🤝｜tech-support

[ornate elk]

so if you start with a 1080 and 16ram now it will work. But upgrade those two parts later on

[wary tulip]

Thank you I appreciate it

[ornate elk]

No problem

Also it depends on what you want to do with Comfyui

[wary tulip]

Yeah I only want to generate images

[ornate elk]

Okay then your good with that.
1.5 based models will work very good with a 1080 and sdxl based models work okay too.

Means 2-6gb models are fine to use

[wary tulip]

Yeah I want to generate anime images and I want to use the control net so I can control them

[ornate elk]

Yea controlnet is nice and works too

I used a 1080 in my last PC

In the beginning of SD

Also try out Forge Webui then.
Its easier to start with and has some performance improvements

[compact salmon]

way agreed lol

begin with forge to understand the fundamentals (there are even tips when you hover over things)

installing the pytorchs per ur guide rn 🙂 @ornate elk

is this all normal errors to be having on amd?

Successfully installed intel-openmp-2021.4.0 mkl-2021.4.0 tbb-2021.11.0 torch-2.3.1+cu118 torchaudio-2.3.1+cu118 torchvision-0.18.1+cu118
Running main.py with --auto-launch --disable-xformers --use-quad-cross-attention --reserve-vram arguments...
Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.
Total VRAM 16304 MB, total RAM 31861 MB
pytorch version: 2.3.1+cu118
Set vram state to: NORMAL_VRAM
Detected ZLUDA, support for it is experimental and comfy may not work properly.
Disabling cuDNN because ZLUDA does currently not support it.
Device: cuda:0 AMD Radeon RX 9070 XT [ZLUDA] : native.

A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.2.4 as it may crash. To support both 1.x and 2.x
versions of NumPy, modules must be compiled with NumPy 2.0.
Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.

[novel owl]

you want the time to generate a video ?

[ornate elk]

Yes you also need to do the numpy fix. Its explained in the guide as well

But everything looks correct so far

[compact salmon]

sweet ok

missed the numpy fix part ur right

[novel owl]

9070 xt are better on ia of an 7900 xt ?

[ornate elk]

Depends. A 9070xt is for sure faster in ai stuff. But if you try to use a heavy workflow that requires more than 16gb of vram the 7900xtx beats it easily

[novel owl]

3 minuts to generate a video with this settings

[ornate elk]

How big is the wan model file?

[novel owl]

wan ?

2.771 Go

[ornate elk]

Ah okay

[novel owl]

its good ?

[ornate elk]

Seems good, your on a 7900xtx?

[novel owl]

no im on 7900 xt

no xtx

[ornate elk]

Ah okay

[novel owl]

but i have a big overclock

[ornate elk]

Nice how much vram did the workflow used?

I can test tomorrow on my xtx, also overclocked a bit

[ornate elk]

Even while generating?

[compact salmon]

what am i doing in here

i dont see where to add weak security

[novel owl]

when i generate

[ornate elk]

Wrong folder

[ornate elk]

Dont go in .git

[compact salmon]

mb

[ornate elk]

Nice that means you could even use a larger wan model

[ornate elk]

Very nice!

[novel owl]

I assembled it myself

[formal ridge]

For better quality, switch to dpmpp_2s_ancestral / ddim_uniform and thank me later

[novel owl]

7800x3d 7900xt b650 32Go ddr5 6000 c36

[ornate elk]

I have the same CPU and also the RAM amount

[novel owl]

i go to change my ram to the future on 7000 c32

G.skill

at 48 Go

[ornate elk]

Ah nice but does 7000 makes a difference? I thought between 6000 and 6400 was like 1%

Because the GPU does the heavy lifting in games

[novel owl]

on star citizen yes TuT

[ornate elk]

Ah xD

[novel owl]

especially on fps drops

[ornate elk]

I also wanted to upgrade to 64gb but AM4 boards cant handle 4x16gb 6000mhz

So I have to change to 2x32

[novel owl]

AM5 no ?

[compact salmon]

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)

followed your steps to a tee

[ornate elk]

Nope it will downclock to like 4600mhz

And therefore its useless

[novel owl]

AM4 dont support ddr5

and dont support ryzen 7000

[ornate elk]

Ohh autocorrect...

I'm on mobile xD sry

I'm on AM5 too of course

[novel owl]

okay XD

where can i found wan model ?

civit ai ?

[formal ridge]

https://huggingface.co/Wan-AI

[ornate elk]

And you replaced the files?

Can you show me your zluda folder?

[compact salmon]

yes i believe i did

[ornate elk]

Did it asked you to overwrite?

[compact salmon]

no

[ornate elk]

Oh than thats the issue

Where are the renamed copies?

[compact salmon]

there was notihng to replace

it just dragged in

[ornate elk]

Can you go into the torch/lib folder and screenshot?

[compact salmon]

just did it again

i went to overwrite

and it just pasted over

thats D:\Graphics\AI\SD-Zluda\ComfyUI\venv\Lib\site-packages\torch\lib

failed to open now

after pasting those in

[ornate elk]

Wrong torch

Its CPU only

[compact salmon]

wym

i followed your guide

[ornate elk]

Make sure you dont run any bat or cmd as admin when you follow the guide

[compact salmon]

do i restart entire thing? delte venv?

[ornate elk]

Delete the venv folder and relaunch the Start-Comfyui

It should install torch 2.3.1

If you then move the 3 renamed zluda files over it should ask to overwrite.
If it doesnt its wrong again

[compact salmon]

i believe in ur guide im just having issues

ok got u

[ornate elk]

Nobody has confirmed that it works with a 9070xt but it should

The guy with the 9070 yesterday got forge with zluda working from my guide.

So if Comfyui won't work, try forge and we can look at comfyui later again

[compact salmon]

it said foudn no nvidia driver and closed immediately

didnt get to the part where i can try to gen to error

no 127.0 pop up

wait

restarted again

lets see

[ornate elk]

And?

I need to sleep now.
Can help tomorrow again.

[compact salmon]

oh im sorry. it seems like stuff is loading

get some sleep homie!

if it doesnt work maybe ill speak with you tomororw

i wanna try getting forge goign anyhow so ill try to get that complete

Requested to load SD1ClipModel
loaded completely 11779.08583984375 235.84423828125 True
Token indices sequence length is longer than the specified maximum sequence length for this model (82 > 77). Running this sequence through the model will result in indexing errors
0%|

no error so

[ornate elk]

Perfect

It will work

Its compiling zluda now in the background

[compact salmon]

dude ur guide is the only one that will work proper

[ornate elk]

Can take 15-40 mins

[compact salmon]

the other one will not

for a 9070

i tried the xpatients for 10 hours every which way with his troubleshooting, and somneone else no luck

ty!

[ornate elk]

No problem 🙂 glad it worked!

And yea forge is much easier to get working than this.
Comfyui is the hardest guide xD

[compact salmon]

amd doesnt help lol. ill see if i can get forge up n going myself!

have a gn dude

he done did it

[ornate elk]

Thx gn and let me know

[compact salmon]

sounds good ill follow ur guide and lyk if it went well haha

shit we at like 18%|██████████████▊ | 5/28 [01:32<07:04, 18.47s/it] SLOW AF

[cursive hazel]

Are you running a Flux model? With Pony, I get what I feel like is pretty fast? But I tried Flux today, and it was horribly slow. I may be doing something wrong though. I've never really used much Flux before.

[compact salmon]

are u on amd card ?

with 1201 rocm or whatever im on 6.2 rocm

and yes im running flux dev

to forge we go @cursive hazel

[dreamy lodge]

Im on an AMD card using Zluda with SD UI. How do I install comfyUI now. The online tutorials are kinda confusing

[compact salmon]

do u mean forgeui

[dreamy lodge]

no

[compact salmon]

sd ui?

u sure lol

sd ui isnt a thing

[compact salmon]

ah a111

[dreamy lodge]

that yeah

[compact salmon]

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

this guide will get you comfy

if you have a111 working just be careful with install following steps and skip stuff already done (rocm unoffiical update swap, python install, cmake install)

just got forge installed

[cursive hazel]

Yup, I'm on an RX 9070. We were chatting the other day, when I got mine working on Forge and Comfy UI.

I tried running... Idk the name of it. Some highly rated Flux model from CivitAI. And my generating times are slloooow on it.
On Pony, it's fast though.

[compact salmon]

ameteph super easy

oh right right

i tried running the most common flux dev pruned 15.9gb

and it just lagged like crazy like u saw there

slower than a 3060 mobile kinda stuff

[cursive hazel]

I might poke at CS1otomorrow, to see if he has any advice about Flux. I have no idea what I'm doing with things, I just follow his guides

[compact salmon]

i tried for countless hours to follow xpatients but truly seems like cs1 guide just has everything you need for it to work with troubleshooting

ill lyk if forge is working okay @cursive hazel

[dreamy lodge]

whats that cmd command that links folders together. Trying to link my models folder from a1111 to ComfUI's model folder

[cursive hazel]

I tried Flux on Forge and Comfy, both are horribly slow for me.

[compact salmon]

oh you tried flux as well?

[compact salmon]

symbolic linking via

MKLINK in cmd i believe?

MKLINK [[/D] | [/H] | [/J]] Link Target but i just use the app

forgot what its called is it in powertoys

[cursive hazel]

8/8 [11:00<00:00, 82.59s/it]
Is what I got on a 12GB Flux model. 896x1152, Euler Simple.

[compact salmon]

oh jesus

[cursive hazel]

That was in Forge, but I got similar numbers in Comfy.

[compact salmon]

i recommend undervolting card on adren at -50mv helps with fans/coil whine and performance overall after watching countless testing's once it was in my hands lol

[dreamy lodge]

So i followed the installation for ComfyUI but whenever it told me to copy the 3 renamed dlls in the ZLUDA folder, the destination it told me to paste into wasnt there. In the site-packages folder there is no torch folder.

[dreamy lodge]

So whenever I run Start-Comfy so it can create the virtual environment, it fails somewhere along the line and doesnt finish but it closes before i can even read it. Anybody know what i can do?

[formal ridge]

extra_model_paths.yaml.something file in the ComfyUI folder. rename it to end in just .yaml, then modify the file.

[solemn sparrow]

hii
im new here , can someone please guide me how i can make some images or small video for my content ?

[tribal epoch]

Can stable-diffusion-xl be used with MagicAnimate?

[ornate elk]

Nope I dont think so. The pretrained model is for 1.5

[ornate elk]

Hey, what's your gpu?

[ornate elk]

Delete the venv folder Abd relaunch the bat.
Also don't launch it as Admin

[ornate elk]

@cursive hazel
Don't run a 16gb flux model on a 16gb GPU.
You still need vram for the actual image generation that also depends on the resolution and prompt length etc.
Best for Flux is to use the smaller GGUF models.
Also use the faster t5 gguf text encoder instead of the fp8 one.
For example get a 6-11gb gguf flux model it will be faster.
Here is the t5 gguf clip:
https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main
Download the Q8_0 (5gb)
For Flux gguf models check here:
https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
Go for Q4-Q8. Maybe Q6 will work the best.
Also NF4 flux models currently won't work on AMD.

I can send you a flux workflow later

[cursive hazel]

The file size itself is close to 12GB. But it might be the NF4 or something, I honestly have no idea. I just picked a highly rated model from CivitAI. It's uhh.. "Acorn is spinning FLUX"
But I will of course try what you've said here.

I did hear a little about GGUF, but have no idea what it all is. It seemed like an alternative method to doing FLUX stuff, and I was trying to keep it simple. Which of course just meant I stumbled into problems for myself, as usual!

And once again, thank you for your help!

[cursive hazel]

I'm off to bed for now, but I did try to set things up. Maybe I failed somewhere again?
10/10 [07:26<00:00, 44.64s/it]
Here's my workflow. I tried using a different prompt at the top too, and nothing really changed. (Leaving the bottom prompt blank when I filled in the top. From my understanding, the top is tag based, and the bottom is more natural language.)

Maybe I did something wrong again? Or maybe the 9070's just hate Flux for now.
It's definitely faster than what I was getting! Still a bit slow I think though? 7 minutes for 10 steps? I tried 4 steps and the image was hazey.

[ornate elk]

Did you monitored the vram usage while it genned in taskmanager?
Would be interesting to see how much it takes.
In any way flux is very resource demanding and the gfx files for 1201 are currently not optimised.

7 minutes is still very long

[cursive hazel]

Vram was at 14.-something out of 15.5gb. I can't remember exactly, but I'd say 14.7?

[ornate elk]

So it never spilled into shared vram? Even at the VAE step at the end?

[cursive hazel]

I didn't pay attention at the VAE step tbh. It was using about 2gb of shared vram iirc. I can run it again tomorrow and check for sure. I'm on my phone atm.

[ornate elk]

Yep check that again then. Because the moment it overflows the vram and uses shared vram it will get slowed down massively

[cursive hazel]

If that is the case, I should step down to the Q4 instead of Q6?

[ornate elk]

Yes you can try that too

And compare

[solemn sparrow]

Rtx 3050

4gb

Should I download stable diffusion or canndo everything on this server?

[ornate elk]

You can use forge webui or reforge with that GPU

On this server image generation isnt frer

[solemn sparrow]

whats the charges ?

[ornate elk]

All infos are here:
#artisan-faq

Idk the prices currently

[solemn sparrow]

Ty

[stiff hawk]

@ornate elk Redoing and documenting entire installation to get Fooocus working with ZLUDA. My documentation so far is uploaded in this post. Got my first warning:

WARNING: The scripts convert-caffe2-to-onnx.exe, convert-onnx-to-caffe2.exe and torchrun.exe are installed in 'M:\AI\Fooocus_win64_2-5-0\python_embeded\Scripts' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.

So I had put in Environment Variables > System variables: M:\AI
But apparently that's not good enough. I could just add the Scripts folder to PATH now; but it could still be wrong for some reason, and either way I don't know what I don't understand

[ornate elk]

No dont add stuff to path thats not stated in my guide

Open up a cmd and run
Python --version
And then
Where python

And show the output

[stiff hawk]

M:\AI is in PATH, deleted
C:\ZLUDA is in path, but that was for Forge I think

[stiff hawk]

(c) Microsoft Corporation. All rights reserved.

C:\Users\orphi>Python --version
Python 3.10.11

C:\Users\orphi>Where python
C:\Users\orphi\AppData\Local\Programs\Python\Python310\python.exe
C:\Users\orphi\AppData\Local\Microsoft\WindowsApps\python.exe

C:\Users\orphi>```

[ornate elk]

Looks okay

[stiff hawk]

Continuing now

[stiff hawk]

• Launched M:\AI\Fooocus_win64_2-5-0\run.bat. Terminal output:

    Fast-forward merge
    Update succeeded.
    [System ARGV] ['Fooocus\\entry_with_update.py']
    Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec  6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
    Fooocus version: 2.5.5
    Error checking version for opencv-contrib-python-headless: No package metadata was found for opencv-contrib-python-headless
    Installing requirements
    [Cleanup] Attempting to delete content of temp dir C:\Users\orphi\AppData\Local\Temp\fooocus
    [Cleanup] Cleanup successful
    Downloading: "https://huggingface.co/mashb1t/misc/resolve/main/xl-to-v1_interposer-v4.0.safetensors" to M:\AI\Fooocus_win64_2-5-0\Fooocus\models\vae_approx\xl-to-v1_interposer-v4.0.safetensors

    100%|█████████████████████████████████████████████████████████████████████████████| 5.40M/5.40M [00:00<00:00, 7.31MB/s]
    Downloading: "https://huggingface.co/lllyasviel/fav_models/resolve/main/fav/juggernautXL_v8Rundiffusion.safetensors" to M:\AI\Fooocus_win64_2-5-0\Fooocus\models\checkpoints\juggernautXL_v8Rundiffusion.safetensors

    100%|█████████████████████████████████████████████████████████████████████████████| 6.62G/6.62G [10:53<00:00, 10.9MB/s]
    Traceback (most recent call last):
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\entry_with_update.py", line 46, in <module>
        from launch import *
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\launch.py", line 152, in <module>
        from webui import *
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\webui.py", line 10, in <module>
        import modules.async_worker as worker
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\modules\async_worker.py", line 3, in <module>
        from extras.inpaint_mask import generate_mask_from_image, SAMOptions
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\extras\inpaint_mask.py", line 6, in <module>
        from extras.GroundingDINO.util.inference import default_groundingdino
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\extras\GroundingDINO\util\inference.py", line 3, in <module>
        import ldm_patched.modules.model_management as model_management
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\ldm_patched\modules\model_management.py", line 121, in <module>
        total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
      File "M:\AI\Fooocus_win64_2-5-0\Fooocus\ldm_patched\modules\model_management.py", line 90, in get_torch_device
        return torch.device(torch.cuda.current_device())
      File "M:\AI\Fooocus_win64_2-5-0\python_embeded\lib\site-packages\torch\cuda\__init__.py", line 778, in current_device
        _lazy_init()
      File "M:\AI\Fooocus_win64_2-5-0\python_embeded\lib\site-packages\torch\cuda\__init__.py", line 293, in _lazy_init
        torch._C._cuda_init()
    RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx

    M:\AI\Fooocus_win64_2-5-0>pause
    Press any key to continue . . .```

Besides the Nvidia GPU error which I just assume is expected on first run, I saw: `Error checking version for opencv-contrib-python-headless: No package metadata was found for opencv-contrib-python-headless`

[ornate elk]

check the model management py again

it seems it got resetet maybe

[stiff hawk]

I'm not in the point in the guide where it asks me to change that. I'm just before that

[ornate elk]

ahhh

okay then proceed

[stiff hawk]

@ornate elk Same error as before, but now I have full documentation of everything I did

I did not do anything that isn't documented here

[wary tulip]

@ornate elk

[ornate elk]

what?

you should only have C:\zluda and not zluda\zluda

also can you copy over the 3 dlls files again and relaunch?

[stiff hawk]

I don't, I said the directory structure is C:\zluda\zluda.exe, that way I could show that I don't have any other folders between C:\zluda and the .exe

I'm not a technical person but I'm a fairly thorough thinker ^^

[ornate elk]

ah ok sry

but the setup looks correct

[stiff hawk]

No worries!

[ornate elk]

the only issue is the dll part

[stiff hawk]

Seems like it

[ornate elk]

can you copy them over again and relaunch?

and after that send me your model_management.py

[stiff hawk]

Recopied, and:

Already up-to-date
Update succeeded.
[System ARGV] ['Fooocus\\entry_with_update.py']
Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec  6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
Fooocus version: 2.5.5
[Cleanup] Attempting to delete content of temp dir C:\Users\orphi\AppData\Local\Temp\fooocus
[Cleanup] Cleanup successful
Traceback (most recent call last):
  File "M:\AI\Fooocus_win64_2-5-0\Fooocus\entry_with_update.py", line 46, in <module>
    from launch import *
  File "M:\AI\Fooocus_win64_2-5-0\Fooocus\launch.py", line 152, in <module>
    from webui import *
  File "M:\AI\Fooocus_win64_2-5-0\Fooocus\webui.py", line 10, in <module>
    import modules.async_worker as worker
  File "M:\AI\Fooocus_win64_2-5-0\Fooocus\modules\async_worker.py", line 3, in <module>
    from extras.inpaint_mask import generate_mask_from_image, SAMOptions
  File "M:\AI\Fooocus_win64_2-5-0\Fooocus\extras\inpaint_mask.py", line 5, in <module>
    import torch
  File "M:\AI\Fooocus_win64_2-5-0\python_embeded\lib\site-packages\torch\__init__.py", line 143, in <module>
    raise err
OSError: [WinError 126] The specified module could not be found. Error loading "M:\AI\Fooocus_win64_2-5-0\python_embeded\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

M:\AI\Fooocus_win64_2-5-0>pause
Press any key to continue . . .```

[wary tulip]

Is tesla gpus good for comfyui

@ornate elk

[ornate elk]

they are old but if they have a lot of vram they can be good. anything with 8gb or less is crap

[stiff hawk]

@ornate elk A few different threads online suggest installing/reinstalling Microsoft Visual C++

Others are saying downgrade PyTorch

[compact salmon]

this is EXACTLY what i suspected was my issue

dude youre the best

[ornate elk]

try open up a cmd and run
pip cache purge
then delete the venv folder and relaunch

[compact salmon]

also your amd zluda forge guide was 100% clear just like the comfy one 🙏

[stiff hawk]

I don't know where the venv folder is in Fooocus

[ornate elk]

oh true foocus doesnt have one

[ornate elk]

thanks 🙂

[stiff hawk]

I found: "C:\Users\orphi\AppData\Local\Programs\Python\Python310\Lib\venv"

[ornate elk]

delete that

[stiff hawk]

Done, now do I run.bat?

[ornate elk]

have you also run pip cache purge?

then yes

[stiff hawk]

Same error

[ornate elk]

okay, open up a cmd and run
sfc /scannow

also show me your path entries

of zluda and hip path bin

[stiff hawk]

Sorry for the screenshot form factor

[ornate elk]

there is the issue

thats not correct

delete both the hip_path and the Zluda folder path

[wary tulip]

What about p40 24gb

[ornate elk]

would be good

[stiff hawk]

There is both HIP_PATH and HIP PATH, so I'm double-checking with you which to delete

[stiff hawk]

Also HIP_PATH_62

[ornate elk]

leave them pls

[stiff hawk]

You told me the wrong one here, good thing I double-checked

[wary tulip]

For me I want only make images . So is that overpower for me

[stiff hawk]

Deleting HIP PATH and ZLUDA

[ornate elk]

and the C: zluda path too

and then please check how its done here:
https://github.com/CS1o/Stable-Diffusion-Info/issues/7#issuecomment-2408541054

[stiff hawk]

Me? Also what am I looking at?

[ornate elk]

at how you correctly add the variables

in path

this

[stiff hawk]

I can't read German. Is it that I put the Variable section wrong?

Variable Name has to be Path or something?

[ornate elk]

you dont need to read german to understand where the buttons are

you click on Path (which is not hip path 62)

then on edit

and then on new

and add the 2 variables

[stiff hawk]

I'm going to guess that Variable name doesn't matter

[ornate elk]

that means your at the wrong place

you cant define a name where they have to go

[wary tulip]

@ornate elk
I asked CHATgpt and he gave me this pc build

1️⃣ Processor (CPU): Intel Xeon E5-2678 v3 (12 Cores / 24 Threads)
2️⃣ Motherboard: Cheapest X99 LGA 2011-3 Version
3️⃣ Primary GPU (GPU 1): NVIDIA Tesla P40 (24GB GDDR5)
4️⃣ Secondary GPU (GPU 2): NVIDIA NVS 315 (1GB DDR3)
5️⃣ Memory (RAM): 16GB DDR4 2133MHz (2×8GB)
6️⃣ Power Supply (PSU): 700W 80+ Bronze

Is it good ?

[ornate elk]

no, its very old stuff

wouldnt use a cpu from 2015

or mainboard

[stiff hawk]

I meant this. It lets me put in a Variable before Value

Does Variable matter?

[ornate elk]

wrong place still

scroll down to path

click on that

and then on edit

[wary tulip]

@ornate elk so what do you recommend

[ornate elk]

sry i dont have the time for a low budget build i already told you a few things about a gpud like 1080 or 3060 and a ryzen cpu

[stiff hawk]

OH! Okay, and the page you linked to for context, now I get it

[ornate elk]

16-32gb ram

[lucid yoke]

Friendly bit of advice, gpt/llm suck at pc builds, install instructions and other errors. i recommend checking Reddit or other tech forms for budget builds

[ornate elk]

and https://pcpartpicker.com/

[stiff hawk]

@ornate elk Hey, if I'm supposed to add these two things to path, your guide is messed up. The way it's structured, it looks like you only update path if you have a GPU unsupported by ROCm 6.2.4

[ornate elk]

why? it says:

and then:

[stiff hawk]

Yeah, but your bullet and numbering structure puts point 4 right after a point 3 in the "unsupported by ROCm" sublist, which is a different list. Plus you have a 4.1 in a different structure right after

[lucid yoke]

First time I've heard someone mention that. Its quite common to do it like that

[ornate elk]

i think it shouldnt be an issue

but im open for changes of course

[stiff hawk]

But point 4 doesn't apply to that sublist only, it applies to the entire instruction list

Basically the auto-numbering disorganized your instructions

Because point 4 is in the sublist only (based on this structure) I skipped it, as almost anyone would

[ornate elk]

list formatting in github is very complex if i have a lot of substeps

[stiff hawk]

Yee, but this is why I avoid auto-ordered-lists if this starts happening. I use roman numeral main points, bolds, honestly a lot of things can work

[ornate elk]

but i know what oyu mean now, i should move it out of the order from the subguide of rocm 6.1

[stiff hawk]

Auto-ordered-lists hate sublists

Yes, amazing thanks if you do

[lucid yoke]

Hmm i really don't see the issue from the screenshot but for non tech people it could be better i suppose

[stiff hawk]

Tons of software struggle with unordered and ordered lists formatting. Scrivener really needs to fix that too

[ornate elk]

but how should i call it then where the user should proceed after doing rocm 6.2 ?

[stiff hawk]

One sec, let me find the easiest and clearest option

[ornate elk]

i also had an idea of removing the complete zluda setup of every guide and make an own Block for it. and then link to it from every guide maybe.
but then people wouldnd find back easily

[stiff hawk]

I take it back, ignore me for a sec

I didn't scroll back far enough

If your GPU is not supported by ROCm 6.2.4 then proceed with the Guide at Step 1 below:
IMPORTANT: If your GPU is higher/newer than a RX6800, then skip Step 3

Change points 1, 2, and 3 in this section to bullets, keep everything else as-is

Then the point 4 won't act as part of this sublist, and it will be fixed

There

Updating my PATH now

[ornate elk]

but that wouldnt make much sense to start at 4.) then if there isnt a 1-3

[stiff hawk]

Only the 1, 2, and 3 in the "if GPU isn't supported" sublist are changed to bullets, because they're not the same 1, 2, and 3 as in the main list

[compact salmon]

what is fp8? compared to q8 i see it uses less vram

[ornate elk]

its half the precision of an fp16 model, you can use these too
compared to fp8 they have a bit worse quality, but mostly not noticable

[stiff hawk]

Oh, you have two point 4s. So adding things to path would actually be point 10, not point 4

Oh, you have two point 4s. So adding things to path would actually be point 10, not point 4

If all of this sounds confusing, it's only because it is xD

[ornate elk]

yes exactly xD

[true sphinx]

hey guys so im swapping from auto 1111 to comfyui and im trying to change the folder for the models but im having a hard time making it work

[ornate elk]

hey, try this:
Linking Models, Loras, etc. from other Webui's or Folders to ComfyUI:
If you want to link all models from an other Webui you can do that by renaming the extra_model_paths.example.yaml to extra_model_paths.yaml
Then right click the file and edit it with Notepad or Notepad++.
If you have Automatic1111 installed you only need to change the base_path line like in my Example that links to the Auto1111 Webui:
base_path: "C:/AI/stable-diffusion-webui/"
Then relaunch the run_nvidia_gpu.bat

[true sphinx]

im using the desktop version

like the app

so idk if this changes things

[ornate elk]

you mean in stability matrix ?

[stiff hawk]

Alright, so I think in conclusion to what I am saying, point 4 (the second one, "add to path") is part of the main list of things to do, which means it's actually after point 9 and should be renamed to point 10. Then "This Step Only for RX580", "Go into the ZLUDA folder", and "Launch the run.bat" at the very end are actually points 11, 12, and 13 or whatever

[true sphinx]

this version
https://www.comfy.org/download

[ornate elk]

ohh then idk sry

[ornate elk]

thats a good idea thanks!

[true sphinx]

should i use the portable version

is that better

[ornate elk]

its the more common used version and portable

[stiff hawk]

Added things to path, now updating my documentation, then testing

It launched!

I'm so happy

[stiff hawk]

Thanks so much for your guide and for being patient with me

It seems to be using my integrated GPU instead of my main. I don't know if I want to completely disable my integrated GPU, but eh

Found a flag. --gpu-device-id DEVICE_ID, looks like I can maybe edit the run.bat

No, I'm not sure where to put the flag in this one

[ornate elk]

yes you can try to add that but if it wotn work, you probably have to disable the igpu

[stiff hawk]

I've gotten that flag working on the other UI's. Just gotta figure out how to use it in Fooocus

Where's that in Fooocus?

[ornate elk]

sry haha i was confused

guiding to much simultaniosly in github, here and in dm xD

[stiff hawk]

Sorry xD

[ornate elk]

ah maybe just add it in the run.bat

in foocus github they say that worked

[stiff hawk]

run.bat has nothing but:

pause```

I could try adding it to the Target in a shortcut

But I don't like that if it's not the ideal way

Oh, looks like Fooocus won't work at all until I fix this. It's thinking I'm using the wrong GPU architecture

[ornate elk]

try:
.\python_embeded\python.exe -s Fooocus\entry_with_update.py --gpu-device-id 0 or 1 pause

[stiff hawk]

Oh, that seems to change the GPU, awesome

Let me see if it tells me what architecture yet before running a generation

Doesn't look like it

Testing generation

Think I still picked the wrong GPU, will test soon

[ornate elk]

it shows that in the cmd

[stiff hawk]

Awesome, picked the right GPU

But it's just not using it

It's using my CPU and onboard memory

[arctic path]

can anyone help me add stable diffusion inside Photoshop?

[ornate elk]

its probably compiling zluda in the background

takes 15-40 minutes for the first run

[stiff hawk]

DUH, I'm dumb

I read that and I forgot

[stiff hawk]

Will I have to redo it every time I launch run.bat?

[ornate elk]

nope

zluda compiles only when torch, zluda or the driver updates

[arctic path]

why people sending other server link here. so annoying

[ornate elk]

these are scam bots

we try to ban them quick, but they are to much

[ornate elk]

i think its not possible anymore, there was only one good plugin for it 2 years ago:
https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin

[ornate elk]

report them to discord and block

[arctic path]

they want crpto lol 😂

[ornate elk]

yea always xD

[hot briar]

@ornate elk do you have suggestion for base model

RealVisXL V5.0 tried this but couldn't load model.

[lucid yoke]

I feel like this is more of a #💬｜general-chat question but why couldn't you load it?

What's the error

[formal ridge]

If you can't load SDXL models, you're going to be stuck with SD1.5 Based models. The absolute best ones for SD1.5 I know of are Duchaiten-Journey (I recommend 5.6.7.2.3 or 6.0), or realisticVision 5.1 (VAE)

If you want an anime model instead, the best one I know of for SD1.5 is Aniverse v4.0

[burnt sundial]

I need help.

[lucid yoke]

And what is it you need help with?

[wary tulip]

Is core in processor important

[lucid yoke]

You mean core count?

[wary tulip]

Ya

[formal ridge]

I wouldn't recommend anything less than 8 core, 16 threads.

[lucid yoke]

Yeah but seeing how your still looking for a budget pc build

Do you wanna focus on AI, or gaming in general

[hot briar]

Exception: could not load model: C:/Users/niceb/a/stable-diffusion-webui-directml/models/Stable-diffusion/realvisxlV50_v50LightningBakedvae.safetensors

[wary tulip]

Which one I need to care more about single core or multiple core

[wary tulip]

I want to use comfy UI with the most low budget PC

[lucid yoke]

What gpu do you have?

[lucid yoke]

You can run comfyUI on a potato, the models are the problem

[hot briar]

i dont know what i can it takes long and i just tried realisticvision that you said and it just creating what i put or close ones in concept folder

[hot briar]

rx 7800 xt

[formal ridge]

I just ran ComfyUI on an Intel Celeron laptop from 5 years ago that only has 4 GB of RAM. It took an hour for a single image, but it doesn't matter much if you run CPU only

[lucid yoke]

Ah did you follow cs1o's guide?

[hot briar]

idk actually

[lucid yoke]

Do you have zluda/directml?

[hot briar]

ah yeah checked

[wary tulip]

Actually I want to generate anime and manga style images

[lucid yoke]

Still very broad. My gtx1650 can run sd1.5 slowly but cant handle sdxl/flux etc

Whats the budget your working with?

[lucid yoke]

Hmm can you post the full cmd log? Maybe cs1o peeks in and sees whats wrong. Im no amd expert

[wary tulip]

It's not my own money but I'm sure it's under $700

[hot briar]

others work but this ain't work

I think the problem is this model

[lucid yoke]

Could be

[formal ridge]

SDXL models are rather large, are you sure your download completed successfully?

[lucid yoke]

Im gonna be completely honest thats a challenge but do you have a pc you can upgrade or are you building from scratch?

[wary tulip]

No I'm building from scratch

[hot briar]

it looks downloaded in downloads tab

[ornate elk]

Ah maybe the model file got corrupted

Try redownload it

[lucid yoke]

But a used 3070 (i used to use this) is like 350 euros here

But it can run illustrious etc fine but it's not gonna last forever

[hot briar]

i'll try generate images with it

if it does it is not corrupted or something right?

[formal ridge]

They could go even older if they wanted to. I could run SDXL based models on my GTX 1080.

[lucid yoke]

Yeah but rip making it last

[wary tulip]

.

[formal ridge]

If you're talking about speed, it's not that important these days. But I wouldn't recommend anything lower than 2.4GHz baseline, and 3.0 Ghz baseline if you can afford it in your build.

It's more about how many cores you have and how well you can utilize them

[wary tulip]

So Ryzen 5 5800X - 8 Cores / 16 Threads is good

[formal ridge]

Yep, that's what I have (or will have once I get off my ass and build my new PC)

4.7 Ghz max, 3.8 Ghz base is a good choice

[wary tulip]

And how about Intel Xeon E5-2678 v3 (12 cores / 24 threads) 2.5GHz

[formal ridge]

Xeons are generally for servers and they're slower

Where threads matter more than anything else

[wary tulip]

Ok ok so I need to get Ryzen 5 5800X

[formal ridge]

Not to mention server board costs are insane, plus you have to get ECC memory (which is also expensive)

[wary tulip]

I didn't know that before

[formal ridge]

Yeah, I built a server before. And the final cost was a bit out there

[wary tulip]

How much ram I need 16 or 32

[lucid yoke]

I used to run a xeon on my desktop. It worked but not optional

Optimal*

[formal ridge]

I would recommend an absolute minimum of 32, 64 or 128 if you can afford it

[wary tulip]

Ddr3 or ddr4

[lucid yoke]

Hes budget so its gonna be 32

Ddr4 ofc

[wary tulip]

Ok

[formal ridge]

The filter should have kicked in if you're on PCPartPicker...

[lucid yoke]

Ddr3 is obsolete and ddr5 is expensive

[hot briar]

i did with realisticvision @formal ridge but my lora generates like what i put. I mean there's a little changes when trying to generate

what's the problem

is it count of images, steps or something

[formal ridge]

If you're trying to change images, sampler & scheduler will change your image more than anything else will.

[wary tulip]

Yeah I saw a GPU called NVIDIA Tesla P40 24gb is ti good

Is *

[formal ridge]

No, outdated as hell and unsupported by Torch

[lucid yoke]

I really recommend for you buying a prebuilt if im being honest

[hot briar]

i don't know what are those -_-

[wary tulip]

so what should I get ?

[formal ridge]

What UI are you using?

[wary tulip]

What do you mean?

[hot briar]

automatic1111

[formal ridge]

I'd recommend an nVidia 3K series or higher, but again I don't think you'll be able to afford all this at $700 or under

[hot briar]

directml

[wary tulip]

Ya

[lucid yoke]

Power supply and that gpu will be around 450 already

[formal ridge]

"Sampling method" and "schedule type" at the left side on the interface

[formal ridge]

That's what I'm saying. I don't think he can do it. If you scoured ebay continuously you might get lucky on parts, but I doubt it.

[lucid yoke]

Maybe a refurbished and throw in a gpu but thats a gamble

[wary tulip]

Yeah can anyone suggest a PC build for me so I can buy it

[lucid yoke]

https://nzxt.com/product/player-one-prime similar to this if your buying new

For ai

At least, but its almost double your budget

[hot briar]

are these bad

for changing

[wary tulip]

I cannot afford that bro I need to sell my kidney for that💀

[lucid yoke]

Well assuming your lower income, i recommend trying out cloud based solutions

[wary tulip]

Yeah the problem is that I'm not from America and I'm from Africa and I cannot use cloud-based Solutions because I don't have credit card

[formal ridge]

You can't afford anything right now. At the absolute cheapest prices I hit 509 and that didn't include PSU, case, or OS.

[lucid yoke]

Hmm have you tried looking on civitai and it's buzz?

Its pretty okay if everything else isn't possible

[formal ridge]

Africa? PC Parts are going to be even more expensive there...

[wary tulip]

That's why I'm looking for the most cheapest PC I can get to run comfy UI with no problems

[wary tulip]

No

[lucid yoke]

Your misunderstanding comfy ui and ai

My phone could run comfyUI. It can't run the ai models

What pc specs do you have right now?

[formal ridge]

I ran ComfyUI on a potato with worse specs than your phone. I'm pretty sure your phone could do it if you rooted the system and were running custom code

[lucid yoke]

Maybe

[wary tulip]

CPU: Intel i3-6100T

RAM: 4GB

Storage: 512 ssd

GPU: No dedicated graphics card

[formal ridge]

ComfyUI can be run CPU only, but it's slow as hell, and not really recommended.

[wary tulip]

This is what I have now

[formal ridge]

Yeah, technically you could run ComfyUI on that, but here's how long it took on my potato.
"Only 512^2 but "Prompt executed in 4781.67 seconds""

[wary tulip]

So for the gpu I need to go to 1080

[formal ridge]

Maybe... Mine worked with SDXL but that was a while ago. I don't know if it's even compatible with the current UIs now.

[lucid yoke]

could use a1111 for that

hasnt been updated in a year

[cursive hazel]

So, I tried The Q4_K_S model, and got
10/10 [01:17<00:00, 7.77s/it], my VRAM was 14.3GB / 15.5GB, with 0.7GB used in Shared Memory.
Using the same workflow I showed before. That seems reasonable to me.

And I think I figured out the Q6 thing. I think I'm at the very limit of it, in terms of VRAM. I ran it once, and got about 11s/it, which seemed reasonable. I ran it again and got 20s/it again.. After some playing around, I found that if I basically close everything (except Discord), I'll get 11s/it. But if I'm doing much of anything else, it goes to 20s/it on Q6. So I'm assuming I'm riiight on the edge of using too much VRAM. Given this model is 9.6GB on my SSD, I'm assuming I need a model that's closer to 8GB at most? (if that's how it works).

Oh, and I did pay attention to the VRAM spikes when doing the VAE part. In Q6, my VRAM is usually 14.5GB with about 1.4GB in Shared Memory. When it hits the VAE, it spikes up to 14.7GB and 6GB in Shared Memory.
In Q4 (with nothing else running), I got 13.7GB VRAM usage, and 0.4GB in Shared Memory. Then the VAE spikes it to 14.8GB and 4GB in Shared Memory. Though, the VAE part only lasts maybe 2 seconds, even on Q6.

[ornate elk]

ah that much better, thanks for testing. Seems like Q4 is the sweetspot for it

@compact salmon some info about flux speed above

[dreamy lodge]

Didn’t work. The cmd shell that opens when using the bat eventually gets errors and crashes before I can read them

[ornate elk]

can you open up a cmd inside the folder and then start the bat wit hthe cmd, that prevents it from closing

[ornate elk]

okay, open up a cmd and run
python --version
then run
where python
and screenshot it

[ornate elk]

yep wrong python version

uninstall 3.13

then delete the venv folder of comfyui and relaunch

[dreamy lodge]

do i just delete the Python313 folder?

[ornate elk]

no pls not

thats not how you correctly uninstall programms

that can cause issues

[dreamy lodge]

Worked thanks bro

Sorry but Im kinda confused in the linking models section. What part of the yaml file do i change to link to my a1111 models?

[ornate elk]

only one line

[dreamy lodge]

So i just change this line to my a1111 directory?

[dreamy lodge]

https://civitai.com/articles/5177/comfyui-or-lucifael-or-all-in-one-workflow-or-sdxlpony-or-veteranbeginner-friendly-tutorial

Im using this workflow and whenever i generate an image, I just get a blank grey image.

Any idea what im doing wrong?

[ornate elk]

Nope but does a simple workflow works?

Like the base comfyui one?

[arctic path]

so I already installed stable diffusion and other staff in Krita. it's about 70GB. Can I just move it to Photoshop somehow and use in Photoshop without downloading them again?

[ornate elk]

How would you integrate it in photoshop?

Did you found a tool?

[arctic path]

there's plugins

that can be used but wonder if I don't have to download all those stuff again

[ornate elk]

Only if the plugin is based on Comfyui too

[dreamy lodge]

Tried a simple one for 1.5 and now my you entire system is going like a frame per 10 seconds. AMD Software said an error occurred. The image did generate in like 40 seconds tho but it looks like grey mixed with a bunch of static.

Restarting my pc now

[ornate elk]

Okay yep restart could help

Also make sure you dont run anything like Wallpaper engine or games in the background

[dreamy lodge]

Gotchu

[wary tulip]

@ornate elk bro can I get less than 8 cores processor

[tribal anvil]

How do I get dark theme on forge tried searching it but didn't show up?

[outer kindle]

is anyone familiar with the anapnoe ux/ui dock plugin? I donated for access, but it won't show up when I enable extensions. there's pretty much no documentation for the new ui either. do I need to load the 1.0 AND 2.0? even when I manage to get it to run the 1.0 version, it just leaves me on the dock settings page and nothing else works. any help is appreciated

[compact salmon]

did you try fp8 and have issues as well?

[cursive hazel]

Nope! No idea how to do that. Seems like models under 9GB work fine though. I think it's just an issue of VRAM.

[compact salmon]

gonna have to mess about and experiment more havent had much time. we can run gguf models just not nf4s right

@cursive hazel https://civitai.com/models/630820?modelVersionId=936565 was gonna try this one (v2 f8) ill lyk results its an all in one

[cursive hazel]

No nf4s, right That's what CS1o said.
Lemme know how the fp8 one goes!

[fair oxide]

has anyone had experience with xformers causing huge spikes in vram usage?
granted, im trying to edit a 4k image, but doggettx can handle it fine with 12gb of vram.
maybe i have misconceptions about what xformers is supposed to do, ive never messed with any sort of image generation stuff until yesterday so im not quite sure whats what yet

some relevant information maybe:
rtx3080ti 12gb
cuda 12.1
python 3.10.6
xformers 0.0.23.post1
pytorch 2.1.2+cu121
using Automatic1111

[fair oxide]

i think my issue was genuinely "turn it off and on again" which is hilarious

for the future if anyone searches this, i had multiple versions of cuda installed, so i had some weird environment variable stuff going on that required a restart to fix.

[fading forge]

Hey! I'm trying to use ReActor but can't seems for the life of me to manage to install insightface, when I launch SD i get Error: The 'insightface==0.7.3' distribution was not found and is required by the application and then ModuleNotFoundError: No module named 'insightface' for each failed attempt to load scripts for ReActor, any ideas?

[eternal river]

What issues buddy

[craggy root]

i don’t know how to cancel the free trial so it took money out my account, that’s nice so how do i cancel it for real

[fading forge]

Nevermind I manage to resolve it, I hadn't the up to date C++ frameworks

[humble parcel]

Hi. i have same issue as 303. Does ComfyUI only work with C drive, can it not work with D drive?

I have this issue

[ornate elk]

Comfyui can work with any drive as long as its not placed in a system created folder

[humble parcel]

I see

[ornate elk]

You have to increase the windows pagefile

How much ram do you have?

[humble parcel]

I have done that, hopefully correct

I have total of 16 gb ram bro

[ornate elk]

And what's your pagefile size in the Settings?

And is it only enabled for the C drive?

Best is to use 16000min and 24000max for C and disabled for any other drjve

Also a PC restart is required after the change

[humble parcel]

It looks like this

[ornate elk]

Disable it for D

And and enable it for C

Set it to 16000min and 24000max or 32000max if you need or have the space

Then reboot

[humble parcel]

D drive is where i downloaded the Comfyui

[ornate elk]

Yea doesn't matter

Pagefile should be on the fastest drive

[humble parcel]

Interesting

Gives me this

[ornate elk]

Yea you need to have 15-20gb free space on your C drive

Pagefile needs the space on C

[humble parcel]

Ok im going to get rid of some stuff

[ornate elk]

A full C drive can cause your whole system to slow down too.
So always have 15gb free minimum

[humble parcel]

What is pagefile precicely bro? Appreciate the help

[ornate elk]

Its like virtual Memory (RAM).
It uses the disk space of an SSD to give programms some place if they need more ram to switch to.
Pagefile is slower than RAM. Thats why it needs to be placed on the fastest drive possible. And also always placed on internal drives, nor external.

[humble parcel]

I know, but idk its so easy to makes it full Recently, i glad bought big SSD with like 2tb but now got another external one. Think gotta transfer some stuff to the external one.

[humble parcel]

Oh i see

[ornate elk]

Tools like WizTree can help you see what's taking up a lot of space

[humble parcel]

I check that out

Ok deleted some stuff

Does this look ok?

[ornate elk]

Yes but you also have to set it to No Paging file for the D drive

You dont want a Pagefile on an external drive.

[humble parcel]

Like this?

[ornate elk]

No you have still selected C drive

Set it back to custom

[humble parcel]

LIke this?

[humble parcel]

Why so?

Seems good

[ornate elk]

Because that can slow down your whole system.
External drives are connected trough USB which is slower than internal mounted Drives.
M2 ssds use PCIE which is why the perform much faster.
The Pagefile needs to be fast.

[humble parcel]

I see

Ok restarted will see now

Lowk felt PC bit faster

[ornate elk]

Comfyui should now work normaly too

[humble parcel]

Hopefully its cooking

This how it looks

[ornate elk]

What's your gpu?

[humble parcel]

rtx 2070

I pressed enter, and the CMD went away. Last time tho, CMD would not open web page but did now after file page reset.

Still no image tho on Comfy

[ornate elk]

I hope you didnt downloaded the 23gb flux file

Because that won't work on an 8gb gpu

[humble parcel]

I did

But wait i have antother one

the 16 gb

Let me cook a sec

[ornate elk]

You need to use smaller flux files

For 8gb GPUs

[ornate elk]

Like flux nf4 or flux Q4

[humble parcel]

The danm tutorials did not say this

Im on it brother 👌

Is this one good_

[ornate elk]

For example here, take the nf4 version at the Files box:
https://civitai.com/models/141592?modelVersionId=992642

[humble parcel]

Whats the best ojne

I click the this button right?

Thank you bro

[ornate elk]

That one if you click on "Files" on the right side

[ornate elk]

Nope click on files

[humble parcel]

Did it

Ok put it in checkpoint right?

Seemed it had in system since showed (1) but dk, but changed the name of it and removed the (1)

Gonna test

[ornate elk]

Alright and delete the larger models to get some disk space back

Also even if nf4 works, it won't be fast as flux requires a lot of resources.
So you will have more fun with sdxl based models

[humble parcel]

Bro it showed this

This is odd

[ornate elk]

What does the cmd shows?

[humble parcel]

Should it be as none?

[ornate elk]

Yep for D thats ok

[ornate elk]

That workflow won't work as you need the clip models and VAE files for Flux too

You can also try to relaunch and check the vram usage in the task manager

Can you try if this model works normaly?
https://civitai.com/models/133005?modelVersionId=782002

Its sdxl based not flux

[humble parcel]

Sure bro

Is it NSFW to or have to download that later?

What to download?

[humble parcel]

I can do that too, i do think i have the VAE files, seem to followed a tutorial and download it, but downloaded some stuff so not sure, but we can check if its fine for you

Like heard this one

v

'

Think was those placements and checkpoint, but might have to get new ones since this file is 5 gb im not sure. What you think?

[humble parcel]

You there bro?

[ornate elk]

Yep

[ornate elk]

That one

[ornate elk]

Could work if its 5gb

But for Flux you still need the clip files

T5 and clip l

But first try Juggernaut XL

It doesnt need anything else

[humble parcel]

Sure

XL looks beautyfull too

Just heard Flux more fuctional since hands and easier to have 2 uniqe characters for example

[ornate elk]

Yea flux is great but its mostly for high end and newer GPUs.

I would recommend sdxl for your GPU as it will work without issues and still has a great quality and good community support (extensions, loras, controlnet)

[humble parcel]

I understand

Heard bit Flux takes more power yeah, i think my card could work ok, its just at cusp of approval. Its fine card, but yeah but bit high recommendation for people. XL probably works good. Used SD and images is like 5 sec or so thats cool. Liked it, but wanted to see what XL and Flux to offer, after heard some good about them.

[humble parcel]

Could you give me the files, i can download it mean while, while watiing for cmd. Thank you for the help bro ❤️

Testing now

I tried Swarma with SD XL and it worked fine, but no NSFW

Flux havent been able yet to see work

Just says

Got prompt with no progress

I have my files on D drive as said earlier, but dont know if it matters or not

Wait it worked bro...

Thank you ❤️

But dont know why it took so long

Worked instantly now

Is there way to fix the format tho? Like to automatic 1111. Really like the A111 format, less organized and can see image better.

[ornate elk]

Storing the files on D is okay. It will just take a bit longer to load them onto the GPU.
But it doesnt change the image generation process speed

[ornate elk]

Swarm ui is comfyui with a better Ui.
But Auto1111 format can't be applied to Comfyui.
Only Auto1111, Forge and ReForge have that UI.

[ornate elk]

Good that it worked.
For sdxl models try to use a resolution like 1024x1024 for better quality

[humble parcel]

I see

It seemed like the easier version. Whats the point of people uinsng Comfy when Swarm seems easier?

Danm this stuff look good

[ornate elk]

Some people like the node based layout

But comfyui is not recommended for beginners

As its more complex than needed.
So thats what swarm tries to fix

[sour moss]

hi, anyone know why i can't get animatediff to work?

[humble parcel]

@ornate elk Bro, how to get the clips for the FLUX, seems only need that

[wary tulip]

@ornate elk how much the best image generator model need vram and say any number

[lucid yoke]

SD1.5 Can run with 4gb vram, sdxl runs best with a minimum of 8 but ive seen peope run it with less (albeit slower)

flux can also run with 8 but it will take like a few mins per image

[wary tulip]

Ok thank you

[wary tulip]

@lucid yoke can I ask you another question

[dreamy lodge]

@ornate elk still happening after restart

[lucid yoke]

i mean sure but i think we pretty much awnsered most stuff last time

[wary tulip]

Sorry for bothering you. I just want to ask for your opinion on a workstation

[lucid yoke]

no worries

[wary tulip]

Are they good

[zinc comet]

hey everyone

quick question

when you generate images, how do you proceed ? like throwing few prompts, try with a lot of different seeds, once you got something not too bad, you keep the seed, try some varation on the prompt/cfg/scale/scheduler/sampler and after you do some "post production" stuff ? which ones ?

[ornate elk]

what does the cmd shows?

[ornate elk]

they can be found here:
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

you need t5 fp8 scaled and clip l

[ocean apex]

want to confirm my reading of the api docs: only Stable Diffusion 3 & 3.5 api calls support image-to-image, not stable image core?

[dreamy lodge]

@ornate elk here ya go

[ornate elk]

hmm looks normal

what you can try is to delete the venv folder and then relaunch, and then replace the zluda files again

[dreamy lodge]

Do i need to reinstall numpy?

[ornate elk]

Yes

[elder hatch]

My locally ran stable diffusion is now only outputting grey images

anyone have a sollution I have tried reinstalling multiple times

[dreamy lodge]

Same thing happening

Should i also mention that A1111 still works perfectly fine

should i try to reinstall the comfyui package entirely?

[ornate elk]

Yes, and only install comfyui manager and no custom nodes until it works

[ornate elk]

What's your gpu?

[dreamy lodge]

k

[elder hatch]

geforce 4080

[dreamy lodge]

I can just delete the comfyui folder right?

[ornate elk]

Which webui do you use?

[elder hatch]

automatic1111

it worked about a week ago

then it stopped

[ornate elk]

Can you show me the full cmd log?

And an example txt2img prompt where its a Gray output

[ornate elk]

Your probably trying to use a lora as Model.
That won't work

if c0ne is the pixelart thing. Then its a lora

Goes into models/lora

Also add --xformers to the webui-user.bat for better performance

At the line commandline_args=

[elder hatch]

what do I do here

[ornate elk]

You move the cone lora from the models/stable-diffusion folder to models/lora

[elder hatch]

ohhh

[ornate elk]

Everything smaller than 2gb is not a model(checkpoint)

[elder hatch]

why did it change to lora though?

it was working off of the stable diffusion folder before

[ornate elk]

Mostly it was a bug that it worked xD

[dreamy lodge]

same stuff again

[elder hatch]

thank you @ornate elk all working again

[ornate elk]

Perfect np

[ornate elk]

Can you try a different model?

[dreamy lodge]

swapped to counterfeit and got the same thing

[ornate elk]

Whitelist the webui in ublock origin and disable or remove privacy badger as its not needed

Then relaunch comfyui again and close Firefox before

To make sure its not an extension that blocks

[dreamy lodge]

same thing

[ornate elk]

Okay, I need to go off now as its to late.
We can check tomorrow again.
Maybe you can screenshot your environment path settings

[plush breach]

https://www.linkedin.com/feed/update/urn:li:ugcPost:7311162600313012225/ what do u think ?

[dreamy lodge]

Fixed it by reinstalling HIP SDK and ZLUDA. I had the version 5 sdk instead of the 6!

[ornate elk]

Ah okay, good that it works now 🙂

[zealous sorrel]

🆘 Hi everyone! Need help with inpaint in Krita / ComfyUI on Mac 🙏

I’m new to Stable Diffusion and ComfyUI, working on a MacBook with an M-chip, and I’m having trouble getting the inpaint function to work (including in the Krita plugin).

🔍 I noticed that ComfyUI/models/ was missing the inpaint and ipadapter folders. I manually created them and added the required model files (from Hugging Face), but it seems like that’s not enough — maybe this functionality needs to be installed properly via Git?

I tried installing comfyui-inpaint-nodes by cloning it into custom_nodes, and I also ran pip install -r requirements.txt. It all seemed to install fine, but nothing new shows up inside ComfyUI, and the Krita plugin still doesn’t work.

💡 Maybe I missed a step or did something wrong. I’d really appreciate any guidance or examples of a working setup! 🙏

[vocal burrow]

unstable internet connection

1/ ditch wifi for a wired connection
2/ squatting your mcdonalds / university connection is not a reliable internet connection
3/ might be vpn issues
4/ might be internet provider issue if you live in "some" country that would require a vpn.

[shadow oyster]

@vocal burrow How can i approch for partnership?

[ornate elk]

hey, can you try setup forge again with the linked guide from the pinned messages?
if you then get an error please show the full cmd log

[shadow oyster]

@ornate elk

[vocal burrow]

Trying to set a partnership with stability.ai ? The best way would probably be to contact them directly through their mail https://stability.ai/contact

[fluid lava]

Hi folk, I'm sure this has been debated extensively, but I can't find the proper place to find previous discussions. I've been a Mac user for a decade, haven´t built a PC rig since then. Looking to build an NVIDIA Rig for VR gaming and AI work

What is the consensus about the model/performance? Is the 5090 really worth it with 32GB ram?
Would a 4090 with 24GB be better or worse than say a 5080 16GB?

Suggestions welcomed !

[ornate elk]

hey a 4090 would be much better for ai stuff than a 5080. Because for ai stuff the more vram the better it is

for vr both will do the job pretty good

[fluid lava]

Intereseting. Even without Blackwell ?

Trying to figure out it its a future proof limitation

[vocal burrow]

still error during download

You do have free space on c: ? You re not doing crazy stuff while this is running ?

[maiden frost]

Any idea how to fix this? I am using this runpod template
with this workflow

[lucid yoke]

Yes. Basically you don't have the models required to run said model

And its saying its not in your model list

[maiden frost]

Yeah so do you think it can be mostly resolved by adding in wgets etc into my startup .ipynb file?

[lucid yoke]

Hmm maybe/no i always download all my models

Manually

[maiden frost]

lol so RunPod I have to do it every time? ugh

or pay for persistence i guess

[zealous sorrel]

hey ...

[drifting geyser]

wan was working before

what happened?

i dont get this 😦

[maiden frost]

Best way in ComfyUI to make faces less plasticy? Like, actual skin texture that's believable.

[ruby yoke]

don't use flux

[meager stirrup]

Can anyone help with this pls? I can't run the update.bat because of this error (I'm following method 1 on window 11 with NVIDIA GPU)

[broken bridge]

I am trying to do image captioning in Kohya using wd14 captioning. I uploaded the file with images I want to do captioning on but when I check Logs it says "found 0 images". Why is that, can someone help please?

[maiden frost]

Thank you.

So SDXL is the way to go?

[ornate elk]

Do you use the guide from the pinned messages?

[meager stirrup]

yeah I did, I'm following this one

[ornate elk]

What's your GPU?

[meager stirrup]

RTX3070TI

[ornate elk]

Okay and what error do you get when you launch the edited webui-user.bat ?

Make sure to run it NOT as admin

[meager stirrup]

I double clicked it

[ornate elk]

Okay seems like a system issue then.
open up a cmd as admin and run
sfc /scannow

[meager stirrup]

ok, what do i do next after running it

[ornate elk]

A pc reboot and then try launching it again

If you get the same error again, check your windows environment path settings

Screenshot them so I can compare them.
But its to late for me rn. So im back tomorrow

[meager stirrup]

ok, i'll try, thx

[ornate elk]

Np

The issue is that the system doesnt know how to launch the bat files.
Either an third party antivirus is blocking it.
Or some of the environment paths are broken or missing.

[meager stirrup]

Do i need to manually download the checkpoint file and place in model folder or it'll download for me?

[ornate elk]

It will download a very basic one

You find much better ones to download on civitai.com

[meager stirrup]

Still got the same problem, already disabled anti virus this is my path settings

[ornate elk]

Can you open up a cmd and type
Python --version
And then
Where python
And screenshot that too?

[ornate elk]

Okay that's good

At which location is your webui folder?

[ornate elk]

Okay good

[ornate elk]

Here in the bottom part click on Path and then on edit

Then screenshot that please

[meager stirrup]

this one?

[ornate elk]

Nope the bottom path one

In the section system variables

Top one is user variables

[ornate elk]

On the first look it seems okay

But I would need to compare it to mine tomorrow

Do you have any third party antivirus tool installed?

[meager stirrup]

I have kaspersky but I already disabled it

[ornate elk]

Hmm okay, and you have restarted the PC after the system check with cmd

[meager stirrup]

yea

something's wrong with this or it's ok

my webui-user file

[ornate elk]

Nope its correct

[ornate elk]

Can you show the entry for compsec ?

[ornate elk]

Looks okay too hmm

[meager stirrup]

can GTX1050 work? Maybe i'll try on my laptop to see if it work

[ornate elk]

Can you rerun the python 3.10.11 installer and there click
Modify
Then click next
And there check "add python to environment variable's"

[ornate elk]

Yes that works

[meager stirrup]

ok i did that

[ornate elk]

So it was not checked before?

[meager stirrup]

it was already checked

[ornate elk]

Oh okay

Also I saw you also have Python 3.13 installed.
Uninstall that if you dont need it for any other stuff

[meager stirrup]

I had it before, uninstalled, reboot PC, then installed the 3.10 and followed the process. It's probably residue

[ornate elk]

Ah okay

[meager stirrup]

i'll try and use my laptop instead and see how it goes, thanks for the help.

[ornate elk]

NP, let me know if it works there

I'm off for now

[fossil dew]

Okay so I want to go from SD 1.5 to using XL or SD3

I asked X's Grok

It says I dont need to redownload anything, but I need to make sure that automatic 111 is fully updated, running xformers and that the VAE is compatable?

Is that all?

[fossil dew]

Is this the dstable diffusion XL base that I need?

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors

well- im just gonna 🤞 and hope I got it right 😂

[meager stirrup]

I got it to work on my main PC, thanks a lot.

[ornate elk]

Nice, and how?

[fickle zealot]

it should look like this?

[ornate elk]

Nope your on user variables

But it has to be system variables

[fickle zealot]

i should tipe just hip_path or type Path to the HIP folder

[fickle zealot]

.

[fickle zealot]

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>

Failed to create model quickly; will retry using slow method.

[ornate elk]

This error can be ignored

[ornate elk]

Just the hip_path, not the actual path

[fickle zealot]

so i should wait?

15 - 40min?

[ornate elk]

Yep looks good. Compiling will take a while.
Just wait until the image is generated

[meager stirrup]

I deleted all the python path in environment, reboot and then it just workrd

[ornate elk]

Ah perfect so it was a broken python path

[fickle zealot]

Is it normal that 1 generation is not using CPU/GPU/RAM?

i mean

maybe a little ram

[ornate elk]

if its still compiling yes, because compiling is done via cpu only

the generation will use gpu vram when its ready to generate

[fickle zealot]

so my xeon is bad for this xd

over 45min

[ornate ruin]

Hi ☺️ trying to turn a group photo of me and 3 friends into us being knights. I'm using in paint with inpainting our faces and setting it to inpaint not masked

Is that the correct way to go? Because the results are... Very bad lol

[ornate elk]

still compiling? then restart and try generate again

[fickle zealot]

yes

reset pc or generation

[ornate elk]

generation

[fickle zealot]

so its good sign?

i havent it before

[ornate elk]

yep looks good

its still compiling

[fickle zealot]

its cancelled or something?

its strange its taking so long

before i had this error

maybe i should try another model
when i put new model there should i do something else?

[ornate elk]

nope everything is normal

still compling
but its nearly complete

[ornate elk]

normal error that can be ignored

[maiden frost]

Can somebody help me understand what CLIP is?
Right now I want to know where to put this in my ComfyUI directory (e.g. /models/clip or something?)

https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main

[fickle zealot]

finally
now it will generate faster?

when i try 2048x2048

its memory from gpu or ram

[ornate elk]

whats your gpu?

and how much RAM do you have

[fickle zealot]

rx5700xt 8b

32

[ornate elk]

which resolution did you tried

[fickle zealot]

.

[ornate elk]

that wont work

1.5 models are trained on 512x512 resolution.
SDXL based models are trained on 1024x1024

so stay near them, and later upscale

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

[fickle zealot]

but when i download model that support for exapmle 1600x1600 i can render it?

on my gpu

or its too slow

[ornate elk]

no it would take to much vram

and there are no models for 1600x1600 or higher

[fickle zealot]

so limit is 1024 x 1024?

[ornate elk]

nope, you can try genning at 800x1200 for example

but dont try with an 1.5 based model

[lucid yoke]

for SDXL/flux youd want to stay on 1024x1024 resolutions and follow these kinda:

[fickle zealot]

so SDXL/flux is better?

[ornate elk]

yep but flux wont work in auto111 and is not worth for your gpu, it needs to much vram

so sdxl is the sweetspot

for anime models checkout Illustrious based ones

[lucid yoke]

Hey Cs1o have you made it so you can acces your UI thru another network?

[fossil dew]

So Idk WHAT I did but uh- my prompt box got some crazy shit going on:

I'm trying to update from SD1.5 to SDXL in Automatic 1111

It keeps saying "FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers"

Is this something I need to fix?

[lucid yoke]

does it work?

[fossil dew]

The web UI opens, but it takes 1093204823 years for a XL model to upload when selected in the checkpoint box

they actually dont load 🥲

[lucid yoke]

hmm Forge is better for SDXL then A1111 but it depends on your hardware

if you can

[ornate elk]

Hey, open up a cmd and run
pip cache purge
then delete the venv folder inside the stable-diffusion-webui folder.

Then relaunch the webui-user.bat

[fossil dew]

it usually crashes my web thing

[lucid yoke]

ahh Cs1o to the rescue

[fossil dew]

i saw somthing about deleting the venv folder I forgot that!

Ill go do the thing brb

[ornate elk]

yep

works with auto, forge, reforge, but havent tried comfyui yet

[lucid yoke]

ooh howd you do it

[fossil dew]

jesus thousands of files

[ornate elk]

add --listen to the webui-user.bat

[lucid yoke]

hmm ah so portforwarding i suppose then

[ornate elk]

then you can access is by the PCs ip adress + port of the webui

[ornate elk]

oh you want an external access ?

[lucid yoke]

yee

[ornate elk]

vpn, or port forward, or using Rustdesk to access the whole PC remotely

[lucid yoke]

like i got a friend (different country) to acces the UI im using so i was thinking of a tool like hamachi

hmm yeah might just ball with a port forward

thanks!

[ornate elk]

there is also the --share launch arg.
It creates a link you can send him and he cann access the webui for 24hours if its open

its made by gradio

[lucid yoke]

hmm ill check it out

[ornate elk]

you can also create logins

--gradio-auth Username:password

so nobody else can access it

[lucid yoke]

ill dive into it. seems to be what im looking for mostly

[fickle zealot]

can i replace location of some files ?

when i tried new model it stared download

[ornate elk]

get a tool like WizTree to check whats taking up the space.
you later can store the stable diffusion models on your D drive and link them by adding --ckpt-dir "D:\SD-Models" to the wbeui-user.bat

[fickle zealot]

my model is already on s drive

[ornate elk]

ah okay

[fickle zealot]

I will try extend c drive if its possible :p

[ornate elk]

you can try disable hibernate to get 5-10gb back on C
in a cmd powercfg /hibernate off
then reboot the PC

[fossil dew]

Downloaded a Illustrious based XL model- Selected it in my Checkpoint box, now we just wait to see if it crashes

[ornate elk]

yea it depends also where they are stored

if they are on a hdd or ssd or external drive

[fossil dew]

i have them in the models\stable diffusion folder 🤔