🤝｜tech-support

[plush vine]

im fix torch bug by reinstalling it to 2.3.1 on AMD. Should I also reinstall it to 2.3.1 on this SD version?

[ornate elk]

nothing, use the models from here:
https://huggingface.co/QuantStack/Wan2.2-TI2V-5B-GGUF/tree/main

[analog finch]

Right, GGUF, it makes sense now

files looked weird, thank you!

[ornate elk]

no problem

[ornate elk]

i dont understand the question sry

which bug?

[ornate elk]

maybe post the completet cmd log here

[ornate elk]

the whole log please

[analog finch]

Do I go q2 first or you think higher ok?

[ornate elk]

i think you can probably easily use Q8_0

[analog finch]

Say less, ill play around settings, see if its more than noise this time

[ornate elk]

5b doesnt need a low noise (refiner), makes it also easier

[analog finch]

yeah, I read that takes vram too

I use the same text encoder from the 14b?

or none

[ornate elk]

the same

[ornate elk]

which hip sdk version do you now have installed?

[plush vine]

6.2.0

[ornate elk]

and can you show which folders you have in C:\Program Files\AMD\ROCm

[ornate elk]

okay, delete the 5.7 rar there,
Then delete the venv folder in forge and relaunch the webui-user.bat

[analog finch]

can all txt2img models do img2img? I dont see a filter for i2i models on civitai

[ornate elk]

can you try a normal image generation first

without controlnet

or does this error appears for default values?

[plush vine]

with controlnet. okay. start without controlnet

same error

[ornate elk]

okay then uninstall python 3.10.6 and install pythin 3.11.9,
then delete the venv folder again and relaunch the webui-user.bat

[analog finch]

CS1o. Model loads and loads stuff up to start generating and then just "AssertionError"

[ornate elk]

Okay, can test the 5b q8 tomorrow

[plush vine]

im think new torch not comparable and i need reinstall it to old version. What you think?

[ornate elk]

The error is that insightface is missing for controlnet Faceswap stuff.

[plush vine]

so they not on build?

[ornate elk]

Insightface not

Here is the guide how to install it:
https://github.com/Gourieff/sd-webui-reactor-sfw?tab=readme-ov-file#viii-for-windows-users-if-you-still-cannot-build-insightface-for-some-reasons-or-just-dont-want-to-install-visual-studio-or-vs-c-build-tools---do-the-following

[plush vine]

ok, thx

(venv) F:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge>pip install insightface-0.7.3-cp310-cp310-win_amd64.whl
ERROR: insightface-0.7.3-cp310-cp310-win_amd64.whl is not a supported wheel on this platform.

[ornate elk]

If you have python 3.11 installed you need insightface for 3.11

[plush vine]

(
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
blendmodes 2022 requires numpy<2,>=1.22.1, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires numpy<2, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires protobuf<5,>=4.25.3, but you have protobuf 3.20.2 which is incompatible.

Installation amd build SD for me...

[plush vine]

im delete everything and start again. Using this guide: https://github.com/vladmandic/sdnext/wiki/ZLUDA
On which part i need donwload basic rocm to get the base folder that I will replace by this: rocm.gfx1031.for.hip.sdk.6.2.4.littlewu.s.logic.7z
?

im instal hip for this guide and my rocm folder empty

[green plank]

Im on comfyui now, and i downloaded a flux nunchaku workflow from a youtuber, im using the exact same settings as him, downloaded the same modes, and his generates a beautiful image and all i get is noise.

[ornate elk]

cant help with sd.next sry, also i wouldnt recommend it

[plush vine]

reforge and focus good?

[green plank]

In the video hes using "svdq-int4_r32-flux1" with the flux hyper lora, clip_l and t5xxl_fp8_e4m3fn, i use the same settings and models and i get nothing

[ornate elk]

Forge with Zluda or auto1111 with Zluda. Are the best working currently

Fooocus isnt good.

[plush vine]

okay. starting reinstall forge..

[analog finch]

Good day CS1o, helpful as ever

ping me when you have had time to test 5b 😇

[plush vine]

same error insight after hard reinstall:
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
blendmodes 2022 requires numpy<2,>=1.22.1, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires numpy<2, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires protobuf<5,>=4.25.3, but you have protobuf 3.20.2 which is incompatible.

[cosmic furnace]

@ornate elk so what ya think about video ive dmed ya?) just an wan video model or something tricky ? 🤔

any updates on theRock ?

[ornate elk]

Will check it later, sry. Nothing major for therock

sry, canthelpt with stability matrix, it causes more errors than normal setups

[ornate elk]

there is a second way on get insightface installing.
Install Visual Studio Build Tools 2022 and then inside visual studio build tools:
make sure Desktop development with C++ is selectet (checked/installed)
Also make sure that MSVC v143 (or newer) and Windows 10 SDK is checked

then delete the venv of forge and relaunch the webui-user.bat

[plush vine]

im on win11

[ornate elk]

yea works too then, then enable win11 sdk instead of 10

[ornate elk]

got an assertion error 😮

[analog finch]

🤔 Im cooked then

[plush vine]

same error. maybe i need something delete before visual studio?

[ornate elk]

the venv folder

[plush vine]

deleted

[ornate elk]

i will open up an issue on forge neo, maaybe the dev knows how to fix

[plush vine]

trying again now

same

Exception Code: 0xC0000005

im need another torch file, im try 2.3.1 2 hours ago but doesn't work

[ornate elk]

very strange

[plush vine]

on old build im use 2.3.1

maybe it was build like your Focus

[carmine ibex]

I have to install Microsoft visual studio on win10 to install cuda 12.8, but which components of it? I'd rather not have to install things i wouldn't need

[analog finch]

appreciate that

[ornate elk]

you don't need vs studio for cuda

[ornate elk]

Done: https://github.com/Haoming02/sd-webui-forge-classic/issues/352

[carmine ibex]

Is it needed for comfyui and nunchaku?

[ornate elk]

nope

[rose apex]

It's needed if you have AMD gpu ig. I think it helps with HIP or ZLuda stuffs. If you have NVidia you dont need it

[carmine ibex]

Got it working 👍

[ornate elk]

@analog finch sadly the issue got closed as not planned.
I then tested 5b on comfyui and there it works with the 2.2 vae.
The Q8_0 used 14GB for the gen and went up to 20 for the vae step, with tiled VAE decode.

[rose apex]

I have 9060 XT 16 GB and 32 GB ram. What should be my paging file size to run Wan2.2? Currently, when I use the i2v 14B fp8 scaled models, my GPU drivers drivers crash after the first K sampler part but the Ksampler part itself doesn't happens properly. This is because it happens instantly ( 1it/s) and I am using a 4 step lora. Should I just give up or try quantized models instead?

[ornate elk]

You need to use quant gguf models. Try Q3_K_M.
If that doesnt work you need to use the Wan 2.2 5b Q8 or Q6

[analog finch]

not planned to fix? ComfyUI is its own webui right?

Is it local like SD

btw can I rename safetensors to organize them?

[ornate elk]

Make subfolders for them

[ornate elk]

Yea it's local

[analog finch]

Will subfolders make them organized in SD itself?

Dont care for the folders

sweet

[analog finch]

It went up to 20 on vae? Isnt that a bit high for me

[ornate elk]

Indeed but maybe it could be lower for you or it will switch in shared memory

[jovial crater]

Hey. I'm trying to install AUTOMATIC1111 on my windows pc (specs are AMD R5 9600X and AMD RX 9060 XT 16gb with 32 gigs of ddr5 RAM) but when I try to install it, it says an error (I'll put a ss of the error). What should I do? I have no idea, since I don't really know what I'm doing (the guide for installation on AMD gpus was texted on an ancient RX 6800 and my gpu is 2 gens newer than that. Should I try switching torch versions, maybe find an experimntal one that supports my gpu, or should I do something like switch to kohya_ss?

[ornate elk]

Hey, you should follow the latest setup guides from the first link of the pinned messages.

The one you followed is outdated

[jovial crater]

alright, thanks

[ornate elk]

Go for Auto1111 or Forge with Zluda.
Or the new Any webui with ROCm/TheRock guide

Dont use DirectML. Its just slow and unoptimised

[jovial crater]

Yeah I was looking at ZLUDA while I was waiting for an answer, I think I'll go with auto1111 with ZLUDA

[jovial crater]

I read a bunch of people on reddit and github saying that lol

[jovial crater]

just one question, sdxl models will work on this, right?

[vocal burrow]

yes

[jovial crater]

nice

[lucid yoke]

Anyone recommending a ui that only runs sd1.5 nowadays would be diabolical

[jovial crater]

ah yes

C:\Program Files\AMD\ROCm\6.4\bin/rocblas/library/TensileLibrary.dat: (?) That can't even be a directory T_T also there is no TensileLibrary.dat even if the directory was correct

wait someone on reddit tells me I should deactivate my iGPU

lemme try

[jovial crater]

from uefi bios

[jovial crater]

it's been taking 4 years and a half to load a model T_T

slow as hell

[ornate elk]

How much ram do you have?

[jovial crater]

32 gigs of ddr5, and my 9060 xt got 16gb of vram

[jovial crater]

ah yes

whatever this means

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

[jovial crater]

tell me that this thing can't run on an amd cpu and we're done

I don't wanna install it on my crusty rtx 3060 12gb pc T_T

[vocal burrow]

just explicitely tell it to run on your gpu set HIP_VISIBLE_DEVICES=xxx with xxx being your gpu id. As you have both a rocm cpu and gpu capable device, just set xxx to 0 or 1 to make it pick either your cpu or gpu.

put it in your .bat

[jovial crater]

In the webui-user.bat?

I'll take that as a yes

[vocal burrow]

yes

[jovial crater]

Alrighty, snack time then I'll go hopefully not blow up my pc

[ornate elk]

Please provide the complete cmd log if you get an error again

And yes disable the AMD Radeon TM Graphics in the device Manager or bios

[plush vine]

thx for your time @ornate elk

Are these the default settings for the runtime process queue?
I think I've found the cause of this issue:
Exception Code: 0xC0000005
In Windows 10, SD waited for the checkpoint to switch before starting generation, but in Windows 11, it simultaneously tried to switch checkpoint and start generation. I just checked, and generation on the model works. It works with ControlNet and Insight. But if I switch checkpoint, the same generation returns an error.
Maybe this SD have this button on setting?

[ornate elk]

Interrupt is only shown when Generate was pressed

[plush vine]

no, you don't get.
Im find why win11 have Exception Code: 0xC0000005
Now for sure cause im starting work with sd now normally.
When SD change checkpoint they don't pause new generation until the checkpoint finally loads.
It's trying to generate a new image, but the checkpoint hasn't loaded yet, and here's the error.
In win10 they have some fix of this, this fix doesn't work at win11

[ornate elk]

so you try to change the checkpoint mid generation?

and that crashes?

