🤝｜tech-support

[slow garden]

That seems to be fine right?

[small quest]

that's the preinstalled one in the folder, which version is on your pc?

[slow garden]

whats the command for that

[small quest]

dunno, I just type in search bar and it pops up

[slow garden]

yep 3.10

[nocturne topaz]

that is the system one the one in the venv would have (venv) C:\Users\bulle> at the front

[small quest]

odd, for me it showed 3.10.6 in the cmd window, but I only had 3.12 on my system

[slow garden]

yea,

I dont see it in here

I deleted the one in the stable folder

[nocturne topaz]

in that case go into the venv folder in the stable diffusion, then in there go into the scripts folder and run activate.bat

[slow garden]

okie\

[nocturne topaz]

that will open a command window with (venv) at the front

send a screenshot at that point

[slow garden]

it just opens and closes extemely fast

when I start that file

[nocturne topaz]

shit.. ok

then go into the address bar at the top and type cmd

and press enter

that should open a command window at the right folder

[slow garden]

in venv?

or venv scripts

[nocturne topaz]

in the scripts folder yep

[slow garden]

okie

done

ive got the cmd open

[nocturne topaz]

type in activate and press enter

[nocturne topaz]

cool - type cd .. and press enter

and then again

[slow garden]

okie

done

[nocturne topaz]

screenshot

[nocturne topaz]

and then do the same command again

[slow garden]

okie

[nocturne topaz]

then type pip install -r requirements.txt

[nocturne topaz]

that should start importing a whole swathe of stuff

oh wow.. it's all there.. in that case pip install keras

[slow garden]

okie

[nocturne topaz]

run.bat

[slow garden]

As a command or?

[nocturne topaz]

yep

[slow garden]

wait, were you talking about the activate.bat?

[nocturne topaz]

nope

exit out of the command window then and find the run.bat file you used last time and run it now

[nocturne topaz]

honestly - at that point I'd be going back to the github where you got this thing from and creating an issue there (or looking through the issues to see if someone else has the same issue)

[slow garden]

alright buddy, its the one that was linked in the chat

[nocturne topaz]

found it.. go to the stable diffusion webui directml folder in windows, then type cmd in the address bar to open a cmd window again

[slow garden]

okie

[nocturne topaz]

then type git checkout d500e58a65d99bfaa9c7bb0da6c3eb5704fadf25 and press enter

[slow garden]

I dont see a folder caled webui directfml folder

[nocturne topaz]

you in the right place now lol

i just didn't put the hypends in

[slow garden]

ima fucking moron sometimes

im sorry bro lol

cmd

[nocturne topaz]

lol

[slow garden]

just a sec

okie

I have the cmd open in the folder

[nocturne topaz]

then type git checkout d500e58a65d99bfaa9c7bb0da6c3eb5704fadf25 and press enter

[nocturne topaz]

cool, now close that and run the run.bat file wherever that got to

[slow garden]

ITS DOWNLOADING SHIT

HOLY FUCK

bro, ive been balls deep in nixos for the last 3 days, and my brain hurts, the cli hurts man lol

[nocturne topaz]

hahaha cli is where millenials grew up lol

[slow garden]

Oh yea, Im not anti cli, but when its all ive done for the last 3 days im getting sick of it 😆

Im guessing this means it should work?

[nocturne topaz]

theoretically

[slow garden]

okie, Ill ping ya if anything goes wacky

I appreciate all the help buddy

[nocturne topaz]

I mean you'll want to go to civitai and get some better models as base 1.5 is yick

[slow garden]

oh yea, ive got a couple models ive been looking at lol

[nocturne topaz]

realism or anime?

[slow garden]

ehhhh lol

You prob dont wanna know buddy

[nocturne topaz]

ps - https://github.com/lshqqytiger/stable-diffusion-webui-directml/issues/386 was where I found the solution

[nocturne topaz]

eh fair enough - I always recommend the ICBINP line of models, but then I am biased 😛

[slow garden]

Hell yea man, And im not sure whats going on but the download hung up, think im safe to close and restart it?

[nocturne topaz]

should be

[slow garden]

ill have to check that model out, it looks awesome

nvm it started back up

[nocturne topaz]

could also hold off, grab the model from civitai and throw it in the models/StableDiffusion folder and then fire it back up

[slow garden]

IT WORKED

[nocturne topaz]

and the fun has just begun 😛

[slow garden]

your a real one dude

[nocturne topaz]

welcome to the rabbit hole!

[slow garden]

is there anywhere I could like give you props or something?

[nocturne topaz]

buymeacoffee.com/residentchiefnz

[slow garden]

I got you big dog

Give me a couple days to get payed and ill buy you a coffee

[neon ruin]

1200 https://pcpartpicker.com/list/LYZqQP

The build in question

[lavish sparrow]

what model r u setting up

[foggy arch]

hi good afternoon, i recently redownloaded SD again but i seem to forgot something/dont know how to fix it because it seems to not be generating any image once i pressed generate for some of my test runs

[nocturne topaz]

Have you restarted the pc?

[foggy arch]

not yet

[autumn thistle]

Anyone have a link to a good sdxl notbook for colab?

[foggy arch]

i restarted and nothing is coming out

idk how to fix

these are my GPUs for my laptop, i think its probably because my laptop cant handle it i think, idk?

what was the extension for Stable diffusion to actually run for lower GPU pcs?

[nocturne topaz]

—lowvram

[foggy arch]

Thanks

Its an extension right? Or a code for cmd?

[nocturne topaz]

That goes in the webui-user.bat file on the line commandline_args=

[foggy arch]

like this?

hmm.. i still have an error, odd

[ornate elk]

Its the Paperclip button on the right

[ornate elk]

Close all browser tabs and Programms. Then restart SD and try again

[foggy arch]

tried, no effect

[ornate elk]

Restart the PC. Also always launch the webui-user.bat not webui.bat

[foggy arch]

alr

[foggy arch]

its still not working

[opal knoll]

how does everyone make their UI dark theme? Mine loads light theme

[ornate elk]

It will be dark if Windows is set to dark theme.
If you want just SD to be dark
Use --theme dark
In the webui-user.bat

[opal knoll]

Thank you!

[nocturne topaz]

Lowvram not medvram for 4gb

Also use a sd1.5 model not sdxl

You wont get sdxl working on 4gb

[ornate elk]

Already solved

Sdxl was the Problem

[nocturne topaz]

Must have gone the dm route. Good to hear its working 🙂 Love your work @ornate elk

[ornate elk]

Thanks! 🙂

[nocturne topaz]

Do they pay the community guides or is it volunteer work??

[ornate elk]

Nope, but we get Nitro

[outer olive]

Hi
I have a problem
**
I tried InstantID on Automatic 1111 but the results are not even close to the subject and the face model i insert ! is there anything i'm doing wrong ? or everyone facing this issue ?
I'm using turboDiffusionXL, ip adaptor + embedding for Unit 1 and instand_id_sdxl and keypoints on Unit 2
CFG Scale on 5
Control weight between 0.5 or0.8
but it just does not work
**

here is my result

[nocturne topaz]

Im not 100% sure its compatible with turbo models

[outer olive]

the result with juggernautXL_v8Rundiffusion

[nocturne topaz]

And with both control weights at 1?

[outer olive]

YEs
I tried 1 - 0.5 - 0.8

nothing changed at all

[nocturne topaz]

Guidance step also needs to finish at 1

[earnest stream]

Did I miss something when doing LCM with LCM lora? I did everything that was said, though I am running in medvram. I decreased sample steps to 8 and below, tried CFG on 1 and 2, and it still gives me semi-coherent images. Got any ideas how do I improve on it?

[nocturne topaz]

Lower lora weight to 0.8 or 0.9

And try 5,6,7 and 8 steps in an x/y plot

[outer olive]

it's 1
i tried it but the result it's still the same !

btw i looked at the cmd and i saw this error

from D:\Ai\ImageGenAi\stable-diffusion-webui\extensions\sd-webui-controlnet\annotator\downloads\insightface\models\antelopev2\glintr100.onnx failed:Protobuf parsing failed.

Could this be the problem?

[nocturne topaz]

Ah - yeah thatd do it

[outer olive]

Oh so what should i do ?!

here is the full log

[nocturne topaz]

Id start with deleting the venv folder and restarting and hopefully it fetches a usable version of protobuf

[earnest stream]

Here's the results:

[nocturne topaz]

6 steps at about 1.8cfg looks like a sweet spot

[outer olive]

I'm trying it now. will share the result. thank you for your support

[nocturne topaz]

Howd you get on??

[outer olive]

Didn't work. i guess the problem is with my COntrol net webui. i'm redownloading it. our internet is hugely corrupted by government so i'm trying to download it from the beginning with VPN to see if it works

[earnest stream]

@nocturne topaz How do I remove the weird contrast thing that's happening in the image generation? I need it flat shaded, not with dramatic shadows.

