🤝｜tech-support

[formal field]

was running musinc in the background but stopped as it turns to noise

[ornate elk]

How much ram do you have?

If you use flux that would explain your freezing issue

[chrome rampart]

Evening all. Anybody know if/how you can use flux controlnet options like pose/reference in Forge?

[formal field]

32/12

i use sd 3.5

[stone coral]

does someone know how i can fix this issue? trying to install searge llm plugin on comfy

try fix has not helped

or if there is any way to manually install it?

[ornate elk]

how large is the model file?

[primal pendant]

@ornate elkhey i'

back

so any ideas since last night

i did some googling last night after u went offline seems alot had that issue around 14th dec

[ornate elk]

ah okay, have you tried installing auto1111 ? to see if that works?

[primal pendant]

yeah doing it now

not been in for long

huh

i deleted venv and ,zluda letting it try again

or was that config issue ?

[formal field]

tried several between i think 11g and 3 or so

[ornate elk]

would need to see the full cmd log

[ornate elk]

does normal sdxl models work?
3.5 and flux are likely to much for your gpu.
So maybe only nf4 or Q8-Q4 ggufs can work

[primal pendant]

yeah can't now it's rebulding lol

no idea you said they shoudl

hmmm

[ornate elk]

looks okay and should work, but this line could be the indicator why forge wont work:

[ornate elk]

something is wrong with the paths to rocm

[primal pendant]

huh how ?

[ornate elk]

idk

but the cmd says it cant find rocm

so its an environment path issue i guess

[primal pendant]

so?

[ornate elk]

can you also show the rocm paths

this

[primal pendant]

how do i extend it so shows?

[ornate elk]

?

[ornate elk]

not there

thats ok

go one step back

[ornate elk]

looks okay

[primal pendant]

ok so what to check for next ?

[ornate elk]

delete the venv folder and relaunch again

[primal pendant]

shouldn't i let it finish compiling first ?

or just close

[ornate elk]

you said you installed auto clean, but you have very old extensions in there??

lycoris is outdated for over a year

[primal pendant]

hmmmm

i will remove again

i'm just gonna delete everythign to do with any evrsion of sd now then

as i had them all

[ornate elk]

save your stuff before

like models, outputs, loras etc

[primal pendant]

to late just deleted all

i didn't have much

like only 3 checkpoints and about 10 lora

as i had already deleted lots of crap before space is precious

[ornate elk]

wow

xD

[formal field]

i'm grabbing fp8 cp to see if flux runs

[ornate elk]

also use the t5_fp8 text encoder

not the fp16

[formal field]

flux dev runs like a chrm, lools betetr too apparently lol

just trying to get sd3.5 to spell a word right was painful

[primal pendant]

still does it even with clean

[ornate elk]

a rocm hip sdk reinstall maybe fixes it

but you would also need to replace the library files again, so better copy the library folders and the library.dll

[primal pendant]

ok can i just install over current one ?

[ornate elk]

no better uninstall and then reboot and then install it again replace the library files again and then reboot again

[primal pendant]

pain

😴

ok guess i will give ita try

[ornate elk]

yep save these

[proud terrace]

will try this, thank you!

[primal pendant]

right i uninstalled all rebooted installed all added bk up rebooted

now what

[ornate elk]

now you can delete the .zluda folder and the venv and relaunch

[primal pendant]

deleted venv yeah t=did that to

ah yes ofc

[ornate elk]

here everything looks correct

[primal pendant]

even with rocm runtime not found ?

[ornate elk]

no but i dont see that in the first screenshot

[primal pendant]

oh

so any ideas ?

[ornate elk]

need the full cmd log after a relaunch

[primal pendant]

oh relauch

oh

ah that's jsut checkpoint

[ornate elk]

yep

[primal pendant]

who the hell uses sd in 2025 lol

[ornate elk]

so i dont know why your rocm path is broken. as i never saw anyone with that issue.
But the webui should still work as the dev made it so it uses the base rocm path

[primal pendant]

ok so it's good now then ?

[ornate elk]

its ok, good would be if everything would be normal xD

but yea it should work.

[primal pendant]

we compiling

[ornate elk]

what you then can try is to install forge webui later too if auto1111 works

[ornate elk]

this can be ignored

its a normal error for auto111

[primal pendant]

hahaha yeah maybe if that help me run xl

[ornate elk]

what was your gpu?

[primal pendant]

6700 xt

[ornate elk]

yep that can run it

[primal pendant]

any tweak needed ?

[ornate elk]

you have 32gb ram right?

[primal pendant]

like adding to aug ?

in system yeah but not on gpu no

[ornate elk]

if sdxl is slow you can add --medvram-sdxl to the args
thats all

and for upscaling tiled vae extension is needed

[primal pendant]

oh i have to have that ?

[ornate elk]

yep you can install it later trough the extensions tab

but first try to generate the first image

as the compiling takes a while

[primal pendant]

yeah i will get rid of the sd afterwards

[ornate elk]

yea the 1.5 checkpoint can be deletet after that xD

[primal pendant]

so u reckon the 6.2 about 20% better than 5.3 or 5.2 can't remember what i had now

yeah gonna be long time yet lol

that wasn't to long compared to before

[ornate elk]

6.1 and 6.2 are more stable than 5.7

and longer supported

[primal pendant]

https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

is this the one ?

[ornate elk]

yes

[primal pendant]

ok nice what else i need?

[ornate elk]

boorutag autocompletion is very handy for anime /pony/illustrious) models

and adetailer is a must have

[primal pendant]

https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

this

oh i still need adetailer for xl's ?

[ornate elk]

Can help yes

[primal pendant]

i thought was more sd 1.5 thing

ok will add to

[ornate elk]

Its always an improvement so doesn't hurt to have it if needed

[primal pendant]

ok cool so now with any luck i can run xl

🙏

ok got this as my chekpoint good ?

😭

[ornate elk]

Can you show the full cmd log pls

Also the settings you used like resolution steps etx

[primal pendant]

full log now as for what i tried it was very basic

lora:Lucy_Heartfilia_-_FairyTail1stSeason_-_IllustriousXL:1lucy_heartfilia(ft), 1girl, blonde hair, cleavage, solo, large breasts, dress, open mouth, pointy ears, closed eyes, side ponytail, purple dress, bare shoulders, ribbon, red dress, choker, couch

worst quality, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, jpeg artifacts, furry, ambiguous form, feral, zoom layer, clone,twin,multiple heads, sound effects, selfcest, sketch

it was the basic promt that was already there

i will try another

[ornate elk]

Can you relaunch the webui and then try sdxl without a lora?

[primal pendant]

ok

[ornate elk]

You forgot to restart the webui after you installed the extensions

Thats always needed

[primal pendant]

hhahahhaha

opps

i had few drinks earlier

yay

sucess nice thank you

[ornate elk]

Perfect, no problem 🙂

[primal pendant]

now do i dare use forge again lol

[ornate elk]

You can try set it up

[primal pendant]

maybe tomorrow

[untold mauve]

hi i have rocm 6.1, should i update to 6.2.4 and update zluda? there are better performance or same? thanks ❤️

[primal pendant]

for me it really was yeah but i wasn't 6.1 i was 5 something

[ashen gate]

oh i have controlnet models for sdxl loaded

[proud terrace]

hello.
I am a bit confused... I am running Stable Diffusions "WebUI" but now everybody is talking about Forge.
But when I google tutorials it looks like WebUI and has the same installation routine.
Can I use models like Flux Schnell now on WebUI? or do I somehow update to this "Forge WebUI"?

[ruby yoke]

forge and auto1111/webui are two different products. and you'd probably be happier running ComfyUI inside Swarm

[proud terrace]

ok, then I ll spare me the trouble and try ComfyUI. thank you

[ruby yoke]

you're welcome. consider grabbing a copy of Swarm and letting it do the techical stuff. we have a #🐝｜swarm-ui channel, for help if you get stuck

[proud terrace]

will do

[ornate elk]

Feel free so screenshot the whole webui when you get the issue

[ornate elk]

Forge can run flux, yep.
If your GPU has low vram then forge can increase the speed.
If you have 12gb or more vram then you can try out flux on it.