[plush vine]

yeAH

[ornate elk]

which webui have you set up?

[plush vine]

which build?

[ornate elk]

forge or auto1111

[plush vine]

old forge

on python 3.10 and git 2.47.
It doesn't matter here, I think the new forge has the same problem, I downloaded all of this

[ornate elk]

yea then the easy solution, dont switch checkpoints until the generation is done ^^
But maybe there is an option in the settings related to model, model ram chaching etc

[plush vine]

it's like this build can't set queue on win 11 correctly:
Generate current img -> loading new checkpoint -> generate new image
and queue:
standby -> loading new checkpoint -> generate new image

[plush vine]

downgrade(
less comfortable.
After changing checkpoint now need to wait, well, probably a minute, for checkpoint to load somewhere in the shadows

[ornate elk]

tested my forge zluda setup, and i can switch models while the image gens

no problem

im on windows 11

[plush vine]

on amd?

maybe you download something this:
https://dotnet.microsoft.com/ru-ru/download/dotnet/thank-you/sdk-9.0.306-windows-x64-installer
and im not

I'd like to find out what exactly you have downloaded that I don't have.

[ornate elk]

I have:
HIP SDK 6.4
Python 3.11.9 64bit
.Net SDK for 8.0.415 and 9.0.306

[plush vine]

like this?

¯_(ツ)_/¯

invisible hand of the Market

[analog finch]

I appreciate you testing, is there a special amd download for comfyui? I had it and it ran on cpu it seemed

[plush vine]

how fix this?
ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied
when im trying instal insoght face by 5.
https://github.com/Gourieff/sd-webui-reactor-sfw?tab=readme-ov-file#viii-for-windows-users-if-you-still-cannot-build-insightface-for-some-reasons-or-just-dont-want-to-install-visual-studio-or-vs-c-build-tools---do-the-following

[ornate elk]

Go to the Comfyui github releases and download the portable_amd ZIP Version.
It should work

[analog finch]

this one yes?

[ornate elk]

Here:
https://github.com/comfyanonymous/ComfyUI/releases

[ornate elk]

Where is your webui located?

[plush vine]

trying C this time

[analog finch]

thanks

[analog finch]

Do I just look up any video on comfy for setup? Saw no guide @ornate elk

[plush vine]

do you now how fix this bag, when downloading insigh?

pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
blendmodes 2022 requires numpy<2,>=1.22.1, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires numpy<2, but you have numpy 2.2.6 which is incompatible.
mediapipe 0.10.21 requires protobuf<5,>=4.25.3, but you have protobuf 3.20.2 which is incompatible.

[half island]

Hoi, does anyone know why my qwen image edit always turns out deepfried?

[onyx solar]

Have you tried turning down your cfg?

[half island]

Yep, just got less artifacts and more og image.

[analog finch]

https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html
@ornate elk people in comfy discord told me to run this

[ornate elk]

Not needed, just download and run the Comfyui amd portable version

It has these files included from AMD

[analog finch]

Can I run that dead file?

[ornate elk]

?

[analog finch]

oh im stupid

its a winrar but looks like unassigned file

[ornate elk]

Its a zip yes

Unpack to C:\ai for example

[analog finch]

Yeah alright, thanks 😛

your SD and comfyui with zluda is a hit in the comfy disc

for a few atleast

[ornate elk]

Glad to hear ^^

I will add a Comfyui with rocm guide soon too.

[analog finch]

It will run faster?

Im using SD with Rocm and im happy with speed of some models/settings

actually its forge zluda, is it sd the rock with rocm

[ornate elk]

Zluda and rocm are different setups

Rocm is faster than zluda but still in development

But it can work and I'm using it primarily currently.
Comfyui portable uses a stable rocm version.
The dev version is faster

[analog finch]

Would be nice to resetup with ROCm

[ornate elk]

Your SD?

[analog finch]

Eventually maybe, if its faster

[ornate elk]

Yea, my "Any Webui with TheRock" Guide covers the steps needed.
But I will make a individual guide for every of them later too

[analog finch]

Sweet, ill watch out for it

[analog finch]

wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors
This is a LORA yes?

[analog finch]

KSampler
at::cuda::blas::getrsBatched: not supported for HIP on Windows

[ornate elk]

When do you get this?

[analog finch]

after a little while of generating

[ornate elk]

but before getting an output?

[analog finch]

Output as in video?

I dont remember where in the process I was

It was hovering the sampler settings

[ornate elk]

ah okay

[analog finch]

Does that mean something?

[ornate elk]

would need to see the workflow and the whole cmd log to say more

[analog finch]

@ornate elk

Im running standard settings on wan2.2 5b

Did a testgen now that im home, just a simple prompt and simple start_image and control video

I dont know if settings are wrong, models downloaded went wrong or what

I didnt touch anything, it said wan 2.2 5b generated 121 frames at 24 fps at 704x704 I think

not 100% sure about resolution

Is the model wan2.2_fun_control_5b_bf16 not right? I didnt find any other ones

[ornate elk]

dont know this workflow but also fun control is not for direct video gen but only for motion control of people and stuff

cant help much with it.
For me the normal 5b Q8 gguf file works:
https://huggingface.co/QuantStack/Wan2.2-TI2V-5B-GGUF/tree/main
and using these comfyui examples:
https://comfyanonymous.github.io/ComfyUI_examples/wan22/#workflows

[cosmic furnace]

Cs1o . What are best & easiest video making tools now? wan? 🤔

Have u looked zi DM links ?)

[analog finch]

It said it didnt want gguf files, am I off?

[ornate elk]

gguf files are needed to not get out of vram

[ornate elk]

currently its wan 2.2 14b or 5b with either Forge Neo or comfyui

[cosmic furnace]

forge neo. ouch) have to wait a lil bit for therock updates i guess. have u looked into link ?) WHat ya think

Although weird but therock -forge neo seem to run better on ssd 🤨

Did they fix this issue about long VAE finish & need of tiled vae? Doubt it -_-

[ornate elk]

nope

but no issues when using the Tiled VAE option

[hushed flame]

Hi, does anyone have any tips for improving this prompt?

"Create a cinematic image of an urban street at dusk, seen as if encapsulated in a fluid space-time cube moving across a vast spatial plane. Two human figures, a man in a dark coat at point A (lower left corner of the cube, near a broken shop window, gazing longingly into the distance), and a woman in a light dress at point B (upper right corner of the cube, several blocks away, near a flickering lamppost, looking back curiously). They sway slightly in their positions, creating subtle vibrations in the fabric of the cube. The camera travels dynamically: begin with a close-up at A, revealing the local setting with shallow depth of field; rotate the cube to transition to B, showing the cube's movement as it rolls across the plane, expanding internal dimensions to reveal contrasting areas of the street—elongated shadows, distant fog, reflections in puddles separated by empty space. Maintain strict spatial separation, with forced perspectives that emphasize distance and multidimensional layers, a realistic style with surreal elements, high resolution, and dramatic lighting."

[ornate elk]

hey, wrong channel, this question should go in #📝｜prompting-help

[hushed flame]

Sorry

[analog finch]

Ok I fixed workflow and got the gguf, told me to put them in unet but it wont load

[ornate elk]

And you used the gguf loader node too?

[analog finch]

No, I saw it in a workflow screenshot

Seperate download or something I load inside comfy? Didnt find

[ornate elk]

You need to have Comfyui manager installed first

Then you can click on install custom node, then you can search for gguf and install it

[analog finch]

ComfyUI-Manager: installing dependencies. (GitPython)

WARNING: The script pygmentize.exe is installed in 'C:\SD-ComfyUI\ComfyUI_windows_portable\python_embeded\Scripts' which is not on PATH.

[DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/buttonGroup.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.

[ornate elk]

Can be ignored

[lone wolf]

Do I have to install comfyui first before using rocm the rock? https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#any-webui-with-therock-by-amd

[analog finch]

I downloaded ComfyUI-GGUF inside manager, or is it Qwen 2.5-VL GGUF Nodes

[ornate elk]

Hey, I will ad the new Comfyui guide soon. Right now its easy.
Download the Comfyui_portable_AMD zip from the comfyui releases.
Install HIP SDK 6.4 and reboot the PC.
Then unzip Comfyui into a folder created on C drive.
Then launch the start bat.
It then uses native ROCm and not TheRock.
I will add steps to make it faster too then.

[ornate elk]

Its Comfyui-gguf or just ggut

[analog finch]

Yeah I figured it out but the GGUF Loader wont find a gguf file 🤔 otherwise it looks fine I think

GGUF Loader = Load Diffusion Model in this workflow right

[analog finch]

It wont find the 5b gguf, does it not go in unet?

[ornate elk]

idk, i linked my models to a seperate folder

try checkpoints, unet or diffusion_models

[ornate elk]

here is the wan 2.2 5b workflow:

[chilly trail]

So I have a weird issue. Comfyui can see models I installed yesterday, like WAN, but today I tried to download some IL models and loras, but it can't see them seemingly?

[ornate elk]

Thats strange in which folder did you put them?

And which loader node do you use?

[chilly trail]

Load checkpoint

In the same folders as the WAN models, models/checkpoints and models/loras

[ornate elk]

Okay sounds right

[chilly trail]

So you got no idea what could be wrong?

[ornate elk]

are all the models in the same folder?

have you restarted comfyui?

[chilly trail]

I have restarted comfyui. I mean, yes. All my checkpoints are located inside models/checkpoint and all loras are located inside models/loras.

[ornate elk]

can you show screenshots?

of the files and the node with the model selection in comfyui

[analog finch]

Much love, will get to it soon

Should I keep "wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors"

Idk when that got downloaded

GGUF created a folder "C:\SD-ComfyUI\ComfyUI_windows_portable\ComfyUI\models\LLM\GGUF"

model probably goes there

[chilly trail]

I figured it out...

