i)', <gradio.routes.Request object at 0x73ec388827d0>, 'couple', '', [], 1, 1, 7, 512, 512, False, 0.7, 2, 'Latent', 0, 0, 0, 'Use same checkpoint', 'Use same sampler', 'Use same scheduler', '', '', [], 0, 20, 'DPM++ 2M', 'Automatic', False, '', 0.8, -1, False, -1, 0, 0, 0, False, False, 'positive', 'comma', 0, False, False, 'start', '', 1, '', [], 0, '', [], 0, '', [], True, False, False, False, False, False, False, 0, False) {}

and ther error is this

OutOfMemoryError: CUDA out of memory. Tried to allocate 10.00 MiB. GPU 0 has a total capacty of 3.81 GiB of which 12.56 MiB is free. Including non-PyTorch memory, this process has 3.38 GiB memory in use. Of the allocated memory 3.25 GiB is allocated by PyTorch, and 53.50 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[ornate elk]

Thats so strange

It should work now

Was that with Dreamshaper?

[ornate elk]

Nice no problem

[terse plaza]

yes, dreamshaper8.safetensors

i have loras in the lora folder, but i understand that they do nothing if they arent referenced in the prompt

[white tide]

Just wanted to say that with this workflow, these are very SOLID consistent results on my 6800 XT! https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/comfy_example_workflows/sd3_medium_example_workflow_basic.json

I will try 2 images in a single batch now, as I haven't done so yet. I am also running Docker Desktop in the background. No hit in performance really. 🙂

[ornate elk]

Ah good to know thx, which sd3 model did you used?

[ornate elk]

So the 5gb one I guess

[white tide]

90s for 2 1024x1024 images.

[white tide]

Which one would be faster to use? I assume the smaller?

[ornate elk]

The 4gb one but it needs the three clip text encoders

[white tide]

All good, I've got them downloaded. It's nice to be back on ComfyUI I suppose, and the start .bat file provided by the fork is nice too. Makes it easy to run it. 🙂

I will just have to remember that the virtual environment is created for me with it.

[ornate elk]

Maybe try --lowvram instead of --medvram

[terse plaza]

ok

what do those arguments mean

[ornate elk]

The start bat is made by me ^^ the fork is from an AMD User im in contact with.

I made the bat file because comfyui doesn't come with a venv by default. And also to let it install the right pytorch version needed for AMD

[ornate elk]

--medvram and --lowvram split models into multiple files to load them into the GPU vram. Because loading the file as one big file can cause the out of vram error

[terse plaza]

ok, that worked. Now i have to see if it stopped freezing or forcing reboot of the machine

and you use medvram o lowvram depending on how much oyou have then?

[ornate elk]

Correct.
GPUs with 4gb can use --medvram but for heavier tasks they need --lowvram

[terse plaza]

what are --xformers --nohalf --no-half-vae --opt-sub-quad-attention

(i been collecting them)

[ornate elk]

Here you find every launch arg available for auto1111:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

With a little info what it does

[terse plaza]

thanks

[ornate elk]

Xformers is needed for lower vram usage and better performance

[terse plaza]

system crash

its not rebboting, its as if i logged out

i only see a black screen, not the logo of the motherboard

[ornate elk]

Oh make sure your Linux os has a large swapfile

[terse plaza]

whats the recommended size?

[ornate elk]

To less RAM available can be an issue

[terse plaza]

btw, this time i was trying a lora with this parameters

lora:mix4:0.7, mix4, lora:20d:0.5, 20d, solo, long_hair, shirt, dress, high_heels,full_body, look_at_viewer, (8k, RAW photo, best quality, masterpiece:1.2), (realistic, photo-realistic:1.37), professional lighting, photon mapping, radiosity, physically-based rendering,
Negative prompt: paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans,extra fingers,fewer fingers,strange fingers,bad hand (low quality, worst quality:1.4), (bad_prompt:0.8), (monochrome), (greyscale)
Steps: 30, CFG scale: 8, Sampler: DPM++ SDE Karras, Seed: 2144095954, Size: 512x1024, Model: chilloutmix_NiPrunedFp32Fix, Model hash: fc2511737a, Face restoration: CodeFormer

[ornate elk]

8-16 GB RAM is needed.
Pagefile (swap file) would be good at 16gb

[terse plaza]

ok, i have 16gb of ram

swap im not sure, i left the ubuntu installer to configure that

[ornate elk]

Don't use SDE karras as its slower. Better try Euler a or DPM++ 2m karras.
Also try 512x768

Chilloutmix is a 4gb model which fills your vram a lot. So using it with loras could be to much

Check your resource usage with htop in terminal

Im off now gn

[upbeat zealot]

hey questiooonnn

im trying to load Pony diffusion Xl with fooocus.
but it wont load the model. it just takes like forever. like wont load at all

and im not quite sure how to fix it.

[upbeat zealot]

nvm. it jusr seems really really slow to load it

[burnt totem]

so ive used to got really bad anatomy and noise color in comfy ui, anyone know what causing those?