[nocturne topaz]

Not sure sorry

[outer olive]

It works noww 😄
Thank you for your help
You put that spark in my mind that the problem might be with my system files
Thank you so much. i really do

[nocturne topaz]

Glad to hear it, have a great weekend!

[foggy arch]

how do i properly use Adetailer to fix the hand of an image i just generated?

[fossil stone]

What is the cheapest way to hosting custom SD models?

I want to build a system that more than one person can use at the same time.

[dusty blade]

interrogate cilip takes almost 5 minutes and then I get <error> does anyone know?

[vocal burrow]

those adetailer settings seem about right, you can tweak its inpainting settings too.

[vocal burrow]

what s the proper error log tho ? what s the full console log ?

[mighty topaz]

why my embeding dosent work?

[ornate elk]

Maybe your model is sdxl

[mighty topaz]

yeh i work on SDXL and i have downloaded SDXL embedding

[earnest basalt]

Does anyone know how to assist me possibly. I am making wallpapers and profile pictures for logos but I am faceswapping my friends faces into them. The problem is when I have a mask or eye patch or anything on face during text to image it messes up the face when the face is swapped in using reactor. Is there anyway to combat this not happening?

[ornate elk]

put mask and eyepatch in the negative prompt

[earnest basalt]

Would that then still apply it?

I am wanting to have things on the face. But it always skews when the faceswap applies

[ornate elk]

ahh so you want to have them in the image

[earnest basalt]

Yeah

[ornate elk]

the faceswap in reactor is applied after the generation so thats not possible. you would need to use IP-Adapter maybe to get that result

[earnest basalt]

Ahhh thanks. I'm not that advanced in SD. Is it difficult to do? I see pics all the time of celebs or actors etc that doesn't mess it up. Like for example John wick aiming down gun sight but his face isn't skewed or messed up by gun obstructing face

[ornate elk]

hmm do you have an example?

Controlnet IP-Adapter is a bit more difficult to setup and use than reactor

[earnest basalt]

So here you can see it tried to apply face tattoo but it's appeared smudged. I was wondering if there was a way to apply things post faceswap without it messing up

[ornate elk]

you could try to inpaint it

in img2img

[earnest basalt]

Would I just select face only and prompt what I want?

[ornate elk]

mask the face, select inpaint area "only masked"
set the denois to 0.5
set the resolution to 512x512

then prompt for tattoo, eyepatch etc

[hollow spade]

does anyone know why my stable diffusion hangs during image generation at times? any help provided would be much appreciated

[ornate elk]

Hey, can you describe it a bit more, what settings, models, what does the cmd shows

[hollow spade]

Anyanimeloramix by Lykon, a custom trained lora, reference-only, 533x832 W-H

The CMD gets stuck at random phases during generation

Sometimes when I hit generate it just doesn't load in the CMD at all

Just at 4% 1/20 etc

Then sometimes it gets stuck at mid way 48% 25/50 in the CMD

Every now and then before my batch of images finish generating it gets stuck at the last one at 98% 50/50 forever until I restart

It's really weird

My PC specifications are 8x2 (16GB) DDR4 3600hz ram, Ryzen 5 5600x, RTX 2060

[ornate elk]

whats in your webui-user.bat?

also do you have any other programms running while generating?

[hollow spade]

no, just discord.

[hollow spade]

isn't that the CMD?

this?

[ornate elk]

nope, when you right click the webui-user.bat and edit it

[hollow spade]

ohh I see

wait

[hollow spade]

pretty much stock aside from xformers

should I get rid of it?

[ornate elk]

no thats good, but add --medvram --no-half-vae there too

[hollow spade]

okay

[hollow spade]

beside xformers, yes?

[ornate elk]

also your using controlnet, does it not freeze when not using controlnet?

[hollow spade]

It does

Whether I use controlnet or not

every now and then

it will randomly freeze during image generation

more prevalent when I am increasing batch size, width-height

[ornate elk]

yea batch size should be 1, otherwise its to much for your vram

better use batch count

[hollow spade]

Weirdly enough I was doing 6 batch sizes fine yesterday

and yeah, it just freezed again

[ornate elk]

depends on the resolution

[hollow spade]

at 533x833

[hollow spade]

600x904 yesterday

yet like right now it froze on 1 image on 533x833

lol

[ornate elk]

do you have hires fix active?

[hollow spade]

nope

[ornate elk]

any extension active?

[hollow spade]

It's not upscaled

[hollow spade]

AIdetailer and controlnet

[ornate elk]

okay how big are your controlnet models?

[hollow spade]

big...?

as in file size

or quantity

if quantity, I have a lot.

[ornate elk]

file size

[hollow spade]

2.5-1.4gb

some are 700mb to 40+mb

it's a wide range

but I got a number within the 2+ gb range

but it shouldn't matter when I am only using one, no?

[ornate elk]

depends on what model your using

the 700mb ones should be the fastest

and recommended

[hollow spade]

Usually use reference only

[ornate elk]

oh ok

that has no model but requires a lot of gpu power

also IP-Adapter is better

[hollow spade]

The point is, regardless whether I am using extensions, controlnet, my resolution is high, or the batch size

It randomly hangs on me every now and then

even on 800p+ generated image with batch size of only 1 lol

[ornate elk]

can you try 512x768 ?

[hollow spade]

Sure, but some other time for now

[ornate elk]

also make sure Wallpaper Engine and Epic Games Launcher are force closed

they will cause lags and freezes

[hollow spade]

Oh is that so?

[ornate elk]

yes

[hollow spade]

I have wallpaper engine open, my bad. I forgot to mention that

Maybe that's the cause

I'll keep it closed from now on, thank you

[ornate elk]

no problem, that thing caused problems for some people, me included, for some the whole pc crashed when generating cause of wallpaper engine xD

[hollow spade]

hahaha

[ornate elk]

it requires vram and gpu power so best is to leave it closed (not paused)

[hollow spade]

I remember a time when I first got this PC, my wallpaper engine was making my valorant go black screen if I would to change wallpaper many times while valorant was ongoing

[hollow spade]

thank you

[ornate elk]

no problem 🙂

[neon ruin]

Still wish I actually understood what was going on under the hood here lol

It's funny, I'm a network guy but hardware just eludes me

I was able to generate some more images last night without crashing for the most part, there was one crash but I couldn't pin it on anything specific. I can't tell if it's simply a memory failure or an actual power supply issue. I would imagine 1200W is more than enough

[ornate elk]

yea 1200 should be more than enough

using 1000W on my 7900XTX

[neon ruin]

I am debating doing a clean install of everything

It did cross my mind that the cable going to the PSU was one of those extensions and wasn't actually the one included with the PSU

But I can't see that being the issue

Or at least, if it is, I'd be surprised.

[tiny abyss]

hello please help

[tiny abyss]

i updated my extensions and now its giving me an error when i use webui user batch file

[ornate elk]

try delete the venv folder and relaunch the webui-usr.bat

[unreal valley]

these 2 models just wont work together

is there a clip vision that works with the ipadapter model

[ornate elk]

are you on amd or nvidia?

[unreal valley]

nvidia

[ornate elk]

hmm

it should work

[unreal valley]

here is the error

[ornate elk]

what model are you using?

[unreal valley]

Anime mix by lykon

the sdxl version

[ornate elk]

yea thats the problem

IP-Adapter plus is made for 1.5

[unreal valley]

I am using the plus face sdxl for the ipadapter

[ornate elk]

you screenshot is low res

which model exactly?

and which preprocessor

[unreal valley]

lemme restart

here

I tried the vit-large-patch14 for the clip vision but had the same issues

once I change the model to this

it works

but will it work well with faces

I mean better than the face plus

[ornate elk]

face plus is only for faces

everything without "face" in the name will use the whole image as input

[unreal valley]

then how do I get face plus to work

[ornate elk]

thats my setting in auto1111

[patent meadow]

Hello can someone tell me why comfy manager wont install please

[ornate elk]

your on AMD?

[patent meadow]

Yes

[unreal valley]

oh boy

[ornate elk]

did you installed it and made the steps for the directml backend?

because Comfyui dont work out of the box with AMD

[patent meadow]

But I thought on linux it did

[ornate elk]

oh your on linux

ok

[ornate elk]

did you made this:?

[patent meadow]

Yes let me double check

Yep

[dense sandal]

suddenly automatic1111 stoped working for some reason? help?

[ornate elk]

update your GPU driver, then maybe delete the venv folder and rerun the wbeui-user.bat

[dense sandal]

ok

[unreal valley]

@ornate elk do you know anyone who could help me with this issue

it would be great if I could get a fix before I sleep tonight

[ornate elk]

have you asked in #🧣｜comfy-ui ?

maybe someone there knows it

[unreal valley]

lol, I didnt know there was a channel

I never scrolled that far down

thanks

[dense sandal]

its giving me some error... its weird yesterday it was working fine and I didn't change a thing

[ornate elk]

open up a cmd and type
pip cache purge
then delete the venv folder again

[dense sandal]

ok thanks

[ornate elk]

also whats your python version?

