🤝｜tech-support

[ornate elk]

but normaly i would say no, not normal

[ornate elk]

your python should be on 3.12.10 and your torch+rocm is old too

with boath updated it should be faster

also euler a is faster as dpm sde by 2x

[fast nova]

Hm, I just updated the rocm recently with pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ torch torchvision torchaudio --no-cache-dir, that actually made image generation work at all. Is there a newer torch+ROCm on Windows?

[ornate elk]

you first have to uninstall the old version, because now your using the old ones mixed with files from the new one

[ornate elk]

Setup ROCm:
While in the webuis folder click in the File Explorer bar (not searchbar) and type cmd then press enter.
8. Now run these commands one by one:
venv\Scripts\activate.bat
pip uninstall torch torchvision torchaudio rocm rocm-sdk-core rocm-sdk-devel rocm-sdk-libraries-custom -y
Now copy and paste only the one command that is for your GPU type:
RX 9070/9060, XT, (GFX120x):
pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/

[fast nova]

I did uninstall before installing, but I'll try again just in case

It didn't changed

[ornate elk]

yep thats strange, maybe you werent inside the venv

or better delete the venv and then relaunch the webui-user.bat, then run the commands from above again

but first update python

[fast nova]

There are cmd while reinstalling and updating

[ornate elk]

oh your in stability matrix

rip

that just adds problems

better do a normal install of forge neo

[fast nova]

Don't worry, I didn't used it. It's just convinient place to store all my UIs)

[ornate elk]

putting it there breaks it still

please use my updated guide for installing forge neo

[fast nova]

it's just path, it breaks nothing

otherwise you'd see corresponding messages in cmd

and it's totaly can't affect anything with pip

especially when it's not even launched

Well and also, since I didn't install it with SM, it don't even see my forge-neo installation and so can't affect it

[ornate elk]

stability matrix messed up a lot in the past with paths and stuff and permissions... so thats why i like it clean

[fast nova]

it just use symlinks for stuff like models that can break some paths if not used correctly, but that actually not SM's problems. Anyway, how do I update python in forge-neo's env?

[ornate elk]

update python, delete venv, relaunch

[fast nova]

Do you mean, system python?

Do I need 3.12.10, not 3.12.12?

[ornate elk]

yes 3.12.10 as 3.12.12 doesnt have a windows installer

[fast nova]

It's updated. Should I launch UI to see if python version changed?

It did changed. What's now?

[ornate elk]

oh you mean python yes

now you have to uninstall torch+rocm again

then open up a cmd and run
pip cache purge
then install the latest version

should be this when relaunching:

[fast nova]

you mean, with the same pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ command?

yeah, it seems it installing this new version

[fast nova]

Hm, UI still shows old version although cmd said it installed new one:

Also, while launching, there's this warning: bitsandbytes library load error: Configured ROCm binary not found at N:\Programs\Stable-Diffusion\_Stability Matrix\Data\Packages\ForgeNeo\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_rocm71.dll (I checked, there's no libbitsandbytes_rocm71.dll, only ones for cuda. Should I reinstall it or it's something that's not used for image generation?)

And also this warning N:\Programs\Stable-Diffusion\_Stability Matrix\Data\Packages\ForgeNeo\modules_forge\initialization.py:40: UserWarning: expandable_segments not supported on this platform (Triggered internally at B:\src\torch\c10/hip/HIPAllocatorConfig.h:36.) torch.zeros((1, 1)).to(device, torch.float32)

[rose apex]

You should try manually installing forge and upgrade to 7.2 there

[fast nova]

Somehow these commands messing with my system's python instead of one in venv

[fast nova]

Oh, my activate.bat was incorrect, it has old path in it

[fast nova]

I updated the rocm, it seems to be better (32s instead of 2 minutes), although tiled upscale (CPU Composite) still takes most of the time. Not sure if normal or not. It also takes the same amount of time and shared memory no matter of upscale factor (1.1x=1.5x=2.5x).
Ignore pink question – it doesn't make a problem anyway, it's probably checkpoint just sits in VRAM even when not generating anything

[plain nymph]

Im using SD Forge Neo and IMG2IMG is only generating black. I am using Juggernaut XL (Ragnarok) as recommended but I am only generating black images

My commandline_args --cuda-stream --sage --fast-fp16

I tried removing --sage and --fast-fp16 as grok said they could be the culprit but thats not the case

Juggernaut claims having built in VAE but I have tried without the SDXL_Vae and with but still black

any idea?

