🤝｜tech-support

[quick forge]

what page? what log?

[ornate elk]

Hires fix slows down.
You have Dreambooth installed meaning your running old torch version

[vocal burrow]

Yup that s what I said.....pinging the wrong person :p

[vocal burrow]

Any of sdwebui webpage

And the console log

[ornate elk]

OK torch looks good

Then yea try without hires fix

[quick forge]

took me a second here to realize im on my windows box lol

[unique bear]

Without hi res I can get 5-7 its/s with loras

[ornate elk]

Do you use sdxl or 1.5 models?

[tranquil vigil]

ok.. the conda issue is solved nnow.. but when I type in .\scripts\gui.bat, it saythe system cannot find the path specified

@vocal burrow

[vocal burrow]

Don't put spaces in your folder names

or use "broken path with spaces"

[tranquil vigil]

something liek this?

[vocal burrow]

yup but then the problem is that their script themselves doesn't account for spaces in folder names

so don't use spaces in name folder

[tranquil vigil]

sorry, I couldnt follow you there

[vocal burrow]

space no good, move to another folder

git pull```
welp ok then whomever I wrote that for

[vocal burrow]

you good now ?

[tranquil vigil]

sorry.. I am stll strugglin

[vocal burrow]

just move draggan in some folder without spaces in its path

c:\AI\DragGan for instance

[tranquil vigil]

okey.. .let me try

[vocal burrow]

every single cmd comman work the same way.
COMMAND ARGUMENT1 ARGUMENT2 ARGUMENT3 etc
command then space then arg1 then space then arg2, etc

so if there s space in your folder path and you use that as an argument.... it s gonna break

[tranquil vigil]

I got it

the CMD worked now

but unfortunately.. I am recieving the same issue as stated in the 1st question..

[vocal burrow]

command c:\something\smart ass folder
is gonna be interpreted as
command,
argument1 = c:\something\smart
argument2 = ass
argument3 = folder

So that s why you put quotes around stuff with spaces in it, so windows know "ok this is supposed to be ONE argument"

[tranquil vigil]

even though I have Conda installed now

[tranquil vigil]

thank you... got it... will definitely keep a note of it

[vocal burrow]

share the whole log

[vocal burrow]

do (or redo) what s stated there now that you ve moved draggan somewhere correct

[tranquil vigil]

I just redo'ed the thing. though.. let me try again

[vocal burrow]

cause right now it s probably trying to use your system's python and not anaconda's venv

[tranquil vigil]

why is it now saying "Couldn't find environemtn" when I try to run it in Anaconda Prompt...

[vocal burrow]

because you didn t create it first

[tranquil vigil]

eitherway, I am reinstalling DragGAN now

[vocal burrow]

conda env create -f environment.yml

it s within the first lines of the readme https://github.com/XingangPan/DragGAN?tab=readme-ov-file#requirements:

[tranquil vigil]

f... I literally forgot that...

[vocal burrow]

so tldr it s

1/ install anaconda, make sure it s set correctly and/or registered in your PATH. All the while keeping in mind that it s adding another python to your system. Therefore other app might be confused and use conda`s python instead of the standard one. You can easily "fix" that for auto1111 by using the 'set PYTHON=' variable in webui-user.bat
2/ open cmd and run those :
cd xxx                   where xxx is some folder without any spaces in its path
git clone https://github.com/XingangPan/DragGAN.git
cd yyy                   where yyy is the newly cloned draggan folder
conda env create -f environment.yml
conda activate stylegan3
pip install -r requirements.txt
.\scripts\gui.bat
3/ enjoy

[tranquil vigil]

I have CUDA toolkit 11.8 installed though

[vocal burrow]

that s your system s cuda not the one in the conda venv

you didn t install the requirements (I also forgot to include it in my tldr but tbf it s still within the first lines of the readme)

there it s edited

conda and venv stuff are bad words employed to mean "virtual environment"

tldr it s gonna create a folder somewhere and install everything necessary to run your program.

that way it won t use any of actual system s binaries/programs

that way it wont mess up anything on your system. worst case it s gonna mess up its own venv and that s it

[tranquil vigil]

just in case, does virtualization needs to be enabled in BIOS?

[vocal burrow]

no

it s not a proper "virtual machine" so there s no need to toggle any bios settings.

[tiny abyss]

Need help, unable to run any generation

[tranquil vigil]

ok... the CUDA toolkit version error was solved though.. but now I am facing another issue...

C:\AI\DragGAN>conda env create -f environment.yml
WARNING: A space was detected in your requested environment path:
'C:\Users\Chandrajit Rout\.conda\envs\stylegan3'
Spaces in paths can sometimes be problematic. To minimize issues,
make sure you activate your environment before running any executables!

Collecting package metadata (repodata.json): done
Solving environment: done


==> WARNING: A newer version of conda exists. <==
  current version: 23.7.4
  latest version: 24.1.2

Please update conda by running

    $ conda update -n base -c defaults conda

Or to minimize the number of packages updated during conda update use

     conda install conda=24.1.2



Downloading and Extracting Packages


CancelledError()
CondaError: Downloaded bytes did not match Content-Length
  url: https://conda.anaconda.org/pytorch/win-64/pytorch-2.2.0-py3.9_cuda12.1_cudnn8_0.tar.bz2
  target_path: C:\Users\Chandrajit Rout\.conda\pkgs\pytorch-2.2.0-py3.9_cuda12.1_cudnn8_0.tar.bz2
  Content-Length: 1324785612
  downloaded bytes: 75805634

CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()
CancelledError()

[neon swallow]

I have encountered an error with DreamStudio while generating an image. The error message displayed is 'Something went wrong, please try later.' I have been experiencing this issue since last Sunday, up until 15:00 UTC. Could anyone provide any suggestions or status updates?

CC @north cedar

[lost crescent]

Cannot get WAS nodes in ComfyUI to work. Can anyone help?

It comes up in a red box in ComfyUI Manager ... !

[vocal burrow]

yes I ve noticed that too. I ve passed the message along.

[vocal burrow]

Seems like there was a network error CondaError: Downloaded bytes did not match Content-Length
Also it s still creating a venv in your 'C:\Users\Chandrajit Rout.conda\envs\stylegan3' .... sooooo there s a space. delete this folder, and use conda env create -f environment.yml -p C:\AI\DragGAN\venv instead.

[tranquil vigil]

okey.. let me check...

[vocal burrow]

there s something wrong with the way you re using controlnet

[tiny abyss]

this is how it looks

[vocal burrow]

how did you manage to have nothing selected for those

[ornate elk]

Maybe browser adblocker or other extension

[tiny abyss]

Ok, it seems to say this now, when i select balanced mode

[vocal burrow]

you still havent selected anything for Control Type and Resize mode

[ornate elk]

Also do you use an sdxl or 1.5 model?

[tiny abyss]

This one i made sure everything's se;ected. The buttons are deselecting themselves so i need to press them again and again, idk why its happening, is it the web-ui ths causing problem too?
Also
i get this error when i run (when everythings selected)

[tiny abyss]

1.5

[north cedar]

@vocal burrow please any help you can provide will be deeply appretiated. This issue has been happening for 5 days

[vocal burrow]

seems like SD is confused about what device to use for controlnet, try to force it to use device 0 with --device-id 0

[tiny abyss]

okay..shall i write that in the arguments of web ui?

[vocal burrow]

There s not much I can do I'm afraid apart from passing the message along through the pipes. I ll let you know if there s any informations I can share. At least for now you know that you re not alone on this one and that the messages has been passed.

[vocal burrow]

webui-user.bat

[tiny abyss]

what does it do? its actually going thru a lot of preocess rn

getting this after putting --device-id 0

[craggy scaffold]

It appears to be working now, at least on my end. Is it working for you?

[north cedar]

Yes thank you!!! Our users at Straico where driving me crazy!!

[lusty mauve]

Is anyone aware of a machine learning model that is capable of generating a Stable Diffusion-compatible prompt from an image description?

[vast knoll]

I'm getting this problem 😟 how to fix it

[lusty mauve]

seems like your gpu isnt powerful enough. May I ask which one do you have?
A possible fix could be running SD with different start arguments. There are some for example --medvram or --lowvram
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

[vast knoll]

I have Nvidia GeForce GTX 1650

[lusty mauve]

Yeah thats deffo the problem.
There are some videos on youtube on how to use different start arguments to increase performance.

[vast knoll]

Any links ? To these videos

[lusty mauve]

This video shows how to add those commandline arguments. I cant tell which ones will work the best for you.
https://www.youtube.com/watch?v=daTo6Y5eHgE

[vast knoll]

Thanks brother 👍

[vocal burrow]

try updating your extensions and screenshot them once you ve got them updated please

[ornate elk]

What's your GPU and whats inside your webui-user.bat?

[vast knoll]

Nvidia GeForce GTX 1650

[vocal burrow]

then --no-half --xformers --medvram

[vocal burrow]

add --no-half

[vast knoll]

It's still messing with me

[vocal burrow]

did you restart it ?

[vocal burrow]

?

[vast knoll]

Yes

[vocal burrow]

so did you restart it after adding --no-half ?

[vast knoll]

Yup

[vocal burrow]

show me your webui-user.bat and the log then

[vocal burrow]

I don t see --no-half in there ..........

[vast knoll]

My bad

Should I put it next to medvram

?

[vocal burrow]

yes

[vast knoll]

Should it be --no-half

[vocal burrow]

yes

[vast knoll]

There is no image being generated

I simply added cat in the prompt and still nothing

@vocal burrow

[vocal burrow]

show your webui-user.bat

and what model are you using ?

[vocal burrow]

ok and what model are you using ?

[vast knoll]

Kawaiitech

-20.safetensors

Ooooooooh dam

I think I found the problem

I was wrong

[vocal burrow]

that s not a model

that would be a lora

[vast knoll]

Still have the problem only now
It generates my image full dark

@vocal burrow where can I find such information

[vocal burrow]

you added --disable-nan-check didnt you..... remove that

[vast knoll]

Yes I did

[vocal burrow]

remove it

SD models are at least 2gb

[desert totem]

I can run a lora training in google colab, with 15 images, takes about 10 minutes. How can I see how many "compute units" it used? Meaning if I was to pay $10 for 100, what is that getting me...