[dense sandal]

3.10.6

[ornate elk]

ok that should work

but you an also update to 3.10.11

[dense sandal]

is there a benefit in doing that?

[ornate elk]

bug fixes, better support for training tools like khoya

[sonic fable]

What is the recommended hardware configurations to try out few stable diffusion models ? I have a spec like 32 GB RAM, 6 GB GPU, CPU - AMD RYZEN 75800H. But still giving memory out of error while trying some specific models or while upscaling

[ornate elk]

whats your GPU?

[sonic fable]

NVIDIA

[ornate elk]

what model?

and whats inside your webui-user.bat?

[sonic fable]

NVIDIA Geforce RTX 3060 Laptop GPU

[dense sandal]

hey. everything went fine it seems. Only this:

[ornate elk]

okay then you need to edit the webui-user.bat
at the line COMMANDLINE_ARGS=
you need to add: --xformers --medvram --no-half-vae
Then save and relaunch the webui-user.bat

[sonic fable]

I am using comfyui and I run run_nvidia_gpu and here is the content inside that file.

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
pause

[ornate elk]

ohhh

comfyui

then it should already using the right args

whats the resolution and upscale size?

[ornate elk]

whats your GPU?

[dense sandal]

rtx 2060 12 GBVRAM

I was hoping to avoid updating python and pytorch I guess

[ornate elk]

okay then remove --disable-nan-check --disable-safe-unpickle -m pip install --upgrade pip
then save.
Open up a cmd and type:
pip cache purge
then delete the venv folder again.
Then launch the webui-user.bat

The problem was that the webui forget to use cuda for your gpu and used cpu instead

[dense sandal]

again

ok

[ornate elk]

then it will reinstall the needed stuff and hopefully useses torch with cuda

[onyx ledge]

Anyone have any idea how to resolve this error I'm getting?

[valid gorge]

I just got my 3090 up and running, and I'm getting quite slow speeds. Currently at around 1.5it/s which is way slower I should be getting. Ideas what to try next? I'm also new Windows haha. Current cmd args are: --no-half --api --no-half-vae --xformers

Also, should the webui.bat be runned as administrator?

[ornate elk]

pls never run the webui.bat
always run the webui-user.bat as non Admin
the slow speed comes from the --no-half
in your webui-user.bat

[valid gorge]

gotcha, will take the --no-half away and try again. thanks!

[unreal valley]

now I have this multiple faces issue

🥹 maybe I wont be able to sleep tonight

[valid gorge]

What does "Style database not found: D:\Stable Diffusion....." mean?

Maybe I should reinstall SD from the ground up

[sonic fable]

I get the below memory error, and having this setting in the workflow.

Error occurred when executing LDSRUpscaler:

Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated : 4.51 GiB
Requested : 1024.00 MiB
Device limit : 6.00 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB

[valid gorge]

Something is definitely crippled, now looks like it's not generating anything anymore

[ornate elk]

can be ignored

[ornate elk]

what do you mean?

[ornate elk]

whats the resolution?

why downsampling?

[valid gorge]

stopped generating all together, just does nothing

I was using an old copy from my old folder

[quick forge]

I'm just about done building my SD rig, is there any guide on how to install SD. I'll be running Arch Manjaro

[ornate elk]

ah okay

[ornate elk]

then you need to delete the venv folder

[quick forge]

I've never played with Linux before

[valid gorge]

Do I need to do a git pull after that?

[dense sandal]

I did everything as you said but I get all the way the first error

maybe I just need to install pytorch 3.10.11 ?

and then pytorch which I can't fint an installer

[ornate elk]

yes if it works, because moving the webui breaks autoupdate function

[valid gorge]

I'll give it a try

[ornate elk]

thats so strange, try restarting your PC too, make sure the gpu driver is updated

also are you on a laptop?

[dense sandal]

desktop

my driver isn't so old

maybe a month old

and I am on teamviewer so can't shut down the pc

I am trying --reinstall-torch --reinstall-xformers

in the webUI user

it didn't work though

[valid gorge]

After deleting the venv folder, I simply clicked the webui-user.bat to see what happens and it just started to download stuff, most likely rebuilding what was inside the venv folder I'm guessing.

[ornate elk]

yes exactly

[ornate elk]

do you have an intel CPU?

[dense sandal]

yes

[ornate elk]

okay, check in the taskmanager if you have 2 GPUs under the performance tab

[ornate elk]

okay, so yea update your gpu driver, run pip cache purge and then delete the venv folder again

[sonic fable]

The main goal is to figure out which upscaling method gives the best picture.. So I was trying different upscaling techniques in the workflow. The default size of the generated image is 512x512 and I am scaling the image by 2 with different upscaling techniques

[cosmic furnace]

Still get error at 93% of generation when usin medvram on 7900 xt =\

[dense sandal]

ok I will try to restar the pc and my girlfriend can start it over

[cosmic furnace]

and it feels like most of theese args doesnt do any good -_-

[valid gorge]

how the heck do you guys get so clean looking comfy workflows, mine are always looking like a noodle soup in a bowl

[cosmic furnace]

cud ive installed smthing wrong? hmgm

[ornate elk]

whats in your webui-user.bat?

and what resolution?

[dense sandal]

what do you mean?

[valid gorge]

What about TensorRT, should I install that at all? If I remember correctly some have got really cool results with it

[ornate elk]

replied to the wrong guy sry

[ornate elk]

i think that could be to much for your gpu

[sonic fable]

Ah you are right, I think the same.

[ornate elk]

if you installed Teamviewer on your PC, it should automaticly start when restarting the PC, so press restart and not shutdown

[cosmic furnace]

seems like writing any of theese --use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling . Making things only worse... With recent update had to use this to launch

https://www.youtube.com/watch?v=TJ98hAIN5io

well updating all shit (pip?) via venv command and install ar requirement...

[sonic fable]

I am not able to add the Node for ultimateSDUpScale in comfyui.. Any thoughts what might be the issue? But I already have it installed under the right path

[cosmic furnace]

the werid thing is that i cant generate anything using medvram just error at 93%

[ornate elk]

do you have the comfyui manager installed?

[sonic fable]

Yes I have

[ornate elk]

by following this stupid tutorial you broke your webui

its so bad

[valid gorge]

Back in up and running after the venv operation. Generating t2i 1024x and getting 1.7it/s at the moment with --api --no-half-vae --xformers, I think I should be gettin much much more than that with 3090

[ornate elk]

i still dont get how someone makes an easy fix that complicated and wrong

[ornate elk]

yes much much more

you can try remove --api
also is your webui updated?

also did you switched the gpu or the whole pc ?

[cosmic furnace]

y, looks i get alot of errors, venv related?o.0

[ornate elk]

thats an RAM error not VRAM

[valid gorge]

This is a new PC, with second hand GPU, fresh install of W11. The webui is from the older folder I was using with my Mac. I just installed python and git before using it.

[ornate elk]

from mac?

[valid gorge]

I need to do that next, good point. With git pull, right?

[ornate elk]

you need a reinstall

[cosmic furnace]

probably ram error lol) yeah i remember tab becoming white and saying not enough memory sometimes 🙂

[valid gorge]

Yea, will do that. Starting from fresh

[cosmic furnace]

wtf happening... shud i just reinstall SD from scratch or smthign? +)

[sonic fable]

I guess, I am hitting the same issue reported here https://github.com/ssitu/ComfyUI_UltimateSDUpscale/issues/23 anythoughts what can be done?

[ornate elk]

pls follow my install guide, its in the pinned messages of this channel,

then copy your models over

easier and faster than fixing this

[valid gorge]

Thanks, will do that!

[ornate elk]

would be the best because you already broke it with modifing some files.
Make a new folder and then follow my install guide, its in the pinned messages of this channel

and if you then get errors, we can easily fix them

[cosmic furnace]

3.10.6 or 3.10.11 python better?)

[ornate elk]

3.10.11

[valid gorge]

I currently have 3.10.6, can I just reinstall and it will update it, or should I first uninstall anything?

[sonic fable]

Anyways, I have fixed the issue. I tried installing Reactor and it was not getting installed which has broken UltimateSDUpscale after removing reactor it started working fine :). Thanks a lot for all the guidance

[valid gorge]

Looks like the installer is smart and ask me if I just want to update 🙂

[ornate elk]

ah good to hear 🙂 np

[valid gorge]

Looks like there's a new pip available, 24.0, I currently have 23.0.1. Should I run the update?

[ornate elk]

not needed

[valid gorge]

Ok, thanks

[dense sandal]

still getting this after updating nvidia driver..

I could update to python 3.10.11 I guess but I don't know how to update pytorch or torch whatever it is

[valid gorge]

How do you guys take a screen shot in Windows? Sorry for being an idiot tonight and not knowing anything about anything. Been on Macs my whole life, my first day in W11 haha

oh I see, alt+prnt screen

[dense sandal]

press and hold windows key + shift + s

[valid gorge]

Thanks, will try!