[ornate elk]

you can keept it if you want

[lone wolf]

Ill wait for the new guide then, cause Im a bit confused as its therock, rocm and Zluda, so several different stuff for AMD; but Im never giving Nvidia money anymore x)

Thank you for all you do tho CS1o ❤️

[analog finch]

Okay I think I download comfyui-gguf and delete gguf, the workflow u sent doesnt seem to use any of the others

nvm then

[analog finch]

Prompt execution failed
Prompt outputs failed validation:
UnetLoaderGGUF:

• Value not in list: unet_name: 'Wan\Wan2.2-TI2V-5B-Q8_0.gguf' not in ['Wan2.2-TI2V-5B-Q6_K.gguf', 'Wan2.2-TI2V-5B-Q8_0.gguf']

I had model in LLM/GGUF and in Unet

I used the same prompt as the preset workflow

even lower resolution also

[ornate elk]

There is a conflicting node, this gguf talking thing

[analog finch]

But I uninstalled, I only have one

How is it conflicting if I only have 1 node 😢

I deleted unet loader in workflow and replaced with same one 🤷‍♂️

its running, we will see

Failed at 3%

Also failed at 3% with Unet Loader GGUF/Advanced

RuntimeError: at::cuda::blas::getrsBatched: not supported for HIP on Windows

Same error message always

[ornate elk]

Strange, is HIP in your path settings?

[ornate elk]

No problem, I'll let you know when its added. If I have time, tomorrow probably.

[analog finch]

Uhh AMD\ROCm\6.4\bin is in system variables path

as a guide stated, thats not same as hip sdk is it

so im just cooked i guess

Im too fresh to have any smart ideas

problem solving skills 1/10

[plush vine]

Sorry for beying annoying.
I rolled back to Windows 10 and tried everything again.
Using now Hip 6.2 AMD chipset 7.06.02.123 AMD Software (adrenaline) 25.10.2 Git 2.51.2 Python 3.9.5 Zluda .NET 9.0.306 SDK
AMD 6700 XT Forge zluda
and have this:
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\launch.py", line 54, in <module>
main()
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\launch.py", line 50, in main
start()
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules\launch_utils.py", line 675, in start
import webui
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\webui.py", line 21, in <module>
initialize_forge()
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\modules_forge\initialization.py", line 58, in initialize_forge
from backend import memory_management
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\backend\memory_management.py", line 133, in <module>
total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
^^^^^^^^^^^^^^^^^^
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\backend\memory_management.py", line 100, in get_torch_device
return torch.device(torch.cuda.current_device())
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\venv\Lib\site-packages\torch\cuda_init_.py", line 1069, in current_device
lazy_init()
File "G:\Stable Diffusion\stable-diffusion-webui-amdgpu-forge\venv\Lib\site-packages\torch\cuda_init.py", line 403, in _lazy_init
raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

My zluda: ZLUDA-windows-rocm6-amd64.zip

[ornate elk]

I need the complete cmd log for troubleshooting

[lyric vector]

ComfyUi do we have a way to put notes in our workflows I can only see making a group and naming it

[vocal burrow]

https://comfyui-wiki.com/en/comfyui-nodes/utils/note

[tough beacon]

Hello

[ornate elk]

Its using CPU torch because zluda isnt correct.
Delete the .zluda folder and the venv folder and relaunch the webui-user.bat

[plush vine]

same error

[ornate elk]

Can you show the log?

[ornate elk]

Hmm something is not correct then.
Maybe hip or zluda are not added in Path

[ornate elk]

Hmm looks correct

[ornate elk]

Try open up a cmd and run
Pip cache purge
Then delete the venv folder and the .zluda folder again and relaunch.

[maiden coral]

Did he get a correct Pytorch

U have CPU and GPU pytorch versions

[plush vine]

how check this?

[ornate elk]

hm it still errors out as it cant load zluda correctly and then installs cpu torch

[maiden coral]

pip show list

python.exe -m pip show list

Thus is the exact command, -m argument for environment of Automatic111 or whatever engine u use

[plush vine]

in the root build folder?

[maiden coral]

And u can check the versions, Lzuda is there only for translating the GPU instructions, so pytorch should be GPU version too, right

In the folder where the environment is (the thing you've been discussing called venv folder)

[ornate elk]

you could try a reinstall, or the last thing i can offer is to look over it via screenshare sometime

[plush vine]

how reisntall it?

[ornate elk]

first try only removing C:\Zluda fro mthe path settings and then delete the .zluda and venv folder again.
then relaunch

[plush vine]

same error. i haven't now Zluda on C"\

[ornate elk]

Very strange

[plush vine]

Somehow i fix this.
Thx for lot of your time.
Maybe my mistake: Python -> Custom installation -> For every users on computer

[midnight roost]

is there anyway to wipe the vram? because i have noticed whenever my gpu uses all of its vram everything slows down

[serene laurel]

ye

[midnight roost]

oh how?

[serene laurel]

inbox me

[midnight roost]

i think i did?

[serene laurel]

alright

[vocal burrow]

They're most likely a scammer if they only want to help in DM. Be careful

[serene laurel]

just help out

[lucid yoke]

why not help out in #🤝｜tech-support so other people struggling with the same problem can easyly find it using the search function

also, previous nitro use, paid profile decor but basic discord pfp? and only joined today??

sussy

[finite robin]

Hello i'm trying to install Stablediffusion using this video -->

https://www.youtube.com/watch?v=onmqbI5XPH8&list=PLVHmAxsj-x2A9tS8Qp4oujpOnu_JXf04Y&index=3

But i get hit with an error i think

[vocal burrow]

This video is really really really old and outdated

Please follow the guide in the pinned comments

[finite robin]

oh ok

[tidal crypt]

Hi everyone!! I would like to start using sd from my phone (it has 16gb ram, while my computer only has 8) is there any way to download and run it on mobile? Thank you!

[remote hornet]

Hey guys, anyone know until when Stable Diffusion 2.1 Base will be down? I was using the CLIP encoder but Idk what variant is used in the actual model.

[odd blaze]

Are there any good Illustrious inpainting models?

[lucid yoke]

Any model can inpaint, rather whats your goal

[lucid yoke]

Maybe a 1.5 sd on an apple device

[odd blaze]

On SD1.5, there's a specific model for inpainting. Isn't there a good one specifically trained for inpainting on Illustrious?

[lucid yoke]

No? Sd 1.5 is a older generation model

Sdxl (illustrious) doesnt really need it

[odd blaze]

I'm not sure if I'm doing it correctly but in painting for illustrious kind of sucks?

[cosmic furnace]

i can see that embeddings are in use list like lora. is this okay or he did something? I mean how do i make sure my embeddings work? I do write lazypos at the start but i dont have them in use list

and now when i see its in list i kinda have doubts mine are working 🙂

[cosmic furnace]

theese errors cud probably be cuz zluda forge bad with 3-11 python or i have not quite fitting torch installed ? hmgm

[ornate elk]

Would need to see the whole cmd log

[fickle herald]

good morning folks
tried to install SD on windows yesterday
but failed several times while getting this beauty of error

i tried everything but didnt work out yet
anyone may assist?

[ornate elk]

Hey, you followed an old guide. please follow the updated setup guides from the pinned messages of this channel. First link.

[fickle herald]

uhh ni ce to know thx will check asap

[ornate elk]

NP, if you have a rtx30-50 GPU you can setup Forge Neo for the best performance.
For older GPUs with 6-8GB vram Forge or Auto1111 are better

[fickle herald]

new 5080 here, bought last week.
need to check what it is and where i can set it up^^ but probably will see

[ornate elk]

@tame cedar I wanted to help but you deleted to fast xD

[ornate elk]

You can find all of the setup guides on the link

For rtx50 Forge Neo is the best to use

[tame cedar]

😄 Sry i wanted to update it, i changed the
Schedule type from Automatic to Karras and somehow now i dont have the problem. 🙃
I only have a different problem with a runtime error ant insuficent GPU space
are there any ways to optimize automatic 1111 for AMD

[ornate elk]

Yea there are. It depends on your GPU model. Also are you on Zluda or directml or ROCm?

[tame cedar]

I’m on Windows 11 with an AMD 7900 XTX
I’m using the DirectML build of Automatic1111
I think im not using ZLUDA or ROCm.
So my setup is currently Windows 11 on a AMD 7900XTX with Automatic1111 DirectML version.
From what I’ve seen the only way to get full native AMD support is to run Linux with ROCm, is that right ?
Im Relatively new to the whole thing

[fickle herald]

pythong giving me a hard time rn...

downloaded the version it says in description

[ornate elk]

Oh damn directml is the worst of the 3 versions xD
I'm on a 7900XTX too.
And nope full support is now possible on windows too but still work in progress from AMD.
Zluda is much faster than Directml and not vram limited.
And ROCm has native support and is even faster but not as stable.
I would recommend following my AMD setup Guides from here:
https://github.com/CS1o/Stable-Diffusion-Info
For Zluda go for Auto1111 or Forge with Zluda.

[ornate elk]

Do you have any other python versions installed currently?

[tame cedar]

OK 😄 Thanks, i will try to follow the guide if i have problems i will ping you if that's ok 🙂

[ornate elk]

Sure! You can also DM me if you want help in German xD

[fickle herald]

i uninstalled everything now and trying to install 3.10.11 again

[tame cedar]

O your also German ? 😄

[fickle herald]

stuck herre for couple seconds then back to the failure

geht auch deutsch

but english fine too

[ornate elk]

Open up a cmd as admin and run
SFC /scannow
Then when its done restart the PC and try again

@fickle herald also for Forge Neo you need python 3.11.9 64bit not 3.10

[fickle herald]

kk

just close on click here

[ornate elk]

Disable Path length limit can't hurt too

But not necessary

[fickle herald]

https://tenor.com/view/done-im-done-finish-family-guy-peter-griffin-gif-15564927

works good so far 😄

thx

[ornate elk]

Nice ^^

[tame cedar]

ZLUDA-nightly-windows-rocm6-amd64.zip triggers Windows Defender 🙁

@ornate elk do i have to deactivate it because it immediately deletes the folder

Sry my bad did not read your guide completely

[fickle herald]

should i run this now?

Copy and paste these four commands one by one:
venv\scripts\activate
pip uninstall torch torchvision torchaudio xformers -y
pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install xformers==0.0.32

still need to add my models

[fickle herald]

something....broke?

tried last step before launch:
pip install xformers==0.0.32

• maybe add post1 after?

or post 2

[ornate elk]

Oh you followed the Comfyui setup guide I guess.
There its better to use my new guide which I haven't added yet instead of the zluda one.
Better try Forge or Auto1111 with Zluda first.

[ornate elk]

Like I said set up Forge Neo, not auto1111, as it doesnt support rtx50 by default.

[fickle herald]

https://tenor.com/view/dumb-huh-stupid-idiot-gif-5042504

i made it work just 1 sec before 😄

going for neo now

😄

[ornate elk]

Yea auto1111 can work with the custom torch and xformers but the webui is outdated for a year now. And Forge Neo has native support with newer torch

[fickle herald]

thx

[tame cedar]

I got the Web UI working but when i try to generate a picture i get
Compilation is in progress. Please wait...