[low basin]

Dumb question- (A1111)- When inpainting (either controlnet or inpaint tab) is there any way to set a "bounding box" instead of being limited to "Only Masked" being blown up to the set resolution or "Whole Picture" which does not play well when inpainting pictures that were not generated and therefore not at a good resolution for inpainting (yes, I know about the "auto detect size from img2img). Hopefully this will help explain what I'm asking:

[solemn moon]

Not dumb
I think inpaint needs more attention on tools

[low basin]

Agreed. It's a bit funny that an IOS app created by a single developer has a significant amount of tools that don't even exist as an extention or feature in regular SD programs

Anyway, does anyone have any input on how to inpaint a photo at a non-standard resolution without cropping the image prior and re-blending the inpainted and original afterwards?

[light maple]

im following this and got the same error when I used Stability Matrix

also how do I uninstall Stability Matrix

it's just there now

[gilded canyon]

A1111 was working fine two days ago, now everytime i open it i get a out of memory error

i have 12GB of vram on a RTX 3060

[low basin]

In case anyone else struggles with this I have a semi-solution. You can paint a few tiny mask dots around the image (cornering off what you want to make a psuedo bounding box) and use the "Mask Only" function and it'll have a lot more understanding of the surroundings to your true mask, and you can then also use a standard resolution to generate. The second solution is a lazy script to crop a folder of images to the nearest resolution divisible by 64 and run the inpaint with the "Whole Image" function. Takeaway - AI image generation ≠ photoshop

[ornate elk]

What's your GPU?

You can check if you can uninstall stability matrix under System settings / Apps. If not just delete its folder

[old bane]

out of cirousity since i already have a installed SD i wondered after being told about this to just to give Kaggle/sagamaker a try.... problem is i cant even navigate though the websites just to make a single image beem hpw they talk more like tech or data centers then me wanting to make a goofy looking photo lol

[kindred meadow]

Are we allowed to ask question of COmfy UI? is this the same mod server for it?

I keep getting this error trying to use stable diffusion comfy AI for Krita's AI generative Ai. I am using Stability Matrix and have installed all the Models needed but it stillk says i am missing it

[ornate elk]

If you can run SD localy its always easier to set it up and use it

[ornate elk]

Yes, but you can also ask in #🧣｜comfy-ui

[ornate elk]

You've got an out of memory error

But with your GPU you should switch back to Auto1111 because Forge is deprecated for normal Users

[viral marsh]

@ornate elk find out someone who could help me with Api

[gilded canyon]

why is forge deprecated?

[light maple]

NVIDIA 2080 Ti, I also installed CUDA 12.5 but it didnt help

[chrome rampart]

hi hope doing well I use forge also. Can you tell what you mean with deprecated ? thks. I did notice but could be wrong no updates past weeks

[ornate elk]

@chrome rampart Because Forge was always experimental and will break with the next Update.
Read the announcement here:
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/801

[ornate elk]

Make sure your GPU driver is updated.
Also then follow my install guide for nvidia

[chrome rampart]

thks cs. Shame I really liked the features of forge compared to normal A111

[ornate elk]

A lot of forges improvements got into the dev build of automatic1111

[gilded canyon]

that sucks, i was having issues with regular a1111 but im reinstalling it rn to see if maybe it works with SDXL models

[ornate elk]

Checkout my install guide from.the pinned messages

Most people just forget to set the correct launch args in Auto1111

And forge sets them automaticly

So thats why most people think they can't use sdxl with Auto xD

[gilded canyon]

What does --no-half-vae do? I've seen a lot of people recommend it but idk what it does exactly

[chrome rampart]

maybe i can try to find things like cut-off for regular a11. super thks again

[ornate elk]

Cut off works Normaly in auto1111

[chrome rampart]

but need extension then ?

[ornate elk]

--no-half-vae loads the VAE file as fp23 instead of fp16.
Thats needed for anime vae files and the sdxl vae

Its for compatibility

[gilded canyon]

ohh okay

thank u for your help 💙

[ornate elk]

No problem 🙂

[gilded canyon]

Just installed regular A111, when I tried generating an image on an SDXL model I got this:

OutOfMemoryError: CUDA out of memory. Tried to allocate 504.00 MiB. GPU 0 has a total capacty of 12.00 GiB of which 2.95 GiB is free. Of the allocated memory 7.71 GiB is allocated by PyTorch, and 265.40 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Any idea why? Forge generations with the same settings would work completely fine

[low basin]

Forge has a more advanced memory managment in the backend. Go to Nvidia Control Panel > Manage 3D Settings > Global Settings > CUDA Sysmem Fallback Policy > Select Prefer System Fallback

[gilded canyon]

Do I restart my PC after this change?

[low basin]

Shouldn't have to. Also if your CPU has an iGPU you should use that as your display when using SD. Shaves off ~1GB of VRAM usage

[ornate elk]

What's in your webui-user.bat?

And make sure you always launch the webui-user.bat and not the webui.bat

[gilded canyon]

set COMMANDLINE_ARGS= --xformers --no-half-vae

and i run the "run.bat" file

am i not supposed to?