[dense sandal]

then use mouse to select area

[valid gorge]

Heeey, this is pretty cool! Working 🙂

[dense sandal]

indeed

[ornate elk]

did you updated your gpu driver?

[dense sandal]

yes

[steady thistle]

@ornate elk Would you recommend zluda yet for AMD? It's supposed to be decent at running Cuda in AMD, but it's in alpha now.

[ornate elk]

nope wouldnt recommend as the pytorch support is minimal and untested

[steady thistle]

Thanks, how about stable cascade, should I try to work with that using AMD?

[sonic fable]

Wanted to check, if there is a way to save each images with a suffix that holds the current timestamp value?

[valid gorge]

hmmm, what heck is going on, I'm back stuck at "Commit hash:xxxxx" been waiting already couple of minutes without nothing

[ornate elk]

idk the hardware specs needed to run it

would wait

[steady thistle]

Okay, thanks

[ornate elk]

can you show a screenshot?

[valid gorge]

Looks like I can't even kill the process

haha

took forever from task manager to kill the process

Maybe something happened when I copied the extensions folder, better to maybe reinstall those manually

[ornate elk]

yes pls xD

now you need to delete the venv folder again

[valid gorge]

doing a clean sweep, just to make sure. I'm hopeless haha

[dense sandal]

why is it downloading an older version?

it asked me to update, I updated but when I run webUI it seems it tries to install 2.0.1 instead of 2.2.0?

and then it gives me again the error

[old parrot]

How much faster would a 3060 12gb VRAM be than 1660S 6gb VRAM? Currently it's about 7 minutes per Hires generation

[ornate elk]

a lot faster

maybe 1 min per hires

[ornate elk]

2.0.1 is the latest for nvidia

2.2.0 is dev

[quick forge]

how would an rx6800 fare?

[ornate elk]

on win or linux?

on linux a bit faster on windows a lot slower

[quick forge]

linux

[valid gorge]

3.48it/s, a bit better but I think it should be much faster than that with this card

[quick forge]

SD doesn't do video does it?

[valid gorge]

SD can do video

[quick forge]

oh sweet

im setting up a rig now to run it

goin gto run arch manjaro

[valid gorge]

Cool!

[quick forge]

I have no idea how to install it though lmao

SD i mean

[valid gorge]

Check the pinned comments of this channel, CS1o has put some install intructions there

[quick forge]

also am I to understand that SD can use multiple video cards, but will only use one video card per generation instead of combing their power ?

I'm only going to be running a single card anyway

[valid gorge]

SD 1.5 at 512x is giving me 17.9it/s and SDXL at 1024x 3.46it/s, sounds too slow for 3090

[ornate elk]

these are for windows only

[sonic fable]

I tried setting the image name to this format in comfyui %date:yyyy-MM-dd%/Default_%date:yyyyMMdd_hhmmss%_%KSampler.seed% but it looks like it is somehow adding _ in the last in the file name.. How can I exclude that underscore? i.e Default_20240217_012040_588185946122308_00001_

[dense sandal]

so... is it okay that it installs 2.0.1 or is it wrong? cause I can't fix this yet.

I upgraded to 3.10.11... suposedly installed pytorch 2.2.0

and still get the message

everything worked fine yesterday which is why is so odd also.

[jolly notch]

can we use stability. ai API with controlnet?

[ornate elk]

yea dont install 2.2.0

[dusty blade]

@ornate elk Hello again, can I use both impaint and upscale?

When I use both, the result is bad

I want to pscale the faces without distorting them. If I upscale the faces selectively, it distorts, if not, the faces change.

[valid gorge]

Are you in A1111 or something else? How do you use inpaint together with upscaling? Are you doing an inpaint for some area with a mask, what are your settings? What model are you using, do you hae VAE enabled and which one?

Depending of the model you're using, use 512x512 or 1024x1024 when inpaining, otherwise you might get weird results.

To me that looks a bit like you're having a problem using an SD VAE with SDXL model?

[dusty blade]

A1111

[ornate elk]

you cant use inpaint and upscale together

[dusty blade]

aw man

faces 😦

[ornate elk]

you need to use --no-half to get inpainting to work on AMD

but that also will use more vram

[valid gorge]

In the inpainting, mask the face, set masked content to original and inpaint area to only masked, depending of your model SD / SDXL set your width and height accoringly 512x512 / 1024x1024, CFG 7 and Denoising at around 0.55. Experiment with a bit lower CFG and a bit higher denoising values. You should get a lot more details this way. Use the same method for the rest of the image; torso, arms, legs etc. other parts that need more details.

Use shorter and preciese prompt when inpainting, like "face of a warlock, long hair" or what ever. When doing other details adjust accordingly. When doing backgrounds etc you can do larger areas at once.

If needed upscale the image first (the extras tab) before starting the inpaint process for the details. Usually keeping the longer side at around 1024 will work fine, then after you've got all the details, you can upscale it to the final size and keep the details.

[old parrot]

Looking to upgrade from a 1660 Super 6gb GPU, any suggestions?

[valid gorge]

Both would make a great fit. If you’re on a budget, go with the 3060. The speed increse between 3060 and 4060 ti is something like 35-40% in SD use.

Between your current card and 3060 it’s maybe at around the same, 30-40%

[old parrot]

4060 it is! Prolly not many better options in that price range? Appreciate you!

[valid gorge]

Yep, the 4060ti is pretty much the only option unless you have a bigger budget, like spendin 300-400 more

[quick forge]

which pinned message is about installing stable diffusion
?

[nocturne topaz]

You gotta find the one that matches your graphics card type and operating system

[quick forge]

am i not looking in the right pl;ace? I have an AMD card and Manjaro linux

and im not seeing a specifc swetup in the pinned comments here

like is it automatic1111 thats being installed?

ummm...also im not sure what to use for a command line.... this is literally my first time ever using linux

ok i can't seem to find a linux version, so i must be looking in the wrong place?

[nocturne topaz]

Terminal for the command line

[nocturne topaz]

I thought there was one there - its be a case of installing RoCM onto the system, then follow the standard install instructions on the a1111 github page

[quick forge]

Im sorry but i have no idea what you're talking about

what is rocm and how do i install it?

im actually really good with computers just never used linux bewfore....ever lol

[nocturne topaz]

Rocm is a bit that plugs in to linux to make amd cards run like nvidia cards - google rocm install linux

[quick forge]

thanks ill look it up

what do i install it as? im using manjaro, is that ubuntu or what?

[nocturne topaz]

No idea sorry - @ornate elk might pop by later and give you more detail than I can

[dense sandal]

automatic1111 stopped working so I am giving Forge a chance...

but when I run update.bat it does nothing

[nocturne topaz]

I dont think forge works with the one click installer?

[dense sandal]

it says so

https://github.com/lllyasviel/stable-diffusion-webui-forge?tab=readme-ov-file

[nocturne topaz]

K cool - what do you get when you run update.bat?

[dense sandal]

Nothing at all

[nocturne topaz]

Like empty cmd window?

[dense sandal]

Oh wait. Maybe because I uninstalled things

Cmd doesnt work maybe I uninstalled what made it work

[nocturne topaz]

Oops

[dense sandal]

Powershell right?

Or did I brake something in my computer

[nocturne topaz]

Command window

Powershell tends to cause weird issue

[dense sandal]

Swont open

Console doesnt work

[nocturne topaz]

Id backup the models and output folders then wipe and start from scratch

[dense sandal]

Yes I am doing that

I am tryig to instal forge

[nocturne topaz]

And take a list of extension too

[dense sandal]

But my console isnt working

[nocturne topaz]

Windows + R then type cmd and press enter

Does that bring up a window?

[dense sandal]

no, but I am doing something to fix it I think

Okay it's in spanish but it's doing some repair

[nocturne topaz]

Yeah somethings not happy then

[dense sandal]

It didt fix it

[nocturne topaz]

Like if sfc doesnt fix it id be looking at system restore

[dense sandal]

Ok

[cursive anchor]

I reinstal everything as said in the pin comment. When I try to run it, my pc froze. This is what I see in the cmd before it happens

And thanks man for the help. Really appreciate it

I added in COMMANDLINE_ARGS= You add: `--xformers --no-half-vae --medvram-sdxl

[nocturne topaz]

Collecting torch can take a while

Especially on slow connections

[cursive anchor]

Ohh then I’ll plug it to Ethernet

Thanks !

[ornate elk]

how much RAM do you have?

let it load, it can take a while

if the pc dont crash, let it run

[cursive anchor]

8gb

Although it crashed

Should I lower the medvram?

[ornate elk]

RAM or vram?

[cursive anchor]

At least is a new one

[cursive anchor]

Let me check

[cursive anchor]

My pc decides to disconnect form wifi. Nice

[ornate elk]

Thats a connection error

[cursive anchor]

I have 2 GPU

[nocturne topaz]

So the 6gb gpu 1 is the one you want to use

[cursive anchor]

Then I’ll delete -sdxl

Any other change I should make? Or I can try to run it again?