[ornate elk]

are you using a negative prompt?

if not, than thats the issue

SDXL always needs a pos and neg prompt in forge neo

also add back --sage and --fast-fp16, they improve the performance

[plain nymph]

He said in his prompt guide to start without negative and then add stuff i want gone

[ornate elk]

yep but the guide doesnt know about this specific thing of forge neo that needs a negative

its due a recent update that its required now, forge neo dev also said he doesnt "fix" it

just type blurry, deformed, as negative

[plain nymph]

I added back sage and fast fp16. And negative prompt was all I needed ty 👍

I played around with inpaint, I had a picture with bright sun and when changing costume/outfit it completely messed up the natural lighting and every gen is noticeable darker, so I just stopped

Any idea? Ive tried all schedulers and samplers and different cfg, denoising etc but nothing seems to help

Even prompts that enhance sun and natural lighting

[rose apex]

Do you face any issues on windows where after the first run the speed goes down dramatically? - RX 9060 XT

Like for a SDXL model with 3-4 loras at 832x1216 -> I get ~2.75its/s then re running it through a second Ksampler after 2 times upscale -> I get ~2s/it.

This speed depreciates on the 2nd run and after. I start getting ~1.3-2s/it on 832x1216 and around ~2-3s/it on the second Ksampler.

[ornate elk]

chaning stuff with sdxl is not good for such stuff

[ornate elk]

can you show the workflow ?

[plain nymph]

Do you have a fix for changes being darker than original picture or do you recommend another model?

Still using Juggernaut XL

[plain nymph]

Can a Lora fix it?

[ornate elk]

can you show an example how darker ?

have you played with the denois?

[rose apex]

The Inpainting model right? Normal models dont do well with inpaint. There is an inpainting model specially for JuggernautXL

[ornate elk]

you always need to use tiled vae for encode and decode for AMD

and the slow part is the esrgan upscale by 2 as comfyui handles that different than forge neo

and why are you using an empty latent image as input?

[rose apex]

The VAE part always happens really fast though, instantaneous for the first one and like 5-6s for the 2nd. Yes the upscale part takes the longest time and i dont know of an alternative yet. I doubt I can directly generate that high resolution without 2 ksamplers. Also, aren't we supposed to use empty latent image as input? I have always done that

Sometimes I use this

[ornate elk]

yea thats okay

just wondered why using a custom node for that

when comfyui has it build in

[rose apex]

I dont have to remember other resolutions because of the list

Although I just use the same resolution 99% of the time lol

[ornate elk]

ah okay

xD

try this workflow from me, its a Latent Hires Fix

[rose apex]

Ok sure

[rose apex]

Wow this works really great indeed. My own workflow used to take around 160-170s. This one is done in 57s for the first run and 30s for the second.

The upscale happens very fast as well but how come it works with just 10 steps on the 2nd sampler?

[ornate elk]

here is the same workflow but with an Esrgan Upscaler with Upsclae by 2 (0.5), let me know how that works

[ornate elk]

10-15 steps, more are not needed for the upscaling part, as its just denoising to get more details

[rose apex]

I see. Do you use anything for better hands ?