[ornate elk]

Nope

Always the webui-user.bat
If you followed my guides

[low basin]

No, thats the right file (you must've done the quick install)

[gilded canyon]

I installed webui using the quick install on the Github repo

[gilded canyon]

okay got it

[low basin]

What was that setting in Nvidia Control Panel set to btw?

[gilded canyon]

It was on something like Default or Automatic, so I switched it to Prefer System Fallback

[ornate elk]

Make sure you close every other heavy programm, like games and Wallpaper Engine

[gilded canyon]

How come I keep running out of memory with 12gb of vram?

[ornate elk]

And then check at the bottom of the webui if it says xformers versoion 0.0.23

[ornate elk]

Mostly wrong settings. Like hires fix with 0 hires steps or to high resolution

[gilded canyon]

I have hires fix disabled, I can't even get a normal generation

[low basin]

Are you just trying to do a standard 1024x?

[ornate elk]

Can you check the bottom of the webui for xformers?

To be sure its enabled

[gilded canyon]

Yea 1024x1024

[ornate elk]

Okay good

[low basin]

Oh also did you click the 'apply' button in the bottom right of NCP after changing it? I only ask bc I forget every single time

[gilded canyon]

Yes I did

[ornate elk]

What does this do?

[ornate elk]

Edit the webui-user.bat again and add
--medvram-sdxl
Then save and relaunch

[gilded canyon]

Okay I'll try that, but I thought medvram is only for systems with <= 10gb vram

[ornate elk]

--medvram-sdxl is only applied for sdxl and can improve the speed and lower the vram usage for sdxl

[low basin]

It just allows the transfer of memory between GPU and regular RAM if you run out of VRAM. If its overflowing it just gets mad slow as opposed to crashing with CUDA Out of Memory error

[gilded canyon]

Okay I'm launching with medvram-sdxl

[low basin]

Does it even start to generate before it crashes or is it crashing at 90+%

[gilded canyon]

crashes at 90%

like the generation will say 100% done sometimes then boom, outofmemoryerror

if it does generate by some miracle, hires fix doesnt work then

[ornate elk]

For upscaling with sdxl you need the tiled vae extension

[low basin]

Haha figured... Yeah that's the VAE crashing it at the end. If you watch Task Manager as its generating youll see the GPU Memory Usage shoots up 2-3GB on top of the models memory

[gilded canyon]

do you have a link?

[gilded canyon]

anyway to fix this?

[low basin]

Is it still crashing even after changing that setting? I think you can choose to downsize the VAE in settings if need be but itll have a pretty significant impact on quality. But if you do have an iGPU I'd 100% be using that for display because it can be the difference between crashing/slow gen speeds and not

[gilded canyon]

no im trying a generation right now with medvram-sdxl

[ornate elk]

First go with the tiled vae extension

[ornate elk]

Its under extensions, available, then click on load from
Then search for Tiled

Then install and restart

[gilded canyon]

okay got it

[gilded canyon]

Is it called TiledDiffusion

[ornate elk]

Yep

And then only enable Tiled VAE, not Tiled Diffusion

[gilded canyon]

okay ill try it and let you know, thank u for all the help so far

[ornate elk]

Np ^^

Also make sure you don't have 10+ browser tabs open

[vivid palm]

Hi this probably wouldn't be considered tech support but I'm not sure where to ask. Can anyone point me to a comfyui workflow that would merge two faces to create one face? Or perhaps tell me what nodes I would need to make it myself ElI5 though because I've never made one before.

Ope I meant to ask on another server but I'll still take help if anyone can

[gilded canyon]

I just tried. I have 2 tabs open (one being Discord), nothing in the background of my PC running, Tiled VAE on, medvram-sdxl, and now it stops at 0% with this:

Model loaded in 349.1s (load weights from disk: 21.0s, create model: 5.7s, apply weights to model: 308.7s, apply half(): 0.8s, apply dtype to VAE: 0.2s, move model to device: 0.1s, hijack: 0.9s, load textual inversion embeddings: 1.4s, calculate empty prompt: 10.1s).
0%| | 0/40 [00:00<?, ?it/s]Press any key to continue . . .

[ornate elk]

Can you show your whole txt2img settings?

[gilded canyon]

My generation settings

[ornate elk]

And the model?

How much RAM do you have?

[gilded canyon]

I have 16GB of ram, 12GB vram, and im using an SDXL model (autismmixSDXL which is a pony model)

web ui forge used to work fine

[ornate elk]

You shoulndt have any issue with auto1111 too.

I know a lot of 3060 users

What does the taskmanager performance tab shows ? When generating?

Also do you have the webui installed on C or an other drive?

