🤝｜tech-support

[calm cairn]

but it should be inverted, it should be the only one on the list, in the new directory

[ornate elk]

Is your webui updated?

Make sure so scroll to the bottom of the model dropdown to check if its there

[calm cairn]

ah good call, it is sorted by path so it is in fact on the bottom

I'll move a few more, and if they show up on the bottom, then thats probably solved the problem

i dont know why it wont lose the old directory but it really doesn't matter for me as long as it uses the new one too

[pure forge]

what's interesting is the first one I run takes 35-45 seconds while each one after is 10-13 seconds

and it doesn't speed it up if it doesn't have that last argument

idk if any of the other arguments are helping actually

[hot spindle]

@karmic crown Recap, strated with basic, just --skip-torch-cuda-test --upcast-sampling , system benchmark and 4 image generation, SDXL img2img with Controlnet Canny , no resize , 4 simultaneous generations, same RES of 1472x640:

In all tests and variations GPU allways works at 100% but memory never goes above 50GB and Python gets the majority but never above 24GB

Same basic arguments

Typical workload for all scenarios:

Browser I'm using Safari for A1111, no memory problems, requires less than Firefox that I'm using for Discrod/WhatsApp/Mail

[calm cairn]

yes, the ones moved to the new directory are showing up and do work when I attempt that. Not really understanding why the default directory is still live, but if it works, it's good enough, ty!

moving all of them over now

[oblique quail]

UnicodeDecodeError: 'utf-8' codec can't decode byte 0xe9 , can someone help ?

[ornate elk]

The main one will still get used as a fallback directory

If the drive isn't available for example

[calm cairn]

thanks, and as a bonus during all this i was googling, and now i am in dark theme too

[vagrant field]

Hello, I keep running into this error when trying to generate an image; DefaultCPUAllocator: not enough memory: you tried to allocate 26214400 bytes. For reference, I have 32 GB of ram

[fading herald]

Just had a CUDA crash using AnimateDiff, tried to restart and now I have this in my cmd

with this error

Any idea what's happening ?

[rugged compass]

ty lots

[vagrant field]

I also don’t know whether or not the program is using my gpu or not

[ornate elk]

Hey, can you restart the webui and show the cmd log?

[grim moon]

Hey how do i check what ersion of sd i have installed

[vagrant field]

venv "C:\Users\ndrew\Downloads\SD\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: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Launching Web UI with arguments: --medvram
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Loading weights [486a245c1d] from C:\Users\ndrew\Downloads\SD\stable-diffusion-webui\models\Stable-diffusion\hentai.safetensors
Press any key to continue . . .

Now its not even opening up a webUI

[ornate elk]

Check the bottom of the webui in browser

[grim moon]

Thanks

[ornate elk]

What's your GPU?

You should add --xformers to the webui-user.bat too
Then delete the venv folder

[vagrant field]

Nvidia RTX 4070

[ornate elk]

Then remove --medvram

And instead add --xformers --no-half-vae
Then delete the venv folder

[vagrant field]

Thank you!

[uncut jewel]

Hey guys, i updated ComfyUI today, when it's loading, it seems to work but i can litterally not interact/import workflow or do anything since then 😭

I can send the full log if that helps

my manager is gone too :/

[ornate elk]

What does the log shows?

[uncut jewel]

I just found someone that told me how to fix that issue after an half day spent doing many stuffs to try to fix that thing lol. Basicly it was related to the Mixlab node extension, i just deleted it and it worked back 🙏

[ornate elk]

Okay good 🙂 and thanks for letting us know here too. If someone has the same problem

[uncut jewel]

No problem! hope that will help!

https://media.discordapp.net/attachments/1145677539738665020/1194354426534711346/mmexport1704826645343.png?ex=65b00c3d&is=659d973d&hm=394d20ec0d57b0ecf45387f76047db8bb5c6ae97eb19981c0c33d52c0c91dbb3&=&format=webp&quality=lossless&width=828&height=636

Basicly that's what he sent me

[fading herald]

I've recreated the venv and I'm still crashing with AnimateDiff...

[ornate elk]

Is your driver updated?

Are you on a laptop?

[fading herald]

Yeah, up to date, and not on a laptop

I recently changed my GPU but I've tested AnimateDiff after that and everything worked well

And besides some optimizations, I've haven't changed anything, so I guess it comes from one of the options I've changed, but can't tell which one

[ornate elk]

whats your gpu? and whats inside your webui-user.bat?

[fading herald]

4070

set SAFETENSORS_FAST_GPU=1
set CUDA_VISIBLE_DEVICES=0

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --opt-channelslast --upcast-sampling --no-half-vae --api --ckpt-dir "F:\Checkpoints" --lora-dir "F:\LoRA" --vae-dir "F:\VAE" --esrgan-models-path "F:\ESRGAN"

git pull

call webui.bat

[ornate elk]

why cuda visibile devices? do you have more than 1 gpu?

[fading herald]

Just trying stuff out, I've seen on a git discussion it would prevent A1111 to use the CPU or something like that, so figured it wouldn't hurt anyway

[ornate elk]

you dont need the two first set lines, did you tried switch to torch2.0 in the animated diff settings?

[fading herald]

Same thing, still crashing

[karmic crown]

which is basically expected, since A1111 almost exclusively use GPU, and just some of CPU resources. regarding the memory usage, as i alredy wronte on github, you probably do not need any of medvram variations, so use them only if you get an out of memory error during image generation, which would probably not happen

[karmic crown]

0xe9 is é
do you have é in some file or folder name?

[fading herald]

txt2img, hires, img2img, controlnet, deforum, everything is working except AnimateDiff, go figure

[hot spindle]

@karmic crown TKS, figured that much, now that I've also figured I can use batch as long it's simultaneous and I don't have memory issues I'm all good. I've also noticed huge improvements on how 1.7 handles loading big images, previous versions would halt browsers, now can easily load and handle big pixel images

[karmic crown]

btw, which torch you are using? 2.1.2 or nightly?

[hot spindle]

2.3 dev

[karmic crown]

as long is not 2.0.2, it shold be faster

not fast, but at least faster than before

🤷🏻‍♂️

[hot spindle]

Not sure exactly but I loaded a 5kx3k image for further upscaling and no stuttering...in 1.5 would halt almost to a stop

[karmic crown]

would you mind to try SDXL-turbo on your machine? https://civitai.com/models/215478/sdxl-turbo

i am really curious how it works on your mac, compared to mine and Slusho's

[hot spindle]

I can make you a tunnelling session if you want

[karmic crown]

i am just interested in one thing

100%|█████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:11<00:00, 11.35s/it]

so image is generated in about 10-12 seconds

but for some strange reason webui does not finish there

Total progress: 100%|█████████████████████████████████████████████████████████████████████| 1/1 [00:18<00:00, 18.45s/it]

so in the end we are getting some insane times like this

Steps: 1, Sampler: Euler a, CFG scale: 1, Seed: 851026603, Size: 512x512, Model hash: e869ac7d69, Model: sdxlTurbo_sdxlTurboPruned, Version: v1.7.0
Time taken: 42.0 sec.

[elder eagle]

Peux tu me créer une montre du futur qui reste réaliste se la marque rolex

[hot spindle]

Out of home but will PM when I get back

[viscid flax]

hi again

i think my cpu is being used instead of my gpu

[karmic crown]

thanks

[karmic crown]

please try with those arguments

export COMMANDLINE_ARGS="--skip-torch-cuda-test --opt-sub-quad-attention --upcast-sampling --no-half-vae --use-cpu interrogate --disable-model-loading-ram-optimization"

spaceman
Steps: 1, Sampler: Euler a, CFG scale: 1, Seed: 2703425679, Size: 512x512, Model hash: e869ac7d69, Model: sdxlTurbo_sdxlTurboPruned, Version: v1.7.0

Time taken: 6.7 sec.

spaceman
Steps: 1, Sampler: Euler a, CFG scale: 1, Seed: 981695187, Size: 512x512, Model hash: e869ac7d69, Model: sdxlTurbo_sdxlTurboPruned, Version: v1.7.0

Time taken: 5.3 sec.

[pure forge]

so you want me to add that line to webui-user.sh?

I only say that because I was watching a video right when you sent that that said to do the exact same thing

[karmic crown]

this exact command line?

[pure forge]

well something similar

no idea what that was about

they're not really explaining it which is a common theme

[karmic crown]

do not use share unelss you want access from internet

please try this command line

export COMMANDLINE_ARGS="--skip-torch-cuda-test --opt-sub-quad-attention --upcast-sampling --no-half-vae --use-cpu interrogate --disable-model-loading-ram-optimization"

spaceman
Steps: 1, Sampler: Euler a, CFG scale: 1, Seed: 1465285502, Size: 512x512, Model hash: e869ac7d69, Model: sdxlTurbo_sdxlTurboPruned, Version: v1.7.0

Time taken: 6.7 sec.

i do not see that slowdown we both noticed

[pure forge]

so you do want me to add that to the sh file?