[ornate elk]

nope

upscaling fixes them mostly

[rose apex]

I'm using the 4xanimesharp model for upscaler but regardless that's what takes most of the time.
For this workflow, the first run took 112s and the second run took 104s

[ornate elk]

yep, its much faster in forge neo using the same model :/

currently now way to improve the esrgan based upscalers in comfyui

[rose apex]

Yeah much faster than what i was doing before. could haved so much time. thanks

I'm satisfied with the details but the hands, i guess i have try a few embedding idk

[ornate elk]

yea esrgan upscalers cant fix the hands sadly, only latent can it in the latent space

[rose apex]

Mhm. I have just recently started learning about inpainting and Seedvr2 stuffs. Will try them one by one

[ornate elk]

SeedVR2 is very nice! but not that great for anime stuff, but for realism, its great

[rose apex]

What would be good for anime?

[ornate elk]

for upscaling?

[rose apex]

Yes.

[ornate elk]

Latent nearest exact or any anime esrgan model

[rose apex]

Oh ok

[ornate elk]

and facedetailer for face fix

[rose apex]

Forge also had adetailer for hands i remember, cant we do that with comfyui impact pack?

[ornate elk]

yes but inpainting hands is always pretty bad imo

[untold path]

Hi ! I'm having some trouble launching the Automatic1111 webui. At some point, launch.py tries to update and git clone from github (https://github.com/Stability-AI/stablediffusion.git - in modules/launch_utils.py) but the project doesn't seem to exist ? Or maybe I'm on the wrong branch ?

[trail monolith]

same issue here I'm trying to redownload it after a fresh reset and it keeps trying to clone something whch doesnt exist any way to bypass that

[untold path]

the master branch seems to not be updated anymore ? I got it to work by going to the dev branch but I have no idea if it is the correct choice...

[carmine ibex]

Getting this strange texture on everything. Using nunchaku qwen edit 2512 to generate, with lighting 8 step lora. Its still there even if i take the lora off and do 20 steps

[glacial cairn]

Anyone know why my flux2 klein edits are always making my images slightly grainier or diertier than the original? even when i tell it to do nothing. Or eventually the result is a pixelated mess or a black box. and i have to restart my comfy

[ornate elk]

@untold path you have to follow the updated guides in the pinned messages of this channel

better use an updated webui like forge neo if you have a newer gpu

[lyric vector]

Using comyfui portable

Already done this
but it error is still showing that I don't have seedvr2

[ornate elk]

Whats in your custom nodes folder?

[unreal talon]

@rose apex i installed forge as you adviced me but when i generate image and check at the data info it says cpu instead of gpu. how do i tell the app to use my gpu?

this looks great but she appears to look younger than the woman on the pic

[ornate elk]

What's your GPU?

[unreal talon]

rtx 3070

[ornate elk]

Then best is to install Forge classic instead of Forge

Its more updated. A setup guide is in the pinned messages of this channel

Also make sure your GPU drivers are updated

[unreal talon]

but i use forge by neo

i read that normal forge isnt maintained anymore

cant i js tell the app to use gpu?

[ornate elk]

Ah well forge neo is even better

[ornate elk]

Can you share the cmd log?

When you relaunch it.
Then we can see why its not using the gpu

[analog finch]

How do I contact someone to ban some1 from this disc

Which role handles that

[unreal talon]

here's what i get when i generate a pic@ornate elk

[analog finch]

💀💀

Who bans these nerds

@pure tusk

Bro think I want a lecture on BYD🤦‍♂️

I didnt click btw

[vocal burrow]

Didn t I already ban them this morning. Maybe it s a new type of spam bots. Anyways, I've dealt with it.

[mild hawk]

Guys can you install A1111 on you CPU only?What command do I use?When I tried to install SD it moved me to git and after that the cloning coulnd no longer work.I get the same error.Can someone guide me and help me?Are there tutorials for A1111?I didnt clone the git all I did was download the file from Github

After I retry the steps I receive the verified sign and the part says it can't find the link

[unreal talon]

does anyone know how to get reactor extension working on forge ?

[ornate elk]

Can you send the complete cmd log?

