🤝｜tech-support

[wet cedar]

I have something called Pinokio...and downloaded ComfyUI

but I can get it working through Wan

^ I get this error

[ornate elk]

whats your gpu?

[wet cedar]

16 GB of Ram, Intel(r) Core(tm) i9-9900 cpu, 3.10ghz

[ornate elk]

i mean graphics card

[wet cedar]

Asus Radeon RX590

[ornate elk]

ah okay

that wont work with video stuff, its to weak

[wet cedar]

What video card do you recommend?

(also thank you so much for your help so far)

[ornate elk]

idk your budget but a 5060ti would be a decent gpu

it has 16gb vram

[wet cedar]

Can I sound really stupid for a moment?

If I haven't already e_e

[ornate elk]

sure xD

[wet cedar]

^ these are my Graphic cards? Correct?

[ornate elk]

2x RX590 ?

[wet cedar]

Yes

Would I be replacing both?

[ornate elk]

yea probably

[wet cedar]

Okay. I may have to do one at a time if they are 140 dollars lol

[ornate elk]

what do you mean with 140$ ?

[wet cedar]

Or wait...that's the cheap one xD

750 dollars @_@

or 1000 lol

If it's that powerful do I really need two of them?

[ornate elk]

no not at all

[ornate elk]

today you only need one gpu

no real benefit in using two

[ornate elk]

check the 5060ti

[wet cedar]

Oh okay 🙂 Thank you

Only 5070 and 5080 are coming up

And my computer can handle this graphics card?

[ornate elk]

yea the 5060 and ti was released yesterday, its not in all stores, but give it a week

[wet cedar]

kk

[ornate elk]

depends on your PSU

[wet cedar]

Where do I find that?

[ornate elk]

you cant, you would have to open up the PC and check whats printed on there

or at the back

like 500watt or 650watts

[wet cedar]

Not seeing any watts

[ornate elk]

hmm

at the backsite where the power plug is?

[wet cedar]

I dont suppose its RM750?

[ornate elk]

sounds like a 750w power supply

[wet cedar]

That's...good?

[ornate elk]

yep, should be good with any midrange gpu

[wet cedar]

is the 5060 midrange?

Ah great! Thank you SO much for your patience and time. Sorry I'm a dummy.

[ornate elk]

no problem,
if you want to use your RX590 for image generation you can follow my Automatic1111 with Zluda setup guide

its the first link of the pinned messages here

[wet cedar]

Oh, I have no problem with image generation. 🙂 I just need to move forward with image to video and I guess having a stronger graphics card

[ornate elk]

what do you use for image generation?
because zluda is much faster than directml

[wet cedar]

ummm Automatic1111?

[ornate elk]

ah true

then probably your using directml

[wet cedar]

I think. It uses half my cpu

[ornate elk]

whats in your webui-user.bat?

[wet cedar]

In it?

I'm not sure?

[ornate elk]

right click, edit, screenshot or copy the text

[wet cedar]

^ this

[ornate elk]

i mean the bat file itself not the cmd

but as i see the launch args your running it on CPU mode only 😮

[wet cedar]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=
git pull
call webui.bat

[ornate elk]

strange

[wet cedar]

Hm?

[ornate elk]

there are no launch args but the cmd shows some

your sure your not launching an other file?

[wet cedar]

^ that's what I click on to launch

I mean, I can set up a launch arg if you want xD

[ornate elk]

im confused because look at your cmd log

normaly such args are added in the webui-user.bat

but yours is empty but the cmd says the opposite

[wet cedar]

Would it be better to add them in it?

[ornate elk]

the thing is right now your not even using your GPUs to generate the images

[wet cedar]

notes his SD is a Frankensteins monster

[ornate elk]

its all running only on CPU

that makes it really slow

[wet cedar]

Oh yeah, it's SUPER slow

But I think it was set up that way because I didn't have a nividia card at the time?

[ornate elk]

yea but it works with AMD now too

[wet cedar]

Ooooh well that is helpful xD

AMD wasn't included at the time when I did all this

[ornate elk]

you now have two options to run it:
Either with DirectML, which is faster than CPU, and easy to setup
or with ZLUDA which is much much faster than CPU and DIrectML but a bit more work to setup

[wet cedar]

Well, if you walk me through it. I'll do what you recommend

Do you have a coffee tip jar thing?

[ornate elk]

i would recommend using Zluda

yea if you click on my Avatar/Username here, and then on "show full Biography" or somehting like that

[wet cedar]

tipping you 10 1 sec

There you go 😄

[ornate elk]

thank you 🙏 appreciate it 🙂

[wet cedar]

Yeah, I appreciate you as well!

So Zluda....I assume I need to download that

[ornate elk]

As im also on AMD i made a lof of Guides to setup the different webuis for Nvidia and AMD
you can find them here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

And the Guide for the RX580 / 590 gpus is here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#rocm-hip-sdk-57-with-zluda-setup

but first check your task manager, if your RX590 has 4 or 8gb of VRAM

[wet cedar]

I see...7.9/15.9 GB

[ornate elk]

perfect

so 8gb dedicated gpu vram?

[wet cedar]

Oh yeah, available 8 .... 16.0/33.6 GB committed

[ornate elk]

and the other one too i guess

we sadly cant combine the vram but you can run 2 instances of the webui at the same time to generate images

[wet cedar]

kk

k editing the webui now

[ornate elk]

the new one? just want to make sure you shouldnt edit your old webui

because thats outdated

[wet cedar]

There's nothing to edit in the new folder

@ornate elk

Getting the Automatic1111 AMD Fork:

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.
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.
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
Press enter and after its done you can close the cmd.
Edit the webui-user.bat (right click), At the line COMMANDLINE_ARGS= You add: --use-zluda --update-check --skip-ort

^ is the steps I've done. On #5

[ornate elk]

alright

[wet cedar]

installed the new python and installing the AMD pro

[ornate elk]

ah nope

alright

you have --medvram-sdxl in your webui-user.bat right?

[wet cedar]

In the old stable diffusion folder I have?

[ornate elk]

in the new one?

[wet cedar]

The new folder is completely empty

[ornate elk]

i thought your on step 5 ?

so you should have the wbeui in the SD-Zluda folder

[wet cedar]

Yup. I created a new folder. I ran the command. But the new folder is still empty

[ornate elk]

ohhh

then you didnt opened the cmd the correct way

[wet cedar]

Ah, okay. So open it from the new folder?

^ file explorer bar?

[ornate elk]

the left one

not the searchbar right

[mystic basin]

does the quadruple clip loader on comfy-ui work with ZLUDA or it only work with cuda for now? i can't seem to find it

[wet cedar]

there we go

[ornate elk]

idk, make sure your comfyui is updated via git pull in a cmd in the comfyui folder

[mystic basin]

already did

[wet cedar]

K moving onto the AMD section

[mystic basin]

i'm going to check again to see if now it appear

[half island]

Hoi, not exactly "a problem", but looking for a solution to add conveniency.
currently expanding the requirements.txt for comfy to make any reinstall a one command, but is it possible to add pytorch to that list somehow?
As i'm installing latest 2.6.0 with torchaudio/vision with cuda, then uninstall pytorch 2.6 for 2.7 dev, how would i add that?

[mystic basin]

nope, still can't use it. I can't find it and I can't use it if I copy someone else workflow

it says the nod is missing

[half island]

Scroll up in the cli, as right before it announced server was ready, you can check the list of loaded nodes, and see what failed. Scroll up to the broken node, and see if you got a module missing like this.

[mystic basin]

oh yeah i understand why i could not find it

i was using the stable version

this quadrupleCLIPLoader is new

[ornate elk]

ahh okay

[mystic basin]

i think that's why

[ornate elk]

i think you cant do it in requirements, but you could create a speerate bat file to do it

[half island]

True Though, i'm not good with making bat from the bottom, and GPT is a 50/50 if it'll work or not lol.

[ornate elk]

venv\Scripts\activate.bat
pip uninstall torch torchvision torchaudio pip install torch==2.7.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu(enter the cuda version here 128 etc)

[ornate elk]

hows it going?

[wet cedar]

Just downloading and installing drivers atm

[half island]

Oh, nothing more? Like "echo" and any other of that stuff?

[ornate elk]

Echo is if you want to output some text for the user like
echo Uninstalling old torch...

[half island]

Aye, meant more the regular "REM" commands and such. Didn't know you could just do as-is commands :P

[wet cedar]

Open the gfx .zip file and drag and drop all files of the library folder into the library folder. Not into the old_library folder.

^ do I skip this step because

For RX 580, 5600/XT, 5700/XT, VEGA 56, VEGA 64 and Radeon VII or RX7700S you need to download the ROCmLibs.7z

^ wasn't a gfx download?

Off to work!

BBL

@ornate elk

[ornate elk]

you need the ROCmLibs.7z

thats the gfx "file" for RX580 / 590

[half island]

Got it working!