[ornate elk]

Yea thats normal

Compilation will take a while

After thats done it will work

[fickle herald]

one question hoefully last one - should i delete the old folder or will it re-write it?

[tame cedar]

10 min ? or what time frame are we talking about

[ornate elk]

15-25 mins probably

[tame cedar]

https://tenor.com/view/disintegrating-funny-thanos-gif-22399978

[ornate elk]

You can delete it

Forge neo will have its own

[odd blaze]

Is there a custom Node that can load multiple Loras at a time (e.g., Power Lora Loader), and also automatically add the trigger words individually?

[fickle herald]

well....i think it normally should not be black...

it wasnt black at the first time but after lowering the "Denoising strength" it went to black

still black no matter what i change

[tame cedar]

works Thanks 🙂

[ornate elk]

perfect, feel free to ask if you have any more questions

[jolly moth]

Hello can someone help me with one thing I'm trying to use Stable Diffusion to color manga panels And I'm looking for best models settings and etc

[lucid yoke]

I feel like a illustrious model (pick your poison) and anime line art controlet for illustrious will do o the trick

I use this controlnet set (its hard to find because they miscatagorized it so here's the link https://civitai.com/models/929685?modelVersionId=1049196

[jolly moth]

Do you have any illustrious model on your mind if i can ask?

[lucid yoke]

Not really, you could go for a popular one like wainsfw

I do recommend using danbooru tags and a artist tag/lora

[jolly moth]

You mean this one?

Or this?

[lucid yoke]

Those are merges but sure why not

[jolly moth]

Yeah like I'm new to this so...

[lucid yoke]

Theres no "best" just preference

Unfortunately I'm on my way to work so I can't give a better breakdown if it rn

[jolly moth]

Yeah no problem

[rose apex]

Can someone please tell me what flux model should be best for 9060 XT 16 GB?

[ornate elk]

I would go for the Q8_gguf version

[rose apex]

Dev or Schnell?

[ornate elk]

Dev

[cosmic furnace]

Had to catch it

[ornate elk]

AnimateDiff error

[valid holly]

Can anybody tell me what i am doing wrong? I used pony in the past but never had results like this before.
The image pre process looks fine and right before it finished it is like someone comes over with a dirty brush and swipes over the face once. ADetailer did not change anything either.

[vocal burrow]

1/ Which VAE ? Model ?
2/ It does not make sense to set Denoiser Strength to 1. Doing that is gonna tell the upscaler "Hey, feel free to modify everything from the image I gave you. Basically use it as random noise."

[valid holly]

prefectPonyXL_v6.safetensors [20fecc2ec6]
Vae is set on automatic.
i just put denoise to 1 in the last test to see if anything changes, but had same results

[vocal burrow]

can you screenshot your whole UI ?

[vocal burrow]

is clip_skip set to 2 ?

I would add a shortcut for the "clip_skip" value if I were you , add "CLIP_stop_at_last_layers" to the Settings -> User Interface -> Quicksettings list

[valid holly]

clip skip is at 2 yes. i will try the setting u suggested

[vocal burrow]

wait .... you're using an XL model and setting the base resolution at 512x512 ?

Those are made for 1024x1024

[valid holly]

omg i am so stupid. i just knew it was something simple i fumbled... thx for pointing it out 😄

still got the same issue tho. images have higher overall quality now. but still mushed faces. i guess i just go back to Illustrious

literally my face, trying to figure this shit out.

[vocal burrow]

it might be because codeformer's face restoration is turned on

that's the kind of artifacts I would expect from it on anime output

[limpid grove]

I'm trying to do simple image to image. Like generate full person from face only. Some web stable diffusion works fine, but official api using sd3.5 gives very bad results. Any help?

Using very basic code

[cosmic furnace]

👀 . Its outdated? I shud uninstall y ? 😑

although i wasnt using it at the error moment 🤔

[ornate elk]

If you dont need it remove it, its old and not good

[woeful knot]

Original vs pony vs illustrious. Both pony and illustrious are inpainted at 0.6 denoise, same settings, cfg of 7, dpm++ 2m sampler. Pony blends in pretty much perfectly, so what is it that makes the illustrious model so bad at blending with the surroundings? I tried multiple vae

[cosmic furnace]

i keep getting error =\

while having lesser python version on prev win. or another torch. i didnt had any errors =\

The Exception Code: 0xC0000005 in Stable Diffusion, also known as an "Access Violation" or "Protection Violation," indicates that the program attempted to access a memory location it was not authorized to use. This can lead to application crashes and system instability.
Here are potential causes and solutions for this error in the context of Stable Diffusion:
Potential Causes:
Corrupted or Incompatible Files: Issues with Python, PyTorch, or Stable Diffusion's own files, including dependencies like c10.dll.
Driver Issues: Outdated or corrupted graphics card drivers.
Hardware Issues: While less common, faulty RAM can sometimes trigger this error.
Software Conflicts: Interference from antivirus software or other background applications.
Data Execution Prevention (DEP): DEP can sometimes incorrectly block legitimate program access, leading to this error.
PyTorch/CUDA Version Mismatch: Incompatible versions of PyTorch and CUDA can cause memory access issues.

indeed i get c10.dll -s file errors . what cud it be ? 🤨

gemini advice to lower pytorch version for usual forge 🤔

something like this -

PyTorch and ZLUDA/ROCm Incompatibility (Most Likely)
You are using PyTorch 2.7.0+cu118. This is a very new version and it may be causing compatibility issues with the current implementation of ZLUDA or ROCm 6.4.
The solution is to try downgrading the PyTorch version. Older, more proven versions, such as 2.0.1 or 2.1.2 with corresponding CUDA packages (e.g., torch==2.0.1+cu118), often work more stably with ZLUDA/ROCm.
You can do this by uninstalling the current Torch and installing the required version in your venv:
Bash
pip uninstall torch torchvision torchaudio
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
Attention (Warning): Ensure that you are using the correct installation command for ZLUDA/ROCm, as it might differ from the standard CUDA command.

[cosmic furnace]

feels like something with hip sdk-torch-python doesnt match up somehow. By what ppl suggest in reddit. Anyway. Any ideas ?)

[steep gate]

may i ask a link shell extension question here?

[ornate elk]

Sure

[cosmic furnace]

how do i reinstall torch to the one that fits? Do you agree with this?) CS1o ?

[ornate elk]

For Forge with Zluda?
You can just delete the venv folder.
Make sure git pull is in the webui-user.bat as in the latest update the dev added support to 2.7.1

[cosmic furnace]

TheRock if i keep same resolution i used once its kinda okay-ish. altough every new one need to generate once long (I mean 3 different ones fot txt2,img2 & inpaint) so usually i have to wait 3 times for them to "compile" . As if it precaches it or smething. Also i saw an error saying pagefile isnt big enough to continue. WHen trying to change model. Didnt happen on old forge-zluda =\

So this & need of tiled vae is things they shud improve... Tell them xD

[cosmic furnace]

Exception Code: 0xC0000005
0x00007FFA6AA660B4 <unknown module>
0x000000CC0DBECE00 <unknown module>
0x0000050645C00000 <unknown module>
0x00007FFA6AACBD48 <unknown module>

o__o what cud this be?

[ornate elk]

Could be out of vram or out of memory

[thorn tinsel]

heya im trying to run the stable difussion webui to run locally on my laptop, but im having trouble, im following h/ttps://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs the guide, runned update.bat, runned run.bat, switched to the dev branch since im using an nvidia 4050 but in the end i get error code 9009 when trying to webui.user.bat "couldnt launch python"

[vocal burrow]

dont use automatic1111 it is very outdated. Use Forge Neo instead.

Tutorial for it in the pinned comment.

[thorn tinsel]

ty, ive been installing it but right at the end, in the last step after launching the webui user bat i get this error msg AssertionError: Torch not compiled with CUDA enabled, then the cmd window closes and when i try to launch it again nothing happens and it just delivers: RuntimeError: PyTorch is not able to access CUDA

i installed python checking the path box and did default installation

[vocal burrow]

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

[thorn tinsel]

where python:
C:\Users\USUARIO\AppData\Local\Programs\Python\Python311\python.exe
C:\Users\USUARIO\AppData\Local\Microsoft\WindowsApps\python.exe

python version:
Python 3.11.9

py list paths:
-V:3.11 * C:\Users\USUARIO\AppData\Local\Programs\Python\Python311\python.exe

also i appreciate the help

[vocal burrow]

looks ok

in your stable diffusion try to delete (or rename to something else) the venv folder and then try again.

[thorn tinsel]

it responds with:
venv "C:\AI\sd-webui-forge-neo\venv\Scripts\Python.exe" Python 3.11.9 (tags/v3.11.9:de54cf5, Apr 2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)] Version: neo Traceback (most recent call last): File "C:\AI\sd-webui-forge-neo\launch.py", line 52, in <module> main() File "C:\AI\sd-webui-forge-neo\launch.py", line 41, in main prepare_environment() File "C:\AI\sd-webui-forge-neo\modules\launch_utils.py", line 314, in prepare_environment raise RuntimeError("PyTorch is not able to access CUDA") RuntimeError: PyTorch is not able to access CUDA

[thorn tinsel]

.

i renamed the dev.txt to test

[vocal burrow]

dev.txt ?

are you running on a laptop or some computer with an integrated graphics ? What are your specs ?

[thorn tinsel]

i am indeed running on a laptop but with a dedicated gpu,
processor: 12th Gen Intel(R) Core(TM) i5-12450HX (2.40 GHz)
16g ram
gpu: NVIDIA geforce rtx 4050 laptop gpu

[ornate elk]

Make sure your nvidia GPU driver is updated

[vocal burrow]

i5-12450HX also has an integrated gpu. Most likely it is trying to use that instead of your gpu.

[thorn tinsel]

uh, ill see if my gpu is updated, if it still happens it is most likely that

[vocal burrow]

in the webui-user.bat, find the set COMMANDLINE_ARGS= line and add --device-id 1 to it then try again

[thorn tinsel]

nvdia drivers were indeed not updated ima do that tooo

[thorn tinsel]

ill add it to the other ones from the tutorial

[thorn tinsel]

god updating the driver is taking forever

[thorn tinsel]

did both changes but in the end still has the same (i think) error msg
Python 3.11.9 (tags/v3.11.9:de54cf5, Apr 2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)] Version: neo Traceback (most recent call last): File "C:\AI\sd-webui-forge-neo\launch.py", line 52, in <module> main() File "C:\AI\sd-webui-forge-neo\launch.py", line 41, in main prepare_environment() File "C:\AI\sd-webui-forge-neo\modules\launch_utils.py", line 314, in prepare_environment raise RuntimeError("PyTorch is not able to access CUDA") RuntimeError: PyTorch is not able to access CUDA

something to do with "CUDA" no idea what it is

[vocal burrow]

can you screenshot the gpu tab of your task scheduler ?

looking for something like that

[thorn tinsel]

uh, its using the itegrated gpu..

alright lets try the device 0

nothing... this is just guessing but...

the --cuda

does it need a space between the = and the --?