[karmic crown]

yes please

you can comment out your current command line with # at the beggining

[pure forge]

well this is my first time adding anything

so I just replaced it

[glad fiber]

do you remember who did that?

[pure forge]

do I have to set the seed to the one you posted or just leave it as -1 which I assume just randomizes it @karmic crown

[karmic crown]

leave it to -1

yes that is a random seed

[pure forge]

okay this time it was 7.9s for the first and each subsequent was 3-4.3s

[karmic crown]

nice

[pure forge]

[karmic crown]

if you do not see any additional errors or warnings, leave it like that

[pure forge]

is it normal for the first run to be longer than the others

[versed sorrel]

How do I solve the error
PermissionError: [WinError 5] Access is denied: 'outputs' ?
Its preventing me from saving my pictures

[karmic crown]

it is the same for me

[pure forge]

okay thanks a ton for the help

[karmic crown]

if you see any error, let me know what error you are getting so i tell you what to add

(or not to add) 🙂

[snow swallow]

Hello to everyone, I'm a big beginner on ai and would be very thankfull for any help setting stable difusion, I dont have a gpu but my amd cpu su´posedly have a graphics card. Downloaded stable difusion, python, git and ran everything to the point when I run

webui-user.bat

I get the message Traceback (most recent call last):
File "F:\AI\stable-diffusion-webui\launch.py", line 47, in <module>
main()
File "F:\AI\stable-diffusion-webui\launch.py", line 38, in main
prepare_environment()
File "F:\AI\stable-diffusion-webui\modules\launch_utils.py", line 384, 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

Can anyone help me with that?

[karmic crown]

i can help you, but it will be sloooow

generation time on CPU only is very slow

my linux server has 10 cores and i have 60GB or RAM and it is painfuly slow

[snow swallow]

No problem, even if I take some time just to generate a single picture it would be able to let me learn how to work with it and afterwards seek for a better pc

[karmic crown]

go to your F drive, than to folder AI, then to stable-diffusion-webui

find and open file webui-user.bat with editor

it should look like this

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

call webui.bat

change it to be like this

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set TORCH_COMMAND=pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
set COMMANDLINE_ARGS=--skip-torch-cuda-test --no-half --use-cpu all --opt-sub-quad-attention

call webui.bat

sorry

find and remove folder called venv inside your stable-diffusion-webui folder

double click webui-user.bat to run it

that should be it

[versed sorrel]

@karmic crown
How do i solve
PermissionError: [WinError 5] Access is denied: 'outputs' ?

[karmic crown]

you need to fix your permissions, but i cannot help too much, since i do not have access to windows at this moment

[versed sorrel]

I've read i need to change the shortcut to administrator, but that option is not available for some reason

[karmic crown]

have you run python or a1111 as admin maybe?

[versed sorrel]

python wont run as admin, but i can try a1111

[karmic crown]

please dont

can you show me the log (what you have in cmd window)

[versed sorrel]

do you need the full console or just the last few lines?

[karmic crown]

as much as you can send please

[versed sorrel]

Heres the full thing. It was working perfectly fine until today. I don't know why it stopped

[snow swallow]

@karmic crown thanks, after some installing I can now acess the stable diffusion web-ui. What would be a lightest prompt I could use just to test if it is working? I tryied a couple of things but I'm not sure its just slow (as expected) or is something broken and I'm waiting for nothing

also my prompt is saying:

Creating model from config: F:\AI\stable-diffusion-webui\configs\v1-inference.yaml
Applying attention optimization: sub-quadratic... done.
Model loaded in 47.3s (calculate hash: 23.9s, load weights from disk: 20.2s, create model: 1.0s, apply weights to model: 1.1s, calculate empty prompt: 0.9s).
15%|████████████▍ | 3/20 [00:41<03:52, 13.69s/it]
Total progress: 15%|██████████ | 3/20 [00:33<03:12, 11.31s/it]
Total progress: 15%|██████████ | 3/20 [00:33<02:46, 9.77s/it]

[karmic crown]

solve this first

go to extensions folder, and remove controlnet

which is not properly installed, and make problems

[versed sorrel]

Thanks, i'll give that a try

[karmic crown]

looks like its working, but as i said it will be slow

change sampler to Euler A (which is the fastest one) and use something like 10 steps (just for test)

[versed sorrel]

Same issue :<
Should I try a clean install rather than trying to find a work around for something i probably broke?

[viscid flax]

@karmic crown i dont think my gpu is being utilized. could you help me fix it?

[karmic crown]

i noticed that some people do not like the idea of reinstall, so i am trying to keep that a last resort, but yes, that would be the best in your case

[versed sorrel]

I hate reinstalling but its whatever; I often spend more time fixing a problem than doing it over. Thanks for your help, though!

[karmic crown]

i would rename the folder with our current installation, and install a new one using one of the guides from pinned messaged in this channel (do you have nvidia or amd)?

then you can move your models from old installation

so you do not need to download them again at least

[versed sorrel]

Im on Nvidia, i was using a youtube tutorial thats probably a little outdated which might be how i broke it

[karmic crown]

please use this one -> #🤝｜tech-support message

[snow swallow]

@karmic crown thank you very much I'm imensely greatfull, I know that from now forward i have to study the methods, and how to feed the IA, do you have any youtube channel you'd recommend?

[versed sorrel]

Thanks again! Will do!

[karmic crown]

i will try to help you, but i have only two hands 🤷🏻‍♂️

[karmic crown]

how much time do you need to get an image with euler a and 10 steps?

[viscid flax]

ohh my bad. i thought you didnt see my message.

[karmic crown]

what is your GPU? what are your command line parameters in webui-user.bat?

[snow swallow]

about 3 or 4 minutes. I typed "Men with a hat" and was something around that 3 minutes euler a, sample steps 10, batch count 1

[viscid flax]

nvidia rtx 3050 ti

how do i check command line parameters?

[karmic crown]

go to https://civitai.com/models and find some nice model that you like and try it. be sure that it shows Checkpoint in the left corner (those are models)

also, be sure that it says this in details

and download the smallest file that exists in download

[karmic crown]

send your webui-user.bat and post it here

and be sure that you are running that one, and not webui.bat

also show me your GPU usage during image generation if you know how

this

[ornate elk]

hey you should follow my AMD install Guide then as it can work with the IGPU of AMD CPUS

its in the Pinned Messages of this channel

[karmic crown]

make sense... but i didnt notice AMD in his original message until now

[karmic crown]

this looks like download of a model

[viscid flax]

umm

this is new

its giving me this error now

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

[viscid flax]

i think thats why

[karmic crown]

wait to download to finish, and then restart webui, and see if it works

[viscid flax]

yeah it worked

[karmic crown]

i need to go now, my eyes hurt badly, too much time spent in front of the screen today

bye people

[ornate elk]

you should try the AMD install then, it should be faster

[snow swallow]

I was trying to follow your posts instructions but my pc froze after this:

[ornate elk]

oh just let it load

if its not completely frozen just wait ^^

that torch step can take a while

[snow swallow]

no problem on that first message of unrecognized comandline_args

[ornate elk]

oh, thats not good, also it says nvidia driver found

thats strange

[snow swallow]

I have a AMD Ryzen 5 PRO 4650G with Radeon Graphics 3.70 GHz

[ornate elk]

you need to edit the webui-user.bat like in my guide

you can also show me how you edited it

[snow swallow]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
COMMANDLINE_ARGS= You add: --use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling

call webui.bat

[ornate elk]

you need to remove the "You add:"

[snow swallow]

sorry, how stupid of me

[ornate elk]

no problem 🙂

you can close the cmd

and edit the bat
then open up a cmd and type
pip cache purge

then delete the venv folder

and relaunch the webui-user.bat

Then just wait for the install to finish, the webui will open automaticly

[snow swallow]

I did this steps, it went way faster but still got the same error

How the file is now:

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
COMMANDLINE_ARGS=--use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling

call webui.bat

"not recognized as a command internal or external, a probrem operable or a file in groups"

[ornate elk]

okay, maybe there is a hidden (invisible) space in the commandline args

maybe at the end after sampling

[snow swallow]

doesnt seem like it, I tryied to deleted it but got the same result, though as I sair now its running very smoothly and withouth freezing

[ornate elk]

but now it downloads the wrong file

because the commandline args are not recognized

[snow swallow]

the other version is working and the .bat is like this:

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set TORCH_COMMAND=pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu
set COMMANDLINE_ARGS=--skip-torch-cuda-test --no-half --use-cpu all --opt-sub-quad-attention

call webui.bat

[ornate elk]

replace the whole text in the webui-user.bat of the AMD install to this:
`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --upcast-sampling
git pull
call webui.bat`

then open up a cmd and type
pip cache purge
then delete the venv folder
and relaunch the webui-user.bat

Im sure then it will work

[steady thistle]

@ornate elk Does it hurt to do pip cache purge every so often to keep things cleared out?