[unreal talon]

so i created several pics and this is the msg i get

pic info: Steps: 20, Sampler: Euler a, Schedule type: Automatic, CFG scale: 4, Seed: 2755311351, Size: 896x1152, Model hash: 6a2e0c8dd7, Model: novaAnimeXL_ilV150, Denoising strength: 0.6, Clip skip: 2, RNG: CPU,

how do i make it use the gpu?

[vocal burrow]

it already is

[unreal talon]

my bad

where do you see that?

[vocal burrow]

1/ "The current free memory for GPU is xxxx"
2/ you have iteration in "it/s", usually with cpu it would be ""s/it"

[unreal talon]

okay thanks for the info, thats reassuring
i have another question: do you happen to know how to get reactor extension working? i worked fine on A1111 but it keep having an issue with forge

[vocal burrow]

no sorry, I don t use reactor or any kind of face swap thingies. It s really not my thing.

[unreal talon]

[ornate elk]

Install it trough the extensions tab with this URL:
https://github.com/Gourieff/sd-webui-reactor-sfw

[ornate elk]

Which launch args do you have in your webui-user.bat?

[unreal talon]

i get thisthis error:
A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.2.6 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'

[unreal talon]

the other user told me my gpu is being ised actually😅

this message shows up @ornate elk

[ornate elk]

ah yea reactor is not updated and needs old stuff to work, you would have to manually downgrade numpy

[unreal talon]

how do i quickly downgrade? wont it cause any issues with my current version of forge?

[ornate elk]

go into forge neo folder and there click into the explorer address bar (not searchbar).
Then type cmd and hit enter.
Then copy and paste these commands one by one:
venv\scripts\activate
pip uninstall numpy -y
pip install numpy==1.26.4

idk if it can cause issues with other librarys

[unreal talon]

f thiss im done, rip reactor if only the dude could make itg compatible with newest version of numpy

i'll stick to what i already know

[paper ridge]

hello 🙂 would someone be able to guide me on the best way to install SD with support for video generation on AMD?

[ornate elk]

Hey, whats your GPU?

[paper ridge]

hey! its the 7900 xtx - 24gb vram

[ornate elk]

ah nice, then you want to setup comfyui

its the best for video gen currently

i have the same gpu, it works okay but not as fast as nvidia

You can checkout the pinnes messages of this channel for the comfyui with ROCm setup Guide

[willow venture]

Hey everyone, I need some expert advice on my ComfyUI workflow for a realistic ai character.

Current Setup:

• SDXL LoRA (self-trained)
• Checkpoint: Juggernaut XL
• Face Detailer with a separate prompt

I'm struggling with facial consistency across different prompts (the face looks slightly different depending on the setting). I have a few specific questions:

1. Trigger Word in Face Detailer: Do you recommend using the LoRA's trigger word inside the Face Detailer's separate prompt as well? Or is it better to keep the Detailer prompt generic (e.g., "high quality skin, sharp eyes")?

2. PuLID for Consistency: Does it make sense to add PuLID into the mix to "lock in" the identity? If so, does PuLID usually play nice with an active LoRA, or do they tend to fight each other for control?

3. Face Detailer "Guide Size": I’ve seen conflicting info online. Some suggest 512, others say 1024 for SDXL. What is the sweet spot for Juggernaut XL?

4. Denoise Settings: What's your go-to denoise strength for the Face Detailer when trying to maintain a very specific trained identity?

Any insights or workflow tips would be greatly appreciated! Thanks in advance! 🙏

[paper ridge]

awesome thank you! sorry for the delayed reply 🙂 so do you think ROCm is better than ZLUDA?

[ornate elk]

Yes its better!

[paper ridge]

awesome 😍 last time i think i only setup zluda, but that was a year ago or so

i should specify im on windows also!

if that makes a diff

[ornate elk]

Yea rocm improved a lot in that time

Comfyui with rocm works on windows too. That's how I use it

For video stuff you can use WAN 14b Q4 models

It takes around