Should I put the python path?

[silk estuary]

I am trying to setup SD forge so that it uses checkpoints and LoRAs from another SD instance (1.6.0) folders. I use --embeddings-dir, --ckpt-dir and --lora-dir for that.
However, when opening SD forge, the tabs are empty
I tried doing the same with an indepentdent instance (1.7.0) and it has the same issue. It isnt finding embeddings, LoRA or checkpints; or it is not searching them in the correct directories.
Any idea what I could do to fix this?

[vocal relic]

Not necessary. When you build the virtual environment, it downloads a copy of Python.

[nocturne topaz]

You are setting the lora directory before setting the a1111 home variable - try putting the full path in instead

[glad wadi]

  out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)```
does this matter? it still runs

Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:08<00:00,  2.39it/s] ```
2.39 it/s is slow for a RTX 3080 @ 512x512 on SD 1.5 tho right?

[nocturne topaz]

Is that using a Euler A sampler

[glad wadi]

yes

[nocturne topaz]

If so, then yeah somethibg is very wrong

I would delete the venv and restart comfy

3080 should be getting at least 10-15it/s for that

[flat olive]

hii, i need to know if someone has an idea about amd drivers on linux, which system is more recommendable to use with sd?, and which this os can support amd drivers?

[lime basin]

Would the Stable Diffusion 1.5-trained Loras work on the SDXL, or would I need to retrain the Loras on the SDXL model?

[nocturne topaz]

Need retraining until x-adapter comes out

[nocturne topaz]

@ornate elk

[drifting aspen]

Hello I have a rtx 3060 with 12 GB and 16 GB ram. He wants to do images in 1080 on 828 model sahastrakoti refiner sdxl. The problem is that I have a message out from memory. I want to generate images in batch. I have in txt file saved prompts load them in automatic1111 and at 2nd or 3rd image I have this message what to do to make batch file generation work ?

[nocturne topaz]

Use batch count instead of batch size

And by 1080 do you mean 1920x1080??

[drifting aspen]

No 1080x828

I tested and at lower resolutions the same thing happens

[nocturne topaz]

Try 1088x836

[drifting aspen]

I used 540x414. Test to see if the whole process will come. I set step 30

[nocturne topaz]

Goes best in multiples of 64 eg 512x384

[drifting aspen]

And after the 4th image the error not enough memory you Tried allocate 2614400 bytes

[nocturne topaz]

Send a screenshot of the A1111 page?

[drifting aspen]

I will check other resolutions. If I have this message then the ram is 100%

Ok is 16 GB ram not enough

[nocturne topaz]

I thought you had a 3060 - it shouldnt be using the cpu??

16gb wont be enough for windows and sdxl

[drifting aspen]

I have a 3060 rtx 12 GB. While working I can see that the rtx graphics memory is taking up as well as the ddr4 ram

[nocturne topaz]

Something is wrong then as it shouldnt be using cpu at all

Send a pic of the contents of the webui-user.bat file

[drifting aspen]

I have not changed the file I have by default

How much ram should it have?

[nocturne topaz]

Close the command window then. Go into the webui-user.bat file and on the line commandline_args= add

—xformers —medvram

To the line then run that file

[drifting aspen]

So far it's working

I understand that the more steps the more accurate the image?

[nocturne topaz]

Not necessarily - depends on sampler but generally 20-30 is optimal

[drifting aspen]

And what Linux do you recommend for automatic1111. Debian? Or any other?

[nocturne topaz]

Debian or ubuntu derivatives

[ornate elk]

Hey

[ornate elk]

Why do you want to use Linux with nvidia ? xD

With a 3060 you shouldn't have any problems in SD

[flat olive]

cs1o

can you answer my question? pls

[ornate elk]

Which OS supports AMD Drivers?

Linux and Windows

[flat olive]

yep but i tried with ubuntu and i cant install the drivers

and with kali

the same thing

[ornate elk]

You normaly dont need to install the drivers if you installed the OS with 'free drivers"

A good Linux OS for AMD would be Ubuntu or Manjaro

[flat olive]

some guy in another server recomended arch fedora ubuntu

[ornate elk]

Thats okay too

The only thing you need to install by yourself are the Rocm drivers

[flat olive]

alright

so

for sd i don't need to install the amd drivers?

on linux

like kali

[ornate elk]

Yea normaly they are included in the install

Make sure you update everything after installing the OS

[flat olive]

everything? like what?

[ornate elk]

Every package

[flat olive]

bios update?

[ornate elk]

No

Linux packages

[flat olive]

okok

im gonna try

[ornate elk]

You should try to increase your windows pagefile

[flat olive]

cs1o

i use webui-user.bat for the commandsargs? for the skip cuda test parameter

[ornate elk]

In Linux its webui-user.sh

[civic thunder]

How to generate image?

[ornate elk]

And dont use --skip-torch-cuda-test

[ornate elk]

The bots are currently under maintenance. So you need to use other services or install SD localy on your PC

[civic thunder]

Thank)

[flat olive]

im supposed to launch it with launch.py right?

[fallow dove]

#🎥｜animation A boy wearing 90's japanese dress hanging to a rod above the ground in playground.

[drifting aspen]

I will still buy ram because I only have 16 GB. 32 GB of ram should get the job done. Eventually automatic1111 will install on Debian. In windows that's what I wanted. Check how it works. Now I let the series go to 576x448 resolution.

[ornate elk]

Yes

[flat olive]

still throwing the skip cuda error

i removed the # from export COMMANDLINE_ARGS=

still nothing

[ornate elk]

Have you installed rocm?

[drifting aspen]

Unfortunately, it's still the same

[ornate elk]

Dont use the refiner and it will work

Its also not needed

[flat olive]

its okey if it show me this when i install rocm?

[drifting aspen]

I found easy difusion is better than automatic1111?

[ornate elk]

Easy Diffusion has less feature and support than auto1111

[ornate elk]

Your still on Kali?

I thought you switched

Doesn't look good

[flat olive]

yep

[ornate elk]

Have you installed python3 ?

[flat olive]

yep

[drifting aspen]

This refiner does what? Since it is unnecessary?

Without the refiner, the images are not so accurate. What should I do to use the refiner?

[ornate elk]

It should smoothen the image but upscaling does the job better

Nobody I know uses refiner + sdxl
Fooocus webui uses it sometimes

[flat olive]

when i installed ubuntu it just doesn't let me install amd drivers so i decided to go with kali to see what happen

[ornate elk]

Ah okay

Can't help much with linux

On the Shark DC is a "help installing Linux channel"

[drifting aspen]

What model and how to set automatic1111 to get good images for coloring?

[ornate elk]

You should use a vae

[flat olive]

the dead server?

[ornate elk]

Its not dead

[drifting aspen]

https://aiconomist.cc/how-to-create-ai-coloring-books-for-kids-free-method/

The video shows how he selects the refiner and generates images

Without the refiner, the memory does not reach 16 GB only 13 GB and the message does not appear so far.

[crystal lynx]

Where can i create pictures?

[drifting aspen]

Automatic1111

I downloaded the vae files. But I don't know how to use them

[vocal burrow]

locally or use online services as for now bots are #1047610792226340935

[ornate elk]

Go into Settings, User Interface, Quicksettings, there add sd_vae
Then hit apply and reload

Then you should have a VAE dropdown

[vocal burrow]

VAE quick guide :
1/ What is a VAE?
It's a part of the stable-diffusion pipeline that encode/decode information from tokens to latent space and from latent spaces to pixels. Aka it transforms math statistics in pictures
2/ Where do I put my VAE ?

• VAE with .vae.pt, .vae.ckpt, .vae.safetensors extensions go into the models\Stable-diffusion folder
• VAE with .pt, .ckpt, .safetensors go into models\VAE
  3/ How do I use my VAE ? Three possibilities :
• Either you name it similar to another one of your model (eg : Anything-V3.0.safetensors + Anything-V3.0.vae.pt), by doing that it should automatically load the VAE when you load the associated model.
• You manually load your VAE by going to Settings -> Stable-Diffusion -> sd_vae and selecting your VAE
• You add an easily accessible VAE dropdown at the top of your page to quickly switch back VAE by adding sd_vae to your Settings -> User Interface -> Quicksettings list

[drifting aspen]

Thank you

[reef smelt]

why does this happen?
"tank taco:10", unless I'm mistaken it should generate a tank and then after 10 steps it should add taco to the prompt?

it gets totally deepfried

but "tank taco" works

except it's not a tank

it's incoherent at that point

[nocturne topaz]

No, that has multiplied the weight for taco by 10 which is way too high

[vocal burrow]

No ? that's not how it works.
Plus there's not even some barrier to limit which prompt should be replaced by what.
With your logic if the prompt is a man walking his dog cat:10 what should it be interpreted as ? a man walking his dog until step 10 then it turns into a man walking his cat ? Why would it limit itself to one word ? Why wouldn't it simply turn into cat at step 10 ?