Make sure flash attention version correspond with your cuda/zluda, torch and python version, and update comfy to dev channel

[spark shadow]

ive tried looking online and haven't found much/whatever i found hasn't quite been helpful, my computer has a 6900xt and a quadro p1000, stable diffusion keeps using the quadro even if i tell it to use -device-id=0 and to skip cuda checking, is there something im missing? the only thing i could think of is overwriting where it looks for an nvidia driver?

ive installed the amd version/followed the amd guide here and tried to follow some troubleshooting but to no avail

[ornate elk]

Hey, make sure you follow one of the Zluda setup Guides from the first link of the pinned messages of this channel.

[spark shadow]

zluda > directml? i havent seen these yet bet lemme try

[ornate elk]

Yea zluda is much faster and uses less vram than directml

[spark shadow]

oh man goated

alr lemme mess with this and see if i can get it to work

[ornate elk]

Sure, let me know if you have any questions.
Would recommend starting with Auto1111 with Zluda

[spark shadow]

is it reccomended to uninstall older versions/other versions of python if this guide specifies one different than the one i have

[ornate elk]

If you don't need them, yes

If you need them, then you have to specify the path to Python in the webui-user.bat

[spark shadow]

nah i dont need em ill wipe em and reinstall the version this one wants

ty ty

[ornate elk]

No problem 🙂 I'm off now to sleep. Can help tomorrow.
The dual GPU setup is tricky but we should get it to work

[spark shadow]

Thanks for the help! yeah i didnt even think it'd be an issue till i got started hahaha

[balmy plover]

anyone got a list of models and control points i should install install? outside of the ones which came with cs10's startup ones?vidia 32gb ram

[spark shadow]

wanted to say thank you so much, ran into a weird runtime error but i looked in the chat history and followed what you told someone else and that worked for me too. thank you!! 🙏

[balmy plover]

warnings.warn(f"Importing from {name} is deprecated, please import via timm.layers", FutureWarning)

is this an issue?

[ornate elk]

Would need to see the full cmd log

[ornate elk]

No problem 🙂 which runtime error?

[formal trout]

Still OOM, I thought it was the image size I was using but I re-scaled and it was the same

I'm on Ubuntu, I don't have much luck with Comfy

Requested to load WanTEModel
loaded partially 10748.8 10739.4765625 0
0 models unloaded.
loaded partially 10739.476562595368 10739.4765625 0
Requested to load WanVAE
loaded completely 3548.3 484.05659103393555 True
Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.
!!! Exception during processing !!! HIP out of memory. Tried to allocate 2.53 GiB. GPU 0 has a total capacity of 11.98 GiB of which 2.28 GiB is free. Of the allocated memory 7.04 GiB is allocated by PyTorch, and 2.32 GiB is reserved by PyTorch but unallocated.

[ornate elk]

You should look at LTX video then

0.9.6 was release a few days ago

Needs less vram than wan

[balmy plover]

is it just me or is there no settings on this?

[ornate elk]

What do you mean?

There is a settings tab

[balmy plover]

hmmm

oooh got it ty, spent 10 minutes looking at that screen

oooh got it, spent 15 minutes staring trying to see it, right at top bar!

ty

[half island]

Has anyone experienced that either confyui or when training with ai-toolkit will just stop processing out of nowhere? just ceases GPU activity, and when you press control C to cancel, it instead ramps up to 100% usage again?

[mellow coyote]

i had use automatic 111. but my vram is always not enough. is there any suggestion or alternative?

[wary tulip]

Can anyone help me building a workforce in comfyui

[balmy plover]

i currently have 8 gig ram, do i need to upgrade it for this?

[lucid yoke]

Ram or video ram?

[balmy plover]

Operating System
Windows 10 Home 64-bit
CPU
AMD Ryzen 7 2700X 61 °C
Pinnacle Ridge 12nm Technology
RAM
16.0GB Dual-Channel DDR4 @ 1197MHz (16-16-16-39)
Motherboard
ASUSTeK COMPUTER INC. PRIME A320M-K (AM4) 52 °C
Graphics
S24D590 (1920x1080@60Hz)
4095MB NVIDIA GeForce RTX 2070 (ZOTAC International) 45 °C
Storage
1863GB Seagate ST2000DM008-2FR102 (SATA (SSD)) 44 °C
238GB ADATA SU800NS38 (SATA (SSD)) 50 °C
Optical Drives
No optical disk drives detected
Audio
High Definition Audio Device

[lucid yoke]

Ahh 4gb video ram, hmm i think you should be able to run stable-diffusion 1.5 models but it won't be the fastest

[balmy plover]

looks like it is time ot buy a new pc

ty

whats specs do i need?

[lucid yoke]

If just for AI, graphics card

And probably psu

[balmy plover]

and video ram?

[lucid yoke]

Not sure about your cpu though

[lucid yoke]

Thats on your graphics card

[balmy plover]

gotya

[lucid yoke]

You have a rtx 2070 = 4095Mb vram

Youd be looking for 8gb minimum for sdxl models but if you want to future proof, a nvidia card with 12-32gb vram is best

But if your unsure i don't recommend investing so much before you've tried it and liked it

[balmy plover]

vidia 32gb ram

[lucid yoke]

You'd be investing a lot

[balmy plover]

3k pounds?

or do i need more

i'm lucky I have a decent decent business i dont mind spending the dosh.

[ornate elk]

You need to add --xformers --medvram to the webui-user.bat
And remove --medvram-sdxl

[balmy plover]

ty

[warped canyon]

any way to add a shortcut to switch between these two?

[formal trout]

great ty, do you know any installation guide? I'm looking I havent found yet

[left jackal]

tried onnx version 1.15, tried uninstalling and installing onnx, tried removing python and installed 3.10.9, tried removing and cleaning install kohya. cant able to fix this. please help its related Kohya-ss

its related to WD14 captioning

[graceful oar]

hey does anyone know why it takes so long to generate an image with the dpm++ 2m sampler? its stuck at 99% for a relatively long time (1-2 minutes), only then its done.

im using a1111 with zluda (ive got an rx6600, 8gb vram)

[ornate elk]

Kohya wants python 3.10.11 64bit

[ornate elk]

Hey, what are your txt2img settings?

[pliant arrow]

Hello! Can't install SD

[ornate elk]

Hey, that's a timeout error, related to your internet connection

[left jackal]

how come it never warns me python is out of date. installing 3.10.11 will get back.

[pliant arrow]

So there's nothing that I can do?

[ornate elk]

You can relaunch and the download should go further

[pliant arrow]

Is there specific command, or I just have to run .bat file another time?

[ornate elk]

Just run the webui-user.bat again

[left jackal]

even with 3.10.11 same error

packages

cant figure out what am i missing. :<

[pliant arrow]

It seems like it started downloading from zero... :(

Again...

[ornate elk]

Make sure your connected with ethernet

[pliant arrow]

Already was

[ornate elk]

You can also try use a VPN

[pliant arrow]

It's a problem with my provider I guess

Anyway, thank you for your help. The problem has been localized at least

[left jackal]

still doesnt fix anything, updated everything and it shows onnx exists and yet it keeps saying onnx models not found. stuck at this for an entire day. 😦

[ornate elk]

You can manually download the model file thats missing:
https://huggingface.co/SmilingWolf

[left jackal]

ohh the time i wasted figuring it out. bloody 12 hrs of time.

THANK YOU DUDE. @ornate elk

WORKS!!

[ornate elk]

no problem 🙂

[left jackal]

I should have asked earliest. 🙂 thanks man. somehow it was downloading but then deleting it automatically hence the model missing issue.

[ornate elk]

yea sometimes its an easier fix than it looks ^^

[graceful oar]

i tried dpm++ 2m, 100 steps, 512x512 resolution, cfg 6. batch size, count 1. no hires. fix or refiner.

no face restoration.

[lucid yoke]

100 steps?

why 100

try using 30

[ornate elk]

can you try with 30 steps ?

and which model do you use?

[graceful oar]

dreamshaper 8

[graceful oar]

its still happening but maybe cuz im running mc at the same time

alr no, its still hapenning

nothing else running in the background

[ornate elk]

its normal for 1.5 models to stuck a bit at the last step with automatic

if you try an sdxl based model it should not stuck that long

and dont run anything in the background, that makes it slower.
Also restart SD after you closed MC

[graceful oar]

where can i get one?

[graceful oar]

got it

[ornate elk]

civitai.com

also make sure you have --medvram-sdxl in the webui-user.bat

[graceful oar]

alr, ty

[ornate elk]

np

[inland portal]

does anyone knows how to fix this?

[ornate elk]

Hey, what's your GPU?

Also don't join "support discord Servers" they are scam

Some bots try to invite users to scam them.

The error you got is a internet connection timeout error.
Make sure to use a stable Ethernet connection and then retry

[inland portal]

920 mx

2gb

[mellow coyote]

are anyone can teach me how to use automatic 111 with google colab?