[gilded canyon]

I have it on my D drive, my C drive is my SSD so I use it for just my windows installation

[gilded canyon]

yea i have no idea why im having issues, i upgraded from a 1660ti to fix this issue and its still here

[ornate elk]

Ah interesting. Then you need to check your Windows Pagefile settings

It needs to be enabled only for the C drive and Disabled for any other drive

[gilded canyon]

how do I do that?

not familiar with windows pagefile

[gilded canyon]

I clicked generate, then a few seconds letter "connection error"

this is so frustrating

idk why a 12gb vram system is having issues with this

[low basin]

Yeah something is definitely not right. Also in the error message you sent earlier it said it tried to allocate 5XX MB of memory and there was 2.XX GB free yet it still gave you OOM error. When you get a chance, if you could go to Command Prompt and type in " nvidia-smi " and post here

[gilded canyon]

Will do in a few minutes

I suspect something is up with my system since nobody else is facing these issues on the same hardware

[low basin]

I suspect something may have gotten screwed up during install or config because your 3060 is not at all out of memory. I just wanted to see the smi to rule out a driver issue

[low basin]

Yep, not a driver issue. I'm at a loss here... Personally I'd nuke the whole A1111 install and try it again 😦

[gilded canyon]

its a fresh install...

[vocal burrow]

What does the log says ?

[gilded canyon]

Model loaded in 349.1s (load weights from disk: 21.0s, create model: 5.7s, apply weights to model: 308.7s, apply half(): 0.8s, apply dtype to VAE: 0.2s, move model to device: 0.1s, hijack: 0.9s, load textual inversion embeddings: 1.4s, calculate empty prompt: 10.1s).
0%| | 0/40 [00:00<?, ?it/s]Press any key to continue . . .

when the press any key to continue pops up, thats when the connection error appears

would reinstalling pytorch help or is that unrelated

im lost on what to do

[vocal burrow]

The Full log

[gilded canyon]

i closed that command prompt but ill do another generation rn

[tame yoke]

is there a setting i can change to make a model that generates normally on 1024x1024
work on 512x512 canvas?
on 512x512 its kinda distorted and wrong in every way

[gilded canyon]

if you're using an SDXL model then 1024x1024 will always generate better results than 512x512, 512x512 will look weird

sdxl models are trained on 1024x1024 not 512x512

[vocal burrow]

not really, you can increase/decrease the resolution about 50%. But more than that and you will get cursed artifacts.

You can also upscale/downscale afterwards but that's probably not what you're looking for

[gilded canyon]

Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.9.4
Commit hash: feee37d75f1b168768014e4634dcb156ee649c05
Launching Web UI with arguments: --xformers --no-half-vae --medvram --opt-split-attention
Loading weights [821aa5537f] from D:\Downloads\STABLE DIFFUSION\webui\models\Stable-diffusion\autismmixSDXL_autismmixPony.safetensors
Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Creating model from config: D:\Downloads\STABLE DIFFUSION\webui\repositories\generative-models\configs\inference\sd_xl_base.yaml
Startup time: 70.3s (prepare environment: 22.1s, import torch: 17.4s, import gradio: 7.6s, setup paths: 10.8s, initialize shared: 0.8s, other imports: 5.0s, list SD models: 0.4s, load scripts: 2.3s, initialize extra networks: 0.4s, create ui: 1.5s, gradio launch: 2.8s).
D:\Downloads\STABLE DIFFUSION\system\python\lib\site-packages\huggingface_hub\file_download.py:1132: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True.
warnings.warn(
Applying attention optimization: xformers... done.
Model loaded in 220.1s (load weights from disk: 4.4s, create model: 4.6s, apply weights to model: 201.2s, apply half(): 0.2s, hijack: 0.3s, load textual inversion embeddings: 0.3s, calculate empty prompt: 8.9s).
100%|██████████████████████████████████████████████████████████████████████████████████| 40/40 [00:29<00:00, 1.36it/s]
Press any key to continue . . . ███████████████████████████████████████████████████████| 40/40 [00:13<00:00, 3.01it/s]

[tame yoke]

i see, thank you

[solemn atlas]

hi, is it already possible to get sd 3?

[vocal burrow]

https://stability.ai/news/stable-diffusion-3-medium

[vocal burrow]

don't see any errors there, what's the problem exactly ?

[gilded canyon]

nope, it says 100% but it crashes right as it reaches 100%

[vocal burrow]

how does it crash ?

it crashes your whole computer ?

anything showing up during generation ?

[gilded canyon]

crashes a1111

and nope, nothing showing up

[vocal burrow]

nvidia drivers updated ?

[gilded canyon]

yea all updated

my nvidia-smi

[vocal burrow]

try with just --xformers

[solemn atlas]

I'm a little bit overwhelmed, which one do i need to download to rund SD3?

[gilded canyon]

no difference

im just gonna download an old version of forge and use that

[vocal burrow]

no overclock ?

[gilded canyon]

nothing

i bought the gpu brand-new two months ago

and it hasnt worked a single time with a1111, it only ever worked with forge

[vocal burrow]

• Stable diffusion 3 is out, where do I download it ?
  • https://huggingface.co/stabilityai/stable-diffusion-3-medium
• There's so many stuff, where do I click ?
  • sd3_medium.safetensors : basic model, does not include any text encoders, you will need to download them additionally from there https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders.
  • sd3_medium_incl_clips_t5xxlfp8.safetensors : basic model + including fp8 version of T5XXL (mostly used to render text) => should be a good balance between quality and resources.
  • sd3_medium_incl_clips.safetensors : basic models WITHOUT the T5XXL text encoder. => it should be even lighter but results will differ
• Does it work with Automatic1111 ?
  • Not yet, it only works with ComfyUI (or stableswarm with ComfyUI backend obviously) and sd.next https://github.com/vladmandic/automatic/wiki/SD3, make sure to use "Default" Sampling method

[vocal burrow]

that's really strange... Try to switch to the dev branch of A1111 maybe. git checkout dev

[ornate elk]

Did you reinstalled nvidia drivers after the GPU upgrade?