[ornate elk]

it wont hurt 🙂 you will free up multiple GB after Updates or Reinstalls

[steady thistle]

I feel like I get uneven speeds on SDXL, sometimes they pop right out and other times it is slow as christmas.

[snow swallow]

Thanks, it seem to be working:

[ornate elk]

😄 perfect

[ornate elk]

oh i forgot your on Igpu
so if you get an out of memory error, then you need to change the --medvram to --lowvram in the webui-user.bat
you can do that later after testing it

[snow swallow]

Great, I'll keep an eye to that, now it strangely is stuck on download torch though my connection is fine, I'm waiting to see if it gets unstuck

[ornate elk]

oh if you have a little white block in the cmd, then press space

to unpause the cmd

[steady thistle]

I'm trying to speed things up. Are there any args I need to add or remove from this? I use: --use-directml --lowvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling (and --no-half when I'm using img2img)

[snow swallow]

I'm not sure what hapened but I was stuck in this 173.3mb file, now I closed the window opened it again and is downloading another one of 2,6gb though now clearly says "nvidia driver found"

[ornate elk]

you paused the webui before by clickin in it

so you may need to do the steps again:
in a cmd:
pip cache purge
and then delete the venv

[snow swallow]

ok, living and learning, had no idea that the prompts could pause and unpause

[ornate elk]

😄 yea its not well known, also very stupid that it pauses just by a missclick

[steady thistle]

I didn't know that either, might have been an issue I had before.

[ornate elk]

what was your gpu?

[steady thistle]

Processor AMD Ryzen 9 5980HX with Radeon Graphics, 3301 Mhz, 8 Core(s), 16 Logical Processor(s), Radeon RX 6800M GPU with 12GB GDDR6

[ornate elk]

you could switch back to --medvram

thats the only speedup

[steady thistle]

I was afraid of that. I tried earlier but it crashed, but might have been the model being too big. I'll try smaller models.

[ornate elk]

yea sdxl might not be working with that

[steady thistle]

Well, for this entire conversation, I've been processing one 512x768 image.

[ornate elk]

oh thats slow

[steady thistle]

It just finished

I really love SDXL but damn. I should also stop using Chrome for this.

[snow swallow]

@ornate elk thank you very much for your patience and your wisdom, its all working now

[steady thistle]

Interesting! Chrome was 6 seconds faster processing, with tabs being unused but present and Discord running, compared with Edge with absolutely nothing else going.

Exact same image and process.

[ornate elk]

No problem 🙂 idk how much RAM your Laptop has but in BIOS you can define RAM as GPU Vram.
So if it has more than 16gb its a good idea to do so.

[ornate elk]

Your using discord in browser mode?

[snow swallow]

guess would be cheaper to buy more ram than a new gpu

[ornate elk]

Probably xD

[ornate elk]

For now stick to 2gb models, these are the fastest and you won't get out of memory (vram) errors

[steady thistle]

Yep

[ornate elk]

For example Dreamshaper v8

[ornate elk]

Ah okay thats very interesting

But edge is a bloatware Browser

Im using Firefox 🙂

[steady thistle]

Seems to work well. Discord is using 80mb memory

Have you tried Opera?

I haven't loaded either one, but I used Firefox back when it was Navigator. lmao

[ornate elk]

No and I never want to xD opera primarily owned by a Hongkong based company, also opera is bloated with unneeded features or their "free VPN".

[steady thistle]

Good to know! I don't trust google, but I trust China slightly less. lol

[ornate elk]

Yes thats why I use Firefox. Because they have so much privacy and security features and dont spy on you like google

[hot spindle]

Not that I feel any need to use anything but Safari, but now curious to try if a bare bones browser improves anything, something Like Mini Browser or SeaMonkey

[steady thistle]

Or DuckDuckGo

I use that on my mobile

[ornate elk]

And the customisation of firefox is superior. Also google activly works against adblockers

[steady thistle]

I think I'll grab DuckDuckGo actuallly! They are heavy into privacy

[ornate elk]

I didn't tried their browser but heard good stuff of their mobile one

[versed sorrel]

It seems making a shortcut to the launch.py file on my desktop permentently breaks the entire thing, even if the file is deleted. Can anyone tell whats wrong in my console? I've reinstalled twice

[ornate elk]

Let us know ^^

[hot spindle]

Dont really know how to tell, using Safari with A1111 uses less memory than Firefox opened with Discord/WhatsApp/Gmail

[ornate elk]

It tries to use the wrong python.

You need to edit the webui-user.bat and at the line Python=
You need to add the Path to the python.exe from 3.10.11

You get the path by typing
where python
In cmd

[steady thistle]

I'm using a stopwatch on my phone with the exact same settings and comparing speeds.

[versed sorrel]

Sadly, that did not work. It was working perfectly fine until i tried to make a desktop shortcut. I dont get why

[ornate elk]

You shouldn't make a shortcut to launch.py

Always ever use the webui-user.bat

To launch

[hot spindle]

Not sure if it will make any diference, my Studio just chewes everything I throw at it..its not ubber ultra fast, but no memory issues so just keeps going whatever I throw at it

[versed sorrel]

My bad. It seems to be working perfectly fine as it is and was. Thank you!

[ornate elk]

Np ^^

[ornate elk]

Dont forget to add the needed COMMANDLINE_ARGS for your GPU into the webui-user.bat

[versed sorrel]

I did that already, but i almost did forget

[steady thistle]

@ornate elk Sorry, I have infinite questions. Do you know anything about using Conda and Olive to get better speeds from AMD as per this article: https://community.amd.com/t5/ai/updated-how-to-running-optimized-automatic1111-stable-diffusion/ba-p/630252

I'm going to try the process again... with reluctance... but I really want that boost.

[hot spindle]

I'm not a AMD person, but my son runs his 7800XT on Ubuntu whenever needs AI , mostly because its one of the few supported linux distros for AMD, and he says the gains are massive

[steady thistle]

I just can't with Ubuntu... I can't face it

[hot spindle]

Yeah, unfortunately AMD doesn't officially support many distros , its Ubuntu,REdHat and some other I cant remember, even Fedora is a PITA

[ornate elk]

Yes its supported in the directml webui.
To use it you need to add --onnx --backend directml
But the performance you gain isn't worth the incompatibility of the webui.
Your stuck on a few onnx models. And these only work on locked/fixed resolutions. Extensions and other stuff won't work or break.

Olive is like the Shark webui but much worse.
So you should try out Shark if you want to have more speed.

[ornate elk]

Works on arch too 🙂

[hot spindle]

Its Suse and now apparently CentOS

[hot spindle]

Good luck with that, from experience anything not officially supported is pain

[steady thistle]

Thanks, I don't want to lose compatability, so you saved me a lot of heartache right there.

As for the webui, I am perfectly fine with automatic1111 out of the box in terms of the UI, I may have to look at Shark, haven't heard of that one.

[hot spindle]

I once resorted to Fedora for my home iMac (AMD) also tried Arch , mostly because there's a limitation in Apple drivers not supporting more than 4GB in OpenCL, at the time runnning LuxCore Render , long story short needed linux to acess 8GB , it was very painfull to get AMD drivers to work properly

[halcyon knoll]

Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

Got this after trying to generate a image via stable diffusion (first time)

I suspect it has something to do with my cpu having integrated graphics maybe?

[ornate elk]

Shark was the first AMD SD webui when there was no auto1111 directml fork.
Since then the Shark team works daily on the webui. Its not just for AMD. They support Nvidia and Intel Arc cards too.
Also they added LLM support into the webui.
Its the fastest webui you can use for AMD rn.
But your also stuck on low resolution. Sdxl support is beta.
And you can only use 1 lora at a time.
But yea they work on it.
Last month Shark got bought of by AMD

[steady thistle]

Ooh, crossing fingers that AMD is going to fix their stuff!

[ornate elk]

They push forward with them and supporting other open source projects to implement amd machine learning stuff and driver support

[ornate elk]

Are you on a laptop?

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

[halcyon knoll]

No sir, rtx 4070.

[ornate elk]

After the COMMANDLINE_ARGS=
Add: --xformers --no-half-vae
Then save and relaunch

[halcyon knoll]

OH I restarted stable diffusion and it works now!

I did not consider doing that, thanks for the help though. I'll add that just in case

[ornate elk]

Yea I thought a restart will help so a took the chance and gave you the commands for the better performance 😄

If you add them your ready to generate

[steady thistle]

I restart often. Seems to help sometimes

[halcyon knoll]

jesus the eyes on these faces look hidious

https://tenor.com/view/squidward-squidward-tentacles-spongebob-look-around-awkward-gif-496727432057798195

[ornate elk]

I work in IT, and a restart is the problem solver number one

[ornate elk]

Prompt for portrait or close up portrait

[steady thistle]

If you make any YouTube, I'll watch and subscribe.