[little bough]

HI, any idea how to resolve this?

[flat yarrow]

This is tying me up in knots as I can't seem to work it out, but I think am halfway there. Am using ComfyUI and the Impact Pack and Face Detailer, when I generate may main image the face features are OK but when I pass it to the Face Detailer node the face gets scubbed out (messed up) and also added painted lines like hand drawn. I think its something to do with the Checkpoints as when I use a Checkpoint with a VAE it works fine, when when I use one without a VAE and use the standard SDXL VAE that's when things get messed up, I've tried various settings but I think if it works on some checkpoint and not others. My logic it must be an issue around the Checkpoint and VAE. Does any one know of this issue and what I can do to get the other Checkpoints to work, any guidance would be great Thank You!

[warm turret]

@ornate elk ı got issue while installing rocm in ubuntu

[ornate elk]

Oh your on Linux?

[warm turret]

no

windows

using linux subsysten

m

ım getting several errors all time

now this

[ornate elk]

Why not just following my guide xD

Can't help with WSL

[warm turret]

ı was following

wait

how to check rocm is installed

[ornate elk]

If you see it installed in the system control

Or just reinstall it and then its installed

What's your GPU?

[warm turret]

6700xt

[ornate elk]

That works with the latest hip SDK 6.2.4

And the latest zluda version

[warm turret]

its installing some stuff now

but ı got this error ZLUDA device failed to pass basic operation test: index=None, device_name=AMD Radeon RX 6700 XT [ZLUDA]

@ornate elk

[warm turret]

ı have installed stable difusion

but got this issue RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

when loading the model

[midnight roost]

hi i have found out if i restart my pc or disable/enable gpu the generation speeds doesn't slow down, so that must mean its a vram issue is there a way to clear it with stable diffusion?

[ornate elk]

Can you show the full cmd log?

Have you replaced the correct library files for your GPU (gfx1031) ?

[warm turret]

Stable diffusion model failed to load

[ornate elk]

Not in Auto1111 or Forge but in Comfyui there is a node for that

[warm turret]

its still using the cpu

@ornate elk

File "C:\Users\user\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\torch\nn\modules\module.py", line 1532, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\user\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\torch\nn\modules\module.py", line 1541, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\user\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\torch\nn\modules\sparse.py", line 163, in forward
return F.embedding(
File "C:\Users\user\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\torch\nn\functional.py", line 2264, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument index in method wrapper_CUDA__index_select)

[ornate elk]

Show me the full cmd log

Restart the PC.
Then delete the venv and the .zluda folder and relaunch.

[regal mason]

I'm trying to write captions for some images for LoRA training.

What is the best way to describe an image of a person whose face is at an angle? Like not a side profile, but not a front portrait either. Somewhere in between?

Like this angle:

[ornate elk]

Looks okay. Relaunch the webui-user.bat
And then try to generate an image with an 1.5 based model

[warm turret]

RuntimeError: invalid argument to memory_allocated

@ornate elk this error appears

[shut geyser]

launched fooocus after a while yes i disabled the fooocus v2 thingy
and it still crashes
upon generation

[ornate elk]

how much ram do you have?

[warm turret]

16

there was an issue with cuda and pytorch version

ım fixing it now

then ı ll try again

[knotty turret]

@ornate elk Hi again forgot to talk with u again lol

.

[ornate elk]

then you need to increase the windows pagefile. set it to 16000min and 24000max for the C drive.
ANd disable it for any other drive. then restart the PC.
Here is a guide: https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html

[knotty turret]

what should i do i really want to use comfyui

[tawny saddle]

i say

download it

and give it a go

:3

[warm turret]

done

still same error RuntimeError: invalid argument to memory_allocated

import torch

print(torch.cuda.is_available())
False
print(torch.cuda.device_count())
0

[ornate elk]

can you relaunch and try to generate a 512x512 image with an 1.5 (2gb) model?
then send me the cmd log

[warm turret]

@ornate elk

[ornate elk]

It can't load the model properly

What's the model you try to use?

[warm turret]

default one

v1-5-pruned-emaonly.safetensors

and dreamshaper xl loaded

[ornate elk]

Delete the default one and try this:
https://civitai.com/models/4384/dreamshaper

Thats the 1.5 variant

[warm turret]

le "C:\Users\user\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\torch\nn\functional.py", line 2264, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument index in method wrapper_CUDA__index_select)

this error appearing

[knotty turret]

@ornate elk idk if u can u help me now i just ping you to remember

[ornate elk]

Also make sure your GPU driver is updated

[warm turret]

it is

model loaded

RuntimeError: invalid argument to memory_allocated

this error still appears

CS1o
on Jul 4, 2024
Hey, your missing the steps where you have to replace the rocm library files with custom ones.
Follow my AMD Automatic1111 with Zluda install guide from here:
https:/github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

can u tutorial this step

[ornate elk]

Its the step where you Download the correct gfx library files for your gpu

[warm turret]

realized ı have skipped one step on ur zluda tutorial

fixed it now

[ornate elk]

So it works?

[warm turret]

generating rn

Failed to create model quickly; will retry using slow method.
Running on local URL: http://127.0.0.1:7860

To create a public link, set share=True in launch().
Startup time: 150.8s (prepare environment: 13.1s, initialize shared: 140.8s, other imports: 0.4s, load scripts: 0.5s, create ui: 0.7s, gradio launch: 0.3s).
Applying attention optimization: Doggettx... done.
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...
Compilation is in progress. Please wait...

[ornate elk]

Compiling?

Ah perfect

Which step did you skipped?

[warm turret]

installing gfx1031

ım trying to install this stable difusion stuff for 2 days

really tiring

0%| | 0/20 [00:00<?, ?it/s]Compilation is in progress. Please wait...
Compilation is in progress. Please wait...

is this the one gonna take 15 20 minutes

[ornate elk]

Yep but I asked about the gfx step a few hours ago and you didnt answered my question xD

[ornate elk]

The compiling to launch is shorter than the image gen.
So if the webui opens in browser you just have to type
Cute Cat into the prompt and hit generate.
Then it will compile again.
And this compilation can take 15-40 mins.
But mostly its shorter

Then just wait until you get the image

[warm turret]

is in progress. Please wait...
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [02:43<00:00, 8.20s/it]
Compilation is in progress. Please wait...██████████████████████████████████████▍ | 18/20 [00:00<00:00, 20.95it/s]

Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:17<00:00, 20.95it/s]

finally it used the gpu at 99 percent

got the image

[ornate elk]

Nice

Then your done

The next image should take seconds

[warm turret]

yes

creating something in 20 sampling step and 512 to 512 takes 3 or 5 seconds

it still long tho

thank u btw

[ornate elk]

1.5 is a bit bugged in the display of how long it takes

You can try sdxl models too now.

[warm turret]

which sampling method should ı use on dreamshaper xl

to create some portrait or realistic visuals

[ornate elk]

DPM++ 2m or Euler a are good overall

For realism I would use juggernaut XL or realvisxl

[warm turret]

btw ı have installed linux subsystem docker and other unrelated stuff to download these

ı really dont understand software things not even surface level

thats something horribly boring

[ornate elk]

You can uninstall that stuff.
To do it correctly check this:
https://itsfoss.com/uninstall-wsl/

[warm turret]

much appreciated bro

[ornate elk]

No problem

You only need to have:
Python 3.10.11
Git
HIP SDK (ROCm)
For the webui to work

[warm turret]

no need to cuda right

ı have installed that also as chatgpt told me

and do u have command line argument recommendation for optimization

[ornate elk]

If sdxl is to slow you can try add --medvram-sdxl
and --opt-sub-quad-attention

[warm turret]

can ı use upcast sampling instead of medvram

ı have 12gb vram

[ornate elk]

Upcast is not needed

And --medvram too

Only the ones I mentioned can help

[warm turret]

thx

[ornate elk]

No problem.
And if you want to upscale images with hires fix.
Make sure to First install the Tiled Diffusion Extension

And then only enable Tiled VAE

[primal pendant]

forge working no issue

that fix yesterday did it all

so thanks

[ornate elk]

Perfect, no problem 🙂

[primal pendant]

should i keep on normal for 12 gb ?