[thorn tinsel]

same error msg btw

[vocal burrow]

let s try something else. remove --device-id 0 and instead in NVidia Control pannel make sure that the GPU is set as the prefered graphic processor. Then try again

[thorn tinsel]

sorry about the delay i was doing a workout because this was stressing me out, im back, the preffered graphics is already nvidia gpu

[thorn tinsel]

so far no luck, im going to try to reinstall pytorch

update, added --skip-torch-cuda-test to command args, result: still nothing

Launching Web UI with arguments: --cuda-stream --sage --fast-fp16 --xformers --skip-torch-cuda-test C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find cuobjdump.exe warnings.warn(f"Failed to find {binary}") C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find nvdisasm.exe warnings.warn(f"Failed to find {binary}") Traceback (most recent call last): File "C:\AI\sd-webui-forge-neo\launch.py", line 52, in <module> main() File "C:\AI\sd-webui-forge-neo\launch.py", line 48, in main start() File "C:\AI\sd-webui-forge-neo\modules\launch_utils.py", line 487, in start import webui File "C:\AI\sd-webui-forge-neo\webui.py", line 23, in <module> initialize_forge(startup_timer) File "C:\AI\sd-webui-forge-neo\modules_forge\initialization.py", line 67, in initialize_forge from backend import memory_management File "C:\AI\sd-webui-forge-neo\backend\memory_management.py", line 132, in <module> total_vram = get_total_memory(get_torch_device()) / (1024 * 1024) ^^^^^^^^^^^^^^^^^^ File "C:\AI\sd-webui-forge-neo\backend\memory_management.py", line 99, in get_torch_device return torch.device(torch.cuda.current_device()) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\torch\cuda\__init__.py", line 1071, in current_device _lazy_init() File "C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\torch\cuda\__init__.py", line 403, in _lazy_init raise AssertionError("Torch not compiled with CUDA enabled") AssertionError: Torch not compiled with CUDA enabled

im gonna try a solution i found on for a different webui called confyui

[ornate elk]

remove --skip-torch-cuda-test and delete the venv folder and relaunch

[fair oxide]

My stable diffusion folders deleted themselves completely is there a spot they go or are they just gone? They’re not in my recycle bin

Also I can’t figure out how to do text to video.

[ornate elk]

thats not normal, but then you have to reinstall it

[ornate elk]

either with comfyui or forge neo

[thorn tinsel]

still nothing, gives this error msg
Launching Web UI with arguments: --cuda-stream --sage --fast-fp16 --xformers C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find cuobjdump.exe warnings.warn(f"Failed to find {binary}") C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find nvdisasm.exe warnings.warn(f"Failed to find {binary}") Traceback (most recent call last): File "C:\AI\sd-webui-forge-neo\launch.py", line 52, in <module> main() File "C:\AI\sd-webui-forge-neo\launch.py", line 48, in main start() File "C:\AI\sd-webui-forge-neo\modules\launch_utils.py", line 487, in start import webui File "C:\AI\sd-webui-forge-neo\webui.py", line 23, in <module> initialize_forge(startup_timer) File "C:\AI\sd-webui-forge-neo\modules_forge\initialization.py", line 67, in initialize_forge from backend import memory_management File "C:\AI\sd-webui-forge-neo\backend\memory_management.py", line 132, in <module> total_vram = get_total_memory(get_torch_device()) / (1024 * 1024) ^^^^^^^^^^^^^^^^^^ File "C:\AI\sd-webui-forge-neo\backend\memory_management.py", line 99, in get_torch_device return torch.device(torch.cuda.current_device()) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\torch\cuda\__init__.py", line 1071, in current_device _lazy_init() File "C:\AI\sd-webui-forge-neo\venv\Lib\site-packages\torch\cuda\__init__.py", line 403, in _lazy_init raise AssertionError("Torch not compiled with CUDA enabled") AssertionError: Torch not compiled with CUDA enabled

[pallid berry]

Fai un video

[ornate elk]

Can you try follow the Forge setup guide?

That should work

[thorn tinsel]

it works... so far

[tribal anvil]

Is there a way to use adetailer on civit?

[chilly trail]

When using Depth controlnet, how do you avoid getting it "cutoff" when generating a new image based on it? Is my only solution to scale down the depth image to match the resolution i'm genning at?

[brave heron]

Hey I'm using this video as a guide

https://www.youtube.com/watch?v=RXq5lRSwXqo

And I got this error

I'm going to be honest, I have 0 clue on what I'm doing

[vocal burrow]

1/ you're using a video tutorial. They are bound to be outdated quite fast as they can't be edited once uploaded.
2/ you're trying to install automatic1111 which hasn't been updated for more than a year. Use forge or forge neo instead
3/ you're installing it on your desktop. Don't install stuff on your desktop that's not the place to do it because Onedrive is trying to sync it and doing so it's causing issues.
4/ use the tutorial from the pinned comment instead

[brave heron]

What pinned comment? There isn't one on the video.

Oh you mean the pinned message in THIS server

[brave heron]

Also if i can't do it on my desktop where can i do it?

[chilly trail]

Does anyone know why I can't download the qwen-image edit 2509 Q8_0 model from hugginglabs?

[ornate elk]

make a new folder on C

[cosmic furnace]

#44 0x00007ffbf35cf8fe (G:\SD\Python311\python311.dll+0x2f8fe)
#45 0x00007ffbf35ec13b (G:\SD\Python311\python311.dll+0x4c13b)
#46 0x00007ffbf35ed423 (G:\SD\Python311\python311.dll+0x4d423)
#47 0x00007ffbf35eae64 (G:\SD\Python311\python311.dll+0x4ae64)
#48 0x00007ffbf3694cf3 (G:\SD\Python311\python311.dll+0xf4cf3)
#49 0x00007ffbf35f1f7f (G:\SD\Python311\python311.dll+0x51f7f)
#50 0x00007ffbf35c7feb (G:\SD\Python311\python311.dll+0x27feb)
#51 0x00007ffbf3692533 (G:\SD\Python311\python311.dll+0xf2533)
#52 0x00007ffbf36927ca (G:\SD\Python311\python311.dll+0xf27ca)
#53 0x00007ffbf3692746 (G:\SD\Python311\python311.dll+0xf2746)
#54 0x00007ffbf378ed26 (G:\SD\Python311\python311.dll+0x1eed26)
#55 0x00007ffbf36ff0c9 (G:\SD\Python311\python311.dll+0x15f0c9)
#56 0x00007ffbf36fee38 (G:\SD\Python311\python311.dll+0x15ee38)
#57 0x00007ffbf36fea87 (G:\SD\Python311\python311.dll+0x15ea87)
#58 0x00007ffbf36fe943 (G:\SD\Python311\python311.dll+0x15e943)
#59 0x00007ffbf35ba4ec (G:\SD\Python311\python311.dll+0x1a4ec)
#60 0x00007ffbf35ba37d (G:\SD\Python311\python311.dll+0x1a37d)
#61 0x00007ffbf35b880d (G:\SD\Python311\python311.dll+0x1880d)
#62 0x00007ff7542c1230 (G:\SD\Python311\python.exe+0x1230)
#63 0x00007ffc80fd7374 (C:\Windows\System32\KERNEL32.DLL+0x17374)
#64 0x00007ffc81ddcc91 (C:\Windows\SYSTEM32\ntdll.dll+0x4cc91)
& now only this kind of error

[past vortex]

Trying to learn controlnets in sdnext... I'm getting black outputs despite following the doc pretty much exactly....

anyoneg ot any idea why this is happening?

[brave heron]

Im trying to install [AMD] Forge with ZLUDA, I think I followed everything Correctly but it comes up with this error.

should I just be installing Amd automatic1111 instead?

this is why I try and use videos because otherwise I have no idea if I'm doing anything correctly

[brave heron]

nevermind i got it working

[ornate elk]

At the end of the amd forge with Zluda guide there is this error + fix mentioned

[cosmic furnace]

why dyu think python 311 crashes on forge zluda? mb i do need lower version ? 🤔

[ornate elk]

i think its out of vram or out of memory system error

[cosmic furnace]

32 gb ram + 20 vram + 15 to 26gb pagefile 🤔

dyu think i shud ramp up max pagefiles size to up to 32 gb or whuht? 🤨

sometimes i get this error very quickly tho. kinda from start

[ornate elk]

strange

i dont get it at all :/

[cosmic furnace]

y ikr)

• i didnt had theese at all on previous windows

with almost same page file (maybe slightly bigger) + lil bit lower python & hip sdk version

[ornate elk]

hmm

try open up a cmd and run sfc /scannow

[cosmic furnace]

found some damaged files & restored them o.0

[cosmic furnace]

nah errors are all same -

probab happens when looking full screen at img during genering but not sure... i guess shud try pump that page file even more but bleh

Ill try to update zluda folder. it seems to have october 24 files there 🤔

pretty sure i did that with previous zluda forge . the one was working without errors. but not sure it ll help -_-

The rock version runs without any errors btw 🤔

[ornate elk]

when do you get this error and with which settings?

[ornate elk]

larger pagefile wont help there

[cosmic furnace]

Usually when upscaling or inpainting , this is inpainting now

forget to smitch from resize by to resize to & sometimes starting generating like this res (& ofc it can clog up memory to max) - but error happens in sessions i dont do this fail

Anyways default .zluda had somewhat weird files this ones was 0 kb & looked like shortcut files. dunno

[ornate elk]

these are symlink files

[cosmic furnace]

oh) . dunno . too early to tell but probab updating zluda folder worked

what samplers u recommend btw? for anime-toon -art ? & for realism ? Or depends on model?) Overwhelmingly popular is illustr as of now

[ornate elk]

euler a for illustrious is the best

[cosmic furnace]

hmmm intresting o.0 . a friend whos generating alot lately, said kinda same thing) and what about dpm 2m ? or dpm 2m heun?

[ornate elk]

dpm++2m is okay too, heun is bad,

[cosmic furnace]

r u sure forge -zluda runs fine on python 3.11 ?🤨

[ornate elk]

yes

totally fine

im using it like a year or so on 3.11

[cosmic furnace]

Can it happen because of dynam promts now? kinda doubt it but ...

