🤝｜tech-support

[rose apex]

6.2.4 windows 11/10

[rose apex]

In ComfyUI, the installation part worked smoothly but I get this error while generating image.

[round elk]

can anyone help? when run infinite talk workflow

[lime pivot]

@vocal burrow yep dw been here long enough to know haha. just found a solution online so deleted my msg

[ornate elk]

Your GPU needs hip SDK 6.4 not 6.2

[rose apex]

But the guide that I'm following recommends to install 6.2

[lapis raven]

Where does “full” come from?

[ornate elk]

Which guide? If its mine, it needs an update because hip SDK 6.4 got released 2 days ago xD

[ornate elk]

I mean the whole cmd log (all of it)

[rose apex]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-automatic1111-with-zluda

Damn that's really you? 💀

I didn't even read the name until this embedded picture showed up just now

Anyways, I am currently trying a clean install with 6.4.2. Then I'll see what happens

[ornate elk]

HIP SDK 6.4 includes native support for the 90xx cards. so you can follow the guide but dont to the step 5

I will update the guide soon.
Let me know if it worked

[rose apex]

It didn't. I'm getting a missing cudnn64_9.dll error with comfyui. I'll just wait for the upgraded guide ig, I'm not very good with these kinds of stuff

[ornate elk]

can you try the Auto1111 with zluda guide? it already has hip sdk 6.4 support

[rose apex]

Ok

[ornate elk]

you can skip the zluda steps

just setup the webui should be enough

[rose apex]

I skipped the Zluda part but when installing the webui there was an error at the end which I couldn't read because I closed cmd by mistake. However, when I ran the webuser.bat file again, it somehow worked and launched the webui. It's currently being compiled.

There's also this timm.layers warning thing and a pytorch lightning message

[ornate elk]

The warnings can be ignored

[rose apex]

Yes it's working now 🔥 thanks

Everything seems to be working, loras, upscalers and all.

The compilation took exactly an hour

[lapis raven]

Sorry for the delay. Here it is.

[ornate elk]

np and thanks for the info, so i can edit the guides and it will work

[ornate elk]

add --skip-ort to the launch args too then delete the venv folder and relaunch the webui-user.bat
alos which extensions have you additional installed?

[lapis raven]

You've installed a lot of things.

[ornate elk]

has it worked before installing those?

[lapis raven]

I haven't touched it, so I don't know...****

[ornate elk]

can you screenshots your extensions folder?

[lapis raven]

ok

[hot thorn]

@ornate elk can you help me?
i got new gpu now, and i want to increase quality of my pics. so I went ahead and increased upscaler to 2.0. but it created distorted torsos and sometimes other body parts and even parts of the image (15 high res steps). so i lowered upscaler to 1.8 and increased resolution to 1152x1152, but the issue still persists

[ornate elk]

you installed those before even testing if the webui works? xD

[lapis raven]

I think it's better to do things quickly lol

[ornate elk]

delete these and then delete the venv folder, dont forget to add -skip-ort to the webui-user.bat
then relaunch it

[lapis raven]

← The OS is also a guinea pig without thinking about the consequences.

Will it recognize it if I put it back?

There was an error.

[ornate elk]

can you try to generate an image without a lora or anything special checked?

also can you try to generate with this model, put it into models/stable-diffusion and select it in the dropdown
https://civitai.com/models/4384/dreamshaper

[lapis raven]

By the way, is it normal that the “rocminfo” command does not work?

[cloud monolith]

anyone able to help with set up stable difsuion? im using a 5090. and.. i followed normal steps as per many guides and i keep getting an error

[ornate elk]

hey, checkout the first link of the pinned messages, there are guides with the steps for rtx50 series too

[cloud monolith]

was trying to use A1111

[lapis raven]

error…

[cloud monolith]

i.. already tried those and that didnt fix anything as already said?

[ornate elk]

can you provide more info?

like a cmd log

[cloud monolith]

im trying to use stable difusioon and get this error: RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

in the web broswer ui when it ry to use

[ornate elk]

complete cmd log pls

[cloud monolith]

...how do i even do that

[ornate elk]

you can copy the content of the cmd window after you got the error

[ornate elk]

oh i see, you have hip sdk 6.1 installed but there are gfx files for hip sdk 6.2
you should unintall hip sdk 6.1, then install hip sdk 6.2 and then download the updated gfx files for gfx1010 and follow the guide how to include them to rocm again.
Then reboot the PC.
Then delete the venv and the .zluda folder (boath are in stable-diffusion-webui-admgpu)
then relaunch the webui-user.bat

[cloud monolith]

so..

[ornate elk]

yea you clearly didnt followed the pinned guide

[cloud monolith]

i did though....

[lapis raven]

Thank you, I'll give it a try.

[cloud monolith]

do i need to uninstal everything and start over?

[ornate elk]

no, you have the wrong python version,
you installed it in the folder which is mentioned not to be installed to!
and you didnt made the steps needed for the rtx50 series
Here is the guide: (automatic1111 nvidia)
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[cloud monolith]

but i followed a youtube guide that told me that was ok?

[ornate elk]

yea but they arent experts, and sometimes give false information

[cloud monolith]

so do i delete the whole thing and start over using just that guide?

[ornate elk]

you can delete the webui folder and start with the guide

[cloud monolith]

are you? 🤨

[ornate elk]

im doing this since 2022 and im a community guide here so i have a pretty good knowledge, i also made all the guides you find on the link xD
and im running this channel here mostly as one man support xD

[cloud monolith]

https://tenor.com/view/oh-dear-uhoh-uh-oops-gif-22478082

[hot thorn]

u can ignore him for being a sausage and help me instead xD

[ornate elk]

oh 1 sec need to scroll up

[ornate elk]

what was your old gpu and what is your new one? and which model got used?

[hot thorn]

had gtx 1070, got rtx 3060, by model do u mean checkpoint?

[cloud monolith]

since im starting from sctach and changing python, do i also need toreinstal github since i uninstalled python and am isntalling the one from the guide? or will github be unaffected since im reinstalling the guides python?

[ornate elk]

if you mean Git and not Github then yes its unaffected

but even if you installed github gui it should be fine probably

[cloud monolith]

gitbash?

sorry they all look the same to me >.<

[ornate elk]

should be this (https://git-scm.com/downloads)

[hot thorn]

@ornate elk i was trying to use chat gpt for help, and it suggested using controlnet for openpose and depth, but something seems to not work either

[ornate elk]

can you show the settings you used and the output?

[cloud monolith]

using ai.. to fix the ai. 😂

[hot thorn]

output is nsfw so cant share here. but basically distorted torso, fucked up feet and hands, sometimes messed up body parts

[hot thorn]

its a ciiiiircle of life

[cloud monolith]

https://tenor.com/view/darktide-adeptus-mechanicus-mechanicus-warhammer-tech-priest-gif-14919800332216808310

[ornate elk]

can you try 832x1216 and then upscale by 2, hires steps 15, denois 0.5 ?

[hot thorn]

not rly, im using embeddings and for them pictures need to be square

[ornate elk]

that should normaly output a good image

oh which embedding does require that?

[hot thorn]

ones from HDA guy on civitai

[ornate elk]

can you show me an output in dm ?

[cloud monolith]

why do guides not have pictures

https://tenor.com/view/books-the-book-read-belle-gaston-gif-7494310

(also am following ugide as best as i understand it. just letting ot do its instal thingy now)

[ornate elk]

didnt had the time and space to add them to every guide there, and i think its mostly understandable written, but im always open for suggestions

[ornate elk]

let me know if it worked or if any step isnt clear

[cloud monolith]

Will do. It's like 55% done a 3.97 GB download

I have one of the world's best graphics cards, but one of the world's worst download speeds.

[cloud monolith]

ok i got to step 6

and i know it sasy for 500 series i need to do aditonal steps

but do i close the cmd i had pen for those steps?

[ornate elk]

yes

[cloud monolith]

after step 7 it should run good right?

installing torch right now. but want to know if theres anything i should be keeping aneye out for that might not be int he guide

[ornate elk]

nope you should be fine, but if xformers doesnt get used then you could add --opt-sdp-attention to the webui-user.bat
we will see if you generated the first image

[cloud monolith]

it generated. i used a single word promt. so image is kind ascuffed but looks like its running now

[hot thorn]

yay step 1 out of 10000 finally behind u xD

[cloud monolith]

do you recoemnd any extensions for A11111?

[ornate elk]

can you show me the cmd log again to check?

[cloud monolith]

venv "C:\AI\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr 5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Launching Web UI with arguments: --xformers
W0821 17:23:33.438000 11380 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\AI\stable-diffusion-webui\venv\lib\site-packages\timm\models\layers_init_.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
warnings.warn(f"Importing from {name} is deprecated, please import via timm.layers", FutureWarning)
Loading weights [6ce0161689] from C:\AI\stable-diffusion-webui\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors
Creating model from config: C:\AI\stable-diffusion-webui\configs\v1-inference.yaml
C:\AI\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub\file_download.py:945: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True.
warnings.warn(
Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Startup time: 79.2s (prepare environment: 50.5s, import torch: 3.8s, import gradio: 0.7s, setup paths: 22.9s, initialize shared: 0.2s, other imports: 0.2s, load scripts: 0.2s, create ui: 0.4s, gradio launch: 0.2s).
Applying attention optimization: Doggettx... done.
Model loaded in 2.0s (create model: 0.4s, apply weights to model: 1.3s, calculate empty prompt: 0.2s).
Reusing loaded model v1-5-pruned-emaonly.safetensors [6ce0161689] to load XD\prefectIllustriousXL_v3.safetensors
Calculating sha256 for C:\AI\stable-diffusion-webui\models\Stable-diffusion\XD\prefectIllustriousXL_v3.safetensors: 1a66b7e7f5411f317ffb0e809eb6c78cd80a1ec173a09dc55709581ae44ceaa2
Loading weights [1a66b7e7f5] from C:\AI\stable-diffusion-webui\models\Stable-diffusion\XD\prefectIllustriousXL_v3.safetensors
Creating model from config: C:\AI\stable-diffusion-webui\repositories\generative-models\configs\inference\sd_xl_base.yaml
C:\AI\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub\file_download.py:945: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True.
warnings.warn(
Applying attention optimization: Doggettx... done.
Model loaded in 3.1s (create model: 0.2s, apply weights to model: 2.7s, apply half(): 0.1s).
Downloading VAEApprox model to: C:\AI\stable-diffusion-webui\models\VAE-approx\vaeapprox-sdxl.pt
100%|███████████████████████████████████████████████████████████████████████████████| 209k/209k [00:00<00:00, 1.64MB/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 9.55it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.64it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.16it/s]

[ornate elk]

yes, there are:
Booru Tag Autocompletion
Adetailer
Controlnet

[ornate elk]

ah seems like xformers isnt used, so you can edit the webui-user.bat again and remove --xformers and instead add --opt-sdp-attention

[cloud monolith]

trying that now

aaaaa

[cloud monolith]

now i get an error when i try to open it

[ornate elk]

okay then remove --opt-sdp-attention again and add --xformers back xD and we leave it as it is

[cloud monolith]

[ornate elk]

the rtx50 gpus are not that good supported currently by the webuis

auto1111 is also a bit outdated, but still fully usable

[cloud monolith]

i got the 5090 becuase i thought it would make using ai stuffs super better

[cloud monolith]

do you have a recomendation for something else thats easy to use and up to date?

iv onyl every really used a1111

[ornate elk]

There is Forge webui, which is based on Auto1111 and a bit more updated and supports Flux which your GPU can use easily too

[cloud monolith]

im technologically ilterate

[ornate elk]

you find an install guide for it at my guides

its as easy to install as auto

probably easier

[hot thorn]

im using forge too.