[nocturne topaz]

You need to find the correct syntax

There is a prompt swapping syntax in a1111 for swapping tokens mid generation

I think it uses square brackets and pipes

[vocal burrow]

you can "alternate" prompt with a man walking his [dog|cat] https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words

so something like a man walking his [dog|dog|dog|dog|dog|cat|cat|cat|cat|cat] will do dog for the first 5 steps and then cat

[nocturne topaz]

Its the prompt editing in the section above it

[vocal burrow]

also that yes that was my transition :p

As just mentioned you can also type a man walking his [dog:cat:5]

[vocal burrow]

That's why I was taking so long to answer but it seems like it's not applying a weight of 10 tho ? For that you'd need the parenthesis. So I don't know ? Undefined behavior ?

But it definitely has the typical look of an overweight image tho.

weird

[nocturne topaz]

The default is to multiply the whole prompt by 10 in that scenario

So above it would be tank 1 taco 1 all times 10

A1111 syntax is a funny beast

[reef smelt]

op- I see

[nocturne topaz]

So yeah ‘[tank: taco:10]is what is needed there

[ornate elk]

okay wtf, Zluda for AMD is fast af! (16it/s on 7900XTX)

19 seconds for, 512x768, 30 steps, DPM++ 2M Karras, Hires Fix upscaled by 2, 10 Hires steps, and Controlnet Openpose

[reef smelt]

what's zluda

[nocturne topaz]

Piece that allows amd cards to do the same processing commands as nvidia ones (cuda)

[reef smelt]

oh okay

[nocturne topaz]

Its good news that its going as that means nvidia hasnt got artists by the short and curlies and will have to start upping vram on cards to match amd

[reef smelt]

I get like 1.2it/s on rtx 2060 but other people with the same gpu get 13it/s?

[nocturne topaz]

Are you using a1111?

[reef smelt]

yes

[vocal burrow]

Can I get source on that ? Cause that's not what the code seems to hint at ? So it might legitimately be an UB.

[nocturne topaz]

Try using sd-webui-forge in a seperate install to test speed

[reef smelt]

separate install?

[nocturne topaz]

As opposed to cloning over the top like in the instructions

[nocturne topaz]

I remember testing it a few months ago.. will look it up

[vocal burrow]

Who's getting 13 it/s on a 2060 ?!? Even when I was using SD1.5 model converted for a usage with TensorRT I had around 10it/s on my 2070S

[reef smelt]

I'm gonna look

I saw it tho

[ornate elk]

hey, i tested it and it works amazingly fast!
i modified a new directml install to use zluda and it works, but some extensions not and sdxl models wont load.
but some people tried SD.next there sdxl works but vae hangs
so if you want fast 1.5, you should give it a try

[vocal burrow]

As a reference without using any speed hacks or dark magic optimization I'm having 1.7 it/s when using SDXL models on my 2070s (with --medvram-sdxl)

[nocturne topaz]

You tried forge??

[vocal burrow]

I plan on it but honestly giving that's is basically auto1111 with a few tweak, I just hope that they manage to write a a proper PR for their stuff and for it to be accepted in the near future.

[nocturne topaz]

I can get 13it/s on a rtx5000 using forge, but thats got 3000 cuda cores vs 2300 in 2070

[vocal burrow]

rtx 5000 ?

[nocturne topaz]

Forge wont get merged into a1111, its rewritten wayyyy too much code

[nocturne topaz]

On paperspace

[ornate elk]

yea its better to contribute to the original if the changes are minor

[nocturne topaz]

Its a pro level card

[reef smelt]

I suppose medvram does have an impact...

[vocal burrow]

Definitely yes.

[reef smelt]

well even when I tested without medvram I still got values that are subpar

[vocal burrow]

I wouldn't say the changes are "minor". And that's probably part of why it's not "just a PR"... We'll see how it goes.

[nocturne topaz]

Given the cuda core count and the slower nature of the older card, id say 6-7it/s for 512x512 euler a is a good target

[nocturne topaz]

I went from 7.8 to 13it/s on the rtx5000 and from 9 to 16it/s on the a4000

It really is worth a look

[vocal burrow]

I might have time for it this evening or tomorrow.

I also want to take a look at what they've actually done under the hood.

or at the very least a quick glimpse.

But for now I have some IRL stuff to do. Please @ me if you ever find an explanation about that weight thing.

[nocturne topaz]

Tl;dr, got rid of monkey patching, split the unet out to allow seperate modification, cleaned up spaghetti code

[reef smelt]

how does controlnet reference only work?

[nocturne topaz]

Do you want a eli5 version or a technical one?

[reef smelt]

😔 how do I use i

it

[valid gorge]

Wondering what's going on. Installed my extensions and after that I get this.

Just added the --no-half, let's see if that helps and how much slower things get

[ornate elk]

What's your GPU?

[valid gorge]

RTX 3090

[ornate elk]

Dont use --no-half

What's your model and VAE?

[valid gorge]

JuggernautXL v7, VAE automatic

[ornate elk]

Set VAE to none or sdxl vae

Also use --no-half-vae

[nocturne topaz]

Go download the fp16 fix vae

Both solutions will work, btw

[valid gorge]

Thanks, will test those. Wondering what caused that in the first place. I currently have only --xformers only in my bat, worked fine earlier

[nocturne topaz]

Sdxl models occasionally end up breaking the stock vae as they exceed the limits of 16 bit mathematics, and so you need a vae that stays within the limits or allow it to go to 32 (ie no-half)

[valid gorge]

Ok, gotcha. Just wondering what happened as everything was ok just a minute ago before installing my extensions haha

[nocturne topaz]

Itll only do it on some prompts with some models, but using the fixed vae sorts that out

[valid gorge]

Ok, thanks! So far haven't really bothered using separate VAE at all

[grand fable]

hi, can you give me some hint how to install a custom node from GitHub to ComfyUI? I work on MacBook.. https://gist.github.com/laksjdjf/487a28ceda7f0853094933d2e138e3c6

[vocal burrow]

check pinned comments

[quick forge]

Is there a guide somewhere on how to install SD on a linux machine with an AMD card, but like for complete idiots? I've never used linux before and I'm a bit confused

[steady thistle]

Thank you so much for testing this! I'll try to install it, but do you have any install suggestions?

[ornate elk]

depends on what you want (the normal auto1111 ui or the SD.next fork)

[steady thistle]

A1111 is what I'm used to. I want the most simple way that won't crash everything. lol

[ornate elk]

Here is a guide for Linux:
https://civitai.com/articles/1503/how-to-install-rocm-on-any-form-of-archlinux-or-archlinux-variant

[steady thistle]

I meant with Windows. I don't run Linux

[ornate elk]

okay, i can summarize the steps needed

[ornate elk]

this message wasnt for you

[steady thistle]

oh, heh

Good, because if this fixed it I might have to go down the Linux rabbit hole

[tiny abyss]

Launching Web UI with arguments: --xformers --medvram
Traceback (most recent call last):
File "C:\NEW SD_stable-diffusion-webui\launch.py", line 353, in <module>
start()
File "C:\NEW SD_stable-diffusion-webui\launch.py", line 344, in start
import webui
File "C:\NEW SD_stable-diffusion-webui\webui.py", line 29, in <module>
import gradio
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\gradio_init_.py", line 3, in <module>
import gradio.components as components
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\gradio\components.py", line 55, in <module>
from gradio import processing_utils, utils
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\gradio\utils.py", line 494, in <module>
class AsyncRequest:
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\gradio\utils.py", line 513, in AsyncRequest
client = httpx.AsyncClient()
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\httpx_client.py", line 1397, in init
self._transport = self._init_transport(
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\httpx_client.py", line 1445, in _init_transport
return AsyncHTTPTransport(
File "C:\NEW SD_stable-diffusion-webui\venv\lib\site-packages\httpx_transports\default.py", line 275, in init
self._pool = httpcore.AsyncConnectionPool(
TypeError: AsyncConnectionPool.init() got an unexpected keyword argument 'socket_options'
Press any key to continue . . .

Having this error after reinstallig venv folder via Webui batch file.

[ornate elk]

make sure the wbeui is whitelisted in any browser adblocker

oh wait is your webui updated

[tiny abyss]

ahh, how can i do that?

[ornate elk]

@tiny abyss You go into the NEW SD_stable-diffusion-webui folder, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
then type git pull and hit enter

then show me what it says

[chrome rampart]

hello all. Can you in SD Forge also set somewere that it should pick up my models/lora etc from me normal SD A1111?

[ornate elk]

okay looks good

now delete the venv folder again

and relaunch the webui-user.bat

[ornate elk]

havent looked at it but Symlinks should always work

[chrome rampart]

thks as always

[tiny abyss]

i also have another venv folder in the extention of the sad talker i recently installed, after that the sad talker ran for two days or so, but then i updated everything from Webui extention tab

after which the main web ui gives me this error

[ornate elk]

do you still get the error now?