[vast knoll]

My ram is 8gb

@vocal burrow so remove the model

?

[vocal burrow]

Use a proper model

and use lora as a lora, not as a SD model

LoRa/Lycoris quick guide :
1/ Where do I put my LoRa : stable-diffusion-webui\models\Lora, Loras are NOT stable-diffusion model but rather patches for said models.
2/ How do I use my LoRa, you can either :

• Invoke it directly in your prompt by using the following syntax <lora:filename:weight> (or <lora:alias:weight> depending of the setting in setting->extra network). eg : <lora:animeTarotCardArtStyleLora_v20Offset:1> for using the Lora animeTarotCardArtStyleLora_v20Offset.safetensors
• Use the UI "Extra Networks" / card system by doing as described in my screenshot. Clicking an associated lora card should automatically insert the lora prompt liked mentioned above. https://cdn.discordapp.com/attachments/1002602742667280404/1191779742534348870/loraHowTo17.jpg
  3/ Be careful, LoRa are trained specifically for some SD version. Lora trained for SD1.5 (and derivatives) models will not load on SD2 models and vice versa (same goes for SDXL). Also even though Lora trained on SD1.5 models should work for every SD1.5 derivatives, they'll probably work best with the model that has been used for its training. Therefore, (by default) only Loras compatible with the model you currently have loaded will show up.
  4/ LoRa, LoHa, LoCon, what are these ? They're all commonly refered as "Lora" but differ slightly technically :
• LoHa : Lora with hadamard decomposition
• LoCon : Lora extended to redisual blocks.
• More explanations : https://stable-diffusion-art.com/lora/, https://www.sdcompendium.com/doku.php?id=lora_0014, https://www.sdcompendium.com/doku.php?id=locon_training_0071, etc

[tranquil vigil]

@vocal burrow after a long work, I was able to install environment.yml file,but in the end, I am facing the same runtime error

[vast knoll]

@vocal burrow thanks brother for all the help

[vocal burrow]

you did install somewhere else with no space as I indicated ?

[tranquil vigil]

yeah...

[vocal burrow]

what s the error now ?

[tranquil vigil]

in C drive only

[tranquil vigil]

@vocal burrow

[vocal burrow]

conda activate c:\AI\DragGAN\stylegan3

or conda activate c:\AI\DragGAN\stylegan3

you get the idea

conda activate fullPathOfTheCondaVenvYouCreated

[tranquil vigil]

tried the same.. seeing the same result

[vocal burrow]

show your try and the associated directory

[neon ruin]

Curious, went to reproduce an image and send the PNG info to text-to-image, however it says it can't find the VAE. The VAE itself is in the correct folder and at first I thought maybe it didn't like the name not being the exact same formatting so I even made a copy and placed the copy with the exact name the prompt was expecting to see in the folder but it just does nothing

[noble belfry]

first: can you select the VAE manually in your editor?

[cyan belfry]

Does anyone know a fix for the dubious ownership problem?

[ornate elk]

Did you launched it sometime as admin?

[vocal burrow]

use a proper filesystem (aka not fat32) or suffer and use the command it tells you to

[noble belfry]

get stability AI out of users and put it on C:

i dont think dragging and dropping will help, i had a similar issue early on when i put it in program files. HOLY GOD did i have to change so many permissions

[vocal burrow]

and yes don t install stuff in users / desktop / any other funky folder than is sync'd with dropbox, onedrive, etc

[noble belfry]

you should not use frankly anything involving programming at user level, too much red tape that coders didnt think of. you personally might be able to click the "ok run as admin" bits, but the code cant and will silently fail.

[cyan belfry]

alright I'll delete it an reinstall to my C drive

[noble belfry]

i quick clean install might be better. I keep all my ai stuff on windows in C:\ai\ if that helps.

well, dont delete the models and stuff you've downloaded

you can pull them over to your new install after. dont throw baby out with the bathwater 🙂

[cyan belfry]

I only downloaded SDXL and 1.5 but I moved them back to my downloads temporarily

[noble belfry]

thats fine, you can move those over. windows is super overprotective of stuff, but if its anything running code and your own machine, where possible allow for all users.

[ornate elk]

follow my install guide in the pinned message of this channel for a clean install

[neon ruin]

Yeah, I tried that as well. Still gets hung up when I go to try and generate

[noble belfry]

^ when in doubt follow CS1o

[ornate elk]

what vae is it?

[neon ruin]

so this gets placed in the override settings, there's no vae selected at the top of the UI at the moment though, seems like regardless of if I select it or not I see the same issue

[ornate elk]

go into settings, user interface, quicksettings, there add sd_vae
then hit apply and reload ui

then you get a dropdown to select the vae

[neon ruin]

I did that already yeah

[vocal burrow]

what s in that folder ?

[ornate elk]

ok then just select the vae and then click on the x at the vae at the bottom at (override settings)

[neon ruin]

I've tried this both ways, where I manually select here from the vae drop down in the UI, then remove like you said yeah

and now it gets hung up

[ornate elk]

maybe cause of the token merging

its zluda so idk if tokenmerging will work

[vocal burrow]

I ll be afk commuting home. in the meantime, screenshot that folder and try source activate c:\AI\DragGAN\stylegan3

[neon ruin]

Ah

Oh well

Yeah I was wondering what I was doing wrong there

[cyan belfry]

Worked flawlessly, thanks

[pseudo summit]

Hello, i returned after a while to SD and i decided to update it. Now i'm unable to use hi-res module anymore cos i get error atached to this msg, my gpu:RTX 3060 12gb

[ornate elk]

it says 1.6.0-2 but the latest version is 1.7.0

so pls show the cmd after you tried git pull to update it

[mighty topaz]

do you know how turn on this dir fodlers? in normal a1111 we have button to show folders but in forge dosent...or forge dont add this button?

[pseudo summit]

i thought updating meant deleting and downloading again vae, how do u do the update again?

[ornate elk]

running git pull inside a cmd inside the stable-diffusion-webui folder

[vocal burrow]

git pull

[pseudo summit]

thx

[winged stone]

what could be the reason of getting images irrelevant with the image on controlnet? i posted about it on #🏞｜general-with-images i would greatly appreciate any help

[pseudo summit]

my output

now i start the webui or should i delete venv again?

[pseudo summit]

now i start the webui or should i delete venv again?

[vocal burrow]

no need to delete venv when you update

[pseudo summit]

k

[dense sandal]

How can I load the deforum settings? it gives me error..

[vocal burrow]

pretty sure you just created a file named deforum_settings.txt.txt

because windows hide file extensions by default for some absurd reason.

[pseudo summit]

back to the original error when i try using hi rez

[dense sandal]

I just removed the .txt and it still gives me error

nevermind. its fixes once I opened stable diffusion on another tab

[vocal burrow]

what settings are you using ? what's the full log ?

[dense sandal]

the settings load weird though, negative and positive prompts are moved into each scheduled prompt.

[pseudo summit]

which setting we talkin about? like start arguments?

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --xformers --autolaunch --no-half-vae

call webui.bat```

[vocal burrow]

no, screenshot your webpage

[pseudo summit]

this?

[vocal burrow]

the whole page

[winged stone]

installing stable diffusion messed up my youtube videos.. am i the only one who encountered this? how do you even fix such thing?

[pseudo summit]

the model is jsut a remix of 2 random models that i made but it work before my attemt to update

[pseudo summit]

maby u run out of vram while watchin yt?

[vocal burrow]

change sampler

[winged stone]

i dont think that is the cause

[pseudo summit]

where's that?

[pseudo summit]

i used to use it all the tiem before withotu any problems though

[vocal burrow]

maybe, still try changing it to see if it fixes it

[pseudo summit]

Seems liek adaptive is out of the window with the hi rez

f

[vocal burrow]

can still generate with it and upscale from it using a different sampler

[junior orbit]

im little bit new to SD i have sd-v1-4. how do i get multiple versions like 1.5 and sd-v1-4 like in check points

[outer adder]

does someone know how to fix the runtime error

[ornate elk]

Hey, on Civitai.com you'll find the most models and loras

Just download them and put them in the correct folders

Models (larger than 2gb) go into models/stable-diffusion
Loras (smaller than 1gb) go into models/lora

[ornate elk]

What's your GPU and whats inside your webui-user.bat?

[slate pulsar]

my images with animatediff + controlnet come out looking like this 1) very ugly and 2) not smooth at all between frames, Any ideas on what I am doing wrong ?

[unique bear]

1.5 models

[pseudo flame]

does anyone know how to setup a good face swap? willling to pay $

[ornate elk]

Do you have a GPU for a local setup?

If you have automatic1111 webui installed you can install Reactor Extension for Faceswap

[quick forge]

I really need some clarification about some things (Linux/AMD)

1. A1111, ComfyUI, Foooocus, etc are all GUI's, so where is Stable Diffusion installed/running from? where is it's install directory?
2. A1111 says version 1.7, is that the A1111 version, or SD version?
3. Stable Diffusion and Stable Diffusion XL..... what is the difference? is one a fork? How does one use SD or SDXL?

[unique bear]

Anyone know if the thin plate spline motion model is limited to just 256x256?

[ornate elk]

1.) mostly on the Home path.
2.) Auto1111 is a GUI to use Stable Diffusion models. The GUI is on version 1.7.0
Stable Diffusion has some base models like 1.5, 2.1 and SDXL.
3.) to use these models you can download them and put them into the models/Stable-diffusion folder

Community made models are based on either 1.5, 2.1 or SDXL

1.5 models got trained on 512x512 resolution, while SDXL got trained on 1024x1024 resolution and better word understanding

[ornate elk]

Yes its limited to 256

[unique bear]

ahh rip

[quick forge]

ok, so I get that they're GUI's, and if SD is mostly installed in the home path, what does that mean?

does this also mean that I could "install" SDXL by default? or is it all about the "models/checkpoints" that is selected in the GUI? Like, the base install of Stable Diffusion is "mostly in the home path" and it's the actual 1.5, 2.1 and XL (safetensor??) files that tell the base install what to do?

[fair oxide]

hey, im using A1111 webui and noticed that it wasnt using my amd gpu.
i had this issue and asked on reddit. i later found that i missed some arch linux specific steps.
but its still not working, does any one have any tip?