[tribal anvil]

I keep getting an error when I try to gen what do I do?

[ornate elk]

post your specs and the full cmd log here

and which webui you use

[ornate elk]

extension error of dynamic prompting

[tribal anvil]

GTX 970m forge

I'll get cmd

Discord wouldn't open on my laptop hope this works too

[ornate elk]

how much ram do you have and which model did you tried?

[tribal anvil]

3gb

Model doesn't matter same thing happens

Tried with and without one

[ornate elk]

then i guess it wont work with this old card :/

if you have 16gb of ram you could try to increase the pagefile and see if that works

[tribal anvil]

Damn

[hollow patrol]

Hi,
Could someone help me setup stable diff locally on a PC with a 9070xt, model keeps saying not enought Vram to alocate

[main rune]

should I update ZLUDA? heard there was an update in october

[ornate elk]

No only the fork from lahqtiger bases on zluda 3 works with Ai stuff right now until it gets implemented in the official one

[ornate elk]

Hey, which webui have you installed?

And which model

[hollow patrol]

I followed the github installation for AMD on the github https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs.

Model is Pony6XL

[ornate elk]

Ah thats an outdated guide.
You have to use the one from the pinned messages of this channel.
First link.
There follow Auto1111 or Forge with Zluda.

[hollow patrol]

Ah ty.

[ornate elk]

NP feel free to ask if any step isnt clear

[hollow patrol]

Works fine now

[main rune]

what about AMD drivers?

[ornate elk]

Should be on 25.11.1 and HIP SDK 6.4

[hybrid python]

hey@all thank you for your great work!! I want to donate a large amount of ethically sourced audio data files (40+ million files + the tools to algorithmically increase the amount available), all types of hq audio assets (songs, loops, instruments, compositions etc) for ai training purposes to the stable diffusion project. Just send dm me

[cosmic furnace]

i didnt had any issues with dynamic promts before. this is weird 🤨 . Can it be because i used yaml not txt wildcard? Not sure i used it while generating last image tho. pretty sure didnt 😑

[ornate elk]

idk but it gets errors in the log

[cosmic furnace]

Do you belive in this "we lack Vr memory" for our GPU's so its gonna be less of them & prices will go up ? I pretty sure ive already heard that before 😄 . Sounds like excuses made up for pricing up -__-. It seems both AMD & Nvidia gonna be more expensive now 😑

[ornate elk]

yea as the memory chips get pricier

ssds, ram and gpus will get more expensive

[cosmic furnace]

Yeah but. Are there any objective reasons for the price increase of these memory chips?

Kinda feels like theyr trying to "scam" ppl with these statements 😑

[ornate elk]

yes, a lot of companys bought all the memory chips for their AI server farms.
so all products that require the basics memory chips are gonna have less and the prices go up

[chilly trail]

So I installed a custom node through the terminal, but after restarting I'm prompted to select a new base path which comfyui refuses to accept both S:\ComfyUI and S:\ComfyUI\env

When I search custom_node for the new node, it doesn't even seem to be there even though I got no errors when I installed it.

[teal fern]

Heyy, are there any repositories holding archival releases of Stable Diffusion? I'm specifically interested in SD 2.1, i've looked through Hugging Face but it doesn't appear available there anymore

^^^ nevermind, found a community mirror https://huggingface.co/sd2-community/stable-diffusion-2-1

[steep gate]

Is there a way to select all my checkpoints at once for XYZ Plot? shift-click doesn't seem to work

[main rune]

@ornate elk should I get HIP SDK 6.4

[ornate elk]

If your on a rx7000 GPU yes

[main rune]

uh oh

[main rune]

any news for us lowly rx6000 GPU users

[ornate elk]

AMD is still working on bringing rocm windows support for the rx6000 cards

If you only use forge with Zluda or Auto1111 with Zluda then you can easily upgrade to hip sdk 6.4 too

[cosmic furnace]

sd can rly stop working cuz of some error with dynamic promts ? o.0

What can i do about this 🤨 , didnt happen before. weird. didnt had any errors for so long. and now i get this. mb shud reload. was running a very long session -_-

[ornate elk]

dos your dynamic prompt end with a comma or with a space?

[cosmic furnace]

space and then coma 🤔

just in case i add space before & after coma 🙂

[ornate elk]

i mean at the end of the prompt

there shouldnt be a , or maybe there should

or the extensions just needs an update xD

[chilly trail]

I made a clean install of comfyui. Set the base path/env. It started just fine. I closed it down, moved my lroas/checkpoints into the correct folders. Now it cannot determine base path?

Anyone knows what up with this?

[cosmic furnace]

so qwen is just more accurate regarding promt than flux? Is it by that much better to be worth it ? tho)

[main rune]

Before I go further I would like to know if HIP SDK 6.2 works with the latest AMD drivers

Redoing everything is a pain

[ornate elk]

HIP SDK 6.2 should work with the latest drivers too

[lone wolf]

Any news on updated AMD ROCM Guide? 😄 Or should I use DirectML with comfy or the like`? ^^ Still havent attempted to download SDXL again 😛

Clueless as always <---

[fair oxide]

Hey it’s making me delete my venv folder and reinstall it each time I launch. Any solutions?

[ornate elk]

Can you share more infos, like GPU, webui, and a cmd log?

[ornate elk]

Oh nope sry, will add the guide later. Hadn't much time the last days

[lone wolf]

Ah cheers! best to just wait for it and setup the propper one instead of any of the current ones IG? 😄

[ornate elk]

do you want to setup comfyui or forge neo? what was your gpu again?

[lone wolf]

7900 XTX 😄 I've used comfyui with 4090 in the past, not sure what would be the best, I'd be happy to learn

[ornate elk]

ah alright, then sure you can install comfyui again, but if you want something easier to use forge neo is also a good recommendation

[ornate elk]

@lone wolf im done!

Updated the Guides, Sorted the different backends and added 2 new guides for Comfyui and Forge Neo working with ROCm

[lone wolf]

TYVM!!! <333 Recommend Forge Neo or Comfy mostly? Which is the easiest or prolly the smoothes one? 😄

[ornate elk]

if you want to do video stuff, go for comfyui, if you want just chilling and creating easily some images go with forge neo.
I have boath and share the models folder with them.

[lone wolf]

I have never ever delved into video creation before x) But it is possible via comfyui with a 7900 XTX? Would it take a lot of time to render both hires images and videos with this card?

7900 xtx probably still better than 9070 XT

[ornate elk]

yes its possible with Wan 2.2 14b and 5b and LTX video. It requires using smaller gguf model variants and uses a lot of vram and takes long for even a 3 second video xD

[ornate elk]

yep the 24gb vram is needed ^^

[lone wolf]

Crossing fingers that AMD release a big VRAM card next year aswell then ^^

[ornate elk]

a 32gb consumer gpu would be so insane

[ornate elk]

reload my guide site, i did some minor improvements to readability

[lost scarab]

Can somebody tell me is it possible to use SDXL models with 6gb vram for image generation at 512 res?

[ornate elk]

yes thats possible with Forge webui

and when you only have 16gb of RAM you need to increase the indows pagefile

[lost scarab]

I got 24gb ram but sometimes it still crashes

Is there a fix for that?

[ornate elk]

yep, increase the windows pagefile

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

[fiery forge]

can anyone help me? i reinstalled forge with zluda, pretty sure i didn't mess anything up, but when i try launching it get an error saying "No module named 'optimum . onnxruntime'

[fiery forge]

fuck was that shi gang why was it asking for my coinbase wallet

[fiery forge]

i already tried what was advised in the guide and its not working

[ornate elk]

Hmm that fix should still work as I tested it yesterday

Someone mentioned he tried it multiple times and in the end it worked. But he probably wasn't correctly in the venv while installing.

[fiery forge]

you helped me before dunno if you remember, but i didnt have this issue last time

was something updated?

[fiery forge]

so i should try again but do the cmd stuff while in venv folder?

[main rune]

I effed up my comfy, guess it's a good a reason to rebuild everything again

It's almost working, can even see that the denoise preview had it going

But it failed near the end

[lost scarab]

im new to forge ui and there are any commandline args that i should use?

[ornate elk]

What's your gpu?

[iron wraith]

Currently, I am experimenting with generating high-quality illustrations in a different art style while preserving the atmosphere and outfit of a specific character by using reference images.
However, because the character is quite obscure, it is difficult to describe them accurately through prompts alone. I have been trying img2img and ControlNet, but I am struggling to achieve the desired results that match the original character’s look and feel.

I would greatly appreciate any advice on effective methods to handle this kind of situation.

[frail raptor]

hello, I'm beginner with local ai related stuff, can somebody help me how I can start to generate images / videos?

[ornate elk]

hey, whats your gpu ?

[frail raptor]

rtx 4070ti

[ornate elk]

ah nice, then checkout hte first link of the pinned messages and go with the Forge Neo setup Guide.
Thats the Webui you can use to generate Images.
For video stuff you would need to setup Comfyui, which is also on the guides list.

[frail raptor]

alrighty, can I later comeback here and ask for models recommendations?

[ornate elk]

sure

here or in #📝｜prompting-help

[frail raptor]

thanks

[main rune]

hey @ornate elk, is your AMD ZLUDA tutorial still up, I need to basically redo everything since I effed up my current setup trying to update stuff

for comfy

[ornate elk]

for comfyui i link to the comfyui-zluda setup guide from patientx, but you can find that on my guides list

[frail raptor]

i was following pinned tutorial, installed python 3.10.11 but it's not working

[ornate elk]

ah you didnt checked the "add python to path" then.
Relaunch the python install, click Modify, click Next, and then select "add python to system environment variables", then Ok

then relaunch

[frail raptor]

oops, rookie mistake

[ornate elk]

np, easy to fix ^^

[frail raptor]

alright i managed to run stable diffusion, which models would be good for pixel arts and creative fictional work (e.g. fantasy designs, map top downs)

[ornate elk]

for pixel art this here is good:
https://civitai.com/models/1180112/miaomiao-pixel

[frail raptor]

i don't think that's the right output 😄 what should i change in settings to make it work?

[ornate elk]

can you show the settings? xD that would help

[frail raptor]

which one, there's a lot of them

[ornate elk]

screenshot the whole page

from txt2img

[ornate elk]

oh you have setup Auto1111 instead of Forge Neo

Auto1111 is outdated and probably has issues with this model type

[frail raptor]

oh, i did it from discord post but now i see whole github page

[ornate elk]

ohh xD

[frail raptor]

xddd is there any way to delete other stuff that has installed with this auto1111 or i just simply delete whole folder

[ornate elk]

move the miao pixel out of auto11111 and then you can simply delete that folder

then uninstall python 3.10.11 and install python 3.11.9

[frail raptor]

will pip packages be uninstalled as well?