[tiny abyss]

trying

still in process

[ornate elk]

@steady thistle are you ready? xD

[steady thistle]

I think so! I'm downloading CMake and looking at Rust next. I installed AMD Adrenaline

[ornate elk]

[AMD] ZLUDA Guide: Install Zluda on AMD Windows with Auto1111 Directml Fork:
For Advanced Users
Works with AMD RX5600XT, 5700XT, RX 6000 series and above.
Requires AMD Adrenalin 24.1.1 Driver or newer.

Preparation:
You need to install Git 64bit: https://git-scm.com/download/win
and Python 3.10.11, 64bit: (any python above or 3.10.6) Not 3.11, 3.12 https://www.python.org/downloads/release/python-31011/
And check "add python to path" when installing Python.

1. Make a new Folder on your Drive (not on Desktop, Downloads, Documents, Programms, Onedrive) and name it SD-Zluda.
   Dont use the same if you had installed the DirectML webui before.
2. You go into the folder you created in this case SD-Zluda, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
3. Then you copy and paste this command:
   git clone https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu && cd stable-diffusion-webui-amdgpu && git submodule init && git submodule update
4. Press enter and after its done you can close the cmd.
5. Edit the webui-user.bat (right click), At the line COMMANDLINE_ARGS= You add: --use-zluda --medvram-sdxl --update-check --skip-ort
6. Optional: Move one 1.5 based model from your other existing directml webui install into the models/stable-diffusion folder. Or skip this if you dont have one.
7. Launch the webui-user.bat, you'll get an error. Just close and go with the next steps.

Getting Zluda to run:

1. Install AMD HIP SDK 5.7 from here: https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html
   Dont check the Pro Driver.

2. Download the latest Stable build of ZLUDA as zip file from here: https://github.com/lshqqytiger/ZLUDA/releases/
   Unzip as Folder and move it onto the C: Drive and rename the Folder to ZLUDA

3. Add the C:\ZLUDA folder and %HIP_PATH%bin to your PATH in the System Variables. Like shown here:
   https://www.wikihow.com/Change-the-PATH-Environment-Variable-on-Windows

4. If your GPU is below a RX 6800, you need to do the following steps:
   For RX 6700 S, XT or 6750XT download Optimised_ROCmLibs_gfx1031.7z
   For RX 6600 or 6600XT download Optimised_ROCmLibs_gfx1032.7z
   For RX580/590, RX5600XT, 5700XT, VEGA 56, VEGA 64 and Radeon VII you need to download the ROCmLibs.7z
   For AMD 780M-APU you need to follow the steps here: https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU-
   File Source: https://github.com/brknsoul/ROCmLibs
   4.5 Go into C:\Program Files\AMD\ROCm\5.7\bin\rocblas folder.
   There copy and rename the library folder to old_library
   Open the .zip file and drag and drop all files of the library folder into the library folder. Not into the old_library folder.
   Important Restart your PC after this Step before proceeding.

5. Launch the webui-user.bat. It will open in your default Browser.

Noteworthy: The first image generation can take from 15 up to 40 minutes. First time only.

If you want to link all models from an other Webui you can do that by adding --ckpt-dir "Path/to/models/stable-diffusion" --lora-dir "Path/to/models/lora" --controlnet-dir "..." --embeddings-dir "..." to the webui-user.bat, at the line set COMMANDLINE_ARGS=

[steady thistle]

Oh, this isn't bad. Do I need CMake or Rust?

[ornate elk]

nope

[steady thistle]

Yay

[ornate elk]

its already compiled for windows, so you only need the zluda zip from the link (not from the main zluda github) and hip sdk from amd, also linked

[steady thistle]

Don't forget to pin this! lol, this is very exciting, thank you so much for all your help here.

[tiny abyss]

its stuck on receiving objects : 11%

[valid gorge]

1024x image 4x ultimate sd upscale takes around an hour with 50 steps on my Mac Studio M1. My new GPU handles that under 4 minutes.

[ornate elk]

can you show a screenshot?

[valid gorge]

wheeee

[ornate elk]

hmm make sure you have a good internet connection

using ethernet at best

[steady thistle]

"6. Move one model from your other existing directml webui install into the models/stable-diffusion folder." Which model do you mean? a 1.5 model or sdxl? Or just...?

[ornate elk]

oh good point 1.5

sdxl wont work rn

[steady thistle]

When adding --medvram-sdxl --update-all-extensions --update-check --skip-ort, do I add these after: --use-directml --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling? (and it's okay to remove --lowvram?)

[ornate elk]

did you edit your exsting install or did you git cloned new?

because it will break your directml webui when doing these steps, thats why i said to git clone a new one

[steady thistle]

Oh! I just realized it made a whole new A1111 folder... Now I get it. THanks!

[ornate elk]

and in the new webui-user.bat there only goes the command args mentioned

depnding on your gpu vram you may need to use --medvram instead of --medvram-sdxl

[steady thistle]

For dummies like me, you might want to edit the pin to say "5. Inside the newly created SD-Zluda folder, edit the webui-user.bat (right click), At the line COMMANDLINE_ARGS= You add: --medvram-sdxl --update-all-extensions --update-check --skip-ort"

how do I find my ROCm version?

[ornate elk]

oh right

[sonic fable]

Any one knows, how to fix the face in the picture?

[chrome rampart]

back lol i must be idiot and do something wrong since i get error when i want froge version to pick up models and stuff from me regular SD @REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
--ckpt-dir e:/a1111/stable-diffusion-webui/models/Stable-diffusion
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
--lora-dir e:/a1111/stable-diffusion-webui/models/lora

[steady thistle]

You can use Adetailer but for me personally, I swap the face using ReActor. I find it's quicker and easier and gives a consistent face, which may or may not be ideal for you.

[tiny abyss]

getting this error now

venv "C:\NEW SD_stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Cloning Stable Diffusion XL into C:\NEW SD_stable-diffusion-webui\repositories\generative-models...
Cloning into 'C:\NEW SD_stable-diffusion-webui\repositories\generative-models'...
remote: Enumerating objects: 860, done.
remote: Counting objects: 100% (481/481), done.
remote: Compressing objects: 100% (212/212), done.
error: RPC failed; curl 56 Recv failure: Connection was reset
error: 1943 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output
Traceback (most recent call last):
File "C:\NEW SD_stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "C:\NEW SD_stable-diffusion-webui\launch.py", line 39, in main
prepare_environment()
File "C:\NEW SD_stable-diffusion-webui\modules\launch_utils.py", line 409, in prepare_environment
git_clone(stable_diffusion_xl_repo, repo_dir('generative-models'), "Stable Diffusion XL", stable_diffusion_xl_commit_hash)
File "C:\NEW SD_stable-diffusion-webui\modules\launch_utils.py", line 192, in git_clone
run(f'"{git}" clone "{url}" "{dir}"', f"Cloning {name} into {dir}...", f"Couldn't clone {name}", live=True)
File "C:\NEW SD_stable-diffusion-webui\modules\launch_utils.py", line 116, in run
raise RuntimeError("\n".join(error_bits))
RuntimeError: Couldn't clone Stable Diffusion XL.
Command: "git" clone "https://github.com/Stability-AI/generative-models.git" "C:\NEW SD_stable-diffusion-webui\repositories\generative-models"
Error code: 128
Press any key to continue . . .
@ornate elk

[dense sandal]

Cmd isnt opening. I tried lts of thingss

Even reinstalling windows preserving my files

[ornate elk]

try relaunch again

[quick forge]

do i run all of those commands for manjaro, and the "both arch and manjaro"

[tiny abyss]

@ornate elk pls help

[quick forge]

ok iverun into an error

(1/1) installing go [####################################] 100%
:: Running post-transaction hooks...
(1/2) Arming ConditionNeedsUpdate...
(2/2) Refreshing PackageKit...
==> Retrieving sources...
-> Cloning yay git repo...
Cloning into bare repository '/home/killavolt/yay-git/yay'...
remote: Enumerating objects: 15396, done.
remote: Counting objects: 100% (2341/2341), done.
remote: Compressing objects: 100% (861/861), done.
remote: Total 15396 (delta 1582), reused 2179 (delta 1452), pack-reused 13055
Receiving objects: 100% (15396/15396), 8.96 MiB | 24.41 MiB/s, done.
Resolving deltas: 100% (9884/9884), done.
==> Validating source files with sha256sums...
yay ... Skipped
==> Extracting sources...
-> Creating working copy of yay git repo...
Cloning into 'yay'...
done.
Switched to a new branch 'makepkg'
==> Starting pkgver()...
==> Updated version: yay-git 12.2.0.r20.g92d7cb0-1
==> Starting build()...
bash: make: command not found
==> ERROR: A failure occurred in build().
Aborting...

[steady thistle]

Seems to be stuck on Installing requirements for Codeformer... I'll just wait but it's been awhile now.

[quick forge]

got

it

had to install make

[ornate elk]

Can take a while let it load

[steady thistle]

I figured but it's good to get confirmation that I need to just chill. heh

Heh, yep, just took more time.