[fair oxide]

@ornate elk Can you look private?

[halcyon knoll]

I've heard people talk about some ai face restoration thing

Is that worth looking into?

[ornate elk]

First try the basics. Play around and try out.
Faces can be fixed in different ways.
For example by using upscaling or by installing the Adetailer Extension.

A good prompt and a good negative prompt can also help to get better faces

Also changing models is a must. The basic 1.5 one isn't as good as community ones

[steady thistle]

I find ReActor to be a good face cheat that works fast.

[halcyon knoll]

I am currently using sd-v1-4.ckpt

[ornate elk]

That ones is very old (2022)

You can delete it and download Dreamshaper v8 from civitai.com

[steady thistle]

I think I prefer EpicRealism, but your milage may vary

[halcyon knoll]

oh 2 years thats crazy old if we're talking ai stuff

[severe prairie]

Please tell me about the problem of installing this program, because I downloaded pip and thor, but the program doesn’t seem to see it and gives an error, please tell me

[halcyon knoll]

Cannot find a "V8" on that site, you mean XL?

[ornate elk]

Nope

If you type Dreamshaper its the non xl

[halcyon knoll]

Okay

[ornate elk]

In Search

Its 2gb, put that into models/stable-diffusion

[halcyon knoll]

Oh dude this dreamsharper is tiny compared to v1-4

I see size of model doesn't really dictate how well it is

[steady thistle]

Old is big and clunky

[ornate elk]

Hopefully we will see 3gb sdxl models soon

If fp8 support comes everything will be sliced in half again

I remember in 2022 when every model was 7gb xD and I merged some of them and they got to 9gb 🥲

[ornate elk]

Hey checkout my Install guide in the Pinned Messages of this Channel

[ornate elk]

I'll likely setup a Website with guides and helpfull information

If I find the time

[halcyon knoll]

Holy this model is very good thanks!

[steady thistle]

Try epicrealism too. https://civitai.com/models/25694?modelVersionId=143906

[halcyon knoll]

Quick question whats the point of negative prompts. Why would anyone want something to be bad in their shot

[ornate elk]

There you type what you DONT want to see in the image

[halcyon knoll]

Hmm so say I DONT want any lights

or like DONT want ears

I see

[ornate elk]

Yes

If the person wears a yellow shirt and you dont want any yellow color

You add yellow to the negatives

Or you just use blue shirt in the positive prompt

[halcyon knoll]

❤️ Thank you so much for the tips

[ornate elk]

No problem, you can also ask in #📝｜prompting-help if you have more questions about how prompting works.

[ornate elk]

NP, also adding stuff like blurry, deformed, mutated, low resolution to the negatives can help get a better output image

Same goes for the positive prompt when adding 4k, highres, photorealistic, detailed

[steady thistle]

Or extra ears or missing legs or ugly faces... For me, I use this. https://civitai.com/models/72437?modelVersionId=77169 baddream and unrealisticdream, makes it much quicker and easier.

[halcyon knoll]

Thanks I'll check it out

[ornate elk]

I'm off now gn

[halcyon knoll]

❤️ good night!

[steady thistle]

Thanks for your help!!

[worldly rain]

I installed roop but I can't see it on txt2img interface

[steady thistle]

Roop isn't supported anymore, just so you know. FaceSwap Lab and ReActor are still supported. Otherwise, did you apply and restart UI?

[worldly rain]

oops

yes I did

[steady thistle]

I don't think any tech people are on rn, and that's about all I can think of.

Did you give it plenty of time to install without cutting it off too soon because it didn't seem to be doing anything?

[worldly rain]

nop it even says installed on UI

if its not supported then it should be normal, im gonna install the alternatives you mentioned, which one would you prefer?

[steady thistle]

I'd go get ReActor, but that's all I have for you. 🙂

[worldly rain]

alright thanks a lot 🤝

[steady thistle]

And ControlNet can help

but I don't know how. lol

I use ReActor constantly. It is an easy way to get good faces.

[worldly rain]

I have controlnet but its not giving me an option to swap faces

is ReActor good enough for photorealism?

like an AI girl

[steady thistle]

Is it?

[worldly rain]

I guess so 😂 😂

[steady thistle]

That's ReActor of a real person's face on an AI image.

[worldly rain]

thanks a lot

[steady thistle]

hehe, np

[worldly rain]

I have a last question tho if you dont mind

[steady thistle]

Sure

[worldly rain]

I saw this AI model yesterday, which checkpoint model do you think is used here?

the hair is as realistic as it can be, the dress and the bunny ears customization and all that, any tips for me to generate content like this?

[steady thistle]

Tons of them could create that. I use EpicRealism a lot. 1.5 even

Try a LoRA with that, search for Playboy maybe

[worldly rain]

as a beginner idek know what lora means i keep seeing it

[steady thistle]

But you won't have any issues once you get your prompts and negative prompts. grab https://civitai.com/models/72437?modelVersionId=77169

That takes care of your negative prompts

[worldly rain]

alright

what about the dress and bunny ears, can I also generate those with prompts only?

[steady thistle]

Yeah, try taking that image and hitting 'Interrogate'

in img2img

It takes a photo and turns it into a prompt

[worldly rain]

I feel like I owe you 😂

thank you so much

[steady thistle]

No worries, I've been helped enormously. I've only done this for a month now. lol

[worldly rain]

lololol

handshake emoji bro

[steady thistle]

But I like the same sort of thing you do, so I knew some stuff

Try using SDXL when you get the chance. Try JuggernautXL

Different models will make a LOT of difference.

And follow their prompts, try copying their examples even, and adjust from there

[worldly rain]

will do!

[steady thistle]

Oh, here, try https://lexica.art/ Find pics there you like, steal the prompts and you can learn a ton.

[worldly rain]

wow alright this will be useful

thank you so much

🙏

[arctic zinc]

Hello, does anyone use adetailer for mac? It runs extremely slow or are there any alternatives for face restoration.

[frosty portal]

Hello! Anyone know if theres a good way to copy anime character style with ip adapter or something? Im using comfyui, lets say i want to recreate the same character but in another pose, im having issues with the faces (facid doesnt seem to work well with anime faces) and with general style, it doesnt quite get the style right, any ideas?

[willow zodiac]

Hey. Is anyone using ComfyUI and A1111? I'm trying to share the models... between the two and it seems super easy with the yaml file, but I now have it named .yaml, but when I open it in Visual Studio it shows up with the .example extension.

Nevermind I got it

I take it back. I don't got it. I've updated the file with my A1111 path. It is a Yaml Source File, but when I open comfy, I get no link

Still need help

[karmic crown]

I have 5 browsers on all my Macs (Safari, Brave, Chrome, Edge, and Firefox), but I use most of them only for testing of our websites 🤣 my primary browser is always Brave

[karmic crown]

Safari on Mac and Edge on Windows have the best memory/energy efficiency (which is very logical)

[delicate urchin]

AssertionError: Torch not compiled with CUDA enabled when launching

When launching ComfyUI on second user of pc.

[karmic crown]

you probably installed the wrong version of PyTorch

https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#nvidia

but first, remove the old one https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#troubleshooting

[delicate urchin]

ye thats it thanks

[ornate elk]

Yes, with IP-Adapter it should work. Or with Reference mode.

[gray leaf]

Does anybody know how to control the color of object to specific color? Is there controlnet for this? inpainting and prompting doesnt work this good

[vocal burrow]

usually using CutOff is enough for that.

https://civitai.com/models/18840/no-more-color-contamination-read-description

[gray leaf]

This is great thank you

[karmic crown]

if you just want to stop color contamination, cutoff does a good job
if you wanted to have a complete picture in some specific tone (like "iridescent gold" for example) then adding that to the prompt is the proper way
and the last thing maybe you wanted this https://www.youtube.com/watch?v=u-jvjP2k-RU

that was not for you, sorry 🙂

i wanted to reply to @gray leaf

hahaha

[vocal burrow]

I got it. I just answered with the easiest solution according to what I understood their problem is. Just like with any artistic stuff there are many workflow to achieve a similar result.

I haven't watched that video (I'm not really fond of Olivio stuff) but you can probably recolor stuff using controlnet's depth or trace model, using inpainting or a combination of both, ip transfer, pix2pix models, basic inpainting methods, etc

[frosty portal]

Can it work for anime faces as well? Like copying a specific chara ter?