[cloud monolith]

is it as user friendly as a1111?

[ornate elk]

its the same ui

with a bit of extra stuff

[cloud monolith]

does this mean i have to reinstal everything again

[ornate elk]

you can move the models over

and the extensions (but not controlnet!)

[hot thorn]

since we're talking forge. is it true that forge is slightly faster than a1111, but has slightly worse detail?

[ornate elk]

nope didnt saw any differences

[hot thorn]

speed wise and quality wise?

[ornate elk]

im on AMD, the speed between auto1111 and forge with it is marginal, quality is the same

on nvidia, if you would have a low vram gpu below 10gb vram, forge is faster than auto1111

[cloud monolith]

so is a 32gb vram good then?

[hot thorn]

i tried to recreate the same image as a friend on a1111 and my image was slightly less sharp than his, even tho settings were exactly the same tho

[hot thorn]

shush man

my 12 gb is listening

u really want to make it feel bad?

[cloud monolith]

how do you do anything with 12?

[ornate elk]

its the best

[cloud monolith]

could they be using an upscaller for more sharpness?

[hot thorn]

as i said, EXACTLY the same settings

[cloud monolith]

https://tenor.com/view/gollum-lord-of-the-rings-im-not-listening-lotr-gif-12832328

[hot thorn]

somebody needs to be thrown into lava

[cloud monolith]

is Booru Tag Autocompletion the one that like suggests words an dtsuff as you type promt?

[hot thorn]

yes

[cloud monolith]

then why no work when installed

oh nvm now it works

[hot thorn]

reset webui

[cloud monolith]

i swear to god stuff dosnt work and then as soona s i ask for help it works

[hot thorn]

give it more than 5 seconds xD

[cloud monolith]

https://tenor.com/view/star-wars-tarkin-fear-fear-will-keep-the-local-systems-in-line-gif-23372956

[cloud monolith]

https://tenor.com/view/peter-griffin-but-i-want-it-now-gif-26307521

[hot thorn]

nah

[cloud monolith]

also why does ai seem to like making big boobs. lol

liek unelss i promt smaller chet or something my anime girls have huge chests

[ornate elk]

it depends on the model you use

but most anime models like big boobs xD

[hot thorn]

ye, even flat chest often makes small boobs

magic of giant knockers i guess

[cloud monolith]

https://tenor.com/view/the-dictator-strong-boobs-crush-break-gif-5325535

does anyone here use civitai?

[hot thorn]

ye

[cloud monolith]

i founbd a model i like, but i want to find loras that work with it. model is Prefect illustrious XL. but idk witch SDXL it is and the website has a bunch of versions

[hot thorn]

if ure going for nsfw then i guess u can dm me, and i can help out, ive been doing it for a while

[cloud monolith]

im just gettign back into stable after liek being away for a year

last time i used it there wernt all these options

[hot thorn]

hit me up if u want some recommendations

[ornate elk]

check the base model in the info box on the right of it

it should say SDXL, Pony, Illustrious, NoobAi, and then you know which loras is compatible. it has to be the same type

[ornate elk]

Left Auto1111, Right Forge.

its ab it different but quality wise its the same

[hot thorn]

a111 from friend

mine on reforge

exactly same settings

and prompt

[ornate elk]

can you share the lora link?, ah found it

would like to try

[hot thorn]

even without lora theyre different man

i dont know if i have something changed in my settings that causes this, but my prompt and generation settings are the same

[ornate elk]

im getting this image

but im on AMD so the hardware seed is different

[hot thorn]

im just gonna give up for now. work on negative prompt, then maybe instal a1111 along my reforge, check shit, then move on to the upscale 2.0 and try more

trial and error

[ornate elk]

oh wait you were on reforge, oh cant test that rn

[hot thorn]

and ye, i deleted about half my negative prompts and the image is MUCH worse

meanwhile, after deleting half of what i had, and using the thing u sent me, its better

but still some things are worse

[rose apex]

I am getting OOM errors with 9060 XT 16 GB on A1111. I tried both 1024x1536 resolution and 512 X 768 with 2x Hires.fix. Do I need some command line arguments for memory management?

[ornate elk]

you need to always set the hires steps to 10 and then for upscaling that highres you need the tiled vae extension, there enabled only tiled vae after installing it

[rose apex]

It's not fixed, I still get the same error. This doesn't happen in SD.next but sdnext Hires.fix doesn't work properly.

This is cmd error log

[ornate elk]

which model did oyu used? sdxl based?

also make sure your not run anything lese in the background like games or wallpaper engine

[rose apex]

Illustrious

[rose apex]

Yes nothing is open in the background

[ornate elk]

alright

you shouldnt get an out of memory with 768x512, thats really strange

can you relaunch the webui and try again?

[rose apex]

I already did a few times

I will try different checkpoints now

But I can generate 1024x1536 images directly in sdnext

[ornate elk]

also try out using this vae file:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
it lowers the vram usage

it goes into models/vae

then you have to go into settings -> User Interface -> Quicksettings and there add sd_vae
then hit apply and reload ui

then you have a vae dropdown to select it

[rose apex]

Ok let me try

[rose apex]

Still the same, even with different illustrious models and this vae

[ornate elk]

hmm how is the vram usage in taskmanager when generating?

[ornate elk]

ohh you only have 16gb of RAM

then you have to increase the windows pagefile

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive. Set to 16000 Min and 24000 Max
Then apply and reboot the PC.

[rose apex]

Ok

This is the vram usage right before crashing.

I will try increasing the page file

[ornate elk]

looks okay

[rose apex]

My page size is already 16 GB min and max btw, should i still increase it

[ornate elk]

you need to have max set to 24000 (24gb)

16gb min is ok
And dont forget to reboot after changing it.

[remote elm]

Anyone familiar with this node/error? Reinstalling does nothing.

[ornate elk]

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

[remote elm]

Thank you.

[rose apex]

No way it started working just like that 🤯

Thanks again. You know too much 😐

[ornate elk]

no problem 🙂 glad it works now

sdxl/illust models need to have 16gb ram + bigger pagefile or better 32gb normal RAM

[rose apex]

Does ram speed matters as well?

[ornate elk]

not that much, but it should be the optimal speed for your CPU depending if you have an intel with ddr4 then it would be 3200MHz or if you have AMD with DDR4 then it would be 3600Mhz and AMD with DDR5 would be 6000MHz

also dont forget to enable XMP or AMD expo in bios

if not your RAM speed is much slower

[rose apex]

Yes, it's enabled

[ornate elk]

you can check that in the taskmanager when you click on the RAM and then on the bottom right

[ornate elk]

also important is if you get more ram that is should have the same speed as the ram already in the PC.

[rose apex]

Yes

[bitter pier]

Can I run SDXL models on AMD Gpus?

[ornate elk]

yes if the gpu has 8gb or more vram

[bitter pier]

Its vram size is 16GB.

[ornate elk]

then its no problem

[bitter pier]

What should I do as an env setup?

[ornate elk]

All AMD Guides can be found here:
Go for Forge webui with Zluda:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

i updated them today as AMD released HIP SDK 6.4

[bitter pier]

Awesome!

Thanks bro!

[ornate elk]

np, let me know if you have any questions

[vocal burrow]

1/ you're using the incorrect python version, SD is only compatible with Python 3.10 versions (3.10.6 and up). Some do accept 3.11 but stick to 3.10 to the latest 3.10 to be safe.
2/ you've installed a1111 which is pretty outdated by now, get forge instead
3/ you have it installed on desktop which could be causing issues because of oneDrive sync features
4/ that's not a complete log so I can't properly troubleshoot.

C:\Users\X\AppData\Local\Programs\Python\Python313\python.exe

Python313

3.13

you probably have both

if you don't need 3.13 just remove it

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

k you should be good to go then (but you can get a more recent version of python 3.10 instead too)

you probably had python3.13 installed when running a1111 the first time

it created its venv with it and that's why it's complaining about not finding python 3.13

yes

checked pinned messages for some guides if needed

[ornate elk]

whats your GPU?

ah okay, yea then go for forge, guide is the first link pinned

you need to update your GPU Driver

oh okay, have you launched the update.bat before running the run.bat?

can you show the full cmd log?

hm thats very strange

can you open up a cmd and run
sfc /scannow

and then open up a cmd and run
where python
and python --version

[vocal burrow]

are you using a laptop ?

[ornate elk]

i dotn think so

[vocal burrow]

Open webui-user.bat
find the
set COMMANDLINE_ARGS= line, add --gpu_device_id 0 to it
and then run webui-user.bat again

let it do its stuff

most likely it was trying to use your AMD iGPU

--gpu_device_id 0 forces him to use gpu 0 => your RTX 4070

nope

meanwhile you can look at civitai, huggingface, etc for some models to use.