[ornate elk]

Yep

[primal pendant]

it took about 3 mins for one xl

[knotty turret]

mine takes that time too

[primal pendant]

6700 xt ?

[ornate elk]

Which resolution?

[knotty turret]

6600

[ornate elk]

And steps and sampler

[primal pendant]

1024 x 1024

20 steps

no sampler

[ornate elk]

No sampler is not possible

xD

[primal pendant]

well default then

[ornate elk]

Okay

[primal pendant]

dpm 2m

[ornate elk]

DPM++ 2M and Euler a are should be preferred.
DPM++ SDE can take double the time

[primal pendant]

ok will try euler a

[ornate elk]

You can also try 768x1024 as resolution

Thats faster for sure. Good for portrait and landscape if switched

[primal pendant]

ok thanks will do i'm just happy it's working

all that pain and suffering paid off

[knotty turret]

make forge work was easy to work comfyui that not helping me

[ornate elk]

We can take a look at it again

Are you on rocm 6.1 or 6.2 ?

[knotty turret]

https://tenor.com/view/crying-the-office-sad-steve-carrel-gif-13099504922146745305

alright

[knotty turret]

6.1.2

[ornate elk]

Okay

And you reinstalled comfyui?

[knotty turret]

Yep

.

[ornate elk]

And you also redownloaded my comfyui start bat?

Just to be sure its not an old version

[knotty turret]

i dont remember

i can try it now

[ornate elk]

Yep, delete the venv before you relaunch it with the new bat

[knotty turret]

ok

[knotty turret]

nothing new

The cmd just install the things

and then close

[ornate elk]

It closes itself?

[knotty turret]

yeah

dont need do nothing

it just close

wanna see record?

[ornate elk]

Maybe you can see what the console shows before it closes?

[knotty turret]

the same

[ornate elk]

Hmm.
What's your python version?

[midnight kiln]

I really need some guidance on where to go in terms of GUI interface. I made myself familiar with automatic1111 first - but it seems like Forge and reForge both are more advanced in various ways. I am thoroughly confused about which to focus on though. Forged seems to be the only one supporting Flux so far - which just seems too good to ignore after trying it.
I have both working in stability matrix - but the main problem comes from all 3 GUIs linking to the same extension library, and apparently very little guidance on what is and isn't compatible.
Or should I be jumping straight to comfy? Halp plox!

[ornate elk]

It reminds me of this error that I couldn't get fixed:
https://github.com/CS1o/Stable-Diffusion-Info/issues/15

[ornate elk]

Hey, Forge development is currently really slow and Auto1111 too.
Auto1111 supports the most extensions. And reforge should work with the same ones too.
Forge works with less extensions but has some own good ones.

If you want new features and active development you should try reforge.
I think it has flux support too but I'm not sure

[midnight kiln]

Your opinion means a lot to me - thanks. Maybe I was wrong about reforge and flux support. I will delve a little deeper.

[knotty turret]

but the cmd different from mine

[noble idol]

hey did you ever get your UI to interrupt faster

[midnight kiln]

Anyone know where to source VAE / CLIP / txt encoder for Flux models? I can't understand why this isn't even mentioned anywhere on the models themselves

[ruby yoke]

https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

[meager quiver]

Sometimes when I'm generating images in Forge my laptop freezes for like 30 minutes before it unfreeze, not sure why is it using too much CPU?

[midnight kiln]

Thanks wiz. I think I actually got all 3 into Forge now. I can just baaaarely squeeze it in on 11GB on some lower resolutions 😄

[midnight kiln]

VRAM overflow? If you're swapping tons of data between VRAM and RAM it can east up a lot of CPU and hog resources

[ruby yoke]

ouch. have you thought of using the ggufs?

https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

[midnight kiln]

I have tested two - and yea - they work much better. I am mostly just testing right now

but damn... Flux is pretty impressive

[ruby yoke]

they're going to work much better in that small amount of vram you have.

[midnight kiln]

Can flux actually not handle 512x512? too low?

[ruby yoke]

it was trained on 1024x1024 - so why would you do 512x512?

[midnight kiln]

well that's why I asked - I didn't know 🙂 just used to falling back to it as a "safemode" for testing

but here it clearly didnt like that

wish there were more NF4's available. They measure up damn well considering the size difference

[thorny iron]

Is Forge still okay to use? It officially turned 1 years old since three days ago. ';d

I'm wondering if during that time major leaps and bounds have been made to AI.

I want to use a good solid interface, but Comfy looks too difficult for me.

Or is Forge literally just a UI and that's it, and all the important things that I'm hoping to not miss out of actually lives under the hood? Assuming that's what the 3GB something file is when first setting up?

.. and which would I download? ':<

[ruby yoke]

forge, auto1111, and comfyUI are literally just interfaces

[thorny iron]

I see, thank you!

Which one would be the one I want to download? :<

[ruby yoke]

i suggest that you start by watching this set of tutorials https://www.youtube.com/watch?v=AbB33AxrcZo&list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x

[thorny iron]

Oh no not Comfy :'c

[tawny bramble]

hey yall, is there any way i can speed up my its? im using sd forge amd, on a 5700xt, flux1-dev-q4 and im getting 70s/it. any help would be appreciated!

[ruby yoke]

other than the fact that it works better than auto1111 or forge, and has support for everything, including all the new stuff, where they don't - why don't you like comfy?

[noble idol]

same reason most people dont like comfy

its daunting, a lot of work, very confusing, especially to new users, even with guides

it is good tho, i like it, i just dont use it much, i still prefer forge's layout

[ruby yoke]

everything is easy once you know what you're doing. everythign is hard when you first encounter it. just install @jagged plover 's really good program, Swarm https://github.com/mcmonkeyprojects/SwarmUI and let it do all the technical stuff. it runs comfy for you

[noble idol]

tried swarm already

still liked the layout of forge better so i dont use it

[ornate elk]

Double click the interrupt button.
If you want it even faster enable never oom for tiled VAE.

[ornate elk]

The middle one

[ornate elk]

Do you use zluda or directml?

[tawny bramble]

zluda

[ornate elk]

Okay. Make sure you also use the t5_q8 Text Encoder.
And if you only have 16gb of RAM then you need to increase the Windows Pagefile

[noble idol]

i just use this setting

[tawny bramble]

yaya already using q8 and i have 48 gigs of ram, i researched a bit and im guessing there arent many ways to speed it up further

[ornate elk]

Nope. Then your out of luck.
But its not worth using it on that card.
You should focus on SDXL based stuff

[tawny bramble]

alr ty for the help anyway!

[ornate elk]

Np!

[midnight kiln]

I'm testing out Flux NF4 models now (the smallest type of Flux), and I can say it works well inside of 11GB on my 1080Ti - but only barely at a reasonable resolution.
So sadly, I don't think that GPU can do it. You will need 11-12GB minimum to get any sort of speed - or otherwise be running it so much in RAM that 70sec/it doesn't seem like it would be abnormal. RAM is easily an order of magnitude slower than VRAM

[cursive wren]

Stable diffusion 2.1 simply stopped working out of nowhere giving this error, help!

[lucid yoke]

Probably on their end, most you could personally do is clearing your cache/cookies

[cursive wren]

Thanks but its still not working

I like this AI cause it gives that imperfect look, also i cant afford paid AI's, for my work being imperfect its actually good

[midnight kiln]

Is there any queuing extension that works for Forge? my usual webui agent scheduler just errors out on Forge. I wish I had some small idea of which ones actually works with Forge and reForge short of just trying everything

[chrome rampart]

Hello all. Any idea why my comfy suddenly does not start anymore? I start run_nvidia_gpu.bat and console opens for moment runs some lines and poof closes again without starting comfy

[ruby yoke]

when did you reboot last and how much space is on your system drive?