[karmic crown]

i am busy at work, so i posted a first thing that popup in search and looked good 🙂

[vocal burrow]

SD started pretty simple with basic tools but as time goes on, more and more add-ons/tools are added, it's difficult to keep track of everything. I imagine in a couple years learning how to use SD will be as daunting as learning the hundreds of photoshop brushes

[vocal burrow]

np, throwing multiple solutions for a problem is not necessarily a bad thing

[karmic crown]

i usually post things that i used (like cutoff or tone prompt), but that one looked good, so i had to posted (i will try it later, it might actually be a solution for my problem)

[azure magnet]

what is the last version of torch?

i see in some tutorial that there is 2.1.0 but i have 2.0.1 So my is outdate?

[karmic crown]

latest stabile version is 2.1.2, but... if you are on AMD do not touch 2.0.1

if you are on NVIDIA stay with 2.0.2

if you are on mac, feel free to insall 2.1.2

🙂

2.1.2 is not tested by a1111 developers

[azure magnet]

i'm on nvidia rtx 3060

[azure magnet]

so no update for me?

https://civitai.com/articles/132/lora-lazy-training-guide here i find to update my torch to 2.10 for training a lora

[karmic crown]

this will probably work without problem, but as i said, it was not tested by a1111 developers

[azure magnet]

ok so i'll wait, but when the update will be out?

[karmic crown]

i am just a user like you, so i have no idea
but I suggested to devs on GitHub to use 2.1.2 for Macs

[azure magnet]

There is a site that i can see to check if is the update out?

[karmic crown]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases

[azure magnet]

thanks

[autumn void]

Hi, I'm coming back here after being suggested yesterday to upscale my images in order to get a higher DPI value, well... It's the same in all cases regardless of the picture count. So I second my question - is there a way to set the DPI value in any way?

[rotund lava]

no,only way to get a crisper img is to generate a 4k image then downscale it to 1080p

[autumn void]

but that still leaves me with a picture that's 72 DPI

Should I just change the DPI in other software and forget about generating a 300dpi image in SD?

[hot spindle]

yes

[autumn void]

And that's not going to mess up the print quality

?

[hot spindle]

DPI by itself doesnt mean anything

You need print size + DPI to know a given resolution and that "translate" to X pixels

[hot spindle]

No , print quality is determined on image by pixel size

[autumn void]

So in this case the print size is 12"x12" which I figured out is 1152x1152 pixels and the DPI is supposed to be 300

got the size down, resolution but the DPI is 72

I changed it to 300 manually but I was told it's supposed to be 300 from the getgo

[hot spindle]

so, aim for 1152x1152 px , forget the rest...DPI is usefull to get to that value, and thats it.

[autumn void]

hence my peppering you with questions

[grand fable]

Hi, I'm new to this AI generated images and trying to start Stable Diffusion on my M1 Macbook. I installed StableSwarmUI but it's been showing the message about loading backends from server for quite a long time.. Any ideas what to do with it? I haven't been able to genereta a single picture..

[hot spindle]

its usefull if you know a printer works at 300 DPI (not the same as ppi tought, but is widely regarded as same) - so if you wanna print 12"x12" you use that to figure out what's optimal resolution , if its embbed in file 72 , 100,150 or 300 DPI doesnt change a thing or pixel resolution

[spark basin]

Certain models will cause my eGPU to disconnect, then hang when I'm trying to shut down, and show this

[autumn void]

I will keep that in mind, thank you @hot spindle you've been a great help in explaining all this

[hot spindle]

just be carefull because 12"x12"@300 dpi is larger resolution than you said

[autumn void]

that's what consuses me too, 12"x12"@72 dpi is that previously mention resolution but at 300 is way bigger

[hot spindle]

Like I said is a reference for print size:

[autumn void]

so I'm guessing I should be aiming at 3600x3600 to have that size in 300 dpi then

[hot spindle]

SAme resolution, but at 72 is good for 50", but 72 is widely regarded as screen resolution (72 DOTS per Inch) not print.Printing at 72 DPI (dots per inch) is very very low quality

[autumn void]

I'll try to figure it out somehow, thanks again boss

[ivory bramble]

Hey, what channel would I go to if I wanted basic help with controlnet?

[karmic crown]

if it is about usage #📝｜prompting-help is a good place

[ivory bramble]

Oh hey Viking! Always good to see you around

Thanks

[karmic crown]

I have an easy solution for you, but you might not gonna like it:
remove that thing from your computer

download my automatic1111 setup script from https://github.com/viking1304/a1111-setup/releases/tag/0.0.5
and run it like this

sh a1111-setup.sh -f all

It will do all the steps I explained in my step-by-step guide for you automatically

when you get your automatic1111 up and running
try something simple like

spaceman or 1girl
in the positive prompt and click generate

[gritty ridge]

Any idea if NIVIDA Ge Force RTX is so superior cause of CUDA versus M3 Max Pro configured on 128 / 8 TB SSD

For Model s Training tasks

[karmic crown]

depends on mac

[hot spindle]

No idea of particulars but I would say a M2 Ultra is probably on par with a 4070, at least for rendering

[karmic crown]

20 images in 13.2 seconds with SDXL turbo on his Mac

[hot spindle]

Blender benchmarks which are now optimized for Metal

[karmic crown]

on my M1, I could go to buy a coffee while wait 🙂

it takes about 8-10 seconds at least for each of those images on my iMac with Radeon

[hot spindle]

M2 with 128 has one singular advantage, unified memory, altought not as fast as VRAM still is a lot

[karmic crown]

that's why ARM Macs works faster with Stabile Diffusion than Intel + Radeon combinations

also, i have a feeling that 24GB+ RAM makes a lot a difference even on M1 (I have 16GB)

but my 16GB version works much faster than for some guys who installed that on 8GB M2 MacBooks

[gritty ridge]

Thanks.. I consider buying M3 _ as it’s the latest and could be upgraded to Neural Engine and 128/8 TB SSD though not sure if so high investment good cause of future CIDA dominance

[karmic crown]

the main question is how much RAM

[gritty ridge]

128 GB

[karmic crown]

that is good

[rotund lava]

nvidia still faster and for 4k usd you could even buy a 40gb nvidia gpu

[gritty ridge]

Yes .. we spoke about Laptop

[karmic crown]

we are still trying to find the combination of params that will give the fastest image generation without errors on MCurto's Mac - my Macs are like snails compared to his
20 512x512 images with SDXL turbo is fast, maybe not as fast as 4070, but still fast

[rotund lava]

a 4090 laptop then

[hot spindle]

I'm ignorant regarding this but I wonder why can't there be a specific version of A1111 that would run with MLS , didn't Apple release some kind of converter?

For models

[karmic crown]

your A1111 runs on Metal 🙂

see MPS device in info

🙂

[hot spindle]

I can see with Blender that Apple invested time and resources to add Metal to it, and in leass than a year went from zero to hero, I wonder why they aren't doing the same for Stable Difussion

[karmic crown]

https://github.com/apple/ml-stable-diffusion

but i do not recommend

you need to convert models

and as i said, our macs are using Metal

[karmic crown]

you can test metal support yourself if you want
#🤝｜tech-support message

scroll a little more back, for some reason it do not lead to my message even i properly copied the link

download test.py, put it inside your sd folder and run it like this
venv/bin/python test.py

[ornate elk]

Yea you need to click multiple times on it

[ripe holly]

hwy guys! I'm trying to use the sdxl typescript rest api, and im getting this error:for stable-diffusion-xl-1024-v0-9 and stable-diffusion-xl-1024-v1-0 the allowed dimensions are 1024x1024, 1152x896, 1216x832, 1344x768, 1536x640, 640x1536, 768x1344, 832x1216, 896x1152, but we received 400x289',

this doesn't happen when i use the sandbox, which uses the same typescript api. Any thoughts?

[karmic crown]

then you should do what the response suggests: use one of those resolutions
that is obviously intentional limitation of that API you are using

[fair oxide]

Hi guys, do you know if there is any way to clone a face to create another body image with the same face?

[ripe holly]

but the sandbox, which literally uses the same exact code, doesn't have the limitation lol

[karmic crown]

https://platform.stability.ai/docs/api-reference#tag/v1generation/operation/textToImage

[ripe holly]

Yea ive seen that, im confuised why the sandbox doesn't seem to follow the same rules.

[strange zephyr]

So basically, my boss wants me to take this project we did in 2010 ish and magically create good looking lakefront residency pictures from it. I've told him this wouldn't work but he is set on doing this so here we go.

One of my biggest problems are:

1. The AI sees what is supposed to be land as water, creating lakes where there are not supposed to be any.

2. The generated houses generally suck. My boss wants them to look all "light" and "friendly" and not as "brutal" as what the AI generates from the blocky geometry in my control net

I'm using SD 1.5 models, I've tried RealisticVision and ArchitectureExterior 4.0

I'm using controlnet, 0,5 -0,75 strength, I've included images of what I use for controlnet input.