That I don't know for sure. I don't have (or at least didn't until recently) have a flashy new GPU so I was mainly using sd1.5 and SDXL stuff.

lol what

what's in your run.bat ?

I meant in the file itself not the log

open run.bat with an editor

yeah remove --gpu_device_id 0 and it should work with webui-user.bat I'm guessing

About to hit the bed. But there seems to be something wrong with your system's python or the one that comes in the prepacked release .zip archive for some reasons.
So you can either :
1/ continue using webui-user.bat instead as it created its own venv (tldr : its own virtual machine specifically for python stuff)
2/ uninstall python and reinstall it (migjht as well go for 3.10.11) and see if it cleaned itself
3/ forget about the forge prepacked .zip stuff. And use the big boy tools instead (really not that hard), tldr being

1/ install git-scm <https://git-scm.com/>
2/ open a cmd and navigate towards whatever folder you want to install forge in (please NOT desktop or documents) with "cd c:\whatever\forge"
3/ still in cmd, download forge with git by running "git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git"
4/ run webui-user.bat
5/ voilà... Now whenver you want to update forge just run "git pull" in forge's directory

yup that's running, cf the line saying Running on local URL: http://127.0.0.1:7860

Suuuuure, it should but it ticks me the wrong way to be honest XD, it's like driving with different sets of tires.
Just forget about the very existence of run.bat.
update.bat should work.... But run.bat should have also worked in the first place so who knows

and with that I'm off, gn

[hot thorn]

if I try to remake the same picture on a different gpu, will it always be different?

[ornate elk]

if the seed was generated by a different gpu maker nvidia7amd or cpu and you dont have the same then yes

thanks for the info!

for anime go with illustrious based models

for realism sdxl or cyberillustrious works good too

here for example:
https://civitai.com/models/1224788/prefect-illustrious-xl?modelVersionId=2081643
https://civitai.com/models/827184/wai-nsfw-illustrious-sdxl?modelVersionId=1761560
https://civitai.com/models/376130/nova-anime-xl?modelVersionId=2100813
https://civitai.com/models/1125067?modelVersionId=2071927

[ashen phoenix]

whyu wont it let me set clip state i apply it and it justs disappears next reload

or does anyone know a way to adjust it in any way

[pale garden]

Any fluxgym users that train flux loras? The program seems really scuffed on 5000 series rtx gpus. Lots of hangs even though the github installation process is followed perfectly

Honestly maybe just flux training numbers etc for a lora for like a tattoo to be very strong

[violet swift]

Qwen Image Edit + ControlNet Openpose is possible?

Basically, I want specific poses for my character, which I have uploaded to Qwen Image Edit, to replicate the poses I could input with OpenPose ControlNet. Is that possible?

I know they can both be used separately, I've tried to make them work together but no luck.

[lunar kiln]

I’m trying to begin my journey doing text to image generations on my own local computer. Are there any guides that I can refer to? Can give my pc specs if requested.

[elder star]

Hi, I’m trying to create a consistent character reference in Stable Diffusion. I’m currently using the Balanced checkpoint together with a LoRA, but the outputs keep coming out distorted

[ornate elk]

Hey, it seems like a VAE issue

Which model and VAE did you use?

[vast crest]

Hi, Colab hasn't been working for a few days (I use this to create Lora). Do you have any easy-to-use alternatives for creating Lora?

[elder star]

Vae-ft-mse-840000-ema-pruned

[ornate elk]

and which model?

that vae is made for 1.5 based ones

[elder star]

LoRa- CassandraDunnArtILL

sd- plantmilkmodelsuite_walnut

[ornate elk]

yea then oyu need an other vae

[elder star]

Could you please suggest some working VAEs?

[ornate elk]

this: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

[elder star]

Thanks, So I just need to download this and paste it into the VAE folder, right?

I’ll give it a try! And… can I ping you here if I run into any issues? I’m new to this SD

[potent sun]

can anyone help me with this? https://github.com/spacepxl/demystifying-sd-finetuning?tab=readme-ov-file

this seems to be a tool to measure loss through the training process

but I don't know how to use it; it seems like it sorts random noise so the default tensorlogs the trainers give aren't enough

[ornate elk]

yea sure!

[pale garden]

wan is so trash

[elder star]

Thanks! I just tried it and it’s working, but it’s taking 20 minutes to generate a single image.

[ornate elk]

whats your gpu?

[tepid crypt]

About to buy a new pc, it has rtx 3060 with 12gb. Finally i'll be able to generate stuff by myself, so i want your recommendation on what software use to generate images. Pls don't recommend comfyui, barely have any knoledge about it and dont have the time or looking forward to learn about it right now.
I'm not into the news about ai generation, so dont know if a1111 or forge keep being good options with fast generation. If there are better version let me know

[elder star]

3060

[ornate elk]

with 12gb vram?

then you have to edit the webui-user.bat with notepad and at the line
set COMMANDLINE_ARGS=
you have to add --xformers
then save and relaunch the webui-user.bat

[ornate elk]

hey, forge will be the best for you

[elder star]

Okay I'll try that again!! Thanks

[ashen phoenix]

I suggest forge neo. It's a fork of forge and it gets updates unlike forge, and has more optimization. https://github.com/Haoming02/sd-webui-forge-classic/tree/neo. I recently discovered a bug with Forge with distorted generations that were fixed only after switching webuis.

[potent sun]

pls anyone

[chrome tangle]

@ornate elk by complete cmd log, do you mean the full error?

||oading Model: {'checkpoint_info': {'filename': 'C:\ZD-Zluda\stable-diffusion-webui-amdgpu-forge\models\Stable-diffusion\novaOrangeXL_reV20.safetensors', 'hash': '3921a30e'}, 'additional_modules': [], 'unet_storage_dtype': None}
[Unload] Trying to free all memory for cuda:0 with 0 models keep loaded ... Done.
StateDict Keys: {'unet': 1680, 'vae': 248, 'text_encoder': 198, 'text_encoder_2': 520, 'ignore': 0}
CLIPTextModel Unexpected: ['transformer.text_model.embeddings.token_embedding.wrapped.weight']
CLIPTextModelWithProjection Unexpected: ['transformer.text_model.embeddings.token_embedding.wrapped.weight']
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
K-Model Created: {'storage_dtype': torch.float16, 'computation_dtype': torch.float16}
Model loaded in 0.6s (unload existing model: 0.2s, forge model load: 0.4s).
[Unload] Trying to free 3051.58 MB for cuda:0 with 0 models keep loaded ... Done.
[Memory Management] Target: JointTextEncoder, Free GPU: 14610.29 MB, Model Require: 1559.68 MB, Previously Loaded: 0.00 MB, Inference Require: 1024.00 MB, Remaining: 12026.61 MB, All loaded to GPU.
Moving model(s) has taken 0.63 seconds
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Interrupted with signal 2 in <frame at 0x00000232DF2CF250, file 'C:\ZD-Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\nn\modules\module.py', line 1762, code _call_impl>||

[ornate elk]

No I mean the whole cmd log

All of it

[chrome tangle]

What command does she use to get it? Sorry if i am a little out of it. I am still loopy from a dental trip.

[ornate elk]

There is no command. Mark the whole cmd text and press Ctrl+c ^^

[chrome tangle]

relayed it will let her know!

So, she started compiling it 15 minutes ago and its still compliling, that doesn't sound good!

is that what you meant by the full cmd, she asks

[ornate elk]

Compiling can take 15-40 minutes

[chrome tangle]

is that the thing you wanted?

[ornate elk]

Wrong python version installed.
3.11.9 is recommended

[chrome tangle]

Thank you so much, youa re a gem!

[ornate elk]

Uninstall 3.10.11 64,bit
Install 3.11.9 64bit
Then delete the venv folder and relaunch

[chrome tangle]

Thank you so much, will let her know!

[chrome tangle]

she asks "does the compiling only happen once"

[ornate elk]

it will need to compile again only if torch, zluda or the AMD Adrenalin Driver updates.
So now its only for the first time. it wont happen when relaunched.

after the compiling it will work normaly

[chrome tangle]

She genned and now its compiling again. not sure what hgappened

[tepid crypt]

Ive received recommendations about swarm ui, ive read it has comfyui as backend, not fond of it but from the videos ive seen looks really intuitive, is that a good option? whats your opinions?

[ornate elk]

it needs to compile sometimes when a different type of model gets loaded the first time

[chrome tangle]

AHH

[lyric vector]

I tried to generate a video the first time it told me connection error this time it did a full on shite screen then went to this and its telling me recconecting

Idk whats going on

[lyric vector]

Yeah it tried recomnecting again and it just got stuck there

[ornate elk]

Your GPU probably has not enough Vram for video gen

[prisma moss]

right so, windows/git updated itself now i cant launch webui

nvm reinstalling git worked

[lyric vector]

I have 16gb of ram so how much do I need? @ornate elk

[prisma moss]

read again slowly

[copper cypress]

RAM is different from VRAM. Go to your task manager and find the dedicated GPU memory, as shown here:

[lyric vector]

Oh um ill look here shortly

Apprently with what im reading online about my pc O should have 8gb of vram

[copper cypress]

https://www.youtube.com/watch?v=1Xaa-5YHq_U

try following the workflow from this video, i also have 8gb of vram but im able to generate videos right now

they are low res but they work

[lunar kiln]

Help anyone?

[mighty hazel]

Hi guys I'm trying out SD webforge UI but I constantly get connection errored out when i click generate, here are the logs when i start run.bat

till now I've tried freeing up disk space made sure 1.5 x (RAM) space is available

tried disabling real time protection as well

[vocal burrow]

you're not closing the cmd window before clicking generate, are you.

[mighty hazel]

oh didn't know, will try and see

i tried closing the cmd window too but the same error occurs

[vocal burrow]

you're supposed to NOT close the window

[mighty hazel]

ok even when the window was open it gave the same error

[vocal burrow]

make sure you have whitelisted the page in your adblocker too

[mighty hazel]

disabled brave shield but the same error, i suppose this has nothing to do with the settings i'm using ??