[pliant arrow]

Hello again! I have a differend kind of problem now. I can't upload any of my SD checkpoints

Upd. This thing also can't use another GPU and somehow eats all of my GPU memory

[lucid yoke]

Hmm your trying to use both gpu*s at once. In what way exactly?

[pliant arrow]

They're connected parallely to different monitors, but some games use both of them and work better

In any case, I dunno why 6GB VRAM is not enough to switch the checkpoint

[ornate elk]

You need to add --medvram --xformers to the webui-user.bat

Then relaunch

[cosmic ocean]

Hello, I just got Automatic 1111 to run on my Windows PC and I got a couple of technical questions about my setup and hardware utilization I hope this is the right place for that on this server

[ornate elk]

Sure feel free to ask

[cosmic ocean]

So I have an RX 7900xt and found out that you actually need zluda to run auto 1111 with an amd, luckily I had my old RTX 3070 8gb still around and got it to work on that card, now I guess my question is how I can make the most out of this hardware and wether it is worth trying to install zluda and how the performance on these specs would compare

It is also worth mentioning that I have both GPU´s installed and the 3070 only has a PCIe gen 3 x1 (!!!) slot available

[ornate elk]

I have the 7900xtx and its really good with zluda. Performance is great!
Check out the install guides from the first link of the pinned messages of this channel.

But a 3070 with 8gb is good to until it runs out of vram

When you use the 3070 with Auto1111 webui, make sure to add --medvram-sdxl --xformers
as launch args to the webui-user.bat for the best performance

[cosmic ocean]

Yeah that is what I thought as well ^^ well thank you, I guess I will give it a try and come back if I have any more questions 👍

[cursive hazel]

Hi. Me again, again...

I was trying to use Wan image to video in ComfyUI, using Zluda on my AMD RX 9070. All my videos turn out to be, more or less, blank. Does Wan work with Zluda? Or AMD cards in general?

In ComfyUI's command prompt, I did get this: loaded partially 4841.86841796875 4841.865478515625 0, which ChatGPT said could be why my videos are coming out messed up. (Otherwise, GPT just kept saying to change the cfg, change the denoise, change etc etc. Which I did, time and time again.)

[ornate elk]

Feel free to ask me if any step in the guide isnt clear

[ornate elk]

Make sure to use wan 1.3b for the lower vram. It should work but I haven't tested it

LTX video works

[cursive hazel]

Originally I tried using wan2.1-i2v-14b-720p-Q5_0.gguf, but when that failed, I switched to wan2.1-i2v-14b-480p-Q4_1.gguf. Both do the same thing.
I assume the part just before 480p is what I should be focusing on, and find one that says 13b instead of 14b?
(So sorry for my noobishness. Always learning, and grateful to those that help!)

I'm looking into LTX video now too though, thank you for the recommendation. ChatGPT told me to use AnimateDiff, and when I looked up videos of that, they uhh.. Very competitive with things like Wan or Hunyuan.

[lucid yoke]

chatgpt is ouytdated as always

[cursive hazel]

Yeaaah, it doesn't even believe that the RX 9070 is out. It "corrects" me, saying "You must mean an RX 7900" >.>

[ornate elk]

LTX Video works good and also has gguf models

AnimateDiff is very old

[cursive hazel]

One of the youtube videos I had watched, recommended to use gguf models. I admit I know almost nothing about them though. I'm assuming they're just better/easier to run on AMD cards?

[orchid forge]

Greetings Tech Support, I'm back with more dumb questions!

Finally got Stable Diffusion to run on my comp, thanks for the assist, but now it's running me out of memory pretty hard and I've tried to fix it to no avail

PyTorch allocated all my GPU memory as follows:

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 11.12 GiB. GPU 0 has a total capacity of 15.98 GiB of which 0 bytes is free. Of the allocated memory 11.59 GiB is allocated by PyTorch, and 10.18 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

I had to google how to set that variable with that value, said to do it as an environmental variable for System, with name as variable and value as value?

That did not fix it, didn't appear to do anything. Running 7600X CPU and 7600 XT GPU with 16GB

There are a lot of other weird sounding errors, not sure if they matter, I'll copypasta

"Failed to create model quickly; will retry using slow method.
Applying attention optimization: Doggettx... done.
Model loaded in 3144.1s (load weights from disk: 0.2s, create model: 22.1s, apply weights to model: 9.4s, apply half(): 0.2s, move model to device: 0.3s, hijack: 3111.7s)."

"OSError: None is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'
If this is a private repository, make sure to pass a token having permission to this repo either by logging in with huggingface-cli login or by passing token=<your_token>"

"huggingface_hub.errors.RepositoryNotFoundError: 401 Client Error. (Request ID: Root=1-6803efd2-3dcdd25d77e84f452f0cd0ae;7ee134f4-e912-4693-b024-08046a4c54e7)

Repository Not Found for url: https://huggingface.co/None/resolve/main/config.json.
Please make sure you specified the correct repo_id and repo_type.
If you are trying to access a private or gated repo, make sure you are authenticated. For more details, see https://huggingface.co/docs/huggingface_hub/authenticationInvalid username or password."

Not sure what else is needed, I'll lurk until I get a response?

[prisma moss]

Do i need or not the space here?

[ornate elk]

Exactly

[ornate elk]

Doesnt matter

[prisma moss]

ty

[orchid forge]

CS1o you are everywhere and all-knowing lol

Like a god

[ornate elk]

Have you followed a zluda webui guide for the install?

If not checkout the first link of the pinned messages

[orchid forge]

Yes, followed both ComfyUI and Automatic1111 for AMD guides

I'm running from e: but I have copies of both on e: and c:

I should probably delete the ones from c:

E: is a faster drive, so I wanted to run it there

[ornate elk]

With which settings do you get the out of vram error?

Upscaling?

[orchid forge]

Recommended on civit for these models

Using Illustrious v13 as SD checkpoint, Illustrious v11 as refiner

Per a comment on civit

[ornate elk]

For upscaling you need to install the tiled VAE extension.
That helps in preventing the out of vram error.
Also only use 10-15 hires steps

And never ever use a refiner xD

[orchid forge]

Ahhhhhhh

So that guy was full of it, sweet. So much bad info out there??

Is it better to just create the image in full res, or upscale from smaller image?

[ornate elk]

Yep refiner was something made for the base sdxl model

Its worse than upscaling

[orchid forge]

Also what is the tiled VAE extension and where can I get it?

[ornate elk]

Ans needs massive amount of vram

[orchid forge]

yeah it allocated 15.98 GB

Ridiculous

[ornate elk]

Yea because its like loading 2 image models at the same time

Won't work

[ornate elk]

Go to Extensions tab,
Click available
Click load from.
Search for tiled
And install tiled diffusion + Tiled vae

Then relaunch the webui

Then only enable Tiled VAE when also using hires fix

[orchid forge]

Okay I think I followed your instructions, checking now. Wish me luck lol

My computer is brand new and designed to be solidly mid, this is so far not what I was expecting

lol

Okay this one will take 2 minutes, instead of 90. Oh my god.

https://tenor.com/view/facepalm-really-stressed-mad-angry-gif-16109475

On behalf of a grateful nation, I would like to thank you for your service. Hope you have a most excellent day, Sir!

[ornate elk]

No problem 🙂 thx you too

[prisma moss]

How, and when, do you use () alongside \ or something? To like define for example a series which the anime cahracter is from?

Whats teh correct syntax

(im not talking about weights)

[ornate elk]

Install the booru tag autocompletion extension.
It does that when you prompt for a character of a series/anime

[prisma moss]

been a while since i used SD in general, how do i go about installing exensions?

[ornate elk]

Go to Extensions tab,
Click available
Click load from.
Search for booru tag auto
And install, then relaunch

[prisma moss]

I see thanks, any other extensions worth using:

?

[prisma moss]

which booru tag, theres multiple, whats the exact name

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

This one?

[ornate elk]

Yes this one

[prisma moss]

Thank you

[ornate elk]

• Tiled Diffusion & tiled VAE
• Controlnet

[prisma moss]

Thank you

[ornate elk]

Np

[prisma moss]

This control net? These multiples

[ornate elk]

Yep that's it

[prisma moss]

thanks chief

[crystal nacelle]

hey everyone, im getting a connection time out issue with img2img gen, and long generation times on txt2img. my console has told me a possible fix but wanted to consult with yall first.

[midnight kiln]

Hey all. Where would be the best place to go to find well optimized Comfy workflows? I finally took the jump. No point in reinventing the wheel. Especially memory-management stuff is important. At a flat baseline Comfy is eating up a lot more than would Forge for example

[crystal nacelle]

jea it sucks. heres the full console recommendation @ornate elk here is the full log.

[sacred osprey]

Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Traceback (most recent call last):
File "C:\Users\berkk\Downloads\sd.webui\webui\launch.py", line 48, in <module>
main()
File "C:\Users\berkk\Downloads\sd.webui\webui\launch.py", line 39, in main
prepare_environment()
File "C:\Users\berkk\Downloads\sd.webui\webui\modules\launch_utils.py", line 387, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Press any key to continue . . .

i tried everything

please help

[sacred osprey]

what

i click runbat and that

in easy diffision i get 10:46:14.574 INFO cuda Session 1745135108288 task 1696478387600 failed! error

[timber current]

Congrats 🍀 🍀 made it in

[somber basin]

Guys i am trying to make some videos like evolution timelapse i tried Animatediff with different models but it didnt give the videos i want. What should i do to make prompt travel videos like evolution timelapses any different suggestions than Animatediff? Or it just about mu settings i am using comfyui

[ornate elk]

Can you show the full CMD log when you launch the webui-user.bat?

[nocturne sonnet]

I'm trying to generate a Wan2.1 video using kijai I2V workflow in comfyui. I'm able to start generating but it seems to get stuck at the 'WanVideo Sampler' node at 71% complete. the last block if text in the cmd window says:

Sampling 81 frames at 624x624 with 25 steps
0%| | 0/25 [00:00<?, ?it/s]

I am running a RTX 4070, 32gb of DDR4 ram and an i7-9700K. I can see on task manager that around 25gb of ram is being used (despite nothing really happening) but only around 3% of my GPU is being utilized with some occasional spikes.

[lucid yoke]

looks like its not using your gpu at all and using ram only

hmm are you using the 480p model or the 720p

i recommend going for a 480x480 gen since its faster

[nocturne sonnet]

im using the 480 model

ill try go for 480x480 gen

[crystal nacelle]

I tagged you above, sent a cmd screenshot

[ornate elk]

Ah but thats not a full cmd log
Need more information

[earnest goblet]

Hello, I am trying to run stable difussion locally. I believe I've installed things correctly up until I am supposed to "Find and download the Stable Diffusion model you wish to run from Hugging Face. These files are large, so the download may take a few minutes."

I go to huggingface but can't figure out how to download the model.

[earnest goblet]

well that wasn't helpful at all.

[lucid yoke]

hmm Tracer i recommend civitAI (website) for stable diffusion models

it will give you examples of models

[chrome garden]

Hello, I want to train a character's Lora, but I need it for the Illustrious XL models. I would like to know which Pretrained Model is good for those Loras. I had made some Loras for Pony models, but I imagine that I have to make some adjustments to the configuration I have for (Pony Loras) so that the Lora (Illustrious) works great with the Illustrious Models.

[earnest goblet]

thank you.

[fossil terrace]

any help is greatly appreciated! like i said, it finds the images just fine, but not the txt files, even though I set it specifically to .txt in the lora settings on koyah

[vast pulsar]

hey guys, a friend wants to get into forge and all but apparently it needs a nvdia gpu.. any way to get it running with an AMD gpu?

[fossil terrace]

could any kind soul pls help with this bc i have been sitting at this for 5 hours and it just wont recognize them no matter what i do

[warped canyon]

is there any shortcut to "skip" or "interrupt" a generation?

[open fox]

did you prepared the dataset before starting the training and after creating the text files?

[fossil terrace]

the text files have been there from the very beginning, i added them at the start. i did not prepare any dataset

[open fox]

Lora Tab scroll down to dataset preparation if this is still a thing. It has been a while...

[fossil terrace]

i did it now but it still gives this error

found directory C:\Kohya_SS\training\duck_lora\40_qwvs duck contains 18 train_util.py:1996
image files
read caption: 100%|████████████████████████████████████████████████████████████████████████████| 18/18 [00:00<?, ?it/s]
WARNING No caption file found for 18 images. Training will continue without train_util.py:2027
captions for these images. If class token exists, it will be used. /

THERE ARE god damned .txt files in there.. with the same fucking name. I am sick of it

[open fox]

As you use windows it should not be a user rights problem. Linux i would guess the files are not readable to the user running the kohya instance...

[open fox]

Can you check your settings: near the top it'll ask for a file format for caption files. The default is .caption, change that to .txt.

[ornate elk]

Hey, forge has a fork which is for AMD GPUs, checkout the first link of the pinned messages for the Guides.
And go for the Zluda version.

[hot thorn]

How do I merge loras?
switched to illustrious and my pony one which was great doesn't work too well there, so I got an illustrious one, and another, and another, but they're all mediocre, tho they work.
Is there a way for me to combine the new working one and old one from pony and get decent results?

[granite bough]

how do I fix this error

[lucid yoke]

Did you upgrade your gpu?

[lucid dagger]

anyone familiar with derrian for lora training? im running into an issue where I can't launch the ui.

  File "D:\LoRA_Easy_Training_Scripts\main.py", line 69, in <module>
    main()
  File "D:\LoRA_Easy_Training_Scripts\main.py", line 50, in main
    apply_stylesheet(
  File "D:\LoRA_Easy_Training_Scripts\venv\lib\site-packages\qt_material\__init__.py", line 268, in apply_stylesheet
    stylesheet = build_stylesheet(theme, invert_secondary, extra, parent)
  File "D:\LoRA_Easy_Training_Scripts\venv\lib\site-packages\qt_material\__init__.py", line 127, in build_stylesheet
    stylesheet.from_string(template)
UnboundLocalError: local variable 'stylesheet' referenced before assignment
Press any key to continue . . .```
this is a fresh windows installation and a fresh installation of derrian.

it randomly started happening yesterday. tried some stuff to fix it, got frustrated, reformatted my c drive, reinstalled python, git, derrian, and got the same error

both the windows install and the manual install produce this error

installing on other drives also produces this error

someone else did actually. im not the only one with this issue but the responses haven't gotten either of us anywhere

oh. i wasnt aware this is what they ment by ticket

?

i thought by "ticket" you meant create an issue on the github

i wasnt aware there was a ticket system or whatever for this server, i dont post here very often

you posted that already

if you had the same issue how did they fix it?

[lucid dagger]

I was able to fix the issue on my own without the help of these weird ticket things or whatever y'all were tryna force on me.

Steps necessary to fix this issue were:

Open CMD in \LoRA_Easy_Training_Scripts\venv\Scripts
Drag Activate.bat to the open CMD context window and hit enter
pip install qt-materials==2.12

[ornate elk]

whats youe GPU?

[nocturne sonnet]

I'm trying to generate a Wan2.1 video using kijai I2V 480p workflow in comfyui. I'm able to start generating but it seems to get stuck at the 'WanVideo Sampler' node at 71% complete. the last block of text in the cmd window says:

  0%|                                                                                           | 0/25 [00:00<?, ?it/s]```

I am running a RTX 4070, 32gb of DDR4 ram and an i7-9700K. I can see on task manager that around 25gb of ram is being used (despite nothing really happening) but only around 3% of my GPU is being utilized with some occasional spikes. I have tried reducing the image quality and the number of steps but it always gets stuck at the very start of the sampling step.

what do you mean?

[granite bough]

When I open the web up bat

[granite bough]

Amd rx6950

[granite bough]

How do I do that

[lucid yoke]

You probably didn't then, no you can ignore my comment

[ornate elk]

then you need to follow the AMD Zluda Guides.
Check the first link of the pinned messages of this channel.
Either go for Auto1111 with Zluda or Forge webui with Zluda.
Feel free to ask me if any step isnt clear

[granite bough]

Ok ty

[chrome rampart]

Hello all. Anybody played with HiDream-I1 model in comfy yet ?

[fossil terrace]

Im actually devestated. Spent the entire weekend training a dreambooth lora for SDXL 1.0 via kohya and these are my results. the last is what I wanted to achieve. I could literally cry right now, 2 working days went into trying fine tuning the lora

Anyone has ANY idea what led to stable diffusion not even being able to display anything coherent anymore?

I might actually like, buy someone a coffee here or smth if they can help me troubleshoot this. I cant believe i spent 2 days on this garbage.

was the mistake using dreambooth in the koyah tab, instead of the regular "lora" tab?

[lucid yoke]

Iirc dreambooth isn't recommend
Thankfully you still have your dataset, how were the images tagged? What checkpoint were you using? What settings did you have

How many epochs, steps etc

@fossil terrace

[fossil terrace]

1400 steps, 2 epochs, i unfortunately closed kohya already because I was so pissed haha.

wym by how were the images tagged? I provided a .txt file with a short explanation for all of them and kohya read them successfully, if that what you mean @lucid yoke

learning rate was set to 0,00001

optimizer was adafactor

[fossil terrace]

and if by checkpoint you mean which pretrained model i used, the model was stable diffusion xl 1.0

[lucid yoke]

i useually have 3000 steps and 20-30 epochs for a style lora

for a finetuned model, not too familliar for base sdxl

[fossil terrace]

claude told me the steps could have actually been too much, because it overfitted or smth. would u recommend me training a normal lora instead of dreambooth? i just want the vintage art style from the pics above

i didnt touch the noise settings too, was this a mistake?

[lucid yoke]

hmm i useually dont mess on koyha since i cant get it to work on my new card but i really recommend following a actual tutorial instead of listening to ai

a lot of the help seekers came due ai "guides"

https://civitai.com/articles/5545/pony-realism-lora-training-and-preset this one seems promising

https://civitai.com/articles/1824/my-lora-workflow

id go for the bottom one since its more detailed

[knotty turret]

anyone can give me tip on how to make face remain the same on i2v using wan 2.1 model

[subtle quarry]

is illustrious heavy, and is 8gb of VRAM enough? running a 3060ti and it's not happy with it.

this is also the first time i've run AI on this computer, so i'm not sure if it's broken in.

@pure tusk

sorry to ping, but hacked account.

no idea who to report it to, so sorry if the wrong person was pinged

[pure tusk]

all good, purged

[subtle quarry]

There also appears to be some from yesterday in the artisan channels

[ornate elk]

You need --xformers --medvram-sdxl to get the best performance with 8gb vram

[subtle quarry]

..well i had --xformers LMAO aight. i have other tags, --no-half-vae

is --no-half-vae redundant here

[ornate elk]

You can leave it in

But add --medvram-sdxl

Its essential for the speed with sdxl/pony/illus models

For 8gb vram only

[subtle quarry]

got it, ty

it got halfway through to the hires part and then shot up to 12 minutes and i'm like "..that don't seem right"

[ornate elk]

With hires enabled make sure to set the hires steps to 10

[subtle quarry]

Gotcha, gotcha

that's niiiiiice.

speedy

[knotty turret]

hey CS1o do you understand about wan 2.1 settings trying to find way to avoid this idk the name "noise" on hair on output

[ornate elk]

And you have to increase the resolution a bit. Sdxl and pony/illus are trained on 1024x1024

768x1024 will work too

[subtle quarry]

Yeah, I was starting out small just to make sure my pc wouldnt explode. 32gb of ram is good though, right?

[ornate elk]

Can't help with wan sry

[knotty turret]

ok

[subtle quarry]

is comfyui better for illustrious? i dont know how to do usdu (sd upscale) in a1111 and it seems to be not having a good time with that

...then again this is just xl, so it probably doesn't need the upscale.

[lucid yoke]

it can use a upscale, A1111 works but its better for sd.15 imo, Forge (while having obfuscated code) is better optimized for XL but i really recommend SwarmUI

[subtle quarry]

I shall give swarm a try... hopefully it's not terrible to learn..

[lucid yoke]

its pretty easy
https://swarmui.net/

[ornate elk]

you should get the tiled vae extension too if you want to upscale sdxl/illust models

it prevents the out of vram errors and fastens the last step

[subtle quarry]

for swarm?

[ornate elk]

for auto1111

[subtle quarry]

oo

i'll give swarm a shot

[gritty hawk]

Hey guys! Just found the channel... wanted some insight here.

I tried setting up SD locally before, I did get it up and... "working"... it was taking about 20+ minutes to generate 2-4 images at a time.

I'm kinda wondering now though if its because I just wasn't doing something right and quit to early. Can someone smarter than me tell me if that experience was accurate (generation time wise) or if its worth another go?

System and Hardware Info:
OS Name: Microsoft Windows 10 Pro
Version: 10.0.19045 Build 19045
System Type: x64-based PC
Processor: AMD Ryzen 9 5900X 12-Core Processor, 3701 MHz, 12 Cores, 24 Logical Processors
Graphics Card: NVIDIA GeForce RTX 3070 8GB
BIOS Version/Date: American Megatrends Inc. 5013, 3/22/2024
BIOS Mode: UEFI
BaseBoard Manufacturer: ASUSTeK COMPUTER INC.
BaseBoard Product: TUF GAMING X570-PLUS (WI-FI)
Installed Physical Memory (RAM): 128 GB
Total Physical Memory: 128 GB
Available Physical Memory: 115 GB
Total Virtual Memory: 147 GB
Available Virtual Memory: 128 GB
Kernel DMA Protection: Off
Virtualization-based Security: Not enabled
Hyper-V - VM Monitor Mode Extensions: Yes
Hyper-V - Second Level Address Translation Extensions: Yes
Hyper-V - Virtualization Enabled in Firmware: No
Hyper-V - Data Execution Protection: Yes

[ornate elk]

hey, which webui did you set up?

[gritty hawk]

ComfyUI if I remember correctly

[ornate elk]

20 minutes means something is wrong xD

which workflow did you tried?

[gritty hawk]

Which is absolutely possible. I thought I was really ontop of tech related stuff, then I got into this and feel uh... special.

workflow...I tried... I... got python to work! and got it installed... and found a base model... and... you know when I type it out it seems a lot more obvious that it couldn't have been setup right, especially after playing with civit for a minute lmao

I had no idea what workflow was at the time (bout a month ago)... I still don't entirely, other than that its a roadmap of sorts of how you're putting the models etc together

[ornate elk]

okay, so the issue was a month ago? what about now?
WIth an 8gb vram gpu you should probably take a look ate Forge webui, its better optimised

and has an easier usability

comfyui at the beginning is like science xD

[gritty hawk]

Ok. I'll start looking stuff up for that. I'm using CivitAI right now, its a web based platform you can piece models together and gen thru them with flux, illuminate, Pony etc.

I usually create stuff with illuminate, sometimes pony, maybe 1 or 2 Loras... is that practical to do with my hardware? What image generation times you think are practical?

[ornate elk]

yea with Forge you can also use these models. Flux is a bit on the heavier site so stick to 6gb (sdxl/pony/illust) models.
Times should be around maybe 30-40 seconds, but it depends on the resolution, upscaling, loras etc

[gritty hawk]

oh... well shit that's not bad at all. Ok.

[late sigil]

Anyone have a solid strategy for refining generated images (improving details, resolution, etc.)? I've seen a lot of people mention img2img and other methods but I'm not sure where to start at

[gritty hawk]

Its wild to be scraping by with this machine, I thought i specced it well for work doing photo and video editing but then here we are

[ornate elk]

its good for normal editing sure. but ai stuff relies on the gpu vram amount mostly. more is always better in that case

[gritty hawk]

How are people accomplishing more VRAM? Duel cards?

[ornate elk]

nope, just buying newer gpus with more vram

like 12-16gb is a good start. Enthusiasts go for 24-32gb

[subtle quarry]

i got 8gb of vram in a 3060ti, a regular 3060 has 12gb of vram.
Depends on if you're doing it for a business, or for fun.

[ornate elk]

hey, which webui do you currently use? img2img is a good way to refine it, but first you should use hires fix in txt2img

[gritty hawk]

Both. 90% personal 10% bizz right now but mostly because I'm too novice

[subtle quarry]

Start out with the smaller models like cs1o said, because the bigger models get complicated. Depending on the model, you will have to prompt it differently. Pony/Illustrious relies on danbooru tags (anime image hoster website, both sfw and nsfw), and xl does that standard sentence prompting.

[gritty hawk]

I've gotten better with prompting, ChatGPT has been a godsend for that (esp with the custom apps it has).

[subtle quarry]

If you're as average/novice as I am, xl sentence prompting is kinda difficult for me to get right, the booru tags are easier--though it requires a really specific term each time and it doesn't know how to deviate from itself very well

and yeah, chatgpt is a godsend in that regard.

[gritty hawk]

I had a client ask if I could restore an old photo, they got a high resolution scan of it, its over exposed but you can still make out the detail (like the shirt is faded as hell but you can still makeout the pattern and that it is supposed to be red).

The realism models from pictures I've seen posted are getting... really good. Is it a practical venture to utilize those models to restore a photo like that?

[ornate elk]

make sure you have the booru tag autocompletion extension installed, it helps a lot

[subtle quarry]

quick question, does illustrious have controlnets for it yet or do i just use the sdxl ones?

I know pony was so awful with sdxl controlnets that it was a loss to even try

[ornate elk]

there are no specific ones currently

cotnrolnet union++ promax worked good with pony

[subtle quarry]

should i just use the sdxl ones then?

for illustrious

[ornate elk]

you can try but they might not work at all

i can do a quick test with union++ brb

[subtle quarry]

you're a goat, ty

[late sigil]

im using sdnext right now

also any online stable diffusion recs cuz this is ramming my pc lol

[ornate elk]

whats your gpu and ram?

[ornate elk]

doesnt work :/

[subtle quarry]

ahhh damn okay. How do i add extra upscale things to SwarmUI? or even...switch to Upscaling

[ornate elk]

no clue sry xD im not using comfyui or swarm

[subtle quarry]

ooof okay that's fair, ty!

[late sigil]

AMD Radeon RX 6700 XT 16 GB

[ornate elk]

with 8gb vram your best with forge or auto

[subtle quarry]

swarm seems to be working great so far, another user recommended it to me

[ornate elk]

yea its good as its using comfyui as backend, but also has a better ui

[ornate elk]

ah okay, do you use zluda or directml as backend?

[late sigil]

directml

[ornate elk]

oh then i would recommend switching to zluda and to switch to Forge or Auto1111

you will have much more fun with that combo

[lucid yoke]

ah refiner upscale is upscaling

[subtle quarry]

oooh got it. How do I add another one? Just drop the file into the 'upscaler models' folder?

upscale models... whatever it's called

[lucid yoke]

i have mine like these for JUST upscaling, no refining

[lucid yoke]

yes! SwarmUI\Models\upscale_models

[subtle quarry]

epiiiiic ty

[ornate elk]

@late sigil you can checkout my AMD Install Guides and feel free to ask for any questions:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
Make sure to follow a Zluda based Guide, as Zluda is like 2-4 times faster than directml and using less vram at the same time. With that you can use sdxl models easily and also upscale images larger

[late sigil]

thank you! 🙏

[ornate elk]

np, with 16gb ram you also need to adjust your windows Pagefile. but you can do that after the webui install

[vast pulsar]

thnx, we got ti working

[ornate elk]

Perfect np 🙂

[crystal nacelle]

hello again everyone. i mentioned i was having some issues with my a1111 the other day. i wanted to post the full log to get help. my issues seem to be localised to img2img tab: generate stops halfway, and then eventually pressing generate does nothing. eventually i get a connection error sometimes. here is the full log, it might be abit long sorry

im seeing connection error only comes up when i use control net. ill try reinstalling control net. i get a "1006" in my cmd when i try use contorl net. img2img gen error hoowever is prevalent before using control net,

[subtle quarry]

..uh?? what am i doing wrong.

..is it because i'm missing the preprocessors...

...yeah that was it.

[subtle quarry]

I found a model for controlnet openpose, but it's not working...
https://civitai.com/models/1359846/illustrious-xl-controlnet-openpose

i also have no idea how to use any of the other controlnets because they are WAY too specific and don't seem to allow for many options.

[prisma moss]

Why is my 3070 taking 12s/it for training? ITs gonna take almost 6 hours to train, 1poch, 2 batch, 1600steps total, 768,768 (altho my training images r 512x512, maybe i should change that back?)

SDXL

Everything is installed correctly

Alrto i havent touched like the configs or other arguments, idk if thats a given?

I also idk if i should select sdxl or 1.5 in the settings, im basimg the training off of perfectillustrousxl

[ornate elk]

Set batch size to 1, you dont have enough Vram for batch size 2

[prisma moss]

working on it as we aspeak

[prisma moss]

got it down to 4.2s/it on sdxl

that should be decent?

[ornate elk]

This error can be ignored.
The model should still load

[ornate elk]

Sounds okay

[prisma moss]

bet

[ornate elk]

How much RAM do you have?

Okay, then you need to increase your Windows Pagefile.

For C: drive to 16000min and 24000max

And disabled for any other drive

Then reboot the PC

[lucid yoke]

Did you manage to fix this?

Every version of sd / flux needs their own controlnet witch you can find on civitAI

Illustrious needs their own one etc

[subtle quarry]

No, but I'm also not sure I even set up the thing right.

Trying to "preview" in swarmUI prints a black image below it, but something tells me it's not supposed to do that.

A black image with the dimensions of what I'm about to generate.

[lucid yoke]

hmm what are you trying to do

gimme like 20 min and i can hop on vc

[subtle quarry]

I aint got a mic and my oc is weird

Pc* jfc

Mobile is a menace

My headphones have a mic but my pc is a split input vs output, so it can't be on the same wire for some reason ig.

[subtle quarry]

I'll get a screenshot here in a bit. Or a workflow if you know how to get one

[lucid yoke]

just a screenshot of what your trying to do is ifne

[delicate urchin]

What is the reason for black video outputs with WAN?

[lucid yoke]

try sampler euler instead of uni_pc?

[delicate urchin]

Euler gives blurry result

pixels other than black for a change

[lucid yoke]

cant help but notice your using the 720p model but having a value at 480p

should be fine i suppose

scheduler normal? not sure

cant check my workflow rn

[delicate urchin]

Strange results with scehduler:normal aswell.

Downloading wan2.1_i2v_480_fp16 to try

[delicate urchin]

same issues persist with using wan2.1 i2v 480 fp16

[lucid yoke]

Hmm personally i use swarmUI for video gen and my workflow looks slightly different

[delicate urchin]

unipc = black

euler = blurry

tried rebooting, few different comfy wan flows and save video nodes

Few on the internet seem to have this problem

Reddit post that has similar issue says to use correct VAE and set dtype to default. I got both the correct vae and set it do default

does sage attention have anything to do with this?

[spring garnet]

hello guys, can someone maybe help me since im a big noob in this? Im trying to use animatediff but It dosent really work. Im unsure which SD Version I need for the newest animatediff & i also dont know which one I have

when I hit generate it just generates another pic

[fierce hollow]

Has anyone dealt with HiDream models? Where besides ComfyUI can I run?

[ornate elk]

Can you show the full cmd log?

Also which model do you try to load and how much RAM do you have?

[prisma moss]

Why am i not seeing the 'show extra networks' button after applying kohya exension?

oh ok its a bit different

[autumn timber]

does anybody know a fix for this? i tried toggling the setting but it didnt help

[ornate elk]

whats your gpu?

and which model do you try to use?

[autumn timber]

3070 ti

[ornate elk]

then your probably trying to use a lora (small file) instead of a model (2gb and larger file)

[autumn timber]

im embarrased to say exactly what model cuz its furry stuff but here

[ornate elk]

yep thats a lora

it goes into models/lora folder

[autumn timber]

oh ok

thx

[ornate elk]

and has to be used together with a base model

its like an addon

[autumn timber]

would a base model be called a checkpoint or an embedding

[ornate elk]

checkpoint = model

[autumn timber]

gotcha

[ornate elk]

always check the type and the base model here

then you know with what its compatible and what it is

[autumn timber]

i only sometimes dabble with ai so im not huge on the verbage

thanks

[ornate elk]

no problem

[ornate elk]

yes found them! Illustrious and NoobAi controlnet models:
https://civitai.com/models/1359846/illustrious-xl-controlnet-openpose
https://civitai.com/models/929685?modelVersionId=1239319

[subtle quarry]

I tried the openpose one and it didn't work for swarm, but I'm sure I was doing it wrong... was gonna get back to euro but the day got unexpectedly occupied

[ornate elk]

i tested them rn and they work

[subtle quarry]

Then yeah, I was doing it wrong with swarm.

@lucid yoke Won't be at computer again for a minute, but here

[lucid yoke]

Wrong controlnet model

[subtle quarry]

No, it's the one from up here

It's just named wrong and I didn't rename it

[lucid yoke]

https://civitai.com/collections/6541109

O cs1o already linked them

But yeah they Generally have the name in its title

[ornate elk]

the image you used is an already "processed" pose. so you cant use a preprocessor on it.
set it to none and it will work

[lucid yoke]

Also that

[subtle quarry]

It won't let me

[lucid yoke]

What cs1o said

?

It should

[subtle quarry]

Unless I'm stupid

I ..I might...be stupid

[lucid yoke]

You don't need to preview as its already processed?

[subtle quarry]

..... I thought the preview worked differently.

[lucid yoke]

Nah preview is just running input image and showing the output of the skeleton so to speak

[ornate elk]

it should be possible but i didnt used swarm so idk

[lucid yoke]

Disable the pre processor and it should be fine

Otherwise i recommend joining the swarm discord as others there are much more knowledgeable then me

[ornate elk]

yea disable it

[subtle quarry]

I'll give it a shot when I'm home long enough but since I got you here, is there any way to keep an outfit in swarm that was unintentional? I've tried asking there and I got nothing.

[lucid yoke]

Hmm maybe thru an ipadapter but im not good with that

But even then its tricky

[subtle quarry]

Might be NSFW so spoilering it

"Green hoodie" was the prompt and it took it in the other direction but dope af

[lucid yoke]

Your best bet is to try and prompt out everything you see in this, green hoodie, open center, sleeveless arm, gloves, bra, (forgot the thing on the arms)

But its gonna be different

[subtle quarry]

That's fair

[subtle quarry]

Good tags btw, forgot half of them

Tyy

[lucid yoke]

Could try bare arms idk

[slender crescent]

I do not know what I am doing here as this is my first time. I am trying to run this on an RX 6700 XT on Fedora 41.

[subtle quarry]

I think it's arm_warmers

Not sure tho

[ornate elk]

why zluda? rocm works native on linux. also try use the Auto1111 repo:
https://github.com/AUTOMATIC1111/stable-diffusion-webui?tab=readme-ov-file#automatic-installation-on-linux
Or this:
https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu

[brittle lantern]

@ornate elk im ready to start from the beginning

[ornate elk]

alright,
here are the Guides:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

Let me know if any step isnt clear.
Try the Forge Zluda or the Automatic1111 zluda guide

[brittle lantern]

I downloaded the zluda file to put in c driver but when i open it it doesnt show library folder its just DLL files

oh im silly

skip that

[azure magnet]

how to fix it:An outdated version(<1.16.9) of the comfyui-frontend-package is installed. It is not compatible with the current version of the Impact Pack.

[slender crescent]

Getting this right now with ROCm:

RuntimeError: HIP error: invalid device function
HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3
Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.

I set these environment variables, gfx1031 as reported by rocminfo:

export PYTORCH_ROCM_ARCH="gfx1031"
export HSA_OVERRIDE_GFX_VERSION=10.3.1
export HIP_VISIBLE_DEVICES=0
export ROCM_PATH=/opt/rocm

full rocminfo report:

[brittle lantern]

@ornate elk i have this now im on the launch bat after adding the stuff to path

[ornate elk]

Can you show the path settings?

[ornate elk]

Okay 2 things wrong here.
Remove the c:zluda

Then click on Path and click edit and screenshot again

Wait

Just remove C:Zluda

[brittle lantern]

Oh

[brittle lantern]

done

[ornate elk]

Then click on Path and click edit and screenshot again

[ornate elk]

Nope close and click on it and then on edit again

It should look different

[brittle lantern]

i do that but it goes back to that same place on the image

oh wait question

u on about for admin variables

or system?

i might have mixed them up so

[ornate elk]

System

[brittle lantern]

yeah when i click on that it comes up with the image one

this is admin and it looks the same how i used to do it back in windows 10

[brittle lantern]

but system one looks like this

[ornate elk]

Yea thats how the bottom one should look like too

[brittle lantern]

i dont know then

[ornate elk]

I guess you replaced the entire Path setting

You have to recreate it

[brittle lantern]

ive clicked on the other ones in system they all show the same as that one

see

[ornate elk]

Do you double click or edit?

[brittle lantern]

both do the same

[ornate elk]

Thats really strange

[brittle lantern]

yeah 😦

for some reason when doing this i have random issues with it

last time i had low vram error when u said it shouldnt be that

and couldn't fix it now its this

[ornate elk]

The thing now is more important to fix as it effects windows

[ornate elk]

is there a second path entry when you scroll upwards?

[brittle lantern]

theres hip_path

and hip_path_62

[ornate elk]

And not single Path?

[brittle lantern]

no

it wasnt there went i went in it

[ornate elk]

Okay then rename the PATH to Path

Dont touch the hip_path entries

[brittle lantern]

the path entry i showed on the first SS is names "PATH" or do u mean something else?

oh i cant read im sorry

[ornate elk]

Then as value add C:\ZLUDA;%HIP_PATH%bin

[brittle lantern]

u mean path without caps

done

[ornate elk]

Then click okay and close. Then reopen the path settings

[brittle lantern]

ah it works now

that menu thing came up

[ornate elk]

Its a list now right?

[brittle lantern]

yes

[ornate elk]

Perfect

[brittle lantern]

what was the issue? i didnt fully understand what happened

[ornate elk]

Okay now you have to add some essential values again.

Add all these values at the exact positions.
But Skip the java path one at the top.
And skip ffmpeg

At the black bar add your windows username

[ornate elk]

The issue was that somehow the windows path settings got removed
And it wasn't a list because it was only 1 entry as value in there.
We added a ; to gave him two values. Then the list appeared again.

Screenshot when your done, ill recheck.
and then click OK at every open window to close them.

If not the settings won't get saved

[ornate elk]

Perfect

Now click always Ok

Then you have to reboot the PC. And then recheck If the variables are still there

[brittle lantern]

ok i do it now

[ornate elk]

Then you have to delete the venv folder in forge and then relaunch the webui-user.bat

@fair furnace no need to delete your messages. If your question is technical then your at the right place

[brittle lantern]

It’s downloading the stuff now

I’ll let yk if anything bad happens or good

bruh

[ornate elk]

Connection timeout

Make sure your on ethernet

[brittle lantern]

i am dk what happened

[ornate elk]

Then try relaunch it

[stable bough]

Total beginner here. I set up ComfyUI but it’s pretty daunting and I’m having trouble getting even moderately good quality results.

I don’t know much about A1111 or Forge, would those be better options for a fairly non-technical person?

[ornate elk]

Hey welcome. What's your GPU?
And yep Auto1111 or Forge are much easier to get into if your new to image generation

[stable bough]

GTX 3060 ti with 8gb

Is there one you’d suggest over the other?

[ruby yoke]

something with more than 8gb vram

[ornate elk]

Forge is more recent than A1111. And works good with 8gb vram. Try that.

[ornate elk]

Ohh Git

Git had also its variable added to Path.
Best is to just uninstall GIT and then reinstall it.
Then it should work when you relaunch

[ornate elk]

Any proxy or vpn running?

[brittle lantern]

myabe proxy

[ornate elk]

Seems like thats a problem for the webui

[brittle lantern]

changed the dns its fine now

so first gen will take a while correct?

[ornate elk]

Exactly. The CMD will show
Compilation, please wait....
It will take 15-40 minutes.

[brittle lantern]

so should i download a 1.5 model and then after first gen del it

or download my normal illus models?

[ornate elk]

For the first gen better use a 1.5 model

But if you only have illus, thats okay too

Shouldn't really matter as long as you have 32gb ram or 16gb with a big pagefile

[brittle lantern]

i need to download my first model

[brittle lantern]

i have 16gb where do i check my pagfile and what do i set it to?

[ornate elk]

Ah okay, for sdxl/pony/illus you then need to increase the pagefile.
Here is a how too:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html

Make sure the Pagefile is enabled only for the C drive. It doesnt matter where the webui is located.
Then set it to custom size
16000min and 24000max.
Make sure its disabled for any other drive.
Then save and reboot the PC.

[vast pulsar]

anybody know what could be causing these? i don't speak tech sadly

[ornate elk]

Can you share the full cmd log?

And what have you tried to do

[vast pulsar]

this is the rest. When at 99% instead of finishing the picture this error shows up, only idea i got is changing arguments arround and try to see if any of hem is causing this

[ornate elk]

What settings did you used?

anything different then the last time it worked?

[vast pulsar]

no, not really

[ornate elk]

Txt2img or img2img?

Also please provide the full cmd log

[brittle lantern]

i downloaded the oldest 1.5 model so i can do the first gen thing

[ornate elk]

Oh know I now. You added --xformers to the webui-user.bat

Thats only for nvidia

[brittle lantern]

thats fine? ill del after it does that

[vast pulsar]

i seeeee

ok

i'll try to

take that off

[ornate elk]

Alright

[ornate elk]

You also have to delete the two xformers folders inside the venv folder.

[vast pulsar]

xformers? what files exactly?

[ornate elk]

Folders

Named xformers

[vast pulsar]

oh ok, thanks

[brittle lantern]

@ornate elk can i set minimum steps and res or i have to use the normal res and steps?

[ornate elk]

Use the default values

For the first image

[brittle lantern]

Yep it’s doing the compilation

After this I’ll del that model and then download my illus with loras and stuff will comp do every time I do a driver update a normal driver update on adrenaline?

[vast pulsar]

got no xformer folders so it's all good right?

[brittle lantern]

I like ur name trickery there XD

[ornate elk]

Yes compilation is done after updating adrenalin drivers, updating zluda files or when torch updates

[brittle lantern]

Torch auto updates when I open up bat or no?

[vast pulsar]

it is a very based name indeed 😌

[ornate elk]

And dont forget to do the Pagefile thing

[brittle lantern]

I did it

[ornate elk]

Nope

[brittle lantern]

Just did restart now

[ornate elk]

Torch version is fixed

[brittle lantern]

And doing the comp thing

[brittle lantern]

Ok will it let me know there’s an update like pip does?

Or is that a manual website check thing

[ornate elk]

Nope. It will only ever update when the devs change it or you manually update it. So dont worry about that

You can add
Git pull (in an empty row)
To the webui-user.bat if you want to Automaticly update If there is an update available

I'm off now as its very late here.
Gn

[brittle lantern]

Good night thanks for the help

[ornate elk]

No problem

[brittle lantern]

this issue its back

I changed the gpu weights to 6037 seems like that fixed the issue

only fixed for the first gen

2nd one decided to do the same

[stone fog]

Gettomg this with no load up

whoops wrong screenie

[ornate elk]

Hey, your missing the launch args in the webui-user.bat.
Also you should uninstall HIP SDK 6.1 and install 6.2 instead for the 7000 series GPUs

[stone fog]

What launch arcs do I need? Also do you have a link to that download? Is it simple as just drag/drop? I followed your guide way back

[ornate elk]

Everything is listed in my Guides.
Checkout the first link of the pinned messages

--use-zluda --skip-ort are needed

[stone fog]

Is this a new addition? I haven’t changed anything since I installed it a while ago, or so I have to start from scratch?

[ornate elk]

Nope not new

[ornate elk]

You dont need to redo it from scratch

[fair grove]

yall question

is this the official comfyui github

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

nvm it is

[ornate elk]

Which settings did you used?

Resolution etc

[keen mulch]

Hi, I've been tearing my hair out trying to get automatic1111 working on an amd Vega 56 on fedora linux, and at this point I want to check if this is even possible at all?

No matter what I do the display turns to confetti when I try to generate an image

And all the guides online seem to get confused themselves as to whether they're for linux or windows, which makes the whole process maddening

[ornate elk]

hey, it should be possible

cant help much on linux but you may need a rocm version that does support your gpu (i guess 5.7). Or you need to add set HSA_OVERRIDE_GFX_VERSION=10.3.0

[keen mulch]

thanks! where do i put this?

[ornate elk]

in the webui-user.sh

[keen mulch]

hm, i've already put some things in there in my attempts, and they use EXPORT instead of set, is that correct?

[ornate elk]

yea then export

[keen mulch]

let's see if this works

hm, now i'm getting this error:

RuntimeError: HIP error: shared object initialization failed
HIP kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3.
Compile with `TORCH_USE_HIP_DSA` to enable device-side assertions.
Stable diffusion model failed to load

...it seems i didn't have rocm installed at all? wtf

at least, i didn't have the fedora rocm-core package installed

welp installing that didn't fix the error anyways

[ornate elk]

damn

[keen mulch]

added "export HIP_VISIBLE_DEVICES=<device-id-for-gpu>", with id from rocm-smi

now it's back to running, gonna try to generate now and see if it crashes

:O it works!

but it seems not noticably faster than running on CPU..

in fact i'm not entirely certain it's using the GPU? how do i check lol

oh yeah, i had to add --skip-torch-cuda-test too, maybe that disables the GPU?

[ornate elk]

hmm i think you shouldnt add that

but im not sure if its needed

whave you tried without it?

[keen mulch]

without it i just get the "model failed to load" error as before

gonna try reinstalling torch to be sure it has rocm support..

[ornate elk]

good idea

[keen mulch]

also newer version

[delicate urchin]

What is ret_mode?

[placid pier]

Hey guys, I'm looking for a laptop that can run a1111 locally. But I don't know what kind of specs to look for. My pc has a GeForce rtx 4060 with 16GB VRAM and I'm looking for something similar to this but as a laptop version.

[random night]

if you have a pc with this then it doesn't matter what specs the laptop is, cuz you can alway stream your pc screen to your laptop via parsec or ultraview

or if you're tech nerd, you can follow this https://www.reddit.com/r/StableDiffusion/comments/xtkovu/is_there_a_way_i_can_share_my_local_automatic1111/

this way even your phone can access to automatic1111, and take the resource of your computer to run

[cyan quarry]

anyone got an idea where i could discuss lipsync swapping techniques?
I tried LivePortrait, it is not too bad but I wanted to explore other options like LatentSync.
I have troubles installing it...

[placid pier]

I need to take the laptop to an event I go and showcase something. I can't take my pc with me, can I still access it from there? And would that mean any laptop works?

[random night]

yes, any laptop works, as long as you have internet

or even tablet

i recommend parsec as your streaming service, it's pretty much free https://parsec.app/

[brittle lantern]

832x1216
30 steps
Euler A
Anime 6b hi-res 10 steps 1.4-1.5x and 0.45 denoising CFG7

The same I use on my MacBook with 16gb ram m2 pro

[placid pier]

Ah I see. So my pc at home as to be turned on and connected and the laptop as well? And parsec installed on both, so I can get access? Does someone have to operate on my pc at home? My mom is probably the only one at home then and she has no idea 🤣

[random night]

you can remotely turn on your computer

https://www.youtube.com/watch?v=03rQdZOlRAc

like this

https://www.pcmag.com/how-to/turn-on-computer-from-across-the-house-with-wake-on-lan

you can follow this guide for wake on wan

[ornate elk]

okay, and your not running games or wallpapere engine in the background? or multiple browser tabs.
what happens when you set the gpu weight to 6750 ?

[placid pier]

Ohhh okay crazy. 😄 thank you 😄

[random night]

if it too much of the hassle, then simply go to control panel > power option, and set your pc to always on

that's way you don't have to do all the stuff below that

[placid pier]

I see. I'll check out what works best.
Remote desktop itself doesn't work for my pc apparently. It says I don't have the version that supports it 😄

[random night]

you can always use parsec

[brittle lantern]

nope all off and only that tab mainly and i set gpu weight to 6000 it worked fine on the first gen then on second gen gave me that error

[red cobalt]

@ornate elk u here?

im tryna use controlnet

how do i use it tho...

[ornate elk]

have you enabled Never OOM for tiled vae at the bottom? you need to do that when using hires fix

[brittle lantern]

Oh no I didn’t do that I’ll check now what should my GPU weights be?

[ornate elk]

try with 6750 first and then try with 6000

[brittle lantern]

Ok ty

I’ll try now

[stone fog]

Where should I pick up on the guide?

[ornate elk]

just uninstall everything from HIP SDK, then install HIP SDK 6.2
and then add the launch args listed on the guide for your webui

delete the venv and the .zluda folder (boath are in stable-diffusion-webui)
Then relaunch the webui-user.bat

[slender crescent]

I am getting this with ROCm on Radeon RX 6700 XT
I have the following environment variables set:

PYTORCH_ROCM_ARCH="gfx1031"
HSA_OVERRIDE_GFX_VERSION=10.3.1
HIP_VISIBLE_DEVICES=0
ROCM_PATH=/opt/rocm

just set the first two to "gfx1030" and 10.3.0 respectively and it is now doing something

Apparently 12GB VRAM isn't enough

[elder raptor]

i need to pay for use ?

[random night]

umm, no, it's free, as long as you have the hardware

[elder raptor]

oh

[slender crescent]

a lot of it is open source

aight, giving up on this for now

[ornate elk]

try add --medvram-sdxl --opt-sub-quad-attention --upcast-sampling to the webui-user.sh commandline args

[knotty turret]

can someone help me wanna add lora for that workflow anyone know how to do it? i know there another workflows that already have lora option but that one worfklow use some new things wanna test

[subtle quarry]

Iirc, you have to plug the lora in beofore the prompt

so whichever one is your main prompt, plug it in before that.

after the checkpoint model, before the prompt

@knotty turret (sorry, forgot to reply ping)

It's been a second since i've used comfy, i've never used it for WAN though.

[knotty turret]

ok let me try

[knotty turret]

hold one got confuse

where i connect model line

[subtle quarry]

You connect the model into the ksampler, i believe.

Again, I've never used WAN so I'm using what I kind of know already.

[knotty turret]

np

maybe?

[subtle quarry]

Lemme try something, hold on

[knotty turret]

ok

[subtle quarry]

tryna use swarmui to make a workflow

[knotty turret]

already have another workflow that have lora option but this one use some stuff to speed process SageAttention + TeaCache + Torch Compile thats why wanna use it

[knotty turret]

ok

[ornate elk]

pls change your "burger" pfp

[subtle quarry]

gave up and looked one up.
It's supposed to look vaguely like this.

A NON WAN workflow.

[knotty turret]

hmm ok

[knotty turret]

why its just burger😭

[subtle quarry]

it's obviously not a burger.

it's one of them stupid face food things.

[knotty turret]

do u think looks cool now?

[subtle quarry]

shoebill birds are neat

favorite zelda animal

[knotty turret]

@ornate elk cant find this node can u help me?

[sacred osprey]

i want to install roop but i get this error Conda environment is empty.

[slender crescent]

I'm using Diffusers directly

I used the CPU offload option but a library WHL could not compile

Forgot which one

I may try this again once I get a PC together I can use as a dedicated server for this sort of thing so I don't have to do it on my main desktop

[ornate elk]

nope sry

[ornate elk]

which roop?

[lone bolt]

i created a lora using forgegym and it gave me 4 different safetensor files each 17mb.. did i do something wrong or?

[stone fog]

Where is the HIP SDK piece?

Sadly I think I've confused myself into a corner

[ornate elk]

Its a programm

Or multiple

Uninstall them normaly like any other programm

In control panel

[stone fog]

Alright the new SDK is installed, and I removed the zluda folder. You said something about arguments?

[ornate elk]

The launch args

--use-zluda --skip-ort

[stone fog]

How do I do that exactly?

[gusty tide]

Having an issue using someones workflow i had installed in the wrong place I guess

A zuda version of comfyui

i was told to come here and talk to CS10

I moved the folder to my cdrive and not sure what im supposed to do from here

[cedar flame]

Need help with trying to make a lora using Derrian's Lora Trainer but I keep getting an error

HTTPConnectionPool(host=): Max retries exceeded with url: /metrics/validate (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x000001D8B7C2C3D0>: Failed to establish a new connection: [WinError 10061] No connection could be made because the target machine actively refused it'))

[ornate elk]

Right click the webui-user.bat and edit it. Add the args to the line commandline_args=

[gusty tide]

yooo