Please, if anyone can help, I'd much appreciate it!

[grand fable]

I tried a1111 but when I changed model to SDXL or Jagernaut it stoped working. Even when I swirched it back to the basic one. Now I installed the ComfyUI

[karmic crown]

please ask this in #📝｜prompting-help
this channel is mainly for techincal questions about installation and problems with installation

[strange zephyr]

k thx

[karmic crown]

how much RAM do you have?

I have M1 MacBook Pro with 16GB of RAM, and I can use SDLX without any problem
it is not fast as i would like it to be, but is very usable

[grand fable]

8 or 16gb not sure now (not at my desktop atm).. SDXL works me well in the ComfyUI so i dont think it’s HW issue.. will try it tomorrow to see the exact error the A1111 shows..

[karmic crown]

can you remember how long it takes to generate an image like 1024x1024 with SDXL in ComfyUI on that Mac?

[upper linden]

Hello, i need help trrying to get my stable diffusion webui to work. Its giving me this error everytime i try to generate a image: AttributeError: 'NoneType' object has no attribute 'lowvram' . which is originating from the sd_models.py

[karmic crown]

show the full log, real error happened earlier

[upper linden]

where would i find the full log?

[karmic crown]

in your cmd window

[upper linden]

so the entire cmd window from start to finish

[karmic crown]

as much as posible

[karmic crown]

*** Error running install.py for extension F:\webui\stable-diffusion-webui\extensions\sd_dreambooth_extension.
*** Command: "F:\webui\stable-diffusion-webui\venv\Scripts\python.exe" "F:\webui\stable-diffusion-webui\extensions\sd_dreambooth_extension\install.py"
*** Error code: 1

first error is at the beginning

remove that folder sd_dreambooth_extension from extensions

and try again

is it working without it?

[upper linden]

nope

i mean i have the webui menu and all

just no image generation due to the lowvram error

its smaller at least

also this is in the bat file for user webui

@echo off

set PYTHON= "C:\Users\crazy\AppData\Local\Programs\Python\Python310\python.exe"
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --deepdanbooru --api --no-half-vae --xformers --skip-torch-cuda-test

call webui.bat

[karmic crown]

File "F:\webui\stable-diffusion-webui\extensions\stable-diffusion-webui-wd14-tagger\preload.py", line 4, in <module>
from modules.shared import models_path
ImportError: cannot import name 'models_path' from partially initialized module 'modules.shared' (most likely due to a circular import) (F:\webui\stable-diffusion-webui\modules\shared.py)

please remove that extension too stable-diffusion-webui-wd14-tagger

did you have some old version of web UI, like 1.5 and then updated to 1.7.0?

[upper linden]

yeah, i didn't want to loose my old outputs and loras, etc

[karmic crown]

do not touch models

just remove this extension

we will see how many you would need to reinstall, but you should be able to use your old models without a problem

[upper linden]

i see its getting smaller

progress i suppose

[karmic crown]

please MOVE all extensions from your extensions folder to some other place
and run your ui again

then restore one by one, and see which one will crash your ui

if that happen, just delete it

when you finish install those that you are missing

[upper linden]

its not crashing, just not letting me generate a image

[karmic crown]

then update all

[ornate elk]

Remove --deepdanbooru --skip-torch-cuda-test and if you dont need it also --api

[karmic crown]

if you can access ui, go to extensions and check for updates

than aplly

and please do what CS1o told you

[upper linden]

lmao

venv "F:\webui\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: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Traceback (most recent call last):
File "F:\webui\stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "F:\webui\stable-diffusion-webui\launch.py", line 39, in main
prepare_environment()
File "F:\webui\stable-diffusion-webui\modules\launch_utils.py", line 384, 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 any key to continue . . .

[upper linden]

how do i bypass the cuda test if it blocks me from the webui

[ornate elk]

What's your GPU?

[upper linden]

7800xt

[ornate elk]

🥲

Then you have the wrong webui version

You used your CPU the whole time

[karmic crown]

fastest and safest way, move the whole installation to some backup folder

follow the proper AMD guid and install everything from scractch

move your models, loras... from backup to new installation

[upper linden]

where would i find the guide for AMD?

i got the drivers installed already, i just do a fresh start for webui no?

[karmic crown]

#🤝｜tech-support message <- here

do not touch the drivers

open cmd and show me what you get with those:

python --version
pip --version
git --version

if all work, you can probably skip those steps too

but let us see

[upper linden]

C:\Windows\System32>python --versionpip --versiongit --version
unknown option --versionpip
usage: python [option] ... [-c cmd | -m mod | file | -] [arg] ...
Try `python -h' for more information.'

[karmic crown]

one by one

python --version

then
pip --version

[karmic crown]

looks fine

please just do this

pip cache purge

[upper linden]

done

removed 42 files

[karmic crown]

good

now follow the install guide from step 1

be sure not use mentioned folders

like onedrive, desktop...

if your internet is fast, it will take just a few minutes

[upper linden]

is there a link to it please?

[karmic crown]

#🤝｜tech-support message <- CLICK THERE 🙂

that was a link 🙂

[upper linden]

my bad

just need to back up my stuff

[karmic crown]

i would just rename the whole folder as a1111-backup for now

[upper linden]

lmao too late

start from scratch is better anyway

also have no space

[upper linden]

welp i got it running again

but it seems like medvram is too much for 16gb

[ornate elk]

depends on what your settings are

[upper linden]

well its a new installl so i guess i got to find out whats happening, before it was super fast

[upper linden]

settings in the webui menu or the bat file?

[kindred tinsel]

Hey guys I;ve been having a constant issue, not an error but a weird issue, with a111 webui. Every time I boot it up nothing works, like I press generate and it doesn't work, I change tabs or update something and it doesn't work, the buttons don't work at all either.

I need to restart the ui one or two times to get it working, it happend just now. I tried to generate the image, the progress bar appeared for a couple of seconds and then went away

Now the Ui doesn't want to show up at all :c idk if the agent scheduler extension could be causing this, because everything was working before I isntalled it

The UI loaded again but it doesn't allow me to touch anything, not even text or buttons on the screen is stuck on this single page image rather than a usable ui

[upper linden]

also is it normal for a 728x728 image to use more than 16gb of vram?

for medvram option

because if so i got to put lowvram

[kindred tinsel]

Hi I've sort of fixed the previous error by restarting the webui .-. . But I had another question, does anyone know how to run IP adapter in comfy and on amd cards. In the ip adapter plus it says its not support but maybe there is another node that allows for it or a way to fix this?

[ornate elk]

What are your txt2img settings?

If you followed my guide you already added the right settings to your webui-user.bat

You need to know that the Directml fork for AMD isn't vram optimized

So its normal that it always takes the full amount of it.

You also need the Tiled VAE Extension If you want to upscale images

[ornate elk]

Make sure the webui is whitelisted in any browser adblocker

[kindred tinsel]

I did that thank you, also do you know anythinga bout ipadapter on comfy for amd users?

[ornate elk]

Nope, I dont use comfyui

[upper linden]

yeah how does it run out of vram for 16gb though

i'm doing a 768x768

[ornate elk]

What are your txt2img settings?
Steps, resolution etc?
You can make a screenshot

[white thorn]

Hello there, I would need help of someone that create Lora on Nvidia RTX4000 card. I fixed my issue with "NansException: A tensor with all NaNs was produced in Unet""
When trying to use my lora. I noticed that my samples image was most all black, so I changed all my settings one by one to see what causing this. And it was adamw8 bit. everything works fine with regular adamw. So if someone have any info or solution ..

[ornate elk]

how many tokens has your prompt?

[upper linden]

how do i check that?

[ornate elk]

also everytime you get an "out of vram" error you need to restart the webui to clear the vram. if not every generation after that will fail too

[upper linden]

yeah but is it normal to happen on 16gb gpu, could be a Amd bug

[ornate elk]

yes im on a 7900xtx 24gb and it happens for me too sometimes

[upper linden]

gg

[ornate elk]

its the unoptimized directml version

[upper linden]

anyway to get a optimized one

[ornate elk]

but i know every setting you need to generate good images

[upper linden]

oh

[ornate elk]

waiting for Rocm support on windows maybe (february-march)

[upper linden]

gg

and my old 2070 was faster than this

rip

[ornate elk]

yea nvidia works better, but me and you still can generate images pretty fast, you just need to know how

[upper linden]

welp i'm all ears

[ornate elk]

first make sure you close/kill Wallpaper Engine and Epic Games Launcher, and all other heavy programms running in the background

[upper linden]

all i got is discord

[ornate elk]

then restart the webui

then you try to generate a 512x768 image, 30 steps,
with prompt: Cat
and negative: Blurry

a test, that should work without problems

[upper linden]

works

1.7 it/s

took 41 secs

[ornate elk]

what model are you using?