Thats something I would do

[gilded canyon]

yup, i uninstalled every old driver and installed all of my new drivers

i used DDU

so it should've been clean

[ornate elk]

Yea thats good then

[ornate elk]

Have you checked the pagefile settings?

[solemn atlas]

thank you, so it doesnt work yet for automatic1111

[gilded canyon]

no idk what that is

[vocal burrow]

pagefile shenanigans :

launch sysdm.cpl from the Start menu search or run box (Win+R) and navigate to Advanced –> Performances -> Settings –> Advanced –> pagefile/virtual ram -> Modify

https://www.thewindowsclub.com/increase-page-file-size-virtual-memory-windows

Although that's not usually how pagefile related issues fail.

Set it to 16gb just to be sure

[ornate elk]

Disbable for D, enable for C

[vocal burrow]

But yeah definitely try the dev branch of A1111 if forge somehow works for you

[white tide]

I could only seem to get SD3 to work on the forked version of ComfyUI with Zluda?

[ornate elk]

Yes would make sense as it needs extremely more vram. So with directml it mostly will error out

[white tide]

What I am saying is that I can't seem to get something like Stable Diffusion Cascade to work with a workflow from ComfyUI examples. It just stalls on step 1/30. 😅

[ornate elk]

Oh okay

I never tried cascade but will test

Sdxl and sd3 should work fine

[white tide]

What would you say runs faster out of the two?

Also, is there a differences chart or something for SD Cascade vs Lightning? Those two are pretty vague to me, and I really haven't got to try out both yet. I have SD Cascade model downloaded, but can't get anything out of it.

[ornate elk]

Fastest models are lightning, turbo, hyper and lcm models

[vocal burrow]

those are very different model architectures

[white tide]

Yes indeed

oooo I kinda want to try LCM again

Now that I have Zluda on ComfyUI

[vocal burrow]

sd cascade is its own thing based upon Würstchen.
While lighting/turbo stuff are stil based on SDXL.

[white tide]

Ahh okay

Is there LCM for SD3?

Okay I understand

[vocal burrow]

no, sd3 has only been out for 3 days and I'm not even sure LCM can be made for it. (probably)

[gilded canyon]

I just uninstalled all of my graphics drivers, like completely free uninstalled them all and reinstalled them

and A1111 is finally working

ty for your help though @ornate elk @low basin

[ornate elk]

Perfect! Np

My guess was that SD still thought it has your old GTX vram amount and not 12gb to use

[gilded canyon]

yes that was probably what was happening 😔

[honest tendon]

asking for a friend

[ornate elk]

Waiting for community models based on SD3 xD

[honest tendon]

i hate the restrictions i think its the main reason for having girl laying on grass look so cursed

[vocal burrow]

Like already mentioned :

• sd3-medium was supposed to be called sd3-medium-beta on release
• sd3-medium is only 2b parameters, 4b and 8b have been mentioned for future release
• In the past there was sd1, then 1.4, 1.5, 1.6, etc. So they could revisit sd3 with sd3.5 or whatever later on, no promise, no timeframe, no nothing
• community model might be able to fix those edge cases. If not, stuff like controlnet will for sure. Is it due to censorship ? can't know for sure.
• it's been out for only 3 days, and as with every new model release, it takes time to relearn how to "speak to the AI". I've managed to get descent picture of woman lying on grass within the hour of """"drama"""".
• The meme has been noticed :p. I personally hope the next release will be featuring an official stability.ai picture of "troll woman going outside to touch grass" or something alike :p.
• is sd3-medium perfect, the one and only model to rule them all for the future and beyond ? Of course no.

That should sum up everything. I'm not gonna engage in endless debate about that as I don't see anything constructive coming out of it. The point has been made in all kind of constructive and not so constructive manners. I'm not stability.ai, moving on.

[fair oxide]

reasonable take. as for my own take on sd3 it somehow feels chunks of its data are missing and the model is incapable of doing certain things when it comes to human form. im not so sure this is a good model to train on. however, i can only hope for the better with 8b variant.

usually to train a model you'd have to have a foundation for it, but with sd3 you'd have to train human figure from the basic

it'd be better to focus on 8b

[honest tendon]

This has been generating for 20 mins now, my pc is working its ass off but still 60c because of how cold it is rn, im practically using my PC as a heater

[honest tendon]

That would be funny asf if they did that lol

[vocal burrow]

cs1o might help you more when they come back. But from my understanding the first generation is always slow when using zluda as it's converting cuda stuff to amd stuff. It's a one time process, kinda like shader compilation in modern console emulators.