[chrome rampart]

i started this afternnon and worked. When i retried let say hour ago I had issue. I also restarted entire PC just to be sure. Space on drive free is 174GB

[ruby yoke]

on the sytem drive? are you on windows or mac?

[chrome rampart]

I use windows on (system drive c: 40 gb free and on install drive comfy (f) 174GB free

[ruby yoke]

40 gig free on your system drive isn't enough - that's going to cause you problems with all sorts of things

you also might need to clear your python cache

[chrome rampart]

how do i clear python cache ?

[ruby yoke]

#💬｜general-chat message follow @ornate elk instructions and also #💬｜general-chat message

[chrome rampart]

ok i will check thks

it says 0 files in cache but will do steps again to be sure> thks

[safe inlet]

but thats not a steam link, its to a different site you just obfuscated it, how is that possible, maybe you made a typo

[tidal thunder]

Hello all! I need some help please, can't fix by running run.bat (stable diffusion forge)

getting this error:
Already up to date.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: f2.0.1v1.10.1-previous-642-gf3672ffb
Commit hash: f3672ffbbe181bd95033283562598574fb377dcd
Traceback (most recent call last):
File "D:\textnewforge\webui\launch.py", line 54, in <module>
main()
File "D:\textnewforge\webui\launch.py", line 42, in main
prepare_environment()
File "D:\textnewforge\webui\modules\launch_utils.py", line 434, in prepare_environment
raise RuntimeError(
RuntimeError: Your device does not support the current version of Torch/CUDA! Consider download another version:
https://github.com/lllyasviel/stable-diffusion-webui-forge/releases/tag/latest
Для продолжения нажмите любую клавишу . . .

I have installed MSVC v143 build tools but this didn't help too

[ornate elk]

Hey, what's your GPU?

[tidal thunder]

RTX 2070 super 8 gb

[ornate elk]

And which version of forge did you downloaded,?

[tidal thunder]

this

should i download another version?

[ornate elk]

No thats the recommended one and should work

But you need to update your nvidia driver

[tidal thunder]

it's 572.16 (latest)

[ornate elk]

Hmm okay

[tidal thunder]

previously i ran stable diffusion forge thru the webui-user.bat but now i realize i did all this time wrong

so i want to run it thru run.bat

[ornate elk]

Does it work if you launch it trough the webui-user.bat?

Because then you can ignore the run.bat

[tidal thunder]

Yes, it start's to download all packages

[tidal thunder]

it seems to be not calling enviroments so that could be bad?

when running webui-user.bat

[ornate elk]

Have you launched update.bat before running run.bat?

[tidal thunder]

Yes

[ornate elk]

Do you have a specific cuda Version installed?

[tidal thunder]

yes, CUDA 12.1

[ornate elk]

You can uninstall that and then reboot the PC.
Then update your python to 3.10.11 64bit
Then delete the venv folder inside the webui folder and relaunch the run.bat

[tidal thunder]

i'll try rn, thanks

[chrome rampart]

After calling to the gods several times (lol) trying to find issue i decided to do full re-install and now it works again. Thank for helping yesterday🫂

[tidal thunder]

didn't help

[ornate elk]

Still the same error?

[tidal thunder]

yes

[ornate elk]

What's in side your webui-user.bat?

[tidal thunder]

@echo off

set CUDA_VISIBLE_DEVICES=0
set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --cuda-stream --lowvram --cuda-malloc --always-offload-from-vram --opt-sub-quad-attention --opt-sdp-attention --theme dark --device-id 0
set SAFETENSORS_FAST_GPU=1

git pull

call webui.bat

[ornate elk]

Are you on a laptop?

[tidal thunder]

no, pc

the main problem is in environment.bat, when i'm add it to webui-user.bat it produces error

[ornate elk]

OK then remove --device-id0, --lowvram --always-offload-from-vram --opt-sub-quad-attention --opt-sdp-attention

Then save and relaunch

[ornate elk]

Yea that bat won't work

Is not needed

[tidal thunder]

so is that fine to load just webui-user.bat instead or run.bat? it won't call an enviroment (i can set in manually) so is that would be okay?

[ornate elk]

Yep and you dont need to set in manually.
If you have a venv folder inside the webui folder your fine

[tidal thunder]

okay, thanks for the help!

[ornate elk]

And dont forget to remove the args I mentioned

Also you can try ReForge if you want a more stable version.
As forge doesn't get updated lately

[steady thistle]

"ufunc_add_CUDA" not implemented for 'Float8_e4m3fn' Error... hmm

[amber prairie]

Yo if anyone can tell me the BEST possible way to make this type of background visuals, it would be greatly appreciated.

I've spent WEEKS trying to use comfyui and several different ai models, following youtube videos and shit, and bc its my first time, I don't really understand much but I have slowly gotten to grasps with the different nodes and models.

So these three are the best 3 videos out of the dozen that I've prompted and they were made with "HunyuanGGUF" which took like 15/20 minutes a clip. I don't got time for that if and its not even that technical/complicated.

i've tried other stuff like nvidia cosmos, took 50 minutes to generate an "ethereal" looking style which i did not want, ltx didnt work, I used the huggingface for EasyAnimate which made a more sketched style which looked nice but it took like 20 minutes queue to even process it.

[lucid yoke]

HunyuanGGUF is the best one for now 👍

LTX gives lesser quality but with the benefit of image2video

and is faster

and if it didnt work (witch it did on my 3070TI) is probably a workflow fault

[regal mason]

How do I remove glasses from a face?

I tried Lama object remover but it keeps giving me errors.

Anyone got a simple/minimal workflow for this?

[warm turret]

ı havent installed anything for optimizations

just models

--use-zluda --medvram-sdxl --opt-sub-quad-attention

commandlines

[ornate elk]

Screenshot your webui when you get the error

With resolution and model and if you have hires fix enabled

[ornate elk]

Set the resolution to 768x1024
And then enable never oom but only for tiled vae

And set your GPU weight slider to 11500

[warm turret]

crashed

[warm turret]

x = block(x, context=context[i], transformer_options=transformer_options)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\nn\modules\module.py", line 1532, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\nn\modules\module.py", line 1541, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\backend\nn\unet.py", line 181, in forward
return checkpoint(self._forward, (x, context, transformer_options), None, self.checkpoint)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\backend\nn\unet.py", line 12, in checkpoint
return f(*args)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\backend\nn\unet.py", line 235, in _forward
n = self.attn1(n, context=context_attn1, value=value_attn1, transformer_options=extra_options)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\nn\modules\module.py", line 1532, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\venv\lib\site-packages\torch\nn\modules\module.py", line 1541, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\backend\nn\unet.py", line 154, in forward
out = attention_function(q, k, v, self.heads, mask)
File "C:\Users\user\SD - Zluda\stable-diffusion-webui-amdgpu-forge\backend\attention.py", line 335, in attention_pytorch
out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 5.62 GiB. GPU
CUDA out of memory. Tried to allocate 5.62 GiB. GPU

[ornate elk]

Restart the webui and make sure the GPU weight is at 11500

Then try again

And dont have anything open in the background like games or Wallpaper engine

[warm turret]

do u recommend flux

[ornate elk]

Not for your GPU

Its very slow and requires a lot of vram even on my card

[warm turret]

out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 5.62 GiB. GPU
CUDA out of memory. Tried to allocate 5.62 GiB. GPU

[Low GPU VRAM Warning] Your current GPU free memory is 831.04 MB for this diffusion iteration.

[warm turret]

get this on the first step of generating

ı turned off hires fix

generating took 10 or 15 seconds

seems fixed

ı guess its not good to use hires fix on realvisxl

[ornate elk]

Hires fix is very resource hungry but with 768x1024 and 10 hires steps + never oom for tiled VAE it should work.