[upper linden]

v1-5-pruned-emaonly.safetensors

[ornate elk]

ah okay

thats old and also slower as its 4gb

try download this one:
https://civitai.com/models/4384/dreamshaper

its a good starter model, and only 2gb in size

[upper linden]

so models do better on certain GPUs?

[ornate elk]

smaller ones are faster

quality isnt affectet by the size

@upper linden As next Step you need to install the Tiled Diffusion Extension. (That will safe you most of the time from out of vram errors).
Go on Extensions, click on "Available", click on "Load From", then you get a List. Search for Tiled Diffusion and click install.
Then click on the Installed Tab, and There click "Apply and restart"

[upper linden]

i have that installed already

[ornate elk]

thats good

you dont need to use the Tiled Diffusion stuff. Whats important is the Tiled Vae Option.
Enable that and set the Encoder to 1024 and the Decoder to 128.
thats important if you want to use hires fix to upscale an image.#

and then try these settings with the cat

that should work on your card, as it works on a 6700xt 12gb too

if that works, you can also try to get a FullHD image by setting the resolution to 540x960

and as last thing to mention, make sure the tokens dont go over 225

Also noteworthy are SDXL models (6gb+) these can be used to but are much slower. You can try them at 768x1024 native res with tiled vae enabled and without hires fix.

[upper linden]

aren't those models meant for high end cards or at least optimized ones

[glad fiber]

I am trying to train a lora, but when I try to use it it doesnt work

I am training with kohya_ss

my captions look like this

trying to train the din djarin from the mandalorian, but my output images has nothing of it

10 images with 30 step training each

I followed this tutorial https://www.youtube.com/watch?v=_F39RbO3tYo

[ornate elk]

Yes, so you shouldn't use them but you can in lower res

[meager carbon]

Heya, I am currently trying to use SD with an AMD GPU, and although it is not officially supported, i maybe thought y'all could have some insight on it.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs
I am using this as a general guide, and I got it up and running, but it refuses to use my 7900XTX, and instead uses my CPU.

Any idea as to what i can try for it to use my GPU?
Windows 11 btw

[ornate elk]

Hey, best you can do is to follow my install Guide for AMD. Its in the Pinned Messages of this Channel.
I have a 7900XTX too and it works 🙂

[meager carbon]

awesome, I really appreciate it.

this one, right?

[ornate elk]

For any questions feel free to ask 🙂

[meager carbon]

Of course, I really appreciate it.

found my first issue

[ornate elk]

Did you switched from nvidia to AMD ?

[meager carbon]

yeah.

I went from a 1080ti to the current one

i should prolly uninstall the gpu driver first

[ornate elk]

Then you need to uninstall all nvidia drivers.

Then you need to restart the PC
And then delete the venv folder inside the stable-diffusion-webui folder

The second issue is your Python 3.11 version

[meager carbon]

that's the one that i was pointing out

[ornate elk]

xD no the first issue was "nvidia drivers found"

[meager carbon]

i know

and i didn't see it

until i posted the ss

im a bit silly.

[prime mulch]

I am having a bit of a problem, I got a 4070 graphics card laptop, nvidia, heck it can run ark survival ascended. I was gonna try to get stable diffusion on this as well, but when I went to check my vram just incase it says i only have about 448 mb, even tho when I bought it, its supposed to have 8 vram. Im not sure if theres a setting issue going on here, or if its just displaying weird. I have 15 gb memory, and this graphics card does say it has 16 gb vram, but since its a laptop thats cut in half

[glad fiber]

windows?

[prime mulch]

Yes

[ornate elk]

where does it shows 448mb ?

[prime mulch]

I have not tried stable diffusion yet, to be clear, im wondering if this is normal, but my other computer did not say this

Huh for some reason right now its saying amd rradeon instead of Nvidia, im not sure how to make it use nvidia here

[glad fiber]

what does it show in the task manager?

[glad fiber]

in the gpu

[ornate elk]

thats the IGPU from your Processor, not the nvidia one

[glad fiber]

i am seeing 2 there, which seems to be right

[prime mulch]

Oh thank you, im not great at this kind of stuff yet, usually im pretty good with computers but the graphics cards and processors stuff confuse me, heres the gpu for nvidia

[spice wing]

Hey guys, I have reinstalled SD after getting an error and now im getting this error:
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

I know that i can fix it by selecting to run it with half, but then it only uses my cpu and not my gpu.
My setup is: RX 6900 XT, Ryzen 7 7800X3D, ASRock B650 PG Lightening, 32 gb 6000 mhz RAM.
My command lines arg are: --lowvram --disable-nan-check --skip-torch-cuda-test

Is there any way, to fix it? It somehow worked before...

[ornate elk]

hey you probably installed the wrong webui. AMD needs a different one if your on Windows.
Follow my Install Guide for AMD. Its in the Pinned Message of this Channel

Looking at your settings, yes your using CPU only mode rn. so you have installed the wrong version

[spice wing]

Oh ok, thanks Ill try that

[kindred tinsel]

Does anyone know what this means on comfy? Is it basically using my cpu to render the image or what?

[glad fiber]

do you have a gpu?

[kindred tinsel]

Yeah an AMD one

its supposedly working rn at 100% but that warning scares me

[ornate elk]

its using your cpu for a specific task

[kindred tinsel]

okay? and what is that :=

[glad fiber]

it is nothing to worry about then

[ornate elk]

idk, but model_sampling indicates something with samplers

[kindred tinsel]

hmmmmm well idk but thanks for the help ^^

[glad fiber]

looks like that thing doesnt have support for amd yet

but nothing important ig

[kindred tinsel]

I guess, it isn;'t like it doesn't work so something internal isn't working yet, but everything kinda works

[glad fiber]

I still need help with this

[prime mulch]

So with 8 gb vram do you guys think I have enough to be able to create my own loras?

[ornate elk]

yes, lora training works with 6gb and more

[glad fiber]

I created one with 4gb

[prime mulch]

nice! My old laptop only have 4 gb so I could not, person above me is lucky

[glad fiber]

I couldnt manage to get it working (skill issue) but it trained

no OOM

[kindred tinsel]

How can you train a lora? Which tutorial did you follow?

[ornate elk]

wow pretty fast too

[glad fiber]

I followed this and another one https://www.youtube.com/watch?v=_F39RbO3tYo but I think I may have done something wrong because I cant get my chracter in the image

I didnt use a SDXL model, of course

[meager carbon]

I'm having trouble getting SD to view the correct python path, I have 311 uninstalled.
how can I point SD at a specific python path? Can I edit the webui-user.bat?

[ornate elk]

you need to delete the venv folder again

[meager carbon]

where would that be located?

[ornate elk]

in the stable-diffusion-webui folder

[ornate elk]

best darkness cmd 10/10

[meager carbon]

ty, it's a gif, and she dances.

[ornate elk]

😄 is it easy to do do that?

[meager carbon]

very

[ornate elk]

perfection ^^

[meager carbon]

Glad you like it.

it's using windows terminal

which i HIGHLY recommend

[ornate elk]

wow thats very very easy

[meager carbon]

mhm.

[ornate elk]

thanks a lot 😄

[meager carbon]

I personally did it in the json

but it's up to you.

"what type of psychopath keeps their terminal backgrounds in their downloads folder"

me.

[ornate elk]

there are people who install their SD webui in Downloads 🥲

[meager carbon]

totally didn't do that

not this time around

but when i first used SD

i did

[ornate elk]

it happens i guess

[meager carbon]

i didn't wanna move it

was afraid it would break

if it had like relative file paths or whatever

[ornate elk]

yes it would break the git

so no updates (but its fixable)

[glad fiber]

I guess it would not

because the a1111 has a venv

and if you move the root directory the venv would follow it

[ornate elk]

the webui itself wouldnt break, only the git path to update it

we had a lot of users here who moved them, and git pull didnt worked after that

@meager carbon does it work now with python?
if not you also can add the path to the python.exe in the webui-user.bat

[meager carbon]

can i change it from medvram to highvram?

[meager carbon]

still working on it

python did work

just doing the rest now

is 24gb high, or is that still med?

[ornate elk]

there is no highvram, and nope, the AMD webui directml version on windows isnt good optimized in terms of vram management.
Its noteworthy to say that you would get the best performance and usability on Linux because there are the Rocm drivers needed for SD.
AMD working on Rocm support for Windows but that will come in maybe 1-2 months.

[meager carbon]

aha I see

delete venv and try again?

[ornate elk]

first run
pip cache purge
in a cmd
then delete the venv again

it mostly found old torch stuff from the nvidia drivers before

[spice wing]

Tysm, it worked. And now even SDXL Works

[ornate elk]

perfect, no problem 🙂
also you should get the Tiled Diffusion extension if your going for upscale later. It helps to not get out of Vram.
Also when you get the out of memory error, you always have to restart the webui.