[quick forge]

like I want to be running SDXL

[fair oxide]

i think all u have to do is to download an XL model and run it

[quick forge]

makes sense, as I currently understand it

so ive got an sdxl base 1.0 safetensor, and an sdxl 1.0 refiner.
Where am I supposed to use the refiner, because it shows up in the checkpoint list in A1111 and I thought a refiner was used "after" the model?

[ornate elk]

Like I said 1.5 and sdxl are models. The webui is Automatic1111 or comfyui or focus. These webuis are processing the models. So you can use every SD model with a webui

[ornate elk]

In auto1111 there is a refiner checkbox where you can select it

But mostly no one uses the refiner

Upscaling provides better result's

[quick forge]

where is this checkbox?

nvm im dumb

[unique bear]

Is there a setting to get a1111 to autodelete generated pictures once you generate a new batch?

[noble belfry]

not after, but you dont have to have it clog up your anything. you can just tell it not to save files

not specifically to not do something after a batch process, no

like i dont have mine save every image

that said, it will be saved in a temp/gradio if you change your mind

[hazy star]

RuntimeError: Couldn't install torch.
Command: "C:\stable-diffusion\stable-diffusion-webui\venv\Scripts\python.exe" -m pip install torch==2.0.1 torchvision==0.15.2 --extra-index-url https://download.pytorch.org/whl/cu118
Error code: 1

how do I get torch

[quick forge]

this might be anal..... in fact it 100% is, is it possible to rename the checkpoint files to not be so .....fucky?

and somehow exclude the refiner(s) models from the list of main checkpoints?

[hazy star]

nvm done

we chillin

[nocturne topaz]

you can rename the files and when you restart the webui they come through properly

[nocturne topaz]

That one is a no.. unfortunately

[cursive hazel]

I'm not sure this is the right place to ask, but it seems the most appropriate.
I'm using Automatic1111, and I'm trying to do some inpainting touch-ups. I want to do a batch count of, let's say 12 images. But I want to adjust the denoising strength between them. So like, the first 6 will have a denoising strength of 0.2, for subtle changes, and the last 6 will have a denoising strength of 0.5 or something, for more prominent changes.
Is there any way to do something like this? Basically, adjusting some of the variables in a batch, without having to just do smaller batches to do it. (Because I often just walk away from my computer when it's doing big batches!)

[quick forge]

well fk

would someone kindly tell me where the option to have the prompt included as a text file beside the image in the output directory?
I've had a few drinks and can't seem to locate it now.

[lucid trench]

hey guys can i ask if anyone here's on forge webui - and got cascade extension to work in there?

[wispy axle]

Hello! I'm trying to use this https://platform.stability.ai/docs/api-reference#tag/v1generation/operation/masking, but it generates a completely different image and it seems to be ignoring the black masking. Any ideas?

[burnt pond]

Hi everyone. I'm trying to play with forge webui as well. but not all of my VAEs are showing up on the dropdowns. Any ideas? The naming should be fine when I run a regular stable diffusion, the VAEs show up. Just in forge they aren't showing up.

[soft fiber]

You're lookin for the x/y/z script, should be near the bottom of the ui i believe

[valid gorge]

@ornate elk Here's an image of the issue I described earlier, ControlNet (Depth, tested both Midas and Zoe) producing an odd shadow.

[vocal blaze]

Hello I am new to stable diffusion and I having an issue with stable diffusion trying to use FIX FP16 Errors SDXL - Lower Memory use! --- sdxl-vae-fp16-fix

where the colors seem to be weird and inverted for some reason

[valid gorge]

What checkpoint are you using?

[vocal blaze]

perfectWorld_v6Baked

[valid gorge]

Sometimes that happens if you mix SD with SDXL or vice versa. Have you tried using another VAE or set it to automatic or none

[vocal blaze]

I tested it and the picture is fine when I turn the VAE off

Oh I see

[valid gorge]

Looks like perfectWorld_v6Baked is SD 1.5 model

and you're using SDXL VAE with it, that might cause the problem

[vocal blaze]

AH! That makes total sense!

Thank you for explaining it

[valid gorge]

You're welcome, hope it fixes your prolem 🙂

[vocal blaze]

I wonder if there is a SD version of that Vae

[valid gorge]

There should be some available, try searching for vae-ft-mse-840000-ema-pruned

You most likely don't need to use VAE at all, if the image turns out fine without. Usually VAEs are baked in the checkpoints

[vocal blaze]

Oh I see thanks a lot for all the information!

[valid gorge]

You're welcome!

[valid gorge]

Is there a way to accidentally close browser tab in windows? Done it twice already today lol 😂

[vital lance]

Is there any way I can by pass saftey check feature that prevent filtered words or NSFW content via REST API?, I want to use SDXL model

[valid gorge]

Is it the API doing the filtering or your SDXL checkpoint?

[ornate elk]

Hey, Linux would be still the best way for speed+usage.
On Windows the best in Usage and support is with the Directml fork of auto1111.
For raw speed the best is Shark webui.
But there is a new experimental way to get SD running with Zluda (tool to emulate cuda for AMD).
Its amazingly fast for 1.5 generations. Only some extensions won't work and sdxl.

[ornate elk]

Hm have you tried lowering the controlnet Weight?

[valid gorge]

Yep, helps only a little to none

Will experiment more later today. I'm using an SDXL checkpoint with diffusers xl depth full controlnet model, could be the model had a bug and is somehow broken. I'll replace that and see what happens.

[ornate elk]

Ah okay, I think there also mid and low controlnets for sdxl which give an other output

[valid gorge]

Goog point, I need to test out those too!

[lusty mauve]

When I finetune a model which is under license xyz.
Is it still the same model with same licensing?

[valid gorge]

That's a good question. I would say yes, the license stays. Have you checked is there a clause of any kind about this kind of use in the licensing?

[lusty mauve]

Well i havent found any clause like this.
IMO the question is, can the model can trace back to the dataset it was trained on.

[valid gorge]

Yeah

Pls let us know if you find more info about this.

[vital lance]

that's not mentioned anywhere I guess they are giving prompt to another model to filter the prompt

[valid gorge]

Tried already with multiple checkpoints, do all have the same filtering?

Depending of what you're trying to create, you might need LoRA some sort

[viral oxide]

Good morning. I hope I'm in the right place.
I installed stable cascade this morning, but I got an error message in the local instance.
The error is : TypeError: Module.eval () missing 1 required positional argument: 'self'
I have no idea of what I'm doing wrong ^^

(I have no idea of what I'm doing at all) I thought I had to create things, not code things ^^

[mighty topaz]

My friend just install Stable Matrix, and A1111 but have GPU from AMD and stable matrix or a1111 use cpu gpu dosent work, can someone share links with tutorial how install amtrix and a1111 for amd?

[ornate elk]

Stability Matrix doesn't support AMD.
Your friend need to follow one of my AMD install guides. They are in the Pinned Messages of this channel.
I suggest to first install the normal Directml fork and if he has a newer GPU, he can later try out the experimental Zluda Version.

[mighty topaz]

Ok, but matrix is only app, so if he install webui a1111 via matrix, sd dont work? because he tell me he install matrix and sd but use only cpu not gpu

so he must leave matrix, and isntall raw local sd without matrix?

[ornate elk]

Yes, matrix won't work

He needs a normal installation for AMD

By following my guide

[mighty topaz]

ok thank you

[ornate elk]

Guide is here:
#🤝｜tech-support message

[hollow crescent]

Hi I can't seem to figure out how to solve this error I've checked the permissions for that file full access is granted. May I get some help please

[proud walrus]

hey guys, i've been working on a render but ı dont't wanted to make it look different . don't wanna change anything, same building but more realistic. i'm using control net but still cant get the similar and more realistic image of my input

[vocal burrow]

try rebooting your computer then delete venv, then run it again

[hollow crescent]

I've tried doing a fresh install 3 times, even upgraded my OS from win10 to win11 just in case

[vocal burrow]

might be more suited for #📝｜prompting-help (and they'd need to the settings used for your generation)

[proud walrus]

oh thanks

[vocal burrow]

ok, run pip cache purge, delete venv and try again

[fathom rune]

old man trying something new. How do I get started in stable diffucsioon. Can I add it to photoshop?

[vocal burrow]

There are some plugin to have some sort of stable diffusion UI/programs interact with photoshop but I wouldn't recommend starting with those.
Rather I'd send you towards one of these :

• a1111's stable-diffusion-webui : https://github.com/AUTOMATIC1111/stable-diffusion-webui/ the most popular solution out there, biggest community and support
• comfyUI https://github.com/comfyanonymous/ComfyUI : second most popular solution, more tweakable but more complex to use
• StableSwarm https://github.com/Stability-AI/StableSwarmUI : a mix of the two above, officialy maintained by stability team
• Fooocus https://github.com/lllyasviel/Fooocus : probably one of the easiest solution out there for people who just want a big button to press and generate "pretty" stuff, but also one of the less tweakable

It also depends on what hardware you have

And then once you get the gist of how one of those work by itself it you'll understand how to actually use it from photohsop using some plugins.

[ornate elk]

Have you edited the webui-user.bat before launching it?

[left iron]

oi, what on earth does latent modifier do? I can't find ANY documentation for it at all despite it being integrated into SD WebUI Forge

I can only find one mention of it on this whole server, by a user who's not online r/n

and, uhhhhhhh

[oblique quail]

is it normal to train 3000 steps in 1.5 hours ? with rtx 3090

sdxl lora

[molten coral]

which py version is best for forge?

[fair oxide]

how can i train loras locally using directml?

or how to train loras online (for both characters and styles)?

[mighty topaz]

on matrix discord they answer like that

also i see forge is supported in matrix

[fair oxide]

how would forge compare in speed to a1111?

[mighty topaz]

Compared to original WebUI (for SDXL inference at 1024px), you can expect the below speed-ups:

If you use common GPU like 8GB vram, you can expect to get about 30~45% speed up in inference speed (it/s), the GPU memory peak (in task manager) will drop about 700MB to 1.3GB, the maximum diffusion resolution (that will not OOM) will increase about 2x to 3x, and the maximum diffusion batch size (that will not OOM) will increase about 4x to 6x.