[vocal burrow]

nothing to do with the settings

try with another browser (even edge should work)

and copy paste the full logs from cmd like you did before

[mighty hazel]

ok i will do after restarting my laptop

i don't have edge but i tried copying the local url into another browser and it refused to connect

also reloading page in brave also doesn't work

and logs are same... this time i changed paging size (virtual memory) minimum set to 1.5 x 16 gb and 3 x 16 gb ram

logs

[vocal burrow]

Not gonna change anything, it's something to do with your network / browser settings.

[mighty hazel]

firewall ?

[vocal burrow]

if you have tweaked it might be

[mighty hazel]

well i never touched it

not sure if i've to allow anything ?

[vocal burrow]

like I said you can try another browser, I suggested edge because it's built-in windows and a pain in the a** to uninstall so basically everyone have it

you can check for error in your browser's console

maybe something failed during the installation (first time you launched the .bat) and you igored the error

[ornate elk]

4-6gb vram + flux?, whats your gpu?

[mighty hazel]

lol, i removed edge long back as part of debloating, ok i tried using chromium browser but got the same error, just to add: before this i tried running inside pinokio but there also i got the same error

[mighty hazel]

nope 4 gb vram

[ornate elk]

that wont work

with flux

[mighty hazel]

i saw people on reddit and youtube say got it to run on gtx 1650 4gb

https://www.youtube.com/watch?v=ha-a8TGzVSk

https://www.reddit.com/r/FluxAI/comments/1eq5b9b/flux1dev_on_rtx3050_mobile_4gb_vram/

[ornate elk]

how much RAM do you have?

[mighty hazel]

16 gb

[ornate elk]

yea that wont help, then you need to increase the windows pagefile

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive. Set to 16000 Min and 24000 Max
Then apply and reboot the PC.

[mighty hazel]

will try again and see

[mighty hazel]

i've one doubt, i've only 7 gigs left on C, will increasing pagefile work ?

[ornate elk]

nope, then you need to make more free space on it

25gb would be good

with a tool like WizTree you can check whats taking up the most space

[mighty hazel]

i've wiztree but choosing what to remove will take a while, welp someone said i need this free space only from the drive I'm running webforge (on D: in my case)

[ornate elk]

pagefile should be on C as its the fastets drive mostly.

D needs also free space for the large models

[mighty hazel]

welp, i might do this later then

[orchid quail]

has anyone tried to run a WAN2.2 GGUF quantisation with Diffusers?

[untold vigil]

Anyone know a good tutorial how to train loras?

[orchid quail]

afraid not, but it's something I'm interested in

[hot thorn]

does order of prompts matter in negative prompt? Im using reforge it that matters at all

[bitter pier]

Of course.

It's the language model and it depends on how the model process data: so it matters.

[hot thorn]

you sure? i heard negative prompts are turned into embeddings unlike positive prompts

[bitter pier]

Definitely sure!

You are right but it is all LM(Language Model)

[hot thorn]

so back to drawing board for me again...

today i tested 6 prompts i had success with for wide generation (same prompt for different character, backgrounds and poses), and i finally got down to 1. now im tackling negative ones

[bitter pier]

And once their embedding is generated they are fed into the LLM to generate image And then all the embedding's order matters difinitely!

[hot thorn]

so i can finally achieve the perfection i want

are u well versed in negative prompts? i need some help, i've been working on my overall prompts and settings for past few days and im slowly losing my mind

at least its easier now with rtx 3060, few days ago i had a gtx 1070 and each prompt was 6-6.5 min long

now theyre below 2 min

[bitter pier]

Sure.
I'm ready anytime.

But how about moving to DM?

[hot thorn]

so i can afford to change or get rid of 1 or 2 things, generate, compare results and continue to revert

sure

[pale garden]

If u have an image with messed up arms, ripped clothes, miscolored back skin etc and just general shoddiness on top of a GOOD base to work with, what is generally the inpainting technique to just mask it all and have it auto fixed

[solemn moon]

Mask what you wanna keep and select “in paint around mask” or summin like that 🤷‍♂️

[ripe nacelle]

Hi dumb question but trying to unsubscribe. Is there a link? Thanks

[vocal burrow]

unsubscribe from what ?

[orchid quail]

Thanks for signing up for Cat Facts! You now will receive fun daily facts about CATS! >o<

[storm knoll]

I updated everything on comfyui and now every single group scoops up nodes in other groups when i drag them over...

please tell me there is a way to fix this lmfao

[next chasm]

I have a rtx 4080 and 32 gb cpu ram. is it normal to oom on 81 frames of wan 2.2 using a GGUF Q4 K m model?

[fiery forge]

where do i go to check how much vram i have? im getting an warning that says im running low and it slows down image generation time

[ornate elk]

You can check in the taskmanager under performance tab and then at the GPU info

[tame cave]

hey guys i wonder if i can install SDForge with a 7800xt GPU ?

[vocal burrow]

yup

check pinned messages

[tame cave]

pinned message ?

[tame cave]

Ohhhhhh thaaaaaaat

Thanks i'll try that

[ornate elk]

Go with the Zluda version

[tame cave]

yeah i saw that, i'm ihalf way of installing it

[haughty pilot]

I have been using a model that says it has VAE built in, now I want to try NoobAI-XL, it says "Directly use the VAE that comes with SDXL" where do I find it? is it preinstalled in forge or something? I don't have anything in the VAE window

[ornate elk]

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

[storm knoll]

I just updated comfy and all my nodes and now I have an issue where groups will steal the nodes from other groups when I drag the group over it to move it
Basically turned all my groups sticky even though I'm just trying to reposition them
is there a solution for this? don't want to play tetris in my workflows or pin/unpin everything constantly

[haughty pilot]

if the model I have been using says it has built in VAE does that mean it will ignore the one you sent if I choose it?

[ornate elk]

If you select the vae in the VAE dropdown, the inbuilt one gets ignored.

[tame cave]

thanks a lot for the tip @ornate elk @vocal burrow , it works

[crimson ruin]

Qwen image always seems to produce weird borders when i specify any sort of sampler and scheduler, why is this? using swarmui btw

random first test, normal borders alongside real image generation attempt and it looks good but it has those stupid borders

just realized how blurry discord makes the images lol

[lusty apex]

hey, for the past few days ive been trying to run SD in python with the StableDiffusionPipeline. I think my code is fine, but i cant seem to find the SD 1.5 model (i only have 6gb vram) anywhere, ive found prune-emaonly ckeckpoint and safetensors files, im new to all of this, can someone explain how to get the base model. (heres my code incase that matters)

import torch
   
# Pipeline laden
pipe = StableDiffusionPipeline.from_pretrained(
    "models/v1-5-pruned-emaonly.ckpt",
    torch_dtype=torch.float16
).to("cuda")

# Prompt für das Panel
prompt = "cat"

# bild generieren
image = pipe(prompt, guidance_scale=7.5).images[0]

# Speichern
image.save("panel1.png")
print("Panel erstellt!")```

[ornate elk]

1.5-emaonly-pruned is SD 1.5 base model

but why not install Forge and use that instead of a selfmade python script? (in case you just want to generate images quickly)
https://github.com/lllyasviel/stable-diffusion-webui-forge

[lusty apex]

mb, i meant i only have 1.5-emaonly-pruned.chkp and 1.5-emaonly-pruned.safetensors, and with both my programm doesnt seem to work so i assume there is a base file. I plan on expanding this programm

[ornate elk]

ah okay, the base file is the checkpoint (preferable .safetensors)

[lusty apex]

ah okay, well then the problem must be my code, im gonna look into it thanks

[ornate elk]

no problem

[upper linden]

Has any of the new drivers been updated to work with amd zula?

[ornate elk]

Hey, yes Adrenalin 25.8.1 works with Zluda.
And if you have a 7000series you can upgrade HIP SDK 6.4

[lofty wedge]

Hi guys, any chance that you know how to fix this error while downloading ComfyUI Packeges?
"pip install failed with code 2: 'Requirement already satisfied: pip in c:\stabilitymatrix1\data\packages\comfyui\venv\lib\site-packages (25.2)\r\nCollecting wheel\r\n"

line 414, in _get_launcher\r\n raise ValueError(msg)\r\nValueError: Unable to find resource t64.exe in package pip._vendor.distlib\r\n'

[upper linden]

Oh okay, so download the latest hip 6.4 and the newest driver, both should work fine

[ornate elk]

yep they work with Auto1111 and Forge, Comfyui should work too but havent tested it

[sage prairie]

Hey all! I'm sure this has been asked before but I couldn't see it anywhere. I'm having to rebuild my system that I had Auto1111 on and I'm trying to see if I should consider another alternative instead. I have a headless system and I put Auto1111 behind a nginx proxy so I could access it from inside/outside of my network. It was a bit of a pain to setup so was hoping somebody might have a better suggestion. It's being hosted on a Windows box, but open to switching it over to some kind of linux distro if that would give me more options.

[lusty apex]

i figured it out, i needed to convert the .ckpt file into a folder... im amazed

[lusty sun]

I keep having connection errors and it just not launching. Any help?

they made an entire discord for helping?

welp ok

ah

was that a scam

Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: f1.0.0v2-v1.10.1RC-latest-2446-ge1dcf9b4
Commit hash: e1dcf9b490673dad1511deb58c3f7e829d4ec603
X:\Program Files\Stable Matrix\Packages\Stable Diffusion WebUI reForge\extensions-builtin\forge_legacy_preprocessors\install.py:2: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html
  import pkg_resources
X:\Program Files\Stable Matrix\Packages\Stable Diffusion WebUI reForge\extensions-builtin\sd_forge_controlnet\install.py:2: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html
  import pkg_resources
Launching Web UI with arguments: --gradio-allowed-path 'X:\Program Files\Stable Matrix\Images'
Total VRAM 8192 MB, total RAM 32683 MB
pytorch version: 2.3.1+cu121
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 2070 : native
Hint: your device supports --pin-shared-memory for speed improvements if having 16GB or more VRAM. It may cause issues with 12GB VRAM or less.
Hint: your device supports --cuda-malloc for a bit less vram usage and small speed improvement
Hint: your device supports --cuda-stream for potential speed improvements but it may cause issues on some GPUs that have 6-8GB VRAM or less.
CUDA Stream Activated:  False
Using pytorch attention
ControlNet preprocessor location: X:\Program Files\Stable Matrix\Packages\Stable Diffusion WebUI reForge\models\ControlNetPreprocessor
Memory change: 1.90 MB (0.80 GB total)
Loading model sd\indigoFurryMixXL_noobaiEPS3.safetensors [d17eaa4b4d] (1 of 1)
Loading weights [d17eaa4b4d] from X:\Program Files\Stable Matrix\Packages\Stable Diffusion WebUI reForge\models\Stable-diffusion\sd\indigoFurryMixXL_noobaiEPS3.safetensors```

here are my outputs and so on, if anyone knows how to help me

[ornate elk]

Hey, I would recommend not using Stability matrix, and instead use the updated setup guides from the pinned messages to setup Forge or ReForge

[chrome tangle]

has anyone used latent modifier? do you have any good settings for it.