[glad fiber]

@ornate elk do you know about lora training?

[ornate elk]

i did trained a year ago, with khoya_ss

so im not updated on that

[glad fiber]

I trained a lora here, but when I try to use it, it is like if the lora wasnt there

I get that kind of output

[meager carbon]

holy FUCK that's fast

thank you

@ornate elk really appreciate all your help

works great now.

she's having fun

[ornate elk]

no problem 🙂 glad it works, for information like i said to BasisZock above too, you want to get the Tiled Diffusion Extension later if you want to upscale images. But you can @ me if you need the exact settings or how to steps.

[meager carbon]

oh i think you will also like my gpu, cs1o

a bit off topic

[ornate elk]

thats epic! love it

what brand is it?

[meager carbon]

yeston

[meager carbon]

can't i use the upscalers?

[ornate elk]

lol thats a brand i didnt know the make custom amd cards, but seems not like they sell in EU

[meager carbon]

you can buy them anywhere iirc

they made a great 4080 also

[ornate elk]

yea these are needed for upscale, but with the wrong settings you will run out of vram, (an issue of the unoptimized directml version)

thats where the Tiled Vae from Tiled DIffusion extension comes to help

[meager carbon]

I see.

ill be fine i think

[ornate elk]

you should be able to upscale 512x768 by 2 with 0.5 denois and 10 hires steps just fine

[meager carbon]

uhm

i upscale 512x1024 by 2 with 10 hires steps and 0.7 denoise just fine rn

[ornate elk]

and it worked?

[meager carbon]

mhm

the anime GPU strikes again

[ornate elk]

yeston magic gpu?!

[meager carbon]

the yeston gpu is not actually that good, the average to hotspot delta is like 18 deg

[ornate elk]

oh, okay xD that was my thought too. because i have the Spahire Nitro+ Vapor and that should be the best custom version out there

[meager carbon]

i got this one for 900

so like

yk

i put liquid metal on it too

dropped delta by 2 deg

[ornate elk]

ah nice

[kindred tinsel]

wait yeston not good? I was going to buy a 4080 from them :=====

[meager carbon]

i mean

in terms of temps they are... eh

also 4080? cursed.

amd on top 💯 💯 💯 💯

[kindred tinsel]

meh 4080 better with more vram than 4060 or 4070

[kindred tinsel]

haha no

nothing is optimized for us

[meager carbon]

linux actually works though :3

[kindred tinsel]

linux? 🤮

[meager carbon]

you realize like 99% of the things you do online are run in linux

[kindred tinsel]

As long as I dont have to install it, im gud

[ornate elk]

Linux is nice but i still hope for the upcoming Rocm Windows support, and yes AMD best

[meager carbon]

hopefully.

[ornate elk]

got an Ryzen 7 7800X3D too

[meager carbon]

based!

there you really can't argue that intel is better

[ornate elk]

absolutly

[meager carbon]

i think i have like 8 epyc cpus

and im never going back to xeon

atleast for now

[ornate elk]

how large is the lora file?

[ivory bramble]

What kind of model should I use to re-create this kind of style wallpaper?

[ornate elk]

figured out, i had --no-half in the webui-user.bat (That is needed for AMD if you want to use Inpaint in img2img). But that slows down a lot and uses more vram.

[meager carbon]

I see.

well i have it too

so yk

[meager carbon]

i would try anything ink, with the correct prompt

but honestly idk

[ornate elk]

your sure? in my guide i only mention --no-half-vae and --no-half for RX 5xxx cards

[ivory bramble]

Hmm maybe I should integrator?

[hazy merlin]

I cant use xformers, an error happens, despite me using python 3.10

[ornate elk]

Hey, can you show your webui-user.bat content?

Seems like a typo

[hazy merlin]

im pretty sure i did it correct, im not sure why its doing this

[ornate elk]

Hm try to delete the space between = and --

[hazy merlin]

same error

[ornate elk]

You need python 3.10.6 or up to 3.10.11

Not 3.10.0

[hazy merlin]

Im using 3.10.6

should i install 3.10.11?

[ornate elk]

That wouldn't help

[hazy merlin]

oof, I have no clue whats happening then

[ornate elk]

whats your GPU?

[hazy merlin]

A nvidia 4070

[ornate elk]

replace everything in there with this:
`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --no-half-vae
git pull
call webui.bat`

then save and try again

[hazy merlin]

That seems to be working, not sure what I did differently, but thank you

[ornate elk]

no problem, maybe a invisible Tab or Space

[karmic crown]

finally someone who knows how to use windows terminal 🙂

mine look the same as my iTerm on mac

[steady thistle]

How can I get insightface? It's not loading with ReActor

[ornate elk]

Follow this carefully:
https://github.com/Gourieff/sd-webui-reactor#installation

[steady thistle]

Okay, will do. It's probably a c++ thing. Thanks!

Yeah, I looked at it again and I /think/ I had clicked the wrong option previously with Desktop Development with C++

As I work on my potato with better internet speed. lol

[steady thistle]

That got me past the errors. I then tried to use the suggested args and it didn't recognize --use-directml... I'm thinking I did something wrong in the install.

(I did this install before I discovered this wonderful server 🙂 )

[ornate elk]

oh okay that means your webui version isnt updates to 1.7.0

[steady thistle]

Ah... hmm... should I do a git pull?

it says version 1.7.0 actually

I think I installed from a different SD fork or whatever.

[ornate elk]

yea if you didnt install the directml fork it wont work

[steady thistle]

What does directml do for me?

[ornate elk]

what was your gpu again?

[steady thistle]

This is the work one... let me see...

[ornate elk]

the directml webui version is made for AMD

there is only one other webui fork of auto1111 called sd.next that works with AMD

[steady thistle]

Ryzen 5 4600H

So it is working, but maybe not as good as a directml install?

[ornate elk]

idk what you used

if you used the official auto1111 then it would run on CPU only

if you use SD.Next or auto1111 directml then it would use the IGPU with directml

my guide is for the auto1111 directml version

[steady thistle]

I'll redo it later. It's an easy enough install particuarly as my work requires low cpu usage.

[ornate elk]

ah now i get it, you tried to install the webui for on the work pc?

[steady thistle]

Yep!

I'm at work now.

[ornate elk]

if it has an AMD GPU you can follow my install guide

[steady thistle]

Yep, it is pretty easy with your install guide.

[ornate elk]

thats what i hoped 🙂

i tried to make it easy to understand

[steady thistle]

It's great. That pip cache purge was genius. That was what everyone else missed.

[ornate elk]

yea people tried to get it running on amd for hours or days while the installation always took the cached nvidia stuff xD

[steady thistle]

I watched SO MANY install videos on youtube and searched for guides for 2 weeks after I broke my previous install.

Yours was the only one that brough up the cache issue

ttyl!

[pine canyon]

hi, how to generate “art” images and have access permission?

[ornate elk]

Hey, where?

In the #1080261341362786384 channels?

[pine canyon]

in all bots

[paper viper]

@ornate elkDo you use stablestudio?

[ornate elk]

Nope sry

[ornate elk]

With the / dream command

Without the space

[pine canyon]

THANKS ...

[astral comet]

@ornate elk hey! I guess it's normal to have long generations overall with a 1070?

[glad fiber]

151mb

[hazy merlin]

im not quite so sure how to determine steps for lora, i have 75 images

[steady thistle]

@ornate elk Do you have any info on https://huggingface.co/latent-consistency/lcm-lora-sdxl ? It's supposed to be a LoRA that increases speeds by reducing needed steps, but it has a whole thing on installing. Thoughts?

[zealous cargo]

is this the ultimate performance boost?

okay do not use sdp it actually made it worse

[leaden needle]

hey so tbh I'm fairly new with stable diffusion but I came across it recently and found it fascinating so I downloaded stable diffusion xl, along with comfy ui, but everytime I try to que a prompt using nvidia I run into a wall of red errors along the lines of "CUDA error: device kernel image is invalid" does anyone have a potential fix for this? or can help me out in any sorta way?

[willow zodiac]

Can anyone help me figure out why ComfyUI is not getting my checkpoints loaded in my a1111 path?

[thorn goblet]

im sorry,im quite intoxicated so im gonna require some help understanding things. i wanna download the whole ai image generation thjing, but im not sure where to start.... faq is very overwhelming rn so if anyone can help break iot down i would greatly appreciate

[willow zodiac]

Are you using a Nvidia gpu?

[thorn goblet]

im not sure where to check that...

give me a moment

its anm intel core

Intel(R) Core(TM) i5-6500 CPU @ 3.20 GHZ

[willow zodiac]

That is your CPU. If you go to your Device Manager>Display Adapters, that should show you what your GPU is.

[thorn goblet]

oh im sorry

[willow zodiac]

no worries

[thorn goblet]

Intel(R) HD Graphics 530