If you use less powerful GPU like 6GB vram, you can expect to get about 60~75% speed up in inference speed (it/s), the GPU memory peak (in task manager) will drop about 800MB to 1.5GB, the maximum diffusion resolution (that will not OOM) will increase about 3x, and the maximum diffusion batch size (that will not OOM) will increase about 4x.

If you use powerful GPU like 4090 with 24GB vram, you can expect to get about 3~6% speed up in inference speed (it/s), the GPU memory peak (in task manager) will drop about 1GB to 1.4GB, the maximum diffusion resolution (that will not OOM) will increase about 1.6x, and the maximum diffusion batch size (that will not OOM) will increase about 2x.

If you use ControlNet for SDXL, the maximum ControlNet count (that will not OOM) will increase about 2x, the speed with SDXL+ControlNet will speed up about 30~45%.

[fair oxide]

like, if one image woukd take 5 minutes to generate, i could reasonably expect the same generation in forge to be roughtly the same as a mid-level card (roughtly a minute and a half)?

[mighty topaz]

something like that? i dont know exacly i have 12gb, but in forge is faster then usulally in a1111

[ornate elk]

Yes, and my guide is exactly for the Directml fork of auto1111 made by lshqqtiger. But my guide features the needed performance steps to get the best speed on it too.
And my other AMD guide uses the new Zluda experimental Backend. Which Isnt supported by matrix

[ornate elk]

I think the real increase is not that much

What's your GPU?

[mighty topaz]

@ornate elk but if i install forge and pick directml in matrix like here in screen, i dont need then instal forke right?

[ornate elk]

Forge won't work with AMD, I've tested it yesterday

Its still bugged

[mighty topaz]

oh ok

[ornate elk]

I mean you could try it. Maybe I did it setup wrongly or matrix does something to make it work.
But I also dont recommend forge for AMD in general

[mighty topaz]

i understand so simply only left forke a1111

without matrix, to bad because matrix has nice options, like civii in it, no need to open more tabs in internet:D And autoinstall everything, folders

[ornate elk]

True, but I prefer having more tabs than actual extra software. And the full control of the Installation

So yea options are Directml fork of auto1111 or directml fork with Zluda.
Or SD.Next (didn't tested it)

[mighty topaz]

Sure, everyone likes something else..but i tell firend nad he isntall fork

[fair oxide]

amd radeon hd 8690m

[viral oxide]

Good afternoon! I hope I'm in the right place.
I installed stable cascade this morning, but I got an error message in the local instance.
The error is : TypeError: Module.eval () missing 1 required positional argument: 'self'
I have no idea of what I'm doing wrong ^^

[ornate elk]

Does your folder is named
Stable-diffusion-webui or stable-diffusion-webui-directml
?

[fair oxide]

stable-diffusion-webui-directml

[ornate elk]

Okay, then I dont think you would see a huge performance increase with forge. For me on AMD it didn't even worked

[fair oxide]

well, is it worth switching forks for a better speed? if yes, what would you suggest? what optimisations would you be able to suggest for me to make generations faster?

[ornate elk]

For that card you have the best webui already

What's in your webui-user.bat?

Maybe you can improve with some commands there

[fair oxide]

--lowvram --no-half --precision full --opt-sub-quad-attention --opt-split-attention --disable-nan-check --skip-torch-cuda-test --theme dark --autolaunch

[ornate elk]

Oh its actually using your CPU instead of GPU.
remove --disable-nan-check --skip-torch-cuda-test

But add --use-directml --skip-ort

[fair oxide]

oooooh

maybe that's why it takes so long

[ornate elk]

Exactly

[fair oxide]

i'll test it with the improvements i have so far (lcm lora, fewer steps)

got this error

Traceback (most recent call last):
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\launch.py", line 48, in <module>
main()
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\launch.py", line 44, in main
start()
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\launch_utils.py", line 677, in start
import webui
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\webui.py", line 13, in <module>
initialize.imports()
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\initialize.py", line 34, in imports
shared_init.initialize()
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\shared_init.py", line 28, in initialize
directml_do_hijack()
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\dml_init_.py", line 76, in directml_do_hijack
if not torch.dml.has_float64_support(device):
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch_init_.py", line 1932, in getattr
raise AttributeError(f"module '{name}' has no attribute '{name}'")
AttributeError: module 'torch' has no attribute 'dml'

maybe i need to install torch-directml?

[ornate elk]

You need to delete the venv folder

Then relaunch

Then it should install torch-directml automatically

[fair oxide]

so far, the avg iteration time has halfed

nvm, gpu video memory error

Traceback (most recent call last):
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\scripts\xyz_grid.py", line 701, in cell
res = process_images(pc)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\processing.py", line 736, in process_images
res = process_images_inner(p)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\processing.py", line 969, in process_images_inner
x_samples_ddim = decode_latent_batch(p.sd_model, samples_ddim, target_device=devices.cpu, check_for_nans=True)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\processing.py", line 597, in decode_latent_batch
sample = decode_first_stage(model, batch[i:i + 1])[0]
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\sd_samplers_common.py", line 76, in decode_first_stage
return samples_to_images_tensor(x, approx_index, model)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\sd_samplers_common.py", line 58, in samples_to_images_tensor
x_sample = model.decode_first_stage(sample.to(model.first_stage_model.dtype))

File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\sd_hijack_utils.py", line 17, in <lambda>
setattr(resolved_obj, func_path[-1], lambda *args, **kwargs: self(*args, **kwargs))
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\sd_hijack_utils.py", line 28, in call
return self.__orig_func(*args, **kwargs)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\utils_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\repositories\stable-diffusion-stability-ai\ldm\models\diffusion\ddpm.py", line 826, in decode_first_stage
return self.first_stage_model.decode(z)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\modules\lowvram.py", line 71, in first_stage_model_decode_wrap
return first_stage_model_decode(z)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\repositories\stable-diffusion-stability-ai\ldm\models\autoencoder.py", line 90, in decode
dec = self.decoder(z)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\repositories\stable-diffusion-stability-ai\ldm\modules\diffusionmodules\model.py", line 637, in forward
h = self.up[i_level].block[i_block](h, temb)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\repositories\stable-diffusion-stability-ai\ldm\modules\diffusionmodules\model.py", line 131, in forward
h = self.norm1(h)

File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\extensions-builtin\Lora\networks.py", line 516, in network_GroupNorm_forward
return originals.GroupNorm_forward(self, input)
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\modules\normalization.py", line 273, in forward
return F.group_norm(
File "Z:\AI SHIT\SD\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\functional.py", line 2530, in group_norm
return torch.group_norm(input, num_groups, weight, bias, eps, torch.backends.cudnn.enabled)
RuntimeError: Could not allocate tensor with 201326592 bytes. There is not enough GPU video memory available!

yet, it still continued making the images (i was running an x plot with 5 images, 5 steps per image with lcm lora, euler a, cfg 1)

like idk, it stops, but then it tries again

it's weird

[ornate elk]

Thats to much for the vram

[ornate elk]

If your on a laptop with a lot of RAM you can give the GPU more vram in the BIOS settings.

But first restart SD and try a normal one image generation without xyz

[fair oxide]

i mean, i have 12 gigs

so maybe 4 more gigs wouldn't hurt

[ornate elk]

You can try sure

Also use 2gb models

And Euler a as sampler

[fair oxide]

wdym by that?

[ornate elk]

Most newer 1.5 based models are 2gb in size.
Using models that are larger will take much longer or even fail

For example sdxl models are 6gb and wont run on Directml with that gpu

[fair oxide]

i have photon with 2.08 gb and colormix with 2.3 gb

i'll try to install some smaller models

[ornate elk]

There are no smaller models. Around 1.98gb is the smallest I've seen.

Everything below that is a lora, embedding etc and not a model

[fair oxide]

ig i should just find models that are around 2gb, right?

[ornate elk]

Dreamshaper v8 for example or juggernaut reborn

Or realistic vision

[fair oxide]

is using lcm sampler with lcm lora going to make a huge difference when generating images?

in speed as well as quality?

[ornate elk]

You dont need LCM sampler when your using the LCM lora

[fair oxide]

alright then

i've seen how you can copy a style from an image using ip-adapter. how about characters? can it do that? how well?

[violet inlet]

guys? can you help me with finding and installing am offline chatbot? i want try to use it in translation and learning other languages

[fair oxide]

increased vram using regedit. tried to do another pic. i got the same memory allocation error

[violet inlet]

seriously? the lora don't go in tandem with the sampler?

you can still gen with 4 steps well, using other samplers?

[west rune]

im getting errors everytime i try training a hypernetwork. Is it broken rn or is it my stuff thats broken?
anyone wanna help me resolve my training issues?

[ornate elk]

Yes

There are also models thst have LCM included

Can be used with any sampler

[ornate elk]

GPT4all is very nice and easy to setup

[violet inlet]

ok, that's new, i suppose i need to work on the lora weight, to make it work better with other samplers

[ornate elk]

The LCM lora needs 0.5 mostly

[violet inlet]

i stiked with lcm sampler, because the others gave me rough results, at low steps

0.5 0.7

[ornate elk]

Yes thats possible too with IP-Adapter, you can also copy the face with it

[violet inlet]

i played a lot with lcm, because it's very good to set a pose

[cosmic furnace]

What r best & most used upscalers?)

[violet inlet]

bery rough on details, but very responsive on character poses

[fair oxide]

trying forge webui now, since the normal a1111 showed errors even after vram i crease via regedit

[ornate elk]

I dont think regedit can increase vram 😮

[fair oxide]

well, i think i used it at one point, since i didn't seem to find an option to oncrease vram from bios

maybe i didn't look in the right place

[cosmic furnace]

sounds like magic)

r u having somekind of error, while system not showing your real vram? Or u being sirious about just increasing vram via bios-regedit hm

U cant alter physical objects via soft)

Well u can make some turn off-on light and stuff...

[ornate elk]

REsrgan 4x, 4x UltraSharp

Nmkd 200x siax

Are often used

Remacri too

And Fatal Anime 5000

[fair oxide]

https://huggingface.co/uwg/upscaler/tree/main

@ornate elk what resolution do you usually generate images in?

[ornate elk]

512x768

[fair oxide]

i found out 600x912 resolutions gives nice enough space for image content w/o distorting anything, and works great for both SD and XL

[cosmic furnace]

What will happen if try sdxl with zluda? 😄 This?

https://tenor.com/view/explosion-mushroom-cloud-atomic-bomb-bomb-boom-gif-4464831

[left iron]

@nocturne topaz Hey, you're the only person I've found who seems to know anything about LatentModifier
Can you give me any tips/explanations?