[warm turret]

ı turned off just upscaling

now

[ornate elk]

Ok

[warm turret]

generating took long again

can u recommend upsclaer

good as esrgan4x maybe less but more optimized

[ornate elk]

The upscalers aren't much different in performance

For realism 4x UltraSharp is good

Can be found here:
https://openmodeldb.info/
The .PTH file has to go into models/ESRGAN folder

[warm turret]

image got cooked when ı just turned off upscaler

[sand pasture]

helo

[lucid yoke]

Are you on a laptop?

[cold junco]

yes

[lucid yoke]

Just to be sure since i dont know french, 421mo = 421mb vram

Right?

[cold junco]

i think

wait 2sec i switch to en

[lucid yoke]

No worries, ill compare it sec

[lucid yoke]

To be completely honest im not sure if you can run it locally

[cold junco]

oh 😦

how can I try?

[lucid yoke]

Well if you insist on trying its going to be very slow or run out of memory while loading it

[cold junco]

yea it cost nothing to try

[lucid yoke]

im looking at the specifics for a AMD card but zluda is not supported for your gpu and DirectML could but it would use a lot of vram

[cold junco]

okay so its not possible

[ruby yoke]

you need to actually read through the AMD GPU guides that @ornate elk wrote and which are in messages pinned in this channel

and then ask him if you get stuck once he's back online

[lucid yoke]

not with directML or zluda since they really want 8gb minimum but i recommend civitAI for online generation

[ruby yoke]

mage.space is a lot better than civit...

[lucid yoke]

he has less then 500mb on his gpu

[lucid yoke]

ah didnt know that one yet. checking it out for future recommendations

[ruby yoke]

he probably wants to rent gpus on vast.ai or runpod then

[cold junco]

okay thanks

[zenith seal]

hi i have this issue while instaling forge

[lucid yoke]

can you try fixing this one? launch.py: error: unrecognized arguments: --skip-ort

[zenith seal]

i just did

let me see

[zenith seal]

the ui lauched but i got this issue when genning

also i don't see the zluda folder

[lucid yoke]

im not too familliar with zluda issues. sorry.

[zenith seal]

alright it's okey thanks tho

[cold junco]

Bruh it rebooted my computer

Ok I think my computer doesn't want to generate images 😂

[ruby yoke]

if you only have 500meg of vram, like he said, no - your computer cant' do this work. you can rent gpus from vast.ai or runpod though

[zenith seal]

maybe the issue is from zluda

[ornate elk]

You can only use the very slow CPU mode on this laptop

[ornate elk]

Yep thats an issue
Your missing the path to hip rocm in environment variable's

[cold junco]

Is it on your GitHub profile?

[ornate elk]

Nope but you can follow the one for Automatic1111 on Nvidia.
And then use --use-cpu all --precision full --no-half --opt-sub-quad-attention
As commandline_args instead of the one for nvidia

[cold junco]

Okay thanks I'll try that tomorrow

[zenith seal]

but i have the path

[ornate elk]

Which guide did you followed?
It seems like you installed the webui version from likelovewant.

Would advise deleting it and following my install guide instead.
Its linked in the pinned messages of this channel

[sand pasture]

How would you train a lora for a character with a 3d model?

in comfyui

is there some set up/json I could use for this?

[sand pasture]

nvm I found something

[hazy pumice]

Hello~
I have been attempting to install and run SD Auto1111 for the past 24 hours and I have tried everything and checked all kinds of other forums and even asked GPT for help. Nothing has worked and I'm about to pull my hair out. Can anyone help me understand what is going wrong?

Please and thank you~

[ruby yoke]

what sort of hardware are you trying to install on?

[hazy pumice]

Nvidia 3070 Laptop GPU

I have not

[ruby yoke]

try that or try installing SwarmUI and let it deal with the technical stuff

[hazy pumice]

okay let me go check SwarmUI, be back soon

[ruby yoke]

there's a #🐝｜swarm-ui channel here if you get stuck

[hazy pumice]

so I just run the code in cmd in the folder?

[ruby yoke]

no. here:

https://github.com/mcmonkeyprojects/SwarmUI go there, scroll down below all the code and folders, and read through the instructions

[hazy pumice]

or download zip and then extract to where?

[hazy pumice]

alright let me go read~

[ruby yoke]

there's a one click installer you download

[hazy pumice]

it says I shouldn't install it under program files, what would be the ideal folder?

[keen gorge]

guy why does my genrate looks low quality how to improve it

ping me if you know how to help me

[hazy pumice]

I'm back again...I started from scratch trying to install and run the Automatic1111 webui. Can anyone help me understand and resolve the following errors below? It seems to be a connection issue? My internet isn't the best but it's not that bad either

error: RPC failed; curl 92 HTTP/2 stream 5 was not closed cleanly: CANCEL (err 8)
error: 6925 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output

[ruby yoke]

how much disk space did you say you had?

[hazy pumice]

8GB GPU and my HD is 500GB

Simple laptop~

[ruby yoke]

but how much of that 500gig is empty?

[lucid yoke]

My swarmUI uses around 260gb (most of it are models tho) be sure you have enough space

[ruby yoke]

okay, found this on that error https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/9074

make sure you're using the latest level of auto1111, that's from 2023

the suggested fix is: export COMMANDLINE_ARGS="--no-gradio-queue"

to add that to your config file

[lucid yoke]

Wait i remember something about that, didnt a older branch link to now no longer maintained gits?

Or no longer existing*

[ruby yoke]

i dunno. possibly. auto1111 isn't all that well maintained any more

[lucid yoke]

Web-ui forge would be the better option i think in general

Especially with sdxl

[ruby yoke]

also - turn off any anti-virus applications yoiu might have running

[ruby yoke]

on a laptop with that little bit of sytem resources?

[lucid yoke]

8gb vram, why not . Might take a bit longer

My 3070ti handles it like a champ. (Desktop)

[ruby yoke]

i'm gonna suspect there's some sort of anti-virus - norton, kaspersky, something - running on that machine and killing the process

[hazy pumice]

I do have Norton installed and running, maybe thats killing it. I'll try shutting it off

[hazy pumice]

Not sure exactly but should be wayyyy more than enough. I barely have any programs/music/games etc installed

[hazy pumice]

I'll check it out thank you!

[ornate elk]

Which model did you used?

[ornate elk]

Checkout my Auto1111 install guide for nvidia.
Its on the first link in the pinned messages of this channel

And yes better disable Norton for the process

[waxen pasture]

Hey guys, how the heck do you use control net properly?
im struggling to give a face back realistic texture on InstructP2P, the eyes have different sizes, and i cant get it to generate a hecking smile

i also got some blurry hair accidentally in the result im using currently, any prompt i can do to fix it?

also, should i stick to 512 resolution or go higher?

[ornate elk]

Hey, can you show an example of what you try?

Instructp2p is not great for changing stuff

[waxen pasture]

its not?

what does control unit 1 and 2 do?

[ornate elk]

You can load up multiple Controlnet models at the same time

[waxen pasture]

is that any useful?

[ornate elk]

So for example change the pose in unit 1 with a openpose model and then in unit 2 use IP-Adapter to change the style

[waxen pasture]

ooo

what sort of language should i use?

should i use action verbs or just describe features?

[ornate elk]

Im still not sure what you try to achieve xD

Are you in txt2img or img2img ?

[waxen pasture]

img2img

im trying to use this thing as a quicker photoshop, but its taking ages!

i want to learn how to change features so i can use ai for work

im using realism engine

[ornate elk]

What's your GPU and did you used inpaint?

[waxen pasture]

yeah!

gpu?

[ornate elk]

Graphics card

[waxen pasture]

computer stuff? its powerful, maybe

its a new computer

[ornate elk]

Depending on what's your GPU. You have to adjust the webui to get the best speed

[waxen pasture]

wont that affect the results?

[ornate elk]

Nope

[waxen pasture]

no???

ok how do i adjust it??

[ornate elk]

First check how much vram your GPU has

In the taskmanager

Also do you use Automatic1111 Webui or Forge Webui?