[grand fable]

hi, I'm getting this error saying I can modify some environment variables. Could you tell me how to do it?

The operator 'aten::upsample_bicubic2d.out' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable PYTORCH_ENABLE_MPS_FALLBACK=1 to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.

[cosmic furnace]

mmmkay. made a clean reinstall.feels like running more smothly but... problem stays . It seems like 32gb of ram is not enough for some reason? o.0

ive looked into it & indeed it says dedicated memory used 40-41gb ish. the error says tried to allocate 47gb of memory

[quick forge]

ok im stuck at webui.sh

not sure how to run that

[cosmic furnace]

the process shows that only python itself used about all 32-34gb of ram

[quick forge]

after that finishes run webui.sh remember first gen takes a minute

[ornate elk]

What model are you using?

[cosmic furnace]

is it okay to have such a huge ram usage? my friend also noticed he had a leaps of usage to up to 48 gb ram. but he has total 64 so no crash. But for me its 2-3 generated images on some models and i get error... or 10-15. Depends on model & how much long promt is. But is seems this sdxl with 2-3 loras and long promt is huge ram eater -_- espec if i choose smthing to generate higher than 512-768

pony diffusion turbo dpo

[ornate elk]

The cause for this huge RAM usage is SDXL

[cosmic furnace]

the one is not pruned, or pruned need to look dont rememb, pretty sure dpo is not pruned one

[steady thistle]

oops, I did put it in the right spot. heh, my bad

[cosmic furnace]

yeah sdxl) 6.46 gb model)

[ornate elk]

Yea if you use 1.5 model it should not take more than 16gb

Sdxl uses 30gb for me too

[cosmic furnace]

so i shud just avoid using sdxl or there might be some tricks?))

[ornate elk]

Yea wouldnt use it, you have more fun with 1.5 models + upscaling and controlnet etc

[cosmic furnace]

hm try increasing swap file mine have only 8 gb max but not sure & blah kinda 😦

[ornate elk]

I mean sdxl still works, its just slow

[ornate elk]

Make sure the pagefile is enabled for C drive and not for any other drive

[cosmic furnace]

and i was thinking its amd on windows so much errors -_-

but its not enough ram :((

[ornate elk]

Your on Linux?

[cosmic furnace]

no win 10

[ornate elk]

Ok

[cosmic furnace]

it might help ?

will inux eat less ram? i mean sd there

[ornate elk]

Linux works much better with auto1111 for AMD

But it will also take a lot of RAM

[cosmic furnace]

i have a feeling it have more less decent speed even on widows alrdy o.0 . but yeah i see ppl very happy when linux)

[tiny abyss]

tried this, getting an error

[steady thistle]

Is SD 1.5 trained on 768x768 as well as 512?

[tiny abyss]

this

[cosmic furnace]

ive also read news that rocm soon coming to windows

[fierce hollow]

Hello, sorry for my English, I'm writing through a translator. So, the question is: What promoters/settings do I need to enter to achieve the following result? I want to work with one image, improve it, change it, and so on. The essence is this: I created a photo of a girl (selfie). It created perfectly, after which I uploaded the created photo to an img2img generator and asked to make it full-length - the result - it just starts drawing her face again. Exactly the same situation if initially to draw a girl in full height and then ask img2img to make a selfie.

[ornate elk]

If you want fast speed on 1.5 you can try the new guide for Zluda. Its experimental and sdxl won't work with it.
But much faster for 1.5

[ornate elk]

I'll take a look later, eating lunch rn

[tiny abyss]

@ornate elk

IT launched !! thanks!
But one issue, it isn't showing me Deforum

[ornate elk]

You need to update the Extensions

Then restart

If it won't show. Reinstall deforum

[cosmic furnace]

what is this zluda? o.0 i dont feel like using neither comfy or shark cuz. well they cant do stuff that automatic do?) models-loras-controlnet ?

[ornate elk]

Zluds emulates cuda for AMD.
It works with Controlnet but (depth, mlsd, IP-Adapter won't work)

And it works with models and loras

[cosmic furnace]

installing olive and miniconda felt frustrating & in the end i didnt train model & thought its not working -_- . Is itany easier now?)

I can see olive tab now in sd hmmmm)

with this zluds i can use everything of automatic111 xcept of sdxl ?o.0

so control net will have... half'ish of its features?

[ornate elk]

Yes

I tested it yesterday and today

Getting 16 it/s on my 7900xtx

[cosmic furnace]

o.0

is it easy to install-use?)

is medvram btw rly works better with 7900 xt&xtx? =\

[ornate elk]

Yes

[ornate elk]

Install is not as easy but usage is the same as the Directml

[cosmic furnace]

Where is this Zluda guide & download?)

[mighty topaz]

Guys, i need help with vae to connect to all save image, i want generate anime with vae anime because without vae anime colors are flat, so when i connect vae to first decoder firs output image have colors, but i want also the same quality colors after upscaleing and 3th step after adetailer, this work only for 1st save image, how connect vae loader with anime to all process

[muted flame]

I'm trying to install stable diffusion on my PC, but when I run webui batch files, I run into an error with it installing the torch.

I saw online that the solution is to install the torch first, so I went to pytorch to get the code.

I got: pip3 install torch torchvision torchaudio --index-url pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

How do I run this in python as I keep getting "SyntaxError: invalid syntax" when I try to run it

[ornate elk]

In the Pinned Messages of this channel.

[ornate elk]

Which guide did you followed? Do you have a screenshot of the error?

Which python version?

[maiden haven]

hi I'm doing inpainting right now on a ton of images, i've never done anything sd before. i have an rtx 4090 but it's just chilling, fans not really ocme in at all in a fe whours as far as i noticed. python processes are at 90% so something is going on. does this just not use my gpu? is something misconfigured? i'm getting26 s/it Total progress: 42%|████▎ | 68/160 [29:49<40:15, 26.25s/it]

also everything seems to haven oticed my gpu in any info/etc thing ive seen

[ornate elk]

Hey, what's in your webui-user.bat?

[maiden haven]

I'm using stability matrix to start things (its just a gui manager for sd) so it might inject these vars at another spot but the webui-user.bat is currently ```@echo off

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

call webui.bat

oh yeah i have some custom args in that startup command

[ornate elk]

Oh which ones?

[maiden haven]

--api --skip-torch-cuda-test --skiop-python-version --no-half --precision full

a few things complained trhat my torch wasnt compiled with cuda but it was and other stuff seems to work?

[muted flame]

I was following this guide up to 6:21 when my prompt is much different (see screenshots) https://www.youtube.com/watch?v=Po-ykkCLE6M

I then went to investigate how to instal the torch in advance as a work around and came upon this post https://www.reddit.com/r/StableDiffusion/comments/17cax3x/webuiuserbat_file_is_giving_me_error_code_2_it/?rdt=37958

That then took me to https://pytorch.org/get-started/locally/ where I got the string: ip3 install torch torchvision torchaudio --index-url pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

[maiden haven]

its hard to tell whatr erros are red herrings or not

[ornate elk]

yes these arguments make it run on CPU Mode
best is to not use stbility matrix and instead install the webuii the "normal" way, by following my install guide in the Pinned Messages of this Channel

[muted flame]

3.10 64 bit

[maiden haven]

dang, figures, i just kept adding commands until it started working. guess i need to dig into stability matrxies venvs

thank you

will the quality/etc change when it uses gpu too? Or just speed?

[ornate elk]

no there is no need to, you can just remove all the args and then only add --xformers --no-half-vae

and then it will work

[maiden haven]

iirc i had to take out xformers last night for some reason, ive got a job running right now but ill try again

[ornate elk]

yea try it again, and if you get errors make a clean reinstall with my guide

[ornate elk]

okay and whats your gpu ?

[fierce hollow]

Hello, sorry for my English, I'm writing through a translator. So, the question is: What promoters/settings do I need to enter to achieve the following result? I want to work with one image, improve it, change it, and so on. The essence is this: I created a photo of a girl (selfie). It created perfectly, after which I uploaded the created photo to an img2img generator and asked to make it full-length - the result - it just starts drawing her face again. Exactly the same situation if initially to draw a girl in full height and then ask img2img to make a selfie.

[ornate elk]

hey, for that you need to use Controlnet + IP-Adapter model, with this you can transfer a style or face onto the new generated image

[maiden haven]

Is there any good way to troubleshoot extensions in web-ui? I copied my exts from another install and it is a lot of extensions and now my UI is running rough

[ornate elk]

check the cmd

is the best option

[maiden haven]

yea there is a startup time somewhere thx

if i just remove the extensiosn form the dir it wont keep anytrhing about them around right, or does some config file point tot hem?

[muted flame]

amd radeon (tm) rx 570, last updated January 4th of this year

[fierce hollow]

I found information about Controlnet, I will study it, thank you

[ornate elk]

big extensions can cause issues even when deletet. because they install stuff into the venv folder. Deleting the venv after removing big extensions like Dreambooth, Deforum, AnimatedDiff etc is a good point