[ornate elk]

Yes xD driver crash

[ornate elk]

Ohh nice, I'm gonna try that later!

[fair oxide]

yeah and hires fix by 2x

[west rune]

I need some tech support for training a hypernetwork

It just throws errors at me

Im using a1111 on stability matrix

and im getting this:

any other info needed?

[fair oxide]

error on forge webui for amd:

[fair oxide]

what is the fastest webui for low end gpus?

[upbeat atlas]

Hi, I have a question regarding stable diffusion for AMD. I'm currently running automatic1111 directml, and while it does work, it's very slow and has so many errors. I was wondering what would be a better option for AMD windows instead of automatic1111 directml?

[ornate elk]

Directml is the most compatible one with any extensions.

What errors do you get?

Also what's your GPU?

[nocturne topaz]

The github has a pretty good explainer on what most if em do. I quite often use the Sharpness one and the extra noise one

[left iron]

I can't find the github

I know this sounds like a strange problem, but I'm struggling

Due to it being built into forge, it doesn't have a link in the extensions

[nocturne topaz]

https://github.com/Clybius/ComfyUI-Latent-Modifiers

[left iron]

Okay, phew
See, I was searching for "sd-webui-latent-modifier" since in forge it was called "sd-forge-latent-modifier"

I did not know it originated as ComfyUI nodes

Thereby I had just no clue where on Earth latent modifier came from and where to find docs

[winged stone]

i remember someone saying that i should put model downloads to different folders depending on their size. is that true or should i just put all model/chekcpoint stuff to models folder?

[left iron]

Thank you @nocturne topaz
I've also managed to find tips from people on how to use FreeU

right now I run B1:1, B2:1.05, S1:0.5, S2:0.2 on SDXL

[nocturne topaz]

Not dissimilar here (1,1,0.9,0.2 for sdxl)

[left iron]

latent modifier does seem to be a bit not-what-i-need rn, tried to run it with none of its stuff enabled and it deepfried the image
already running dynamic thresholding so the high cfg stuff is already handled

[nocturne topaz]

Ah yeah, dont use em together 😛

But there is similar stuff in there to achieve the same effect

[left iron]

also, is latent always the best upscaler to use with hires fix? are some better for certain samplers/art styles/SDXL?

[nocturne topaz]

No, it tends to introduce a lot of noise whic is good for getting more details but needs more denoise and steps to get back to a sharp image

[left iron]

ahh yeah i'm noticing a lot of additional random stuff added with latent+freeu if i dont VERY explicitly spell out the background

[low tiger]

hey guys, just installed SD with a anime model (AAM_XL_Anime_Mix). when i make images they are totally fine while they are being made, but when it hits the end and shows the final product they all look like this (images attached)

[nocturne topaz]

Change the VAE

[left iron]

yeah that kind of noise is 100% a sign of a mismatch between model and vae

or, occasionally, too strong of a lora that has been trained on a totally different style

[nocturne topaz]

Yeah the real esrgan ones do a good job for hires fix

[low tiger]

oh ok, i got the VAE from the yt guide he said itd work with all. how would i know which will work with these specifically?

ty for the help btw 🙂

[left iron]

like R-ESRGAN 4x+/4x+ anime6B?

[nocturne topaz]

You are looking for a sdxl specific vae (search foe “fp16 fix” on civitai)

[low tiger]

thank you so much, downloading one now 🙂

[snow swallow]

Is sdxl more demanding in therms of graphics? I have an AMD, not sure it would work but I'm curious if this model necesserely uses more machine

[fair oxide]

what webui is the fastest and easiest to install for low end amd gpus?

[gray oyster]

not related to stable diffusion but does anybody know how to fix these errors?

tried installing with requirements in oobabooga since the installation failed because of the exact same module

[elder raven]

Haven't logged into Stable Audio in a while and got this, anyone know whats up?

[nocturne topaz]

delete the cached file at that location and try again?

[gray oyster]

that might be an idea, i will see if that works, i just wonder why it happens in the first place though

[nocturne topaz]

file corruption, or wrong whl for the system

[gray oyster]

fair

[gray oyster]

probably the latter

I also keep getting 'PyTorch not compiled with CUDA' despite entering A for nvidia gpu and N for CUDA 12.1 which is what I have installed

I did have it working at some point but after auto-updating it it broke everything

awesome, finally works

[gray oyster]

thanks a lot

[nocturne topaz]

no worries 🙂 pytorch not compiled with cuda usually means that the index-url parameter wasn't supplied when installing torch so it didn't go and get the +cu121 version of torch

[silent sparrow]

what gpu has the fastest training time for dreambooth fine tuning?

[nocturne topaz]

whichever gpu has the most cuda cores - so in descending order h100, a100, v100, a6000, 4090....

[silent sparrow]

sorry. So h100 is the best?

[nocturne topaz]

that is available to us (via runpod) then yip

[silent sparrow]

thanks

[nocturne topaz]

a 4090 is more than fast enough though (unless you have a couple of hundred thousand images in your dataset!)

[silent sparrow]

I'm training an xl model with 70 images and it says an hour and a half

thats on kohya_ss

[fair oxide]

what do i use in order to run stable diffusion and get decent features with unlimited generations?

what mobile nvidia gpu would be a good budget option for ai?

or more specifically what laptops?

[nocturne topaz]

Dont bother with buying a laptop for SD unless you were already but generally vram is key (3060,3090,4080,4090)

There are enough cloud services out there to keep you going in the meantime: runpod, vast.ai, paperspace

[fair oxide]

then, what to use? online options are crap, they limit you in severe ways

so much that it's just a trial

the only option is using a1111 locally using my cpu

[nocturne topaz]

Those 3 above are cloud pcs so running it like you would on your machine unrestricted

[fair oxide]

or to somehow use more vram, by sharing ram with dedicated video ram

[fair oxide]

for free?

[nocturne topaz]

Cheap, not free

Like 0.20c/hr cheap

(Runpod, community cloud, 3090 spot pricing)

[fair oxide]

i mean, it still costs. i do think there are options out there that still allow you to have decent features (like lora and custom models and all that stuff) with unlimited generations

it's a matter of preference for me

do you know some alternatives i could use?

[mighty topaz]

can someone help me with error? when i want generate animation in a1111 i got

[nocturne topaz]

Happyaccidents is close

[fair oxide]

are there other options that don't use tokens?

[nocturne topaz]

For free, probably not, gpu compute isnt cheap

[fair oxide]

or with decent free tiers?

[nocturne topaz]

Civitai (but does have daily limit of like 3000 images)

And no upscaler/img2img

[fair oxide]

do you know how you could get those?

[ornate elk]

yes, much more on amd windows

[ornate elk]

Stable Horde, but it has limited features

[upbeat atlas]

I wanted to use ONNX with automatic1111, so I tried to re install according to what the official article said.
I just keep getting "torch is not able to use GPU"
When I add "--use-directml" with "--onnx" it doesn't recognize onnx as a command

[ornate elk]

yea the official guide is outdated

i dont recommend onnx

you would have to convert models everytime you switch them, also its locked on specific resolutions

it doesnt support the most extensions

You would be better to use Shark if you just want Speed, but it has the same drawbacks as onnx.

What i recommend is using the directml fork but with Zluda backend. Its crazy fast, and compatible with most of the stuff (but not all)

[upbeat atlas]

Wish I knew this 3 hours ago 🗿

[upbeat atlas]

Oh oki thanks, gotta check out how to use it

[ornate elk]

you only have needed to reply to my question 😛

whats your GPU?

because Zluda doesnt support RX 5xxx cards

[upbeat atlas]

RX 6900 Xt

[ornate elk]

i have a install guide in the Pinned Messages of this channel

[upbeat atlas]

Fair enough 😭

[upbeat atlas]

Oh thanks then I'll just check that out then

[ornate elk]

let me say what doesnt work with it (sdxl models, Adetailer, and depth, ip-adapter controlnets), but for 1.5 its so fast you can generate at any resolution without getting out of vram and with controlnet and loras

[upbeat atlas]

That's absolutely cool then, cause I don't use sdxl anyway and adetailer the others I barely ever use. I just hope it works fine

[ornate elk]

yes, i tested it and it works pretty good!

from 4it/s with directml to 16/its with 7900xtx

an upscaled image with 10 hires steps takes now 20 seconds instead of 1:20

If you have any questions about the setup feel free to ask

[upbeat atlas]

Damn that's SO FAST, thanks again for the help, can't wait to have a go at it 😁

[mighty topaz]

@ornate elk i have problem with animation in sd a1111 this is fram animated by other guy and he share this image to others with data to try do the same animation

and when i try generate with this same data everything i got

something is with colors contrast and stuff

[ornate elk]

whats your vae file?

[marble ore]

can someone tell me what the score_9, score_8_up, score_7_up, score_6_up tags do?
ponydiffusion model*

[mighty topaz]

i try automatic, ft-mse-840000-ema-pruned also kl-f8-aniame2vae

with module mm_sd_v15v2 and others

[mighty topaz]

u have description in civii in this model by author

[marble ore]

?

its a model based on pony diffusion

[mighty topaz]

https://civitai.com/models/257749 read description here

[marble ore]

thank you!

[fair oxide]

does zluda support any amd gpu?

[ornate elk]

nope

only hip (rocm) compatible ones

[mighty topaz]

i try automatic, ft-mse-840000-ema-pruned also kl-f8-aniame2vae, do you know what happen? please 😐

[ornate elk]

hmm for me it seems the vae doesnt get applied

[mighty topaz]

yeh, but why vae dont want to animation applied? because on image worsk always good hmm

[ornate elk]

idk :/ what extensions are you using

[mighty topaz]

https://github.com/continue-revolution/sd-forge-animatediff

i dont do this because i dont know where is this terminal an whre put this commands but i thinks this is not because of it i think

[nocturne topaz]

do you have forge installed already, or a1111?

if you already have a1111, then you want the sd-webui-animatediff extension

[upbeat atlas]

Hi, I tried installing Zluda, in the pin it said to run it with "--medvram-sdxl --update-all-extensions --update-check --skip-ort"
I added the arguments but it says torch is not able to use gpu, how do I fix that? I even added --use-directml but that didn't work either

[nocturne topaz]

have you reinstalled torch since installing zluda?