It can take from 15 to 40 minutes

[honest tendon]

he'll be back by like Sunday right

[vocal burrow]

no idea, probably earlier, I don't have his agenda :p.

[honest tendon]

interesting never had that happen the first time I downloaded comfyui

im generating an AI generated video for context

Unless previous installs of it somehow stayed intact

[steady thistle]

I'm running Radeon, and I've been using A1111. Should I consider SDNext or Comfy?

[vocal burrow]

check the pinned comments for guide for each of them

[ornate elk]

I Wouldn't recommended sdnext

For me Auto1111 is still the best experience with AMD

[ornate elk]

Let it load. It can take a while

First gen after Installing zluda?

Or did you updated drivers etc?

[steady thistle]

Thank you very much, as always. A1111 is working decently for me but it's not a great experience, so I keep checking every few months to see if it's worth a change.

[ornate elk]

What issues have you experienced?

[lethal cobalt]

Hi, I'm getting a "ImportError: cannot import name 'c_internal_utils' from partially initialized module 'matplotlib' (most likely due to a circular import) (G:\Stable Diffusion\ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\matplotlib_init.py)" when trying yo use Controlnet Aux Preprocessors from Fannovel16 in ComfyUI, but I'm greeted by this error instead, anyone knows how to go around fixing it. It happens even if I have only this pack installed.

[steady thistle]

I absolutely gave up on inpainting. I just could not get it to process at all.

I have a working text to image system going so I just made myself satisfied with that.

[ornate elk]

Oh Inpainting is a bit tricky yes.
But you Wouldn't be statisfied with comfyui then too

[steady thistle]

Yeah, I am not going to fiddle with it until there's a big change in the software (or my hardware)

[ornate elk]

I will try to get StableSwarm and Invokeai to work with Zluda next

[steady thistle]

I'll ask again in like a week or two, if that's okay.

[native solar]

Recently had a weird corruption issue with training, narrowed it down to the lastest nvidia 555.99 drivers with a 4090. anyone else having this specific issue? currently reverted back to 531.61 and it resolved the corruption isssues for training, seems to exclusive effect training models but not generation of images.

[vocal burrow]

no blood/gore-ish imagery please

[honest tendon]

OH SHI MB WRONMG SERVER

[valid iron]

help, i took bad advice and deleted the packets in appdata to save space and now sd auto 1111 wont work..i tried reinstalling python but no luck

[ornate elk]

Which folders?

[valid iron]

everything in the packets cache C:\Users\liamd\AppData\Local\Package Cache

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--reinstall-torch --reinstall-xformers --autolaunch --api
git pull
call webui.bat

[vocal burrow]

remove --reinstall-torch --reinstall-xformers

delete venv folder

launch again

[ornate elk]

And add --xformers

[valid iron]

shouldent i reinstall python and add the directory first?

i didnt change anyof that before, just deleted packeges

will SD still run on python 12?

[vocal burrow]

no

python 3.10.6 to 3.10.13

pick any

[valid iron]

ok ill try

no installers except these ? is that just like a normal installer?

idk how to do this stuff...

i used anaconda to set it up years ago, whats the simplest methed to get this fixed? or will i have to install SD from scratch again?

[ornate elk]

dowload the installer

[valid iron]

the gzipped source or XZ compressed?

[ornate elk]

non of them

Here:
https://www.python.org/downloads/release/python-31011/

scroll down

there is the 64bit installer

select "add python to path"

when installing

[valid iron]

ah the page i was on didnt have that thanks

is there a install path i should choose or is the standard c drive ok?

[ornate elk]

standard is best

[valid iron]

add to path correct?

ok ill try that and see if that fixes it

now getting this Already up to date.
venv "D:\Ai\StableDiffusion\stable-diffusion-webui\venv\Scripts\Python.exe"
No Python at '"C:\Users\liamd\AppData\Local\Microsoft\WindowsApps\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\python.exe'
Press any key to continue . . .

do i need to put the new path in the batchfile edit in notepad?

[ornate elk]

it seems you had python installed via the windows tore too. so either uninstall that or edit the webui-user.bat
at the Python= line add the path to the new python.exe

[valid iron]

ok the windows one didnt give me the option to add to path so ill try the edi

[ornate elk]

you need to uninstall the windows one

or add the path to the other one manually

open up a cmd and type
where python
to get the path to 3.10.11

then add an \python.exe at the end

[valid iron]

is it possible to install the version 3.10.11 from the windows store still?

[ornate elk]

it can be but its not recommended

[valid iron]

im getting the same error, can i install it into that directory its saying its pulling from? in the windows fil?

[ornate elk]

can you show the error?

you need to delete the venv folder be fore testing again

do that and relaunch

[valid iron]

Already up to date.
venv "D:\Ai\StableDiffusion\stable-diffusion-webui\venv\Scripts\Python.exe"
No Python at '"C:\Users\liamd\AppData\Local\Microsoft\WindowsApps\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\python.exe'
Press any key to continue . . .

this is the edit @echo off

set PYTHON= C:\Users\liamd\AppData\Local\Programs\Python\Python310\python.exe
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--reinstall-torch --reinstall-xformers --autolaunch --api
git pull
call webui.bat

am i removing the venv_dir or the original windows python path folder?