[ornate elk]

after the forge neo install is done you can open up a cmd and run
pip cache purge
then they get cleaned

I removed the old guides from the pinned messages now, thanks for pointing that out ^^

[frail raptor]

gotta love coil whine while generating, but also now it's only black image

[ornate elk]

the model is based on SDXL so you have to select XL on the top left

and set hires steps to 10

and denois to 0.4

[frail raptor]

🙁

o wait my model is yetted out (but still black screen)

[ornate elk]

can you show the full cmd log ?

[timid echo]

What prompts or models do i need if i want realistic image of anime characters

[frail raptor]

CS, I'll get back to troubleshooting in around 1h, if you belive you know issue, lmk and I'll try that when I'll get home, cheers

[ornate elk]

the log looks okayish.
You can try to generate without hires fix enabled.
You can test a different checkpoint like this one:
https://civitai.com/models/1110783/ilustmix?modelVersionId=2110275
If that still outputs black too, then try remove --fast-fp16 from the webui-user.bat
If that still doesnt work then remove --sage from the wbeui-user.bat and try again.

Let me know how it goes!

[ornate elk]

you would need an Illustrious based realism traiend model

then its easy, as they know most anime chars ny default

[timid echo]

uhhh

yea

thanks

[ornate elk]

this here for example:
https://civitai.com/models/2081804/cyberrealistic-cyberillustrious-anime

[timid echo]

what if i want it to be like

realistic as possible

[ornate elk]

there is also this realism lora:
https://civitai.com/models/2149527/cyberrealistic-cyberillustrious-realism-lora

[timid echo]

what is a lora

sorry i just picked it up like yesterday

[ornate elk]

its like an addon on top o f a model that helps getting a specific style or character

loras are trained on specific stuff to enhance the image. But they only work together with a compatible model

[timid echo]

ohhhh

i see

so like

i need to use the illustrious thingy

[hollow patrol]

Hi,
I want to ask if there is way to improve the speed of generation, well the CMD says that the img is 100% generated but the UI gets stuck on 96% and takes around 30s-60s to finalize

[ornate elk]

hzey, thats caused by the vae step. it can be improved but depends on your webui and gpu

[hollow patrol]

gpu is a 9070xt UI is Automatic1111 with ZLUDA

[ornate elk]

ah okay, then download the fp16 vae, its for AMD.
direct link: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
put it into models/vae and select it in the webui

To enable the vae dropdown in txt2img, go int osettings, User Interface, Quicksettings, there add sd_vae
then hit apply and reload ui

[final vector]

Anyone have an example of a ComfyUI workflow where I can give it several images and then have the prompt be "this subject do X action"?

I want to upload a few images of my dog and convert it to a Christmas type card, but I am unsure how to do it. Nano Banana Pro does it really well but I wanted to automate some stuff into a workflow

If anyone replies, please @ me

[hollow patrol]

after setting the VAE, it still has the same issue

[ornate elk]

okay then go on the extensions tab, click on Available, click on Load from,
then in the list search for Tiled VAE
install that and relaunch the webui. Then at the bottom of txt2img enable Tiled VAE only.
that will work

[hollow patrol]

that worked thanks

[ornate elk]

perfect 🙂 also if you use hires fix, always set the hires steps between 10-15

[frail raptor]

hey, I've tested those options and

w/o hires fix - still don't work
different checkpoint - still don't work (set to xl)
removing --fast-fp16 - still don't work
removing --sage - still don't work

could this be issue?

D:\Stable Diffusion\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find cuobjdump.exe
  warnings.warn(f"Failed to find {binary}")
D:\Stable Diffusion\sd-webui-forge-neo\venv\Lib\site-packages\triton\knobs.py:212: UserWarning: Failed to find nvdisasm.exe```

[ornate elk]

try this:
Go into the forge neo folder then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste this command:
venv\scripts\python -m pip uninstall torch torchvision -y
and then this:
venv\scripts\python -m pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

lt me know if that worked

[frail raptor]

should i get 2.9.1?

or 2.8.0

[ornate elk]

2.8.0 then

changed the comand above

[frail raptor]

thanks, also thanks for showing civitai, it's easier to download there than hugging face lol

[ornate elk]

np, yea civitai is the best for model downloading

let me know if 2.8.0 works, will then edit my guide

[frail raptor]

yeah i will it's installing now

i think i might need to delete other dependencies

[ornate elk]

try run these commands one by one:
venv\scripts\activate
pip uninstall torch torchvision torchaudio xformers -y
pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install xformers==0.0.32.post1

[frail raptor]

.post1 or .post2?

[ornate elk]

.post1

[frail raptor]

boss, something broke even more

[ornate elk]

make sure your nvidia driver is updated then relaunch

[frail raptor]

updated, same stuff even after rebooting pc

[frail raptor]

can cuda version 13.0 be issue? torch is installing with cuda 12.8 if i get it right

[feral perch]

Hey guys, I’m doing some research for a project.
If you rent GPUs for SD or ComfyUI, what’s the most annoying part of the process today?

(Price, uptime, boot time, storage, reliability, anything.

[ornate elk]

Yea you should have cuda 12.8 installed not 13

[frail raptor]

so i need to downgrade my graphic drivers, i tried to install cuda toolkit 12.8 didn't work, tried pytroch and xformers for cuda 13.0 still didn't work

[ornate elk]

No the GPU driver needs to be updated

Just if you have cuda toolkit installed then downgrade that

[frail raptor]

i didn't had that, i installed cuda toolkit 12.8 but nvidia-smi still showed cuda 13.0

[ornate elk]

Thats okay

SMI shows the latest compatible one

[frail raptor]

do i also need visual studio or sole toolkit

[ornate elk]

Not matter what's installed

Visual studio build tools would be enough

[frail raptor]

yeah still black image

[ornate elk]

hmm that makes no sense.
then i would suggest opening up an issue onf the forge neo repo.
Maybe the dev knows whats up. I know that he changed the latest pytorch version to 2.9.1 and probably thats causing the issue

[fiery forge]

is there a way to save what i add to the cmd? it resets everytime i open up the batch file again

[frail raptor]

idk if that's relevant but whenever in console log there is mention of cuda it's always ": 0"

also this stuff is displaying before image generation progress

  x_sample = x_sample.astype(np.uint8)
Total progress: 100%|██████████████████████████████████████████████████████████████████| 40/40 [00:31<00:00,  1.27it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 40/40 [00:31<00:00,  1.21it/s]```

[ornate elk]

hmm strange

[ornate elk]

what you can try is use my guide and setup Forge webui instead

that should work

[frail raptor]

bro

[ornate elk]

it worked?

[frail raptor]

i deadass changed xl to flux

[ornate elk]

lol

[frail raptor]

just to tinker around different presets

[ornate elk]

wtf xD

[ornate elk]

thats strange can you try with XL preset again and then set clip skip to 2 ?

[frail raptor]

ilustmix also works on flux preset

[ornate elk]

can you try euler a as sampler

and can you download this vae file and put it into models/vae and then select in in the dropdown:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

[frail raptor]

first without vae second with

[ornate elk]

hmmm

[frail raptor]

i mean, at least it's working on flux preset tho, but idk if this miaomiaoPixel model is only capable of creating characters or it can generate backgrounds or other pixel art related stuff (tried that but failed)

[ornate elk]

oh okay

[timid echo]

how can i do poses for two people

do i need like a lora

[vocal burrow]

which model ?

[timid echo]

Illustrious

thingy

[final vector]

Does anyone have experience running Flux-Dev FP8, specifically using XLabs workflows?

I am running into an error that says :

UNETLoader

ERROR: Could not detect model type of: E:\Projects\ComfyUI\models\diffusion_models\flux-dev-fp8.safetensors

The problem is, I have that exact model in that exact location. I am using the XLab Workflow directly, and using the Flux-dev-fp8 from Xlab's HF directory.

[fiery forge]

could anyone help me with this 'optimum . onnxruntime' error im not sure what im doing wrong

[heavy estuary]

im having the same issue

[ornate elk]

The fix still works. Maybe can show screenshots of how you do it

[fiery forge]

could you send me some screen shots of what im supposed to do incase im actually just doing it wrong

[frail raptor]

CS, idk what happened but now i can use xl preset to generate images as well xD (just giving thumbs up)

Does anybody know if there is any good model for concept arts of items / landscapes, i mostly see models related to characters but i want more of "dead art" type

[vocal burrow]

not sure what node you're using but if it is the one I believe. This UNETLoader is meant for .gguf files, not .safetensors.

[main rune]

shit, I'm making mistakes after mistakes
what the heck did I do to zluda, it keeps deleting itself

[main rune]

` :: 11:12:21 :: - Patching ZLUDA
tar.exe: Error opening archive: Failed to open 'zluda.zip'
:: 11:12:22 :: - ZLUDA 3.9.5 nightly patched for HIP SDK 6.2.4 / 6.4.2 with miopen and triton-flash attention.

.....................................................
*** Installation is completed in 00:02:52.60 .
*** You can use "comfyui-n.bat" or "comfyui-user.bat" to start the app later.
*** If you want to modify the launcher please use the "comfyui-user.bat" as it is not effected by the updates.
*** You can use -- "--use-pytorch-cross-attention" , "--use-quad-cross-attention" , "--use-flash-attention" or "--use-sage-attention"
.....................................................

*** Starting the Comfyui-ZLUDA for the first time, please be patient...

'.\zluda\zluda.exe' is not recognized as an internal or external command,
operable program or batch file.`

sigh

never mind, it seems to be an overly sensitive antivirus

[rose apex]

Can someone please tell me why it says Torch Vram 0 B? Is it a problem?

[main rune]

does comfy not work?

[rose apex]

It does. I can generate images, my VRAM does fills upto 12 GB when generating

I am actually trying to run wan 2.2 and I am running out of memory even with the Q6 I2V GGUF models. So, I was looking into sage attention and stuff

[main rune]

I'm using a 6600xt and I was able to run wan 2.2, granted I was using the i2v 5b model

[ornate elk]

Thats wrong, you have to run the commands one by one, not together

[fiery forge]

does it automatically save after each input

[ornate elk]

No, dont close the cmd, just copy and paste the first command, hit enter and then the second command

[fiery forge]

does it need to be inside the venv folder before i start

[fiery forge]

like this yeah?