[mighty topaz]

i have forge thats why i install extension for forge animationdiff

[nocturne topaz]

If you already have forge, then you can skip that step

that just installs forge

[mighty topaz]

oh ok,

but colors are bad i dont know why:/ csgo also dont know hmmm its weird

[upbeat atlas]

No I didn't, should I?
Also I'm not sure if it's relevant or not but I'm using AMD gpu

[nocturne topaz]

colours are bad in the end result.. that could be a VAE issue, but more likely that it's just animatediff being animatediff and not very good on a1111/forge

comfyUI gives better results for animatediff, and if animation is something you plan on doing lots of, I'd probably investigate switching - there are way more options available in comfy for animatediff and svd

[mighty topaz]

yeh but look at this i go step by step from tutorial look his amazing aniamtion and also in a1111 i set up you time on timeline

https://youtu.be/Gz9pT2CwdoI?si=wgR8sDmA2VaRt73d&t=27

[ornate elk]

no thats the wrong guide

follow my guide for zluda please

not the one for directml

[upbeat atlas]

Shoot, my bad didn't notice xD

[upbeat atlas]

I'm sorry I got a bit confused, is this not the guide?
#🤝｜tech-support message

[ornate elk]

oh yes, thats it, and yes your right sry, use the commandline args there

have you done the torch reinstall ?

dont add --skip-torch-cuda

[upbeat atlas]

No I didn't, so I just have to reinstall torch?

[fair oxide]

yea, i give up

[mighty topaz]

ehhhh jesus ...

i think maybe reinstall everything helps

[ornate elk]

you need to follow the whole guide

i think you only did the preparation steps

[upbeat atlas]

...
Sorry for the stupid questions yes I only did the preparation steps I didn't notice the following steps 🗿💔

[ornate elk]

editet the guide, launch it one time with --directml added too, and remove it after the first installation, before doing the zluda stuff

[nova ermine]

I was generating stuff like this, however I've returned to SD, got that image in PNG info, using exactly the same parameteres and...

Got this.

How to fix?

[ornate elk]

can oyu show your txt2img settings?

[serene oar]

Hello! Im havng trouble with automatic1111

i get this message every time i try to use it

venv "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 601f7e3704707d09ca88241e663a763a2493b11a
Traceback (most recent call last):
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
main()
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\launch.py", line 39, in main
prepare_environment()
File "C:\Users\Usuario\Downloads\Stabble diffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 560, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Press a key to continue. . .
No names found and torch is not able to use GPU

[ornate elk]

whats your GPU?

[serene oar]

An AMD 6500, but i have the version of automatic that should be compatible

[ornate elk]

yes, but you need addtional steps to make it work

[noble belfry]

also who tells people to skip torch cuda test?

[serene oar]

What is that and how do i enable it?

[serene oar]

Please help!

[noble belfry]

chill. would you like me to link a video to follow?

[patent spear]

Probably not the issue, but I noticed you decreased the number of steps in your second/undesirable image down to 60 (your first image has 100 steps), the resolution is also slightly different. Have you made sure your settings stayed the same too? I know that some optimization settings like prompt padding and xformers also change the results. Oh, also when I changed my hardware (GPU), the results changed too... then again I did a whole new webui installation too.

[noble belfry]

it might be easier than text instructions. and dont worry, the first time installing AI stuff is like childbirth

[serene oar]

Yes please

[ornate elk]

first you need to open up a cmd and type
pip cache purge and hit enter
then delete the venv folder inside the stable-diffusion-webui-directml folder.
then Edit the webui-user.bat (right click), At the line COMMANDLINE_ARGS= You add: --use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling
then save and launch the webui-user.bat

[nova ermine]

I don't know how, but the problem has "magicaly" got solved

[nova ermine]

Before the problem "magically" getting solved I thanged that parameter and the results were the same

[serene oar]

C:\Windows\System32>pip cache purge
WARNING: No matching packages
Files removed: 0
It tells me that it didnt remove anything

[noble belfry]

were you using controlnet? did it stop halfway through?

[nova ermine]

Yes, and I disabled it to check if it was causing the problem, but the problem persisted then

[noble belfry]

looks likes a canny or mlds version of an image, it might have stopped early. do you still have the window open?

[ornate elk]

thats okay

[noble belfry]

it might be crashing, do you have less than 12gb vram?

[ornate elk]

do the other steps then

[mighty topaz]

one thing but not my problem friends problem, sorry for bothering you guys, he also want do aniamtion in forge in diffusion and he dont have problems with vae, have normal colors but when he generate no aniamtion just frames how fix it?

[noble belfry]

if you do have low vram theres a small bandaid you can use with nvidia. if you right click your desktop and open control panel settings for nvidia it should give a load of options

[nova ermine]

How does I do that and what's the aim of that options?

[noble belfry]

in the middle of those, theres one called "CUDA - sysmem fallback policy" you can set that as "prefer sysmem"

now it wont let you make bigger images or anything, it will just stop everything dying in the middle of a job. it will throw the job into RAM from vram. that makes it super super slow, but if an image is close it might be helpful.

should be here

(win 10 version so might look different on win 11)

also small things like if the image is big, close your chrome browser. you wouldnt think but it does eat away at lots of resources

i dont think if you have vram less than 6 you should use that setting. up to you though

its not a full fix, its more like an overflow in a bathtub

[ornate elk]

let me know if it worked

[nova ermine]

I see the option in tab, but I still don't understand what will change if I selects "prefer sysmem"

[noble belfry]

it should stop CUDA crashing as it should load the job into RAM instead of failing. Its an option im just giving you.

you dont have to use it at all

[upbeat atlas]

@ornate elk hi, so I did everything, and now when I open it, it starts without a problem at first, but as soon as the interface comes up, it takes around a minute or two to load the sd model, and when I want to generate it doesn't generate anything and after a while says server disconnected without sending a response. I'm not using any loras or embeddings just plain text with 20 steps of generation. Any idea how to fix it?

[ornate elk]

whats the model?

[upbeat atlas]

popPopcornMix, it's a 1.5 sd model

[low tiger]

hey guys, SD isnt following the controlnet instructions, i am fairly new to this sorry if its something simple i cant figure it out. i will attach screenshots, hopefully someone can spot what im doing wrong 🙂

[upbeat atlas]

You didn't download the models for controlnet that's why xD

[low tiger]

oh really xD i thought i did, i will look into this now thank you!

it is these ones? : https://huggingface.co/lllyasviel/sd_control_collection/tree/main

[upbeat atlas]

Below the preprocessor there's 2 sections, the one on the right is for the downloaded models and are necessary.

[upbeat atlas]

Yes

[low tiger]

thanks i really appreciate that, im a brand new rookie my bad.

[ornate elk]

can you check the taskmanager

[upbeat atlas]

No problem, I've been there as well, just remember to add the models into
Stable diffusion folder-Extensions-controlnet-models

[low tiger]

thanks dude, that wouldve likely been my next question xD

[ornate elk]

the linked models, they are for sdxl

use these : (for 1.5)
https://civitai.com/models/38784/controlnet-11-models

[low tiger]

i think i am using xl

im not 100%

[ornate elk]

ohh right, but in the wrong resolution

got confused of 512x768 because sdxl models are trained on 1024x1024

[low tiger]

what does this mean, so sorry

oh

ok yeah will resize thank you very much!

[ornate elk]

no problem

[nova ermine]

So, that opcion "turns" RAM in VRAM if I run out of it?

[noble belfry]

yeah.

[upbeat atlas]

What exactly should I be looking for in the task manager?

[noble belfry]

its not optimum but its not the worst thing.

[ornate elk]

the Vram usage and RAM usage

[low tiger]

1 more question: if i wanted to use this to make 32:9 super ultrawide wallpapers, is that possible?

[noble belfry]

well it actually turns vram to ram rather. which is super slow

[upbeat atlas]

It's VERY low, as if it's not using anything at all.
Vram is at 4.2/16 which is normally at 15.2/16 when making images ):

[ornate elk]

thats okay

it uses less vram than directml

[ornate elk]

what does the cmd shows

[upbeat atlas]

It's been stuck on this for about 5 minutes now

[ornate elk]

ohhhh

i have to say the first generation takes up a lot of time

like 20-40 minutes max

[upbeat atlas]

And ofc as I sent that it magically started working

[ornate elk]

perfect xD

[upbeat atlas]

Thanks so much for the help, ima check this out abit more incase there's any other problem, ❤️🫡

[ornate elk]

sure, can you generate a 512x512, 30 steps ?

how is your it/s

[upbeat atlas]

Ok so I made a few images and I'm already having some problems
First is that after an image is made it's not shown in the interface right away, after a minute or so it just randomly appears. Second is that after an image is done generating, I can't generate an image right away. For some reason I have to wait a while before making another one.
That being said its it/s is very good, it's very fast with 512×512 and 512×768 (at least compared go automatic1111)

[ornate elk]

okay, strange issues 😮

make sure the webui is whitelisted in any browser adblocker, also maybe restart the PC one time

[upbeat atlas]

Is it by any chance unloading and re uploading the sd model every time I generate an image? While waiting to make another image I made another tab for the ui and it was uploading the sd model again

[ornate elk]

i dont think so

[ornate elk]

how much Ram do you have?
is everything installed on an SSD?

[upbeat atlas]

16 GB for both ram and vram

[ornate elk]

and installed on an SSD?

you may need to increase your windows pagefile, or check if its only active for C. (what it should be)

[upbeat atlas]

Yes on ssd,
Also I just checked

Does that mean there's no pagefile for the D drive?

[ornate elk]

Yes and thats good

[cosmic furnace]

this is weird

[ornate elk]

C pagefile is 21 GB currently, that's okay too

[cosmic furnace]

pretty sure followed all steps 1 by 1 right with zluda install and i get this

[upbeat atlas]

Did you add AMD HIP SDK bin file to the path? That's what was messing it up for me

[cosmic furnace]

this?

[upbeat atlas]

Wait that's what we were supposed to add? How did you find this? This is what I added

[molten coral]

I want to train my own pixel art model/lora but I've never really done this before. Is it better to resize the pixel art textures to 768x768 or keep the size at like 64x64 and make the rest pink or train the model on 64x64px images?