[waxen pasture]

oh woah utilization is 91 percent

i got 23 gbs

of gpu

[waxen pasture]

err, i think AUTOMATIC1111

(https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features

[hazy pumice]

I appreciate it, actually this is the exact guide I have been following. Thank you for posting it. I honestly can't figure out why I keep getting errors but im gonna turn off Norton when I get home and try again

[ornate elk]

Then you only have to edit the webui-user.bat and at the line commandline_args=
You add --xformers
Then save and relaunch

[waxen pasture]

do i have to do that everytime?

[ornate elk]

And when you use Inpaint always
At the setting "Inpaint Mode" set it to Masked only

[ornate elk]

No only one time

[waxen pasture]

where is the webui-user bat?
you mean webui batch file?

its downloading something

Installing collected packages: mpmath, urllib3, typing-extensions, sympy, pillow, numpy, networkx, MarkupSafe, idna, fsspec, filelock, charset-normalizer, certifi, requests, jinja2, torch, torchvision

it wont open

webui-user.bat

nevermind

did it on the notepad

oh i think i have been forgetting to enable the control net unit..

i cant be spared

i wasted so much life on this idiocy

[ornate elk]

Oh xD

[waxen pasture]

Oh indeed.

for photorealism should i use pony realism or realism engine?

[waxen pasture]

its not particulary faster, maybe i did it wrong? maybe its taking longer because this time i did enable the controlnet

[ornate elk]

Yep with controlnet it takes more time

But on your GPU it should be still really fast

[waxen pasture]

thank you :D

[ornate elk]

No problem 🙂

[waxen pasture]

wait one more, so how do you prompt this asshole computer?
with orders: "Change this and that, you useless animal"
or: "Feature A, feature B, feature C"

[ruby yoke]

you mask the area you want to change, and then you just type out what you want to see there.

[waxen pasture]

photorealistic, brown hair, high detail masked area

would this do?

[ruby yoke]

you have gen fill in photoshop, right? if you want to change something in it, you select an area and then just type what you want in that area in the prompt box. it's the same idea, but instead of a dotted line around the area you selected, you'll see it covered with a color

and you want to be specific about what to put in that area. if you select someone's glasses, and you want a VR mask, you would type VR mask

or if you select the blonde hair, you would type brown hair, curly

[waxen pasture]

i just want to see if i can do it in stable diffusion

[waxen pasture]

ooooh

ok, do i take off the photorealistic?

im selecting only the hair

[ruby yoke]

that IS what you'd do in stable diffusion. what you're doing with genfill in photoshop, is the same thing you do in stable. it's just that in photoshop you;re going to see the marching ants around the area and in stable you will see the area covered in a color

[waxen pasture]

im using cany since changecolor didnt work

[waxen pasture]

i see

[ruby yoke]

the terms: photoreal, photorealistic, photorealism - are painting terms. they are used to describe paintings that look like photos. but the AI's all know they are associated with paintings and will give you painted looks. don't use them. use photo, photograph, or photographic

[waxen pasture]

ooooo

[ruby yoke]

and, personally, i'd be using photoshop. it's faster, it's got a much better way of selecting, and it's designed for what you want to do. stable diffusion is klunky

plus photoshop has all the other tools you might need to use such as the clone tool - and you're not doing that with stable

[waxen pasture]

i dont got photoshop

its a paid service.

[ornate elk]

Krita-ai diffusion might fit more
Its like Photoshop with ai in the background

https://github.com/Acly/krita-ai-diffusion

[waxen pasture]

anyway, i just want to see if i can do it

it might come handy in the future

i also want to experiment changing face features

[ruby yoke]

yeah, why don't you try krita?

it's free, it functions as a very good photomanipulation program, a paint program, and it runs stable diffusion inside it

[waxen pasture]

would it generate assets like stable diffusion?

[ruby yoke]

https://krita.org/en/ that's where you can get the main program

[waxen pasture]

im trying to use it for photorealism also

[ruby yoke]

grab krita and krita-ai-diffusion and play around with them

[waxen pasture]

why not

idve rathered to learn to use stable diffusion though, i'll see

If you haven't yet, go and install Krita! Required version: 5.2.0 or newer
Download the plugin.
Start Krita and install the plugin via Tools ▸ Scripts ▸ Import Python Plugin from File...
Point it to the ZIP archive you downloaded in the previous step.
⚠ This will delete any previous install of the plugin. If you are updating from 1.14 or older please read updating to a new version.
Check Krita's official documentation for more options.
Restart Krita and create a new document or open an existing image.
To show the plugin docker: Settings ‣ Dockers ‣ AI Image Generation.
In the plugin docker, click "Configure" to start local server installation or connect.

cant find it in the dockers

em dockers are blocked

oh need a picture for that

doesnt appear in the docker still

Enable the Plugin Docker from the Menu bar: Settings ▸ Dockers ▸ 🗹 AI Image Generation

[ruby yoke]

kritia-ai-diffusion IS stable diffusion. i'ts just a version that'll work inside kritia

[waxen pasture]

aint appearing in dockers

[ruby yoke]

go to youtube and search on "how to use stable diffusion in kirita" - there are a lot of tutorials

[waxen pasture]

got it to work
this seems like a bunch of garbage, honestly

its asking me to connect to some comfyui thing

[ornate elk]

Comfyui is a webui like Automatic1111.
Its the Backend for krita ai stuff

[waxen pasture]

wont it give me generic results?

[ornate elk]

Why?

[waxen pasture]

its comfortable, must be bad

anyway another question, regional text prompt is for the context or to see what you want changed?

[ornate elk]

The extension?

[waxen pasture]

in krita ai yes

[ornate elk]

Oh idk

I didnt used it

[waxen pasture]

damnit, i knew i shouldnt have listened to that guy

[main rune]

@ornate elk I've decided to try Docker and run comfy inside it, any advice?

[ornate elk]

Are you on windows or Linux?

[main rune]

Windows

[ornate elk]

Okay, why docker?
Or do you try WSL?

[main rune]

WSL

[waxen pasture]

hate to admit
but this will do the trick. Thanks.

[shut geyser]

where is clip name 2 folder in comfyui?

[ornate elk]

Ah okay, i tried WSL ones but got nothing to run xD so I can't help much

[shut geyser]

clip and clip vision arent the folder since it still says its not there

[ornate elk]

Have you reselected the clip checkpoints in the node?

[main rune]

I'm just starting (again) so I was fishing for things that might be missed, but it's ok.

Got a guide open right now just hoping for the best.

If it fails it's not like the one I have on my windows suddenly stops working

[ornate elk]

You can also install comfyui with zluda from my guide.

[main rune]

I did!

[ornate elk]

Ah okay

[main rune]

It's the best guide as far as I'm concerned

I'm just trying stuff

-# and hoping it wouldn't cause my pc to explode

[ornate elk]

Yea with docker you can't do much wrong

But its a harder to setup

[keen gorge]

Default and with Lora

[shut geyser]

can i somehow see whats colliding with what in the error log?

[ornate elk]

Then you need an other model because the default one is bad

[hazy pumice]

Okay so I gave up on Automatic1111 and went to clone SwarmUI instead. I got this error:

error: RPC failed; curl 92 HTTP/2 stream 5 was not closed cleanly: CANCEL (err 8)
error: 7901 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output

Is my LAPTOP just too shitty to install and run this? Should I give up?

[ornate elk]

Idk what you try but mat shape error is a conflict of 2 files when they are not compatible.
Like sdxl with an 1.5 lora
Or controlnet 1.5 with an sdxl model

[ornate elk]

Are you on WiFi?

[hazy pumice]

yes

[ornate elk]

How stable is the connection?

[hazy pumice]

I have never had any problems with it as far as I know

but it's not google fiber or anything

[shut geyser]

sampler custom advanced gives an error

[hazy pumice]

@ornate elk I love your PFP by the way. One Piece is my life, been watching it since 2001

[shut geyser]

this is the whole workflow

[shut geyser]

luffy moment

[hazy pumice]

Don't F#$% with Luffy bruh lol

[shut geyser]

(never seen one piece)

[hazy pumice]

BANNED

jk lol

[shut geyser]

U CANT BLAME ME

too many episodes but

i guess it can be fun

:D

[hazy pumice]

bro listen, all the episodes have already been released. You don't have to wait until every Saturday like I did

so you can watch it QUICK

20+ years I have been grinding this show, but it's already laid out for you ya know?