[tired lintel]

When i have an upscaler in safetensor format, where do i put it so that automatic1111 can find it?

[ornate elk]

Models/ESRGAN

[jolly skiff]

hey guys i wonder if i can install SDForge with a 7800xt GPU ?

[ornate elk]

hey, yes you can. Follow the Forge with ZLUDA Guide from the first link of the pinned messages

[remote summit]

@ornate elk hi thanks for taking the time to reply

[ornate elk]

thats not setup correctly

[remote summit]

I’m new to SD and pretty confused. I’ve reinstalled a few times and probably messed up with VAEs/Loras. I can run Juggernaut and Realistic Vision on my 3050 (8GB VRAM), but I keep hitting CUDA out-of-memory errors on hires fix or bigger images. I don’t really know which VAEs or Loras I actually need — I’ve been trusting advice blindly and I think my setup’s flawed. Could you help set up an actual stabke SD and point me to the simplest stable setup for my GPU, with the exact VAE + model versions I should be using

[ornate elk]

would suggest you follow my setup guide from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
You can install Auto1111 or Forge, boath or easy to use with 8gb vram

[ornate elk]

yea now wonder you get oom with these wrong launch args xD

and a not correct setup, wrong python version etc

follow my setup guide and then let me know if you still have issues, then we can adress them easily

[remote summit]

Can't lie I dont really know what im doing is why im here now ill give your guide a go though

Cheers CS1o will do

[ornate elk]

yea no problem at all, but chatgpt often provides wrong info when it comes to stuff like stable diffusion setups

[remote summit]

I learnt this the hard way lmao 🤣

[ornate elk]

also how much RAM do you have?

if you have only 16gb you have an additional step to do after the setup

[remote summit]

Yeah im on 16gb atm upgrading after Christmas

[ornate elk]

ah okay, then you need to increase your windows pagefile to not get your pc to stutter when using sdxl based models

Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set to 16000 Min and 24000 Max
Then apply and reboot the PC.

[remote summit]

Is this in your guide? This step? Because I cant lie, im not an IT guy i can build a computer but dont really know software

[ornate elk]

nope currently not in my guide, but i will add it in the next days

[remote summit]

Ok wouod it be ok to message you when I get whats on your guide setup for help with the last step

[ornate elk]

sure no problem

[remote summit]

Many thanks ill get on it as soon as im home

[haughty pilot]

Is 9070 xt fully supported now or not? Still holding with buying my new pc to see if 9070 xt will at least be comparable to 5070 ti

[ornate elk]

Yes rocm hip SDK 6.4 got released and Auto1111 and forge with Zluda support it

No need for custom gfx files

[haughty pilot]

is there any comparision in speed 9070 xt vs 5070 ti 1024x1024?

[ornate elk]

Idk

But I guess a 5070ti still would be faster and better supported for ai stuff. Until TheRock is stable released from AMD

[livid bolt]

Hi. I'm confused as heck.

I installed python onto my computer. I know fooocus also has python separately.

When I press run.bat, does it matter which python I used to update pytorch?

I'm trying to get fooocus working with 5060ti, but I'm having troubles. I think it might be because of pytorch and Cuda versions but I'm not sure, hence my question above

[ornate elk]

hey, yea fooocus is outdated and doesnt support 50 series by default

you would need to upgrade torch

but i would recommend to use Forge webui instead

much easier to work with and setup

you find a setup guide on the first link of the pinnes messages

[livid bolt]

Okay I'll try that right now, thanks !

[rose apex]

You have an AMD GPU right? Did you try video generation with wan2.2? And do you have any suggestions on what models should I be using? I tried with Q6 GGUF i2v model and lightx2v rank 32 lora but the generation time is very long (12-18 mins)even for 480p resolutions.

Sage attention doesn't seem to work on amd so I can't make use of it either

[ornate elk]

Yea I tried wan2.1 and that was very slow and ltx was faster but quality was bad.
Currently its not worth to create videos. Takes to long and to much resources for low resolution 5 second videos.
We have to wait until TheRock supports all of the attention boosts like flash and sage.

[rose apex]

Are you sure it will be possible in the future? To use the 90 series GPU for video generations?

Or is it going to be a dead end

[ornate elk]

yea i think so

video gen will be more advanced an optimised in the future

[jolly skiff]

HI

[bitter pier]

Halo

[merry wedge]

Can anybody help me set up stable diffusion forge on AMD with ZLUDA? I've almost got it done but Im having an error

Failed to load ZLUDA: Could not find module 'C:\Users\User\Desktop\stable-diffusion-webui-forge-on-amd-main.zluda\cusparse64_11.dll' (or one of its dependencies). Try using the full path with constructor syntax.

I see the file in the right spot, and I have Python 3.10.6. Idk what im missing

[bitter pier]

Have you add those env path?

[merry wedge]

I believe so

[bitter pier]

Just believe?

[merry wedge]

sending a pic now

[bitter pier]

Please

[bitter pier]

NO way

It's just the python path.

[merry wedge]

wait one sec

Ok there

Thats all the relevent ones

[bitter pier]

And this is the actual error you got?

[bitter pier]

???

[bitter pier]

And what is your torch version?

[merry wedge]

how can I check that?

[bitter pier]

pip list

[merry wedge]

put that in cmd?

torch 2.3.1+cu118
torchvision 0.18.1+cu118

It does say "a new release of pip available"

Do I do that?

I was at a point earlier where I had to downgrade python from 3.11 to 3.10.6

[bitter pier]

Not really

It doesn't the matter.

[merry wedge]

ok

[bitter pier]

Can you import torch?

And it works correctly with your gpu?

And what is your gpu spec?

[merry wedge]

I have a 9070 XT

Should be good, the github had it listed as supported

How would I import torch? Im sorry if I'm being dumb

[bitter pier]

OK.

How about using some screen sharing tools?

[merry wedge]

Is there a VC we can use?

[bitter pier]

So that I can check it on my side?

[merry wedge]

Im down

[bitter pier]

Not really. Just chatting is enabled.

[merry wedge]

can I add you?

[bitter pier]

Sure.

[merry wedge]

Anybody else availaible for this?

[ornate elk]

Hey, yes

[merry wedge]

hi

[ornate elk]

You need to follow the setup guides for AMD from the first link of the pinned messages

Forge with Zluda for example

[ornate elk]

Because you tried to installed forge for Nvidia GPUs

[merry wedge]

No, I didnt

Its specifically a github for an AMD workaround

[ornate elk]

Ah I see but thats not the suggested method or guide then

[merry wedge]

but I guess this is a different workaround

[ornate elk]

Thanks for letting me know! We will take care!

Sure, that would be good

[haughty pilot]

I really don't understand if I'm doing something completely wrong or there's something wrong with my setup:
NoobAI v-pred 1.0 model
forge UI (updated)
CFG scale 5
Euler
1024x1024 resolution (or others recommended for the model)
and I can't get a good quality image at all, tried using USNR style lora, rafa-pro lora (with recommended weights) and that's what I get

I really don't understand

[ornate elk]

dont use the base noobai model

all base models are crap xD

[haughty pilot]

why? I have seen people doing amazing things with it

[ornate elk]

they are not refined so they include everything and its harder to get good images with them

PonyV6 for example doesnt output anything usefull when used with the same settings like other pony based models

same goes for noob ai probably

[haughty pilot]

well, what model do you recommend then? I have seen people recommending noobAI so I went with it

I want to create unique artstyle and noobai was recommended for that, dont want generic sh*t

[ornate elk]

This: https://civitai.com/models/520661/2dn?modelVersionId=2025196
Or: https://civitai.com/models/131986/cat-citron-anime-treasure-illustrious-and-noobai?modelVersionId=1756306

[rose scroll]

I haven't updated stable diffusion for amonth, how long will the update take after running git pull?

[ornate elk]

the update will take seconds as it doesnt contains any models

[rose scroll]

it is taking a while

[ornate elk]

ah okay your on AMD

thats not the update, thats zluda compiling

it will take between 15-40 minutes

[rose scroll]

ah crap

[ornate elk]

just let it load, normally i say its about 15-20 mins max

[rose scroll]

so that won't update it?

[ornate elk]

but depends on gpu etc

[ornate elk]

its updated now, its just compiling the new zluda version

then its ready to use again

[rose scroll]

Sounds good.

[ornate elk]

whats your gpu btw?

[rose scroll]

7800xt not sure what brand or vram tho

[ornate elk]

7800xt has 16gb vram

with GPU-Z you can see all the details

[rose scroll]

Will every of my models need to be recompiled?

[ornate elk]

nope its only for zluda

[rose scroll]

when i try to create a sample image (it just stays loading) while the cmd window says "loading compilation" even after the update

*compilation in progress I mean

[ornate elk]

yea it needs to do this again, it can happen again if you switch to a different base model as zluda needs to first compile one of these models again.
But if it "learned" how to process one model of a type, it wont need to compile again for the others

so its only for the first time and not for each model

[jolly skiff]

HELLO

[tepid crypt]

how can i change to an older version of forge? i want to specifically use this one: f2.0.1v1.10.1-previous-659-gc055f2d4

[ornate elk]

but why? xD

you can use git checkout "commithash" to switch to a specific commit

[tepid crypt]

i want to try generating some exact like i did in a site, theyre using that version there so want to try it

[knotty turret]

cannot import name 'sageattn_qk_int8_pv_fp16_cuda' from 'sageattention' (/home/zeus/miniconda3/envs/cloudspace/lib/python3.10/site-packages/sageattention/init.py)

someone know how to solve this

[bitter pier]

https://www.reddit.com/r/StableDiffusion/comments/1mexlnk/update_finally_easy_installation_of_sage/

Have you tried this?