[ornate elk]

Yours should work too

[cosmic furnace]

dunno. but it was in guide o.0

rocm bin looks more specific hmgm

[ornate elk]

Have you run the commands to reinstall torch ?

[cosmic furnace]

yeah 🤔

pretty sure did... might missed it hmgm

[quick forge]

well ive fucked something up and now my computer doesn't have any display, i just get 1 long and 3 short beeps anyone have any insight?

[ornate elk]

Yea please recheck, also What's in your webui-user.bat?

[cosmic furnace]

only thing i did before

is updated pit

or how you called it

pip)

it said update pip cuz theres newer...

prior to reinstalling, mb ill shud use reinstall torch again? hmgm

set COMMANDLINE_ARGS= --medvram-sdxl --update-all-extensions --update-check --skip-ort

[ornate elk]

Okay commands are good

Updating pip shouldn't be an issue

[cosmic furnace]

also zluda iz 3.4 not 3.2 not like in guide

thought it doesnt matter because guide was older than newer zluda's

[ornate elk]

Thanks for mentioning that

I didn't tested the new Zluda version

[cosmic furnace]

🤯

mb i shud repeat the 4 step with instal-uninstall torch?

[ornate elk]

Yes exactly

Run the commands again

[cosmic furnace]

pretty sure did it but might skipped 1 step because got distracted. But pretty sure did them hm. all 3

Run the following commands one by one:
.\venv\Scripts\activate
pip uninstall torch torchvision torch-directml -y
pip install torch==2.2.0 torchvision --index-url https://download.pytorch.org/whl/cu118

seems like it didnt worked indeed

[ornate elk]

The first command takes you into the venv

Then it should say (venv) at the beginning in cmd

When that shows you can paste the next command

[cosmic furnace]

yeah pretty sure i saw venv last time. mb didnt give it enough time last time o.0

[ornate elk]

What's your GPU btw?

[cosmic furnace]

7900 xt

[serene oar]

It worked! But now im getting this error sometimes

RuntimeError: Could not allocate tensor with 268435456 bytes. There is not enough GPU video memory available!

Also, how can i make my images more ...well good

[cosmic furnace]

How long all this uninstall-instal usually takes?

[ornate elk]

What's your GPU?
The error means youve runned out of vram (resolution to high) for example

[nocturne topaz]

a good few minutes

[ornate elk]

Not that long. But let it load

[cosmic furnace]

I might be too hasty last time -___-

[serene oar]

Amd 6500 xt

[nocturne topaz]

what settings are you using, what interface?

more resolution doesn't always equal better image if settings are poor

[serene oar]

Settings as if my pc settings or stabble difusion settings?

[nocturne topaz]

stable diffusion - which model, sampler, steps, resolution, batch count

[serene oar]

I was using the base model of stable diffusion, 20 steps (should i add more?) and default everything else. The error happens when i use hires.fix option

[upbeat atlas]

@ornate elk any idea how to fix this?

[cosmic furnace]

this shud be it y? o.0

[ornate elk]

Hires fix can be to much for your GPU. Restart SD to fix the stuck vram. Then try hires fix upscale by 1.5, and use 10 hires steps.

[serene oar]

An example of what im doing with Novelai for example. How to make him have aface 😂

[serene oar]

Thank you very much

[ornate elk]

Change browser and test again

Brave blocked often in the past

[cosmic furnace]

Still get same error o.0

[upbeat atlas]

Same thing with opera gx and edge

[vast knoll]

https://cdn.discordapp.com/attachments/1002602742667280404/1191779742534348870/loraHowTo17.jpg

Just saving image

[ornate elk]

Hmm very strange

[ornate elk]

And thats strange too.
Have you copied the Zluda files the correct way?

[cosmic furnace]

Yeah theyr overwrited the ones was there hmgm

only thing i didnt check the last amd driver instal when amd hip install it was off. I had newer gpu driver if thats it

mb shud try zluda 3.2 ? 🤔

[ornate elk]

Thats okay the hip install should only install hip stuff and no other driver

[ornate elk]

@cosmic furnace @upbeat atlas
I will make a new clean install with the 3.4 Zluda files to see if I get any of your errors

But Tomorrow, because its 00:15 for me.

[upbeat atlas]

That'd be great, thanks 🙏

[ornate elk]

Will @ you then, gn 🙂

[ornate elk]

I dont think the issue is with zluda

Your one step before it even tries to use it

[cosmic furnace]

y 😦

[ornate elk]

How does your edited share_init.py file looks like?

[ornate elk]

Hmm okay, looks normal.

What you also can try is a complete reinstall of the Directml webui. You dont need to reinstall HIP or zluda

[cosmic furnace]

Wait lol

[ornate elk]

And make sure your on the latest AMD driver

24.1.1

[cosmic furnace]

zluda files are owerwrited after reinstalling

i need to replace them again but hmgm

[ornate elk]

Yes replace them again

[cosmic furnace]

nope still same error

[ornate elk]

Need to go off now sry

We can test again tomorrow

[cosmic furnace]

okay 🙂

o.0

not sure if this can relate tho

[solar karma]

Does anyone have any idea why I can't launch SD without deleting the tmp folder every time? If the folder exists I get this error:

If I delete tmp, it launches fine, every time.

[vocal burrow]

What do where python, python --version and py --list-paths say ? (run those in cmd)

[solar karma]

C:\Users\DXHHH101\AppData\Local\Programs\Python\Python310\python.exe
Python 3.10.6
The third one isn't recognized as a command

[vocal burrow]

what's in your webui-user.bat and where is your sd installed ?

please tell me it's not installed in system32

[maiden basin]

hey guys

i am getting a this error whenever i use open pose on sdxl:

RuntimeError: pixel_unshuffle expects height to be divisible by downscale_factor, but input.size(-2)=1080 is not divisible by 16

for all lineart types somehow

[onyx solar]

It's literally telling you in the error. 1080 is not divisible by 16.

Make the other dimension 1088.

[maiden basin]

oh well

haha

i guess i will go higher and crop

[cosmic furnace]

@ornate elk Dont tell me i had to install this amd display driver? When installing amd hip . cuz ive got newest 24.1.1 . Anyway tried making full clean install of directml & all shit and still same error 😑

files that are replaced from Zluda into SD are named _11 but i can see _10 and _8 . Imon 10 windows not 11 ... but naming them _10 and copying didnt helped 😑

[remote nymph]

Hello, I've been having this issue where whenever I go under "Textual Inversion", "Hypernetworks", "Checkpoints", or "Lora"; it displays the following (I'm new to using Stable Diffusion and have just set it up):

[unique bear]

Anyone here have issues installing extensions when running A1111 SD on runpod?

[fair oxide]

comes the LCM, then comes the Turbo, now here comes SDXL Lightning ...

@ornate elk 600x912 then hiresFix by 2x

[nocturne topaz]

yes - you need to go into the webui-user.sh file (connect to the jupyter notebook) and remove --skip-install from the commandline_args, then restart the pod

[unique bear]

thanks ser

[wanton stirrup]

howdy, anyone knows what this can be: AttributeError: 'DecodedSamples' object has no attribute 'to' ?

[unique bear]

Im trying to generate using animateddiff but it's generating normally instead of the gif local version does it correctly but runpod is not

[short stratus]

Hi
I am new to SD and downloaded forge (after hearing about it's great opt)
but every tutorial is outdated at this point (due to insane speed of development)

I need help with installing the basic stuff, extensions and ControlNet ect.
take my hand and guide me, I am lost

[fair oxide]

every tech industry is racing to get Ai where it needs to go, until then consider the developments now a big beta run

[solar karma]

Sorry for the last response there
There's my batch file, it's not installed in System32. It's currently in a folder on the root of one of my SSDs. It was previously on my desktop, where the same issue also happened

[solar sundial]

anyone get stableswarmui working on a reverse proxy?
This is my apache2 config:
ProxyPass / http://localhost:7801/
ProxyPassReverse / http://localhost:7801/
ProxyPass / ws://localhost:7821/
ProxyPassReverse / ws://localhost:7821/
ProxyPreserveHost On

When I attempt to generate anything on the webpage, i get this...
[Error] [WebAPI] Error handling API request '/API/GenerateText2ImageWS': Invalid request method: GET

[nocturne topaz]

99% of a1111 tutorials should work on forge unless its for a niche extension

[frail pebble]

I am looking to find any resources I can on creating my own diffusion model from scratch. Now before EVERYONE hops on my nuts about how hard it will be, how much time it will take, or how much money I will have to spend, lets get one thing clear to start. I honestly don't care. I don't need to hear about the issues I will have to face, that doesn't help me as I am fully aware of the issues I will face.

All I am asking is if someone here can help me with getting started on training a diffusion model from scratch, I will be training it with 64x64 sprites, to make 64x64 sprites. With a small dataset, very small but that still doesn't mean I want a LORA.

No I don't want a LORA.
No I don't want embeds.
No I don't want hypernetworks
No I don't want FineTuning, or Dreambooth.
No I don't want to use GAN if I can avoid it, as HuggingFaces's "diffusion" technology/code/platform is superior IMO.

I want to make my own very small diffusion model from scratch. Any help in this regard will be very appreciated!

[short stratus]

yeah something like that, most tutorials about ControlNet for example were posted around the time of the release
some methods of installing has changed and some links don't work etc..

[nocturne topaz]

In that case start from scratch using keras in python to build a neural network from the ground up. Get a good well tagged dataset, and the best of luck to you!

[upbeat atlas]

Hi I have a problem with the browser interface. I'm on AMD gpu windows. For some reason it doesn't let me generate images right away after generating an image. If I make another tab it can, but I can't do it on the same tab and it's super frustrating. Can anyone help? I tried direct ml and Zluda with this, both have the same problem. I reinstalled directml and even AMD software but no luck

[fair oxide]

https://x.com/Elliryk_Krypto/status/1759967867192082527?s=20

elliryk NFT.NYC 2024🟧 (@Elliryk_Krypto) on X

New X-Adapter give really cool results (b&w AIPhotography is the original made with SD1.5).