[shut geyser]

20 YEARS?

[hazy pumice]

atleast

[shut geyser]

dedicated..

[hazy pumice]

yeah bro

it's that good though

people don't know

they think it's just some goofy anime about pirates....nah

its DEEP

there is a website

onepace

it cuts out all the filler and nonsense

so its a lot shorter

you can watch it faster without the unneccesary stuff

anyway, I just wanna get my SD up and running....if my wifi is too crappy then I guess I give up? I dunno

[ornate elk]

Any proxy or VPN or antivirus?

Can't wait for the break to end in April for OP

[hazy pumice]

negative. I had Norton guarding me but I handcuffed him to the radiator in the basement and stuff his mouth with an old T-shirt

[hazy pumice]

I know FR. Longest 6 months of MY LIFE

Okay so basically, all I want is to get this installed and start designing some tattoo ideas I have for my next ink, but I keep hitting errors and it's making me want to cry

[ornate elk]

The error is network related.
Did you uninstalled Norton completly?

[hazy pumice]

Yes I did actually completely remove it

[ornate elk]

Also open up a cmd as admin and run:
sfc /scannow

[hazy pumice]

okay

[ornate elk]

And your not located in asia, Russia or Iran?

[hazy pumice]

lol....uhhh

I am in ASIA actually....is that bad?

Beginning system scan. This process will take some time.

Beginning verification phase of system scan.
Verification 100% complete.

Windows Resource Protection did not find any integrity violations.

[ornate elk]

These locations mostly need a VPN for the initial install or DNS change

[hazy pumice]

oh crap...okay

any recommendation for a free VPN?

[ornate elk]

Proton vpn

[hazy pumice]

brb

[gray inlet]

Hello, I was trying this workflow https://openart.ai/workflows/beaver_absolute_68/flux-lora-upscaler/Mk8Yz49ROXjkQObezVMp but I’m having an out of memory error during the upscaling.

I have a RTX 4080 with 16gb vram. I tried with « —low-vram » but in this cause my linux system to freeze 🥲 (likely ram issue even if I have 32gb).

Note my linux system is using intel graphic card not my gpu.

Is anyone able to run the worflow?

[hazy pumice]

That's ALOT of Diddy oil...dang bro

[ornate elk]

How big is your flux model file?

[gray inlet]

I tried with Flux schnell Q4 and Q8. In both case I can generate the image but not upscale it.

https://huggingface.co/city96/FLUX.1-schnell-gguf/blob/main/flux1-schnell-Q4_1.gguf : 7.51 GB

I have noticed in the upscale group that flux dev is in a load checkpoint node . I don't have such model ... Should'nt have an error ? 🤔

[ornate elk]

You need the gguf loader

But I doubt it will work

Upscaling flux causes my PC to freeze with 24gb AMD

[gray inlet]

Stack trace: model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
Requested to load FluxClipModel_
loaded completely 14491.2375 4778.66259765625 True
Token indices sequence length is longer than the specified maximum sequence length for this model (203 > 77). Running this sequence through the model will result in indexing errors
Requested to load Flux
loaded completely 9638.449814605712 7288.3731689453125 True
100%|███████████████████████████████████████████| 20/20 [00:17<00:00, 1.13it/s]
Requested to load AutoencodingEngine
loaded completely 2468.0729110717775 159.87335777282715 True
cache_dir: /tmp
ckpts_dir: /home/sellum/Documents/projets/ai/ai-misc/ComfyUI/custom_nodes/comfyui_controlnet_aux/ckpts
use_symlinks: False
model_path is /home/sellum/Documents/projets/ai/ai-misc/ComfyUI/custom_nodes/comfyui_controlnet_aux/ckpts/lllyasviel/Annotators/sk_model.pth
cache_dir: /tmp
ckpts_dir: /home/sellum/Documents/projets/ai/ai-misc/ComfyUI/custom_nodes/comfyui_controlnet_aux/ckpts
use_symlinks: False
model_path is /home/sellum/Documents/projets/ai/ai-misc/ComfyUI/custom_nodes/comfyui_controlnet_aux/ckpts/lllyasviel/Annotators/sk_model2.pth
!!! Exception during processing !!! Allocation on device

[ornate elk]

What could work is to use the SD upscale script node

It uses tiling for upscaling

And then use tiled VAE node at the end

[gray inlet]

https://github.com/ssitu/ComfyUI_UltimateSDUpscalehttps://github.com/ssitu/ComfyUI_UltimateSDUpscale this one ?

I will give it a try thank you 🙂

By the way I'm trying to produce similar video than this : https://civitai.com/models/832683/flux-pro-11-style-lora-extreme-detailer-for-flux1-dev . I'm able to produce images with the LORA without issues. I can see that for the examples they are using "Kling AI". Do you have any workflow to suggest to do the same kind of image to video locally ?

[ornate elk]

For video generation locally, you would need to use ltx video or hunyuan

Both can to txt2img and img2img

But hunyuan is very resources intensive

[gray inlet]

Ok I will try both. Thank you for your help !

[ornate elk]

Did you got it working with the VPN?

[hazy pumice]

It seems to be working. I am going for Swarm UI instead of Auto1111 and I haven't had any problems yet. It's just a little slow because of my internet

[shut geyser]

can someone tell which models here do not play into eachother?
i get the mat1 and mat 2 shapes cannot be multiplied error

[ornate elk]

maybe on of the loras

but the vae for sure is not flux compatible

flux needs the ae.safetensors

[shut geyser]

changed it still gives the same error on samplercustomadvanced

[shut geyser]

so i skipped all the loras, connected it to the main thing

it still errored the same way

[terse hull]

in the tutorial for installing a1111 for amd on directml, it says this

i have a rx 580 series

do i need the --no-half arg?

[terse hull]

wtf happened here (my prompt was kanye west goes crashout😭)

i used the illustrious xl model from civitai

[ruby yoke]

you seem to have used either the wrong VAE - or the wrong sampler/scheduler

[terse hull]

idk what any of these are i just followed the tutorial for my gpu

[ruby yoke]

the VAE is the VAriable auto encoder. it has to be trained for the model you're using. you're using illustriousXL so you need to use a VAE that works with it.

[ornate elk]

Illustrous needs a 1024x1024 resolution or 768x768

Also don't use a v-pred model

[terse hull]

not sure if my gpu can handle that

[ruby yoke]

what hardware do you have?

[ornate elk]

If you install the zluda variant it could work

Directml won't work likely

With directml you can use SD 1.5 based models (2gb)

Maybe xl if you add --lowvram to the webui-user.bat

[terse hull]

my gpu is an amd rx580 series and my cpu is amd ryzen 5 i think

12 core

if that matters

[terse hull]

i dont really have to use this particular model, i just found it. i wanna have like a nickelodeon, 2d animated comic style

[ruby yoke]

i would take @ornate elk 's advice then. read his AMD gpu guides which are pinned in this channel, install zluda, use it. and perhaps use something other than illustrious

[terse hull]

okay ill try it

do i have to delete my stable diffusion folder i just installed?

[ruby yoke]

are you on mac or windows?

[terse hull]

windows

[ruby yoke]

and this is your first time using comfy, correct?

[brittle lantern]

Is there any Auto1111 or ComfyUI tutorial for ROCm on Linux (Fedora)?

[terse hull]

i use a1111

[ruby yoke]

okay, i wouldn't delete anything just yet. just read his guides and install zluda

[terse hull]

okay ill let u know if it works

[terse hull]

is this a problem? (running on zluda) No ROCm runtime is found, using ROCM_HOME='C:\Program Files\AMD\ROCm\5.7'

[ornate elk]

Should work

[terse hull]

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

[ornate elk]

can you share the full cmd log?