[knotty turret]

nop let me try

[iron lantern]

how do i get this to english text

[fair oxide]

how can i use stable diffusion and how much does it cost? can i make videos as well?

[bitter pier]

It's totally free

[fair oxide]

sounds good and is it only accessible through here or can i get it from the website? sorry for the multiple questions. hope you had a great labor day week

[rose apex]

It's accessible through it's official GitHub page. I suggest you watch a tutorial on YouTube on how to get started cause it's not a simple install and run kind of thing. You might have to install a specific version in some cases.

[fair oxide]

Could you please pm it or where on here is it

[ornate elk]

Checkout the first link of the pinned messages for the setup Guides.

[cosmic furnace]

Inpaint anything on forge. Mouse over image part. Not working. Just not reacting as if i click on usual image. Anyone knows why ? o.0

[vestal spoke]

Hey, if someone could assist me with creating my own lora i would be very happy, i have been trying for 2 days to make my own lora, i have around 150-200 sample images of a experiment person and i have captioned them all, i have downloaded stable diffusion and after many youtube guides and articles i just found out that kohya is awful tool (good, but is not working properly, does not even work for me) and after a lot of trying after downgrading python and around 6 hour of similar download install uninstall remove upgrade downgrade etc im stuck here ready to start punching my monitor.

I would be very happy of anyone who could help me, (btw i have an rtx 3060 with 12gbs of vram)

[ornate elk]

Make sure forge URL is whitelisted in any browser adblocker

[haughty pilot]

I want to test some LORAS using XYZ script (3 loras to be exact), that works well but I wanted to make several examples/lora using batch count but instead of getting rows for each batch (2 batches = 2 rows, each with 3 loras) I get all 6 images in 1 row (2 images for lora1, 2 images for lora2, 2 images for lora3), how can I fix this?

[fair oxide]

Hey all, I was having a great time with the Automatic1111 stable diffusion before and recently switched from windows to linux which is giving me some hard time installing it. I think it might be due to running a newer card (5060 8gb), but not sure. Would anyone know where I can find troubleshooting steps or help me with this?

[muted lodge]

Idk if its the right place but i use stablediffusion webui locally, and for some reason i tried 5-10 loras and none seem to work, i try putting weightings but it does nothing doesnt change the image at all

[vocal burrow]

cf the additional step for RTX50XX series GPU in the top pinned message. Also I would ditch A1111 for Forge at that point, it hasn't been updated in a year.

[vocal burrow]

would need so see some examples of your settings

[fair oxide]

switching to reforge and some other messing with settings worked, thanks!

[unique field]

some loras affect art style, and they can bleed over into each other. The results can be impacted quite a bit. Try turning on or off one at a time and see how each lora impacts the generation process.

I normally start with an image with no lora's attached then add them one at a time.

[muted lodge]

thank you for your answer, im realizing using pony the score_9,score_8_up,score_7_up, positive and score_6, score_5, score_4, negatives prompts are messing with my pictures. Currently trying all my loras one at a time seeing what affects what. 🙂 Still a noob learned all this 2 days ago my head hurts lol

[unique field]

If you are brand new to this, I would actually advise you get one of the illustrious or noobai models as those are just better overall for the most part.

instead of running a pony ai model

Also, which tool are you using for ai images? A1111? ComfyUI?

SwarmUI?

[muted lodge]

Webui Forge

[unique field]

Forge is fine for beginners, if you want to get into the most flexible and customizable AI generation, ComfyUI is the way. That said, there is a rather steep learning curve to using ComfyUI compared to others

Also, which model are you using?

Because I'm thinking maybe that has something to do with what the output you are getting.

[muted lodge]

I see ill check ConfyUi later when i got more practice with forge, i use Cyberrealistic Pony for my model

[unique field]

Yeah, I'd skip the pony model family

go to illustrious or Noobai

if realism is what you're after, flux

[muted lodge]

Illustrious is good for realism ? Because i use Illustrious when i do Anime and it works well and when i do realism i use Cyber

[muted lodge]

Oh okay ill check Flux thanks 😄

[unique field]

that said, I think cyberrealistic just made a version for Illustrious that works pretty well.

(have not used it yet though.)

Stable yogi might be a type to check out as well.

for realism

mainly, just try different things and see what works best for you.

[muted lodge]

Yeah theres a lot to try out and so much little time D: Thank you very much for you help i appreciate it

[muted lodge]

When you start with no Lora, after that do you use the seed and then apply lora to keep the image with new details from the lora ?

[unique field]

sorry for the late reply,

yes, I usually use the same seed and prompt

and I just compare the exact impact the lora is having

[muted lodge]

Been doing exactly that for 2 hours im so happy that my loras works now 😄 I can sleep in peace

[unique field]

glad to hear it.

Though, if you read the documentation on each lora, there will often be a recommended weight range

you may wish to put those down in the notes page of your lora displays

or the trigger keywords

so you dont have to keep searching up the lora page to find that information over and over.

[muted lodge]

yeah i tried weight range , with and without my prompt and with other loras , alone ... theres so much customization lol. I use the trigger words when i click the lora too

[unique field]

Yeah, and comfyui takes that to the next level in terms of customization.

But only do that when you are ready for the extra deep dive.

it gives you more control of HOW the image is generated, saved, and upscaled,

For example, I can do the testing of multiple loras on one single button click and it will que up the rest of the gens all in one go.

[iron lantern]

how hard is control net to use got a new gpu wanna put it to the test

[unique field]

controlnet is pretty easy overall.

The hard part is honestly just finding the type of control net you want to use and then enabling them for your use case.

[iron lantern]

just need it for a pose base since iam pretty good with forge ui now

[stone fog]

I'm having trouble understanding the amd guide https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-automatic1111-with-zluda I have a 7900xtx do I need to be doing the SDK step since the exe was giving errors

[ornate elk]

You need to install HIP SDK 6.4 yes

[ornate elk]

Thats not correct

Check the example images of that step

it looked good now

[stone fog]

I know I was getting errors but I fixed it

Basically just trying to fix python now

[ornate elk]

whats the issue with it? wrong version?

[ornate elk]

delete the venv folder and relaunch the webui-user.bat

[stone fog]

venv?

[ornate elk]

should be in the webuis folder. but if you get this at first launch then you need to rerun the python installer for 3.11.9 and click "modify", "next" and there select "add python to environment variables"

[stone fog]

ah that worked

hard parts over

ooop forgot to add the CMD line but it still seems to be working

[ornate elk]

which cmd line?

if you mean the launch args, they are needed as without them you will have a bad performance or other issues

[rose apex]

Is there a channel I can subscribe to so that I get news of any new updates for Stable diffusion on AMD cards? I don't mean the webui, more like new attention techniques or other performance tweaks. Most of what I see are for nvidia GPUs and it's really difficult to pin point which ones work for AMD or not

[ornate elk]

Nope but you should follow AMDs TheRock project which brings native torch support to windows.
https://github.com/ROCm/TheRock

[haughty pilot]

What is the best way to merge a few styles LORAS with different weights into 1 LORA?

[fierce moat]

hi is there a custom node i can configure my video metadata

[ornate elk]

You have to set the face detector to CPU mode in the settings of live portrait

[stone fog]

Uh, what?

[ornate elk]

the extension you installed, live-portrait thing, doesnt work on AMD gpus, so you have to set it to CPU mode in the settings of the webui

its explained on their github page

Meanwhile the "Compilation is in progress" refers to zluda compiling which is normal and can take a moment

[stone fog]

Oh I thought that error was related to the fact that zluda isn't able to make ANY images

[ornate elk]

nope the error is specific to the extension

Does your webui work? if not feel free to provide the whole cmd log

[stone fog]

The webui launches yes, but no imagery is generated

[open inlet]

Is it possible to enable 2 "checkpoint" models simultaneously?

[ornate elk]

What does the cmd shows? Complete log pls

[ember raven]

hi guys, i have a question, can i train my own LoRA if i have an AMD gpu? if so, how?

i assume that most guides available are for NVIDIAs...

[bitter pier]

It's doable

You can use RoCm platform to use your AMD gpu for training Lora models

[cosmic furnace]

TUrned off adblock on forge tab & still cant sketch on top of masked segment anything image 🙁 . TO show what i need to create mask with or what to unite

Is this extension still working? 🤨 SOme ppl say its outdated but ...not sure

[tepid crypt]

opinions on TensorRT? ive heard is really wonderful tho also read that could overheat the GPU and run into issues, ive read that on a commentary almost 1 year ago, how recommende is that nowadays?

[vocal burrow]

Overheat the gpu ? what ?
SD is no different from games or whatever when it comes to temperature management. Gpu gets hot ? fan go brrrrr, gpu get cooled. If someone cooked their gpu using SD it's either because they messed with their fan curves, didn't clean their gpus in years, was generating stuff for hours long session. Probably a mix of all of the above.

Regarding TensorRT, it worked great last time I tried definitely saw a huge improvement in speed. However it required you to """"translate"'"" your model to a TensorRT model, you couldn't do it on the fly, it would take more space on your disk and more time (5 minutes-ish depending of your config). Plus I had some compatibility issues with other extensions when using TensorRT models.

But otherwise yeah I was generating at 3, 4 times my usual speed.

[pale salmon]

sup why my pc make a crazy wind sound when i prompt with those settings ? using txt2img

got a 4080 super

[onyx solar]

Because heat is hot. What do you want? You're running your graphics card. It's generating heat. It needs to cool down. The fans run. The fans make noise.

[pale salmon]

sorry im just scared that it burns or something like that 💀

[vocal burrow]

It could be formulated better but yes that's it. AI pushes your hardware to the max. So yes, fans will ramp up.