X-Adapter " enable plugins pretrained on old version (e.g. SD1.5) directly work with the upgraded Model (e.g., SDXL) without further retraining" so all ControlNet, LoRA for v1.5 works now…

Twitter · Feb 20, 2024, 3:46 PM

[primal vale]

anyone know how to fix this controlnet error?

[nocturne topaz]

Just waiting for a1111/forge/comfyUI integration

[nocturne topaz]

Turn the controlnet on and off again, and make sure the model and preprocessor are selected

[fair oxide]

@nocturne topaz take a look at this https://civitai.com/models/316681

users can run it on StableCascade extension on a1111

[fierce hollow]

Hello, previously I was told that XL models are not compatible with Lara. Can you tell me realistic non-XL models that can be compatible with Lara and not Japanese? The point is: I found several models, either it generates Japanese people or mutants with three legs, two heads, and so on -_-

[nocturne topaz]

Lara - do you mean Lora?

[fierce hollow]

Yes, yes, I'm sorry)

[grim cliff]

can someone help when I try to make a picture, it gives out when loading the model too ? OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 8.00 GiB total capacity; 6.98 GiB already allocated; 0 bytes free; 7.28 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[nocturne topaz]

Xl models need xl loras, 1.5 needs 1.5 loras

[nocturne topaz]

What resolution are you using, what hires fix?

[fierce hollow]

What does 1.5 mean and where can I see it?🥺 👉 👈

[fair oxide]

@nocturne topaz have you tried the lightning models that dropped few hours ago?

[nocturne topaz]

On civitai under the lora details there will be a base model that tells ya

[fair oxide]

version. aka SD 1.5

[nocturne topaz]

Not yet, I see the lora has dropped and dreamshaper has trained one up as well so that is a good sign

[fair oxide]

i just grabbed the dreamshaper lightning, going to play with it

[fierce hollow]

As I understand it, you need to look at it on the site when I download it, right?

[fair oxide]

for example

[grim cliff]

hi res I don't use, I tried to make a resolution of 1024 by 768, I was able to fix this problem, the model loaded, but when I try to create a picture, it now gives the error RuntimeError: mat1 and mat2 must have the same dtype

[nocturne topaz]

Yup, although that said the turbo models are hit and miss with lora compatibility

[fierce hollow]

Thanks)

[nocturne topaz]

You using a lora?

[fierce hollow]

So the Lora version and the model version just have to match, right?

[grim cliff]

no

[nocturne topaz]

Controlnet?

[grim cliff]

No, this is the first time I've heard about it.

[fair oxide]

for now Loras are trained separately for either SD 1.5 or SDXL 1.0 but they have already released X-Adapter which will allow you to use any SD 1.5 based lora, embeddings, for both SD 1.5, and SDXL .. waiting for webui intergration now

[fierce hollow]

Do I need to install the X-Adapter separately?

[fair oxide]

dunno yet for sure

[fair oxide]

you'd have to install the plugins to work, but until there are official release on the method of using it idk atm https://github.com/showlab/X-Adapter

[grim cliff]

I fixed the error with this command --no-half, but for some reason it appeared again when creating the image. OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 8.00 GiB total capacity; 7.07 GiB already allocated; 0 bytes free; 7.28 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[hallow mural]

画一张蒜苗生长的图片

[fair oxide]

@nocturne topaz 4 steps, 2cfg dreamshaper lightning

[nocturne topaz]

What card have you got, what else have you got open??

[grim cliff]

NVIDIA GeForce RTX 3060 Ti is launched via Yandex and only discord is open, this problem only appears in this model https://civitai.com/models/257749?modelVersionId=290640

[ornate elk]

Remove --no-half
And add: --xformers --medvram-sdxl --no-half-vae

[ornate elk]

No you dont need the AMD Pro driver. Just the HIP stuff.
Also they need to be named 64_11 regardless of the OS version

[grim cliff]

yes, it worked, thanks, but can I find out why these commands worked?

[ornate elk]

The problem was that your vram was to less for the heavy sdxl models.

Xformers is an overall performance boost, that also decrease the vram usage of the model.
--medvram-sdxl splits the sdxl model when used, so it doesn't get processed as one big data block.
--no-half-vae is for the vae compatibility

[grim cliff]

I understand

[ornate elk]

And its on every browser?
Also open up a cmd and type
Sfc /scannow
To fix Windows internal problems.
Make sure your on AMD Adrenalin 24.1.1

Is your D-Drive an SSD?

[woven smelt]

I upgraded from a 20 series gpu to a 40 series. Do i need to change anything in my SDXL installation to make use of the newer architecture?

[ornate elk]

What's in your webui-user.bat?

And how much vram do you have now?

[nocturne topaz]

https://tenor.com/view/borat-borat-very-nice-verynice-thumbs-up-gif-25080066

[woven smelt]

@ornate elk 16gb

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --medvram-sdxl --no-half-vae

call webui.bat

[fierce hollow]

Where should I install this? Can you please tell me which folder?

[fierce hollow]

Does this mean the same thing, as I understand it ? https://huggingface.co/h94/IP-Adapter

[nocturne topaz]

nope.. different

ipadapter is for swapping concepts (ie faces/clothing) into an image

[fierce hollow]

Understood... And where to install the X-Adapter?

[nocturne topaz]

there isn't an install available yet, the clever cookies behind the scenes need to figure out how to bake it

[fierce hollow]

Understood... We are waiting then)

[fierce hollow]

I am currently downloading the IP Adapter, please remind me where to install it? In \models\ControlNet?

[nocturne topaz]

you on forge or a1111?

for a1111 it goes in /extensions/sd-webui-controlnet/models/; on forge you are correct

[fierce hollow]

Initially, I was on the weak version, as I was told, later I was given a link to a1111 - but if you tell me how I would make sure that I was on a1111

[unique bear]

does animatediff work on forge?

getting a type error when I try to use it

[nocturne topaz]

down at the bottom of the screen you would have version 1.7.0 or 1.8.0

[nocturne topaz]

sd-webui-animatediff doesn't; sd-forge-animatediff does but the results aren't great, better off using svd in forge

[fierce hollow]

Should I create a folder here?

[nocturne topaz]

you need to install the sd-webui-controlnet extension first

from the extensions tab

[fierce hollow]

?
https://github.com/lllyasviel/ControlNet

[nocturne topaz]

nope..

https://github.com/Mikubill/sd-webui-controlnet

[fierce hollow]

Okay. Let's go step by step. I've already started downloading the IP-Adapter. Once it's downloaded, I'll start downloading ControlNet. I always download them into separate folders. After downloading, I'll write to you to find out what to do next, okay?

[nocturne topaz]

install the controlnet through the a1111 extensions tab (go to install from Url, plug that in, hit install, then go to the installed tab and Apply and Restart UI)

[ornate elk]

Okay then you can remove --medvram-sdxl

Also delete the venv folder

[lost chasm]

hey quick question, i upgraderd from 3090 to 4090 and now in A1111 SDXL renders seem to choke out all the vram, and gens stall if you use highres fix with like 1024x1024 X1.45 scale, it just gets stuck in a bottleneck for ages and the whole machine becomes unusable

anyone see anythng like this?

should be the same as before, same amount of vram!

[ornate elk]

Would try to delete the venv folder and relaunch the webui-user.bat

[woven smelt]

I get these errors when launching. They have been there a while.

[ornate elk]

Try updating your extensions

Also check if your webui is on 1.7.0

[lost chasm]

yeah its 1.7

deleted venv, updating

[lost chasm]

hmm still doing it

webui.bat hasnt changed, dont see anything weird here?

[ornate elk]

Remove --precision full --no-half --disable-safe-unpickle

Your slowing yourself down with the first two and open a security hole with the third one

[lost chasm]

well the speed was fine its the ram use

and it had these settings working fine with 3090 :/

cant even remember now why i had -disable-safe-unpickle

[ornate elk]

Wow you wasted a lot of performance and time with using --no-half and --precision full :/

[lost chasm]

but does it change the ram use?

[ornate elk]

Yes it should

[lost chasm]

but it wasnt choking the ram previously with these exact settings so i cant blame this for the choking probem i'm experiencing

[ornate elk]

Without them it should process fp16 made models and vae as fp16 which will consume less vram

Did you tried again without them now?

Is it still choking?

[lost chasm]

looks like its not, but that doesnt explain the issue i was havng with same vram and same settings 😦

[ornate elk]

The cards are not the same. The architecture of the cards are not the same. The graphic driver files are not the same.

We can't know anymore why the 3090 didn't choked up

But now you shouldn't get any problems and you get the fastest performance too

[lost chasm]

are there any downsides to the faster precision settings or nohalf

[ornate elk]

Nope

For example --no-half is needed for GPUs that can't process fp16

Like a gtx1660

[lost chasm]

no image quality problems for realistic renders?

[ornate elk]

Or a amd rx580

[ornate elk]

Nope

Not at all

[lost chasm]

kk thanks

[ornate elk]

No problem 🙂

[bleak prism]

hi guys i am new here, i am trying to install stable diffusion for amd supported by zluda, i follow this github methode to do so and i have some issue. can any help me pls : https://github.com/vladmandic/automatic/wiki/ZLUDA

i am blocked on this step : Compilation, Settings, and First Generation

i start from sd.next but i don't if is the webui or the text type file named sd.next, i try running webui.bat but got accès denied, i try to run webui.bat on administration but got oher error

this error

[ornate elk]

Where is your webui located

Also if you want to use the Directml webui with Zluda you can checkout my guide in the Pinned Messages of this channel. (Its not for SD.next, but for the lshqwtiger fork of auto1111)

[bleak prism]

i just changer my folder following this and webui.bat can open now :

but got new probleme

[ornate elk]

Can't help you with SD.next, only with my guide for the Zluda directml fork.
For SD.next you should checkout their Discord

[bleak prism]

wait their a zluda directml ? is it fast or slow ?

[ornate elk]

Its not directml+Zluda,
its lsqhqtigers directml fork of auto1111 but modified to work with Zluda

Its blazing fast on my 7900xtx

Going from 4it/s to 16it/s

[bleak prism]

wow

can i have the link ? i have 7900 xt but with directml i only got 1.81it/s

[ornate elk]

It depends on the resolution as always. (Tested was 512x512, 30 steps)

[ornate elk]

The guide to make it work is in the Pinned Messages of this channel:
#🤝｜tech-support message