[rose apex]

I can run the 5B model as well but the results are not good. Thats why I want to try the 14B model

[ornate elk]

Yes, correct

[main rune]

Fancy meeting you here Mr president

[frail raptor]

mr president gonna create some devious propaganada to get back fr fr

[fiery forge]

there is no meme, take off your clothes

[sudden cloak]

Anyone can help me?
There’s a user who manages to generate images with insanely high detail. The sharpness in the hair strands, the skin texture, the eyes—everything is ridiculously crisp.
If you zoom in just a little on his images, you’ll see exactly what I mean.

I’ve tried replicating his setup the same way: extracted the metadata, copied every setting, matched everything… but I still can’t get anywhere close.
The difference is that he’s using the Neo version of Forge UI:
👉 https://github.com/Haoming02/sd-webui-forge-classic

Meanwhile, I’m using the lllyasviel version:
👉 https://github.com/lllyasviel/stable-diffusion-webui-forge

Could the version difference really be the reason?

Another thing: I know some people might say “just upscale it,” but if you actually look at his images, they aren’t even that high resolution, and the files are lightweight too.
I’ve generated 30–50 MB images with upscale alone, and they still don’t come anywhere near the micro-detail quality he gets.

I’ve searched everywhere, tested tons of variations, but I never reach a similar result.
If anyone who understands Forge/Neo really well could explain what I might be missing, I’d appreciate it a lot

[ornate elk]

hey, are the provided images the ones from the User or yours?

would like to check the meta data if i see something

[sudden cloak]

Hello!
These images were made by the creator himself. You can download them and check the metadata, everything he used is already embedded in the files

[ornate elk]

Ah alright, checked them.
They look normal.
Pony realism model with DPM++ 2m SDE is really good for quality and he also upscaled it

Also adetailer got used

[sudden cloak]

Are you able to replicate that same level of image quality?

[sudden cloak]

I’ve already used the exact same settings, but my results still don’t come out the same

[ornate elk]

I can try later

[sudden cloak]

When I try to replicate it, my images end up looking noisy and not nearly as sharp

[ornate elk]

You should be able too, he didnt even used loras

[ornate elk]

Can you show it?

[sudden cloak]

One moment

I’ll send you some NSFW images in your DMs, ok?

@ornate elk

[sudden cloak]

I sent you a friend request

[lost scarab]

1660ti 6gb

[ornate elk]

Okay

[ornate elk]

With forge and your gpu there is no command for improvement.
But if you only have 16GB of RAM you should increase the Windows Pagefile for a much better performance

[frail grove]

guys, my flux 1 dev fp 8 just stopped working
havent used it for a while, tried and it just doesnt load in vram, overflows usual ram (i think) and my pc just perma frozes
if i set it to using only vram it rans out of vram
i have 5060 ti 16gb
again, it worked fine before
chroma (tuned flux) also stopped working

[lost scarab]

im planning to upgrade my gpu to a 3060 12gb but is there a better gpu with the same price?

[frail grove]

flux loads in vram, but clip rans out of it

[ornate elk]

Which webui?

[frail grove]

comfy

[ornate elk]

Ah okay, thats causes by a change in Comfyui

[frail grove]

im guessing they changed it to like
models loading in ram first, then in vram

[ornate elk]

There is a commandline arg to change it but idk which one rn, can tell when I'm back at the PC

[frail grove]

okay, itll be very cool, thank you, will wait

[ornate elk]

I wouldn't by that old GPU but idk how's the price for you

[frail grove]

i think the best relatively cheap gpu is 5060 ti 16 gb and 5070

everything better costs a lot more

at least thats how it is in russia

[lost scarab]

well its even worst in iran since a lot of people charge you double the real price

[frail grove]

kinda same here lol

5070 cost is like

my paycheck

and my paycheck is average in russia

[main rune]

I think either I got zluda wrong, or torch wrong

man...

[amber obsidian]

hi guys for stable diffusion (comfyui) should i use :
my macbook pro m4
OR
pc with rx 7900xt

[ornate elk]

goos questions, you would need to compare them

[amber obsidian]

im trying to get it on my pc now and it says:

RuntimeError: CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero.

i think it is because the new amd drivers dont work

how do i roll them back safely

[ornate elk]

no im on AMD and the drivers are okay

[amber obsidian]

really

[ornate elk]

you probably just installed the wrong version of comfyui

[amber obsidian]

i did this before tho and they worked with the older drivers

can you tell me

which one i need to get

[ornate elk]

follow the AMD Comfyui with ROCm guide

from the pinned messages

first link

[amber obsidian]

aah thanks but i want to do it with stability matrix because i dont understand computers at all and it seemed simpler

what version drivers do you have? newest?

and this should be fine right?

[ornate elk]

wouldnt recommend using stability matrxix. but yea then go with comfyui zluda. its not as fast as my guide with rocm but works too

[pearl panther]

hello, can anyone help me with this? I got forge ui running back on 4090 pc but on new 5090 pc it always shows this error

[ornate elk]

which webui do you use?

[pearl panther]

its forgeui

I will tell you exactly what I downloaded

[ornate elk]

ah then you need to use Forge Neo now, as forge dont have native support for rtx50 gpus

setup guide can be found on the first link of the pinned messages here

[pearl panther]

ok I found the link (its a wall of text so it will take me some time to go through it)

[ornate elk]

yea but it shouldnt take too long to setup, but feel free to ask if any step isnt clear

just make sure you follow the Nvidia Forge Neo guide

[pearl panther]

yeah sure I will

if I installed some other python, do I just also install new version thats linked in guide?

or do I need to uninstall version I have

[ornate elk]

you need to uninstall the old one

[pearl panther]

oh I did steps with pasting arguments and running that webui-user

had that feeling with uninstall

yeah now it messed me up, even tho I installed new one

[ornate elk]

ahh thats easy to fix

[pearl panther]

there was thing where I had to paste python path

[ornate elk]

rerun the python installer, click modify, click next, check "add python to system environment variables" then ok

then relaunch

[pearl panther]

it kinda still wants old python3.10

[ornate elk]

delete the venv folder

then relaunch

[pearl panther]

oh

thx that helped

now its doing something

[ornate elk]

perfect

[pearl panther]

so now it opened ui, and when I wanted to generate something (used some old sd1.5 model that used to work) it shows this message

[ornate elk]

can you show the complete cmd log?

also make sure your nvidia driver is updated

[pearl panther]

i will check the drivers meanwhile

[ornate elk]

Go into the forge neo folder then click in the File Explorer bar (not searchbar) and type cmd then press enter.
Then you copy and paste this command:
venv\scripts\activate
pip uninstall torch torchvision torchaudio xformers -y
pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install xformers==0.0.32.post1
lt me know if that worked

[pearl panther]

ok, its installing something

oh no, another error

[ornate elk]

relaunch

[pearl panther]

I did 3x

same result

[ornate elk]

ohremove --xformers from the webui-user.bat and delete the xformers folders from the venv folder
then relaunch

[amber obsidian]

i am trying a workflow from someone for wan2.2 but what is this

this list goes on

how do i fix this

[pearl panther]

I did delete xformers from here: both empty rectangle as its deleted and now tried also to remove xformers where arrow is

there was nothing about xformers in venv folder

[ornate elk]

its in venv/lib/sitepackages

[pearl panther]

both these?

ok, now webui in browser launched, but if I want to generate something, I get again new error

[ornate elk]

is your gpu driver updated?

[pearl panther]

yes I updated it 10 min ago

[ornate elk]

maybe reboot the PC and then relaunch

[pearl panther]

ok will do that

same error as last time I think

[ornate elk]

hmm strange

i opened up an issue on forge neo github:
https://github.com/Haoming02/sd-webui-forge-classic/issues/408

[pearl panther]

ok something new I noticed

[ornate elk]

hopefully the dev will fix it fast

[pearl panther]

when I open XL tab instead of sd tab, even if I plug model that worked back on sd1.5, it tries to generate image (its black)

[ornate elk]

ah someone else reported that too yesterday

[amber obsidian]

anyone know how to fix this?

[pearl panther]

I will quicly go and download some sdXL model and try it with sd xl tab

[ornate elk]

nope sry

[amber obsidian]

NP thanks for looking anyways

[pearl panther]

so interesting thing, when I downloaded SDXL model it worked and it even generates images, original sd tab with model I used before didnt work for some reason I will make sure and download different sd1.5 model if it wasnt models fault

[ornate elk]

alright let me know

[pearl panther]

just to be sure, here go sd 1.5 checkpoints, right?

yes, same result even with different checkpoint

I wil try flux, in a bit, those models take a bit to download

[amber obsidian]

@ornate elk maybe you know what that error means?

[ornate elk]

nope sry, cant help much with comfyui :/ but they have an channel here #🧣｜comfy-ui or their seperate discord to ask

[amber obsidian]

aah ok

[pearl panther]

so for flux it says I dont have clip state dict, I remember having to get some other files to get flux work

I got these and placed into textencoders, still I cant find vae that it tells me its missing

@ornate elk ok I got that ae file and now its also working with flux, if you ever figure out what was with original sd1.5 please tag me

but thank you so much with the help, I would have no idea to do without it

[ornate elk]

No problem, good to know it works now.
Will update you if the dev changes something

[main rune]

man...

[steep gate]

Am I allowed ask a question about an nsfw image here?

[main rune]

like if it's against the rules or prompting help? If it's the latter you should probably join the civitai discord, they have a nsfw channel over there

[ruby yoke]

this is the tech channel. you should probably ask about images in the image channels

[heavy estuary]

can somebody please explain for the life of me what I'm doing wrong

[iron lantern]

when ever i am gening a image its fine then the preview image turns to black and stable crashes any tips ?

[ornate elk]

That looks correct

[fiery forge]

mine looks the same as his and doesn’t work, idk if i have something messed up or what

first i thought it was my browser or something because last time i did this my main browser was edge and i started using opera about 2 months ago

[lost scarab]

can somebody tell me how tf can i disable vae here? is there a "none vae" option that im not aware of?

[ornate elk]

whcih model do you use?

[fast nova]

Whenever I try to generate something using ROCm, it VERY often (like 0-3 generations) freezes (monitors images stop changing forever) until reboot. Is it possible to fix? 9070XT
I tryed to use Forge and SD.Next – both have this issue, so I guess that's something to do either with rocm, either with amd drivers

[wicked night]

Hello, I installed Forge Neo to use Z-Image. I downloaded the Z-Image model but when I try to generate a image. It gives me this error: "File "C:\sd-webui-forge-neo\backend\loader.py", line 157, in load_huggingface_component
assert isinstance(state_dict, dict) and len(state_dict) > 16, "You do not have Qwen3 state dict!"
AssertionError: You do not have Qwen3 state dict!
You do not have Qwen3 state dict!"
Any help please..

[ornate elk]

on windows or linux?