[terse hull]

its really really long

i get a huge amount of errors

but ill send it wait

its still compiling something but thats it so far:

[ornate elk]

looks good

compiling means it works

you now have to wait until it opens

[terse hull]

okay

[ornate elk]

then try to generate 1 image with default settings

it will compile again 15-40 mins

then its doen and you can generate image much faster (like a few seconds)

[terse hull]

okay

thank u🙏🙏

[ornate elk]

no problem 🙂

[terse hull]

😭😭

i started it again and this happened

[ornate elk]

just restart it a few times aftert it crashes. Its a bug for RX580 users. As the compiling crashes. But after 10-15 times of relaunching it should work

[terse hull]

okay ill try it

[brittle lantern]

@ornate elk u have a ROCm Linux auto1111 tut for Fedora?

[ornate elk]

hey, nope sorry :/

[brittle lantern]

Yk where to find one that updated or updated enough?

[ornate elk]

hmm nope but on the Shark AMD Server are some linux users
Link to the discord is at the bottom:
https://github.com/nod-ai/SHARK-Studio?tab=readme-ov-file#communication-channels

[sand pasture]

so i am looking to train a lora based off of screenshots of a 3d model character I want,

would it help to have a character in different poses or is a t pose with many screenshots is sufficient

[shut geyser]

how to put --medvram

in comfyui

in which spot?

[ornate elk]

Comfyui doesn't have a --medvram. It does have --normalvram

You can put it behind the --auto-launch at the bottom

[ashen gate]

Why do I always end up with faces like this? Her eyes...

[noble idol]

i looked at the metadata

i got scared ngl

[ashen gate]

why

[noble idol]

i think you should go look at some other people's metadata and see how they prompt and use loras

[ashen gate]

ive tried standard formatting and it still does that

[noble idol]

well youre also using the lora weight on 1

lower that

[ashen gate]

roger roger

[noble idol]

dont weight so many things so high either, i dont think thats a good idea

[ashen gate]

Yeah, so formatting it vanilla did nothing

[noble idol]

can you give me the lora?

i dont see anything else wrong but im not a pro

ah maybe try making the image dimensions 1024x1024 instead of 768x768? that might just help in general

[ashen gate]

https://civitai.com/models/556442/eda-clawthorne-the-owl-house-commission

Just don't scroll down to the images. LOL

[noble idol]

LOL fair enough

ill honor that request

im gonna try it, see if it works

[ashen gate]

Resizing to 1024x1024 did not help either. It's a pony lora for pony checkpoints.

[noble idol]

if it does you can just use my image's metadata and compare them and change stuff until it works

[noble idol]

have you considered using a more popular model, like autismmix?

its a good model

[ashen gate]

From what I see it's majorly cartoony and I wanted a checkpoint that can make things a bit more real. MY problem persisted throughout 3 different Pony checkpoints so I don't think this will work, but I'll try it.

[noble idol]

hmm, ive never tried making a cartoon character 3d before

wonder if its hard

im going to download a pony model, ill dm/ping you when im done, itll take a while. Hopefully someone solves your problem in the meantime

im not great at thsi myself either but i wanna see if i cant help

[ashen gate]

USed AutismMix and not much better:

[noble idol]

i got this-

i used autismmix

ignore the hands 💀

[ashen gate]

So am i wat the hell

Also yours has no metadata. I need a PNG with metadata or I can't get similar results.

[noble idol]

ah right thats my bad, i just copy and paste usually instead of saving images- which means theres no metadata

let me grab it with metadata

ill send it in dms so we dont clog up this channel

[vapid stream]

hey everyone, If i have a error message is this the right place to try to get it fixed ?

[lucid yoke]

Yes, if someone has the time they can help

[vapid stream]

awesome, if someone has time, ive been getting an error while generating or trying to generate an image. The error message is NansException: A tensor with NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[lucid yoke]

There's a solution in the error i think. Have you tried that?

[vapid stream]

yep i turned on the upcast cross attention lauer to float32 and nada. I still get this message 😦

i also tried googling this error and no luck there either

[noble idol]

wait and pray CS1o sees your message 🙏

[vapid stream]

🙏

im getting a green box now, Any help?

[lucid yoke]

Please include more details as its very vague

[ruby yoke]

@ornate elk is more likely to see if it he's pinged

[noble idol]

yeah but being pinged is annoying

[ornate elk]

What's your GPU?

[deft hemlock]

any idea why my images look different even with the same settings ?

only difference is the seed and web ui

no im not using it for porn

firsts one is forge and the other is a1111

images look different now, same checkpoints and prompts

[flat yarrow]

My workflow is using SDXL checkpoints switching between these juggernautXL_juggXI, realismEngineSDXL_v30VA, cyberrealisticXL_v4, photopediaXL_45, when I run the workflow my images are giving the subject elongated arms like an Orangutan. I'm aiming for 9:16 and using 768 x 1344 in SDXL, am sure this is my problem as I can see duplications and it looks to me that I'm on the edge of what SDXL 1024 can do. So the only way I can see is making the image smaller keeping the 9:16 ratio but then that's in SD1.5 @ around 512 and the loss of detail with the latent and decoded pixels. Can anyone help with a solution for this, I have low VRAM @ 8GB on a 4060. Thanks...

[shut geyser]

i put it there and it still uses all of the vram

[ornate elk]

Yep you can't limit it with the commands

But you can try add --reserve-vram 2

[ornate elk]

Sdxl models are trained on 1024x1024 so you need to generate at lower res and then Upscale them to get Wallpaper sizes

[ornate elk]

You used a different model and lora on them

I checked the meta data

[deft hemlock]

those are examples

not exact

but i get a similar result

let me generate new ones

[deft hemlock]

New issue

Sadly

When I change the checkpoints A1111 is super laggy

My entire pc slows down like hell

I had lowvram in the startup

My specs are rtx 2050 laptop version with 4gb of vram

I Should just delete the venv folder and it would be fine right?

Idk if this detail helps but like changing the checkpoint takes a really really long time

And when it reaches 95% the entire pc slows down, so much so that if I have anything else happening in the background

Like switching tabs or opening notes or discord

The pc just crashes and turns off (not blue screen)

From the replies I've read I should just delete venv and I'll be okay

Right?

And tiled vae

But idk how to use it

[flat yarrow]

@ornate elk I'm aware of that and I'm also aware that you can work slightly outside of 1024, but my size inside the 1024 is 576 x 1024, but the width is closer to SD1.5 and the potential loss of detail. I need to maintain the 9:16 ratio and I'm upscaling anyway, but my train of thought (maybe wrong) is that I need more detail than SD1.5 even though I'm using SDXL model. As per my original post I'm trying to get my head around when and where I can work in the fringes, but since am upscaling anyway should I be worried about the detail in the latent space keeping within the 1024 boundries. Sorry about this am new and I've only being using Sd for a few months now, so I still need to get my head around a few things.

[ornate elk]

I don't get the question sry

You shouldn't go down to 512 with sdxl models because that can break the image too

Your initial resolution with 768x1344 seems okay but maybe your denois value was to high when you upscaled it

[ornate elk]

You need to increase the windows pagefile then. If you only have 16gb of RAM and use sdxl models that causes the lag

[zenith seal]

Hello!

I have a small question

is it possible the set this setting in auto when starting the ui ?

[flat yarrow]

@ornate elk I'm after as much detail as possible in SDXL, so keeping @ 1024 and maintaining 9:16 ratio my new image size would be 576 x 1024, I'm just a little concerned that I won't benefit from SDXL and the extra detail that SDXL provides compared to SD1.5. On the otherhand I'm upscaling anyway so should I be worried about the inital generated image and the detail in it. I just think crap in - crap out. That's why I'm using the image size larger than it should be to have extra detail with 768 x 1344 but do I need to be worried about that. It's really about what I'll get after the upscale. This deform is happening in the 1st KSampler before anything else. I do know to keep within the 1024, what I don't know and have lack of experiance is the quality of the images produced before the upscale and if it matters and if it does how to get the best detail.

[deft hemlock]

i got 12 gb, do i still increase the pagefile?