[ornate elk]

remove --reinstall-torch --reinstall-xformers
and add --xformers
in the webui-user.bat

and then only delete the venv folder inside the stable-diffusion-webui folder

and then relaunch the webui-user.bat

[valid iron]

ok doing that now

like this?
@echo off

set PYTHON= C:\Users\liamd\AppData\Local\Programs\Python\Python310\python.exe
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --autolaunch --api
git pull
call webui.bat

[ornate elk]

correct

[valid iron]

things are happening!!!!

[ornate elk]

looks good 🙂

[valid iron]

siik hopefully fixed, its stuck on 01 eta but i guess its working behind

thank you so much again!!!❤️

[ornate elk]

yep let it load and no problem 😄

[valid iron]

ok new error from that

[ornate elk]

try delete the repositories folder and venv folder again

[valid iron]

ok

same problem still, would it be a the virtual environment problem with tqdm.auto

[ornate elk]

did you installed any extension recently?

[valid iron]

no but possible its realated to the ones i had, i read it might be Temporal Kit

oh i tried this and it fixed, im going to run a generation to see if all is good

ok all good now thanks @ornate elk for the help, im greatfull

[ornate elk]

okay good np

make sure to also update your extensions

@white tide so the best SD3 Version is the 5gb on with clip included. its faster and uses alot less vram then the 4gb one with the seperate clip text encoders

[half island]

Hoi, probably a too basic question, but what's the proper command to create a venv that only affects that project?

As i usually do "conda create -n comfyui", and that's it.

Unless which i just remembered, i've seen commands that has "create -n venv venv" for instance, i guess that the venv at the end is what only installs X versions of modules in there needed for that venv? As when i didn't do venv, it'll just say "requirements already satisfied" which is what i don't want, as if it overwrites some stuff, then shit tends to break for other python stuffs.

[fossil dew]

https://tenor.com/view/dj-khaled-another-one-gif-26093316

I downloaded the models like you'd said @ornate elk- and its /working/ but I'm getting this warning now?

oh wait that might be because . . . hold up

its this thing gibing the warning

[ornate elk]

Thats the command for python:
python -m venv venv

[ornate elk]

did you downloaded all models from there?

[fossil dew]

uh i believe so

[ornate elk]

you may need to download the config file for the tile model

[fossil dew]

i saw a whole list of shit yesterday 🤔

hold on im looking to see where tf it went

😂

it shows up for me under uh

preprocessor

when I choose All for the control type

[ornate elk]

what does models show?

[ornate elk]

lol

whats in your models/controlnet folder?

[fossil dew]

did i put them in the wrong folder or something

im a cavewoman tryina do this unga bunga

hold on

😐

then where the fuck did that other stuff go... where did I put it O_o

wth

are you uh... how to say this...

Can you see where I messed up with this by chance?

[ornate elk]

maybe you just downloaede 1 model

[fossil dew]

the stuff under preprocessor changed tho

it has a ton of stuff that wasnt there before

what folder effects the preprocessor?

if any 😂

[ornate elk]

preprocessor is made from controlnet

you only need the models

from here:
https://civitai.com/models/38784/controlnet-11-models

Here a visual guide, then put verything in models/controlnet

[fossil dew]

I swear to god I downloaded this shit last night LMFAOOOOO I'll fuck with it again

oh huh looks like something happened ok

well 🥲

ok so IM CLEAR

I download it

and drag it into the fucking ControlNet folder

RIGHT

😂

[ornate elk]

correct

[fossil dew]

ok

[ornate elk]

https://tenor.com/view/yes-sir-yes-anime-i-have-an-idea-think-gif-16027707

[fossil dew]

😐

replaced

ok welp ima try this see what it do

thank again

[ornate elk]

you downloaded 1 model

the tile one

thats not canny, not openpose not any of the others

just tile

you need all of them for all Functions ^^

and their config files

[fossil dew]

https://tenor.com/view/homer-brain-monkey-gif-11098413

ok im confused 😂 let me have my hubs look at it- he understands this stuff, my brains like:

OHHHHHHHH

WAIT I NEED TO GO CLICK THE ONE I WANT

[ornate elk]

yesssss

[fossil dew]

ooooooohhhhh ok hahahahaha

🥲

on it!

[ornate elk]

😄 your not the first one with this controlnet "problem"

[fossil dew]

heh yeah Im so basic with this stuff, I dont understand alot of the mumbo jumbo jargen for it

so I need to download the one for the one Im using?

or I should download them all

[ornate elk]

best is to download them all

but you dont need to

[fossil dew]

and yeah I wouldnt say "problem" either, more like a lack of understanding 🙂

Okay, well I mean I have the room so I might as well-

[ornate elk]

canny, openpose, depth, tile, lineart are basic ones that are mostly used

[ornate elk]

yea excuse my wording, thats what i meant

[fossil dew]

oh no no i was agreeing ❤️

you're all good n_n

[ornate elk]

thx, sometimes i miss some words that would fit better

[fossil dew]

Eh, I dont think you missed anything, I totally knew what you meant ❤️ You're solid~ Thank you for your help again, I'll be back to haunt if I run into anymore misunderstandings of what Im working with < 3

[ornate elk]

okay, no problem! sure feel free to ask here again for anything SD related 🙂

[fossil dew]

still getting warning

[ornate elk]

have you downloaded the config files too?

[wet magnet]

Anybody have any resources for advanced prompting techniques specifically in regards to SD?

[tame yoke]

should i get webui forge? ive heard its good but theres been some stuff with updates and stuff breaking and idk what to do.

[fossil dew]

SO HELP ME GOD

[ornate elk]

Forge is deprecated

i wouldnt recommend it for anyone right now

maybe only if you have a 4gb vram gpu

[tame yoke]

I do

[fossil dew]

were they in that link you sent me?

[tame yoke]

and not sure what deprecated means

[ornate elk]

it means it wont get any updates more

[fossil dew]

run down, not doing well

synonyms

[ornate elk]

Forge will be changed to an experimental programm

thats not usable like auto1111

[ornate elk]

look at the screenshot here:
#🤝｜tech-support message

[fossil dew]

ffs

lol

ty

where do I put THAT

controlnet?

[ornate elk]

yep

[fossil dew]

👌

[ornate elk]

models/controlnet

[fossil dew]

i swear i didnt have to do all this before 😂

oh well

[ornate elk]

@tame yoke so if you want updates and new features than you should go with Auto1111

[ornate elk]

also check what base model you use. because for sdxl you need different controlnet models

[tame yoke]

isnt it good enough as it is though?

[ornate elk]

there is always stuff to improve, better performance, less vram usage etc

auto1111 will get a lot of forges performance stuff in the next update

[tame yoke]

ohhh

that sounds nice

whens that

[fossil dew]

im going to bash my head in 🥲 jokingly

So were the ones you sent me for SDXL or no?

[ornate elk]

they were for 1.5

[fossil dew]

https://tenor.com/view/internally-screaming-wendy-gif-8501519

[ornate elk]

These here are for SDXL:
https://huggingface.co/lllyasviel/sd_control_collection/tree/main
or
https://civitai.com/models/136070?modelVersionId=267493

[fossil dew]

so for Canny I need:

cuz yes, im SDXL

[ornate elk]

yes one of these or both

[fossil dew]

ok so the stuff I downloaded previously can be deleted?

[ornate elk]

if you dont use 1.5 then yes

[fossil dew]

mmmmk

and thats it?

no extra crap hiding from me?

[ornate elk]

idk when the next update will be released, but i guess soon as SD3 support comes too.

but read this from the forge creator:
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/801

[ornate elk]

nope

thats it xD

[fossil dew]

Gosh youre juggling so much, I appreciate all the work you do to help everyone ❤️ Thank you n_n

[fossil dew]

good >:[ 😂

[tame yoke]

well thank u, you saved me time

[ornate elk]

no problem!

[fossil dew]

https://tenor.com/view/coach-josh-wood-coach-josh-gold-star-gold-star-gif-25330659

[ornate elk]

for install guides checkout my install guides in the pinned messages of this channel

[tame yoke]

install guides for what

i have a1111 already

and uninstalled forge

[ornate elk]

ah alright

[tame yoke]

wait

is it possible to get dev branch from update.bat

[fossil dew]

im about to give up lolol 🥲 It's still not controling my net

lmfao

input vs output

the warning is gone tho!

😂

your talking to a cavewoman- You gotta put it in crayon so I understand

💀

idk what to do with any of this stuff, This website confuses me alot

Im going to give up for now 🥲

Ill come back to this later, nothing I do is working

[ornate elk]

what yre you triyng to do?

[fossil dew]

use my IMVU model (left) as a control to make images in the poses as her-

like this: used the right and made the left

proof I had it working 😭

getting me confused throwing all this extra stuff at me @___@

[tame yoke]

are there any downsides to using dev branch

[dim frigate]

Can anyone help me solve it? I'm trying to use animate diff

[ornate elk]

Mostly bugs

[tame yoke]

im not missing out on any updates, no?
and is the dev branch close to forge in terms of speed

[ornate elk]

Dev gets stuff that goes later into stable

[dim frigate]

AttributeError: 'NoneType' object has no attribute 'batch_size'

[ornate elk]

Set the animatediff batch size to 8 or 16

[nova oracle]

I’m trying to backup an sd card in my Chromebook and make up space on it, and i’ve looked for solutions and can’t seem to find any.

do you guys have any tips?

[ornate elk]

What does the cmd shows?

[ornate elk]

Make sure your extensions are updated and also can you show the Tull cmd log?

[ornate elk]

The replacer extension gives errors

Maybe deactivate it

And test again

[dim frigate]

not working

[nova oracle]

do you know how to backup an sd card on a Chromebook?

[half island]

What's the difference between -n and -m?

[ornate elk]

No clue ^^

[ornate elk]

What's your webui version?