[pale salmon]

ok thx, now i have this error..

embedding_db.load_textual_inversion_embeddings(force_reload=True, sync_with_sd_model=False) TypeError: EmbeddingDatabase.load_textual_inversion_embeddings() got an unexpected keyword argument 'sync_with_sd_model'

[stone fog]

venv "C:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.11.9 (tags/v3.11.9:de54cf5, Apr  2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
Version: v1.10.1-amd-43-g1ad6edf1
Commit hash: 1ad6edf170c2c4307e0d2400f760a149e621dc38
ROCm: agents=['gfx1100', 'gfx1036']
ROCm: version=6.4, using agent gfx1100
ZLUDA support: experimental
ZLUDA load: path='C:\SD-Zluda\stable-diffusion-webui-amdgpu\.zluda' nightly=False
Skipping onnxruntime installation.
Installing sd-webui-live-portrait requirement: mediapipe
Installing sd-webui-live-portrait requirement: changing onnx version from 1.16.2 to 1.16.1
You are up to date with the most recent release.
W0908 00:39:55.057000 34048 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Lib\site-packages\timm\models\layers\__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
C:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: `pytorch_lightning.utilities.distributed.rank_zero_only` has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from `pytorch_lightning.utilities` instead.
  rank_zero_deprecation(
Launching Web UI with arguments: --use-zluda --update-check --skip-ort --skip-python-version-check
▶ SD-Hub: v11
Loading weights [a5e5a941a3] from C:\SD-Zluda\stable-diffusion-webui-amdgpu\models\Stable-diffusion\meinamix_v12Final.safetensors
Creating model from config: C:\SD-Zluda\stable-diffusion-webui-amdgpu\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 57.2s (prepare environment: 76.0s, initialize shared: 1.2s, other imports: 0.5s, load scripts: 1.5s, create ui: 0.7s, gradio launch: 0.5s).
Applying attention optimization: Doggettx... done.
Compilation is in progress. Please wait...
Model loaded in 48.2s (load weights from disk: 0.2s, create model: 0.9s, apply weights to model: 3.3s, load textual inversion embeddings: 0.2s, calculate empty prompt: 43.5s).
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
  0%|                                                                                           | 0/20 [00:00<?, ?it/s]Compilation is in progress. Please wait...

[ornate elk]

Looks okay, but you need to wait until compiling is done

[fair oxide]

hey could you add me I wanna ask for some advice, thanks in advance 🙏

[ornate elk]

Sure

[stone fog]

That's the problem, it never stops compiling

[ornate elk]

Can you show me the content of the extensions folder?

And have you let it compiled at least 40 minutes in a whole session?

[stone fog]

yes, but it also never used to take that long.

I tried this before extensions as wel

[ornate elk]

Can you delete all of these extensions and the venv folder after that.
Then relaunch and let it compile at least 30 minutes?

It should work
You can also prompt for an image and press generate.
It will generate the image after the compiling is done.
For 9070xt users someone reported that the compiling took an hour.
Normaly it's more like 15-40 mins

[stone fog]

sure

[ornate elk]

Just be patient. You could also leave it running for 1 hour after pressing the generate button and you will 100% get an image.
In your log from above I can already see that it prepared the 20 steps for the image process

[stone fog]

WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.11.9 (tags/v3.11.9:de54cf5, Apr  2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
Version: v1.10.1-amd-43-g1ad6edf1
Commit hash: 1ad6edf170c2c4307e0d2400f760a149e621dc38
ROCm: agents=['gfx1100', 'gfx1036']
ROCm: version=6.4, using agent gfx1100
ZLUDA support: experimental
ZLUDA load: path='C:\SD-Zluda\stable-diffusion-webui-amdgpu\.zluda' nightly=False
Installing requirements
Skipping onnxruntime installation.
You are up to date with the most recent release.
W0908 13:57:06.241000 18532 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Lib\site-packages\timm\models\layers\__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
C:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: `pytorch_lightning.utilities.distributed.rank_zero_only` has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from `pytorch_lightning.utilities` instead.
  rank_zero_deprecation(
Launching Web UI with arguments: --use-zluda --update-check --skip-ort --skip-python-version-check
Loading weights [a5e5a941a3] from C:\SD-Zluda\stable-diffusion-webui-amdgpu\models\Stable-diffusion\meinamix_v12Final.safetensors
Creating model from config: C:\SD-Zluda\stable-diffusion-webui-amdgpu\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 39.4s (prepare environment: 57.7s, initialize shared: 0.9s, other imports: 0.4s, load scripts: 0.4s, create ui: 0.4s, gradio launch: 0.2s).
Applying attention optimization: Doggettx... done.
Model loaded in 5.1s (load weights from disk: 0.2s, create model: 0.8s, apply weights to model: 3.0s, load textual inversion embeddings: 0.2s, calculate empty prompt: 0.7s).
Compilation is in progress. Please wait...
  0%|                                                                                           | 0/20 [00:00<?, ?it/s]Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...```

It will go forever and never generate, when I used to use the program it was going in minutes

[ornate elk]

It will just let it compile xD
Zluda compiling is needed and done if:

• Zluda had an update
• torch had an update
• AMD Adrenalin Driver had an update

After that it will be fast again

[stone fog]

No what I'm telling you is I left this on overnight.

[ornate elk]

Make sure you dont click into the cmd. That would pause everything.
Pressing Space will unpause it.

What was your GPU again?

[stone fog]

AMD 7900XTX

I'll play around a bit, and try to refamiliar with the ui while I wait, is there a way to save settings?

[ornate elk]

Like the pre sets of steps ?

[stone fog]

yeah

[ornate elk]

That can be done by editing the config_ui.json

But dont change or click stuff while its compiling.

You can also load an image back into the webui prompt field and then press the blue/white arrow to get all the settings of that image applied

I'm off now but let me know how it went.

[mortal lava]

I have a 4060TI, and settings like that shouldn't cause any special heat or noise for me... I use much higher resolution and step count settings and everything works smoothly..
Check that you don't have dust in the case, etc. as a first step...

Can I get help here to activate ip-adapter_face_id on A1111?
I can't and I don't know what the problem is and what to do...
I'm also not very good at English and I'm using Google Translate to ask you...
The error output is sent as an attachment

[fluid willow]

Is it possible to convert EPS model to V-Pred or merge V-Pred with EPS models?

[rain aurora]

Hi, I have attempted to request a Community License and the message after says I should receive something in my inbox but there is still nothing. Could anyone help me with this?

[iron lantern]

What promps can I use for control net ?

[tepid crypt]

Any news in online lora training for XL models? i havent trained one in almost a year, last i did with a google collab, ive seen is still up and takes exactly an average of 30 mins for 500 or so, please dont mention it. I know theres some pages but farm their coin or energy is really tiring and takes quite a time. Is there any new way or method to train quality and speed? have any recommendations to train art style loras?

[bitter pier]

What you need is just the free plan?

[fair oxide]

Hello everyone this is my very first post here and I've encountered a problem relating to the fact that automatic1111 won't let me install extensions due to some command line flag that's in the way, but the problem is that I doesn't in what file the problem lies since I've opened nearly every relevant file like webui-user.sh, webui.sh, webui.bat etc in the text editor but to no avail. Here's an image of the error message I get when trying to check for updates on Automatic1111:

The only thing that worked for me is when I tried starting the Automatic1111 webui manually in the terminal with the command line flag: --enable-insecure-extension-access
which you can see in the image below:

I'd like to resolve this issue once and for all.

😩

[vocal burrow]

what extension ? what is in your webui-user.sh ?

(also you might want to move to forge as a1111 is no longer active)

[lime wind]

Forge is very Broken and I believe the updates have slowed down. Me and other ppl have moved on from Forge. Same with ReForge. Many features broken with both of them.

[lime wind]

What does your webui-user.bat look like?
Mine is like this:
set COMMANDLINE_ARGS=--force-enable-xformers --xformers --enable-insecure-extension-access
If you are using a Different Computer to connect to a computer to Generate then add --listen --gradio-auth MakeAUserName:MakeAPassword

[vocal burrow]

Maybe if you push it with crazy extensions. For me it works well enough. The next recommendation would be comfyui then.

[lime wind]

Yeah xD True

[vocal burrow]

nope they re using linux, webui-user.bat does not matter.

[lime wind]

Thats where im at now. ComfyUI+Broken A1111 Installs

[fair oxide]

I'm a debian 12 user and this is my webui-user.sh file:

[lime wind]

I get that issue he's having whenever I try to use a different PC to connect to one with generations. Had to add those args to fix it ^, Could be other reason tho

[vocal burrow]

ok and the extension you re trying to install is ?

[lime wind]

Try adding --listen after --enable-insecure-extension-access
Also consider adding Username+Pass with --gradio-auth Username:Password

[fair oxide]

I'm not using Automatic1111 remotely, I'm only using it on the computer I installed it on

[lime wind]

Ah i see

Yeah, again other things may cause that, but thats what Fixed it for me

[fair oxide]

Any extension at all, Automatic1111 doesn't let me install or update any extensions.

[vocal burrow]

that s weird, can you show some logs of you trying to install some extension ?

[fair oxide]

It didn't work with --listen after the first flag.

[fair oxide]

How do i do that, Im a rookie you see

[vocal burrow]

Like you guessed --listen and --gradio-auth is to authorize foreign access to your session. This has nothing to do with your problem atm.

[vocal burrow]

you simply run ./webui.sh, try to install some extension and copy paste in here what s printed in the terminal.

[ornate elk]

For me boath work the same. No issues. But have you tried Forge Neo?

[lime wind]

Ill try it out soon then

I believe Adetailer send to ControlNet is broken in Forge or Reforge, can't remember which one.

[fair oxide]

I tried what you said and it worked when I started it manually in the terminal but the same isn't true for when I'm just starting it by clicking the bookmarked link to the webui since I've set Automatic1111 up as a system service so that it starts Automatically when I fire up my computer.

[vocal burrow]

I would not recommend setting it up as a service as by default a1111 load the last model used on your gpu. So you ll always have a good chunck of your vram reserved for a1111 despite not necessarily using it.

if you do decide to go this route, at least use --ui-debug-mode command line argument to prevent it from loading a model on startup

Also not sure how as it depends of your distro but I would need that service s log then to help you more.

you can probably get it using journalctl -u service-name.service but again that depends of your distro

and how you ve setup the service

[vocal burrow]

don t see anything that would prevent you from installing extension there. But like you said it s not the same instance as your service.

[fair oxide]

The reason I set it up as a system service was because I didn't wan't to hassle with the terminal EVERY SINGLE time I wanted to create AI Image for my business, anyways here is the terminal output from the command: journalctl -u sdwebui.service

[vocal burrow]

The log do suggest you to install tcmalloc btw for your cpu mem. And installing insightface (inside your venv ! not system wide)

[vocal burrow]

depends of your DE but most allow you to setup a "start in terminal on double click" for .sh files when you right click the file and go to properties

[fair oxide]

I used this guide (NetworkChuck) to install everything and he got the TCMalloc "error" also

https://www.youtube.com/watch?v=Wjrdr0NU4Sk

[vocal burrow]

what s in your /home/joackim/Dokument/start-sdwedui.sh ?

wait why is it starting with --listen

that would prevent you from installing extension yes.

Allowing other to access your sd instance AND to install whatever extension they want would be a huge security risk. So yes by default if you use any sort of --listen, --gradio-auth, it is gonna turn off extension installation privileges.

[fair oxide]

I set it as --listen because that is what the tutorial / tinkering back and forth led me to do

but I can try deleting It again and see if anything change

[vocal burrow]

Just because someone else is careless does nt mean you have to :p.
It is not an error indeed but it s free performances on the table. Same as for the xformers warning.

[vocal burrow]

welp there you go. then

remove it (the --listen part)

[fair oxide]

OOOOOOOhhhh sheeit😩

[vocal burrow]

and once you ve fixed that. take a look at tcmalloc, take a look at what is xformers and how to activate it.

Also I see a lot of "file / folder not found" in the log of your service. So that s probably something else you want to fix

[fair oxide]

Removing --listen didn't chang anything unfortunately

[vocal burrow]

you restarted the service ?

[fair oxide]

How do I do that?

[vocal burrow]

systemctl restart "serviceNameGoesHere"

but again that depends of your distro

you can also just restart your computer

you might want to read a bit about how to manage linux install.

This is not basic stuff.

I have to go back to work for now

[fair oxide]

I'm maybe not a complete noob to linux but I didn't install TCmalloc or xformers since I didn't want to screw things up and crash things, I know how to install linux from usb allthough I regret using Unetbootin to do so.

[fair oxide]

Restarting the system service crashed the whole webui, now I can't acess it anymore

[vocal burrow]

same thing, would need logs to help you.

[fair oxide]

here are they:

[vocal burrow]

Launching Web UI with arguments: --listen it s still there

and you still have a bunch of "file not found" error before even calling the webui.sh

[fair oxide]

Yes and it crashed when I removed --listen and then restarted the system service sdwebui.servise

[vocal burrow]

So once again, what s in your /home/joackim/Dokument/start-sdwedui.sh ?

[fair oxide]

This is what is in start-sdwebui.sh:

[vocal burrow]

put it in between quotes like so.
```
stuff goes there
```

[fair oxide]

Like this? the text became green which means bashrc is reacting to it atleast.

This is what the webui looks like now.

Put it inbetween qoutes like this?

[vocal burrow]

no i meant you copy paste the content here using quotes to avoid the weird discord reformat

[vocal burrow]

remove --share --listen

(and the quotes you ve added)

[fair oxide]

for clarification this is exactly what start-sdwebui.sh and sdwebui.service looks like:

[vocal burrow]

looks ok now

[fair oxide]

well the webui is still down when I restarted the system service for the first time and I have now done so several times.

[vocal burrow]

most likely your last line is incorrect then

[vocal burrow]

should probably be /home/joackim/Dokument/AIverktyg/stable-diffusion-webui/webui.sh based on the time you ran it without service.

[fair oxide]

well I'm restarting the system service sdwebui.service now, lets HOPE it maybe works now

didn't work but the path for the webui.sh is correct

[vocal burrow]

log ?

[fair oxide]

Logs!

[vocal burrow]

looks like it s still starting

give it some time

[fair oxide]

the webui is BACK! but the extension downloading and updating problem remains.

[vocal burrow]

guess what.

logs

[fair oxide]

I removed again --listen and --share from start-sdwebui.sh but I don't dare restart the system service since I'm in fear that everything might crash again.

[vocal burrow]

and what s in your webui-user.sh this time ?

[vocal burrow]

cause the logs seem incomplete :/ I don t see anything new since the last one

[fair oxide]

here's a copy of webui-user.sh

[vocal burrow]

can t tell what s going there :/ don t know why the log's service is incomplete

[fair oxide]

I don't know but maybe I should restart the system service, but my feet is trmbeling

[vocal burrow]

there s no reasons for it to explode

you had the wrong adress for webui.sh before that s why it failed

[fair oxide]

here's a more complete log file, the thing I copied was just a small part of the readout, there were big gaps between the blocks of text in the logs that were read out.

[vocal burrow]

jul 06 12:52:19 Joackimsdator start-sdwedui.sh[436506]: Launching Web UI with arguments: --listen

I still see --listen in here XD

[fair oxide]

Well not in here:

Maybe --listen will disapear if I restart the system service

Here goes I'm going to restart the system service!

[vocal burrow]

sudo systemctl daemon-reload to reload the service files

or restart computer, it should work

[fair oxide]

I'm restarting the computer just for good measure

[mortal lava]

?

[fair oxide]

Guess who can install extensions now!!!!!!

well now it's time to install tcmalloc

[fair oxide]

Do you have good tutorial / guide on the internet that you know of so that I don't follow something else?

[fair oxide]

@vocal burrow Hello Aryetis I've installed TCMalloc now which automatic1111 recognizes but I've also installed Pytorch and x-formers which automatic1111 doesn't recognize:

I used these websites as my guide:

https://huggingface.co/docs/diffusers/v0.17.1/optimization/xformers

https://pytorch.org/get-started/locally/

[vocal burrow]

juste add --xformers in the command line arguments

don t try to install it manually

[fair oxide]

well I installed it via the terminal using:

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126

pip install xformers

[vocal burrow]

kinda useless tbh, you installed it system wide if you did it like that

but your a1111 install is using a venv... a venv that was created before you installed xformers on your system so it won t even be able to access it.

[fair oxide]

I know what I venv is and to me they're a living hassle

[vocal burrow]

welp doesn t change the fact that you re using one for your a1111 install

[fair oxide]

well then how do I make it so that it can acess the x-formers or do I have to uninstall what i installed via the terminal?

[vocal burrow]

best plan is to just uninstall what you have installed. Instead just use the --xformers argument and let a1111 do its stuff and get everything for you in its venv

keep all the ai stuff seperated from your system setup as it often does not work with the latests updated versions of python and its pip dependencies

[fair oxide]

ok, I've already added --xformers in the webui-user.sh file.

[fair oxide]

ok

[vocal burrow]

ok just restart your service and you should be good to go

[fair oxide]

Gulp 😩

[fair oxide]

Should I really restar the service or just the deamon?

[vocal burrow]

restart the service with systemctl restart blablablaService

[fair oxide]

done

well this time it didn't say anything about xformers when i launched webui.sh, Thanks!

[prisma moss]

@ornate elk Hey, could you perhaps explain why does this randomly happen? I was inpainting, and then after some time, i just get overfilled in vram and shit stops (the spike at the beggining is the inpainting doing its job)

and then outta nowwhere, vram gets fked

and then after 2 3 min, it recovered it self and proceeded with the next inpaint (nothing in the cmd log to suggest something happened)

[ornate elk]

Hey, make sure when you use Inpainting always set the Inpaint area to "only masked".
If not the webui tries to recreate the whole image and not just the Inpaint.
The larger the image the more vram it needs and it will crash.

[prisma moss]

it is on only masked

[ornate elk]

Oh which model do you use?

And other settings

[prisma moss]

hyphoria illu

image being hrfixed so its x2 the sdxl size

[ornate elk]

Okay seems good

[prisma moss]

idk if this should be checked or not

or this

[ornate elk]

Hmm are these default or did you checked them?

[prisma moss]

i dont remember

[ornate elk]

And which Webui?

[prisma moss]

forge

[ornate elk]

Okay, can't compare to my settings rn

[prisma moss]

u also on forge?

[ornate elk]

I have multiple setups, Auto1111, forge, reforge, comfyui

[prisma moss]

wildin

[ornate elk]

How much ram do you have?

[prisma moss]

32

[ornate elk]

If you have 32 or more then you can disable the "keep models in vram" and then apply and restart forge

[prisma moss]

uuh

r they supposed to be in ram?

( i know i checked keep them in vram, but i cant remember if i previously unchecked them)

[ornate elk]

Yea the normal process is that only the model that is currently used gets first loaded into RAM and then into VRAM when it generates an image

[prisma moss]

is that slower or?

[ornate elk]

No

The generation process is only done in vram

[prisma moss]

ooh

that may be the issue then

ill ask someone i know uses forge

[ornate elk]

Yea I guess thats the issue. Because you can store more models into RAM to quickly switch between them (if enough ram is available) but the vram shouldn't be used as a storage for them.

[prisma moss]

i see

[bitter pier]

What is the issue man?

[prisma moss]

ive unchecked the check, and i guess ill find out in the future if its been resolved

[bitter pier]

So now working?

[prisma moss]

idk

[prisma moss]

so it happened again lol

during detailing process (after hrfix)

and then after a minute or two it comes back to its senses

[prisma moss]

Also, is it normal for a model to still load into vram even with that option disabled? Like it always takes up 6+gb.

[gleaming lodge]

Any help yet ?

[prisma moss]

idk u tell me

so this time it came dangerously fkin close during detailing (a hrfixed image)

so i assume that somehow overflow and simply cucks the card

[gleaming lodge]

Can we talk more about it private?

[prisma moss]

sure

[prisma moss]

@ornate elk so, i think the issue is that i stacked too many loras (weights?) and as i removed 1 by 1, and on the same point where it overflows, the usage on that spike was lower

could that have something to do with it?

[ornate elk]

Yes one model (the one selected) is always loaded into VRAM until you select an other one. And pytorch reserves some vram too.

[ornate elk]

Yea loras also need vram at the image generation.
Also using more tokens can use more vram too

[lime wind]

Help Pls? My ControlNet latent upscale destroys the images. Can anyone help me fix?

[acoustic chasm]

How do I fix this?

I am guessing it should be really easy right

GRAHHH

[vocal burrow]

what s your gpu ?

[acoustic chasm]

Uhhh

How do I see that

[bitter pier]

You can use command to see

Type "nvidia-smi" on your command

[acoustic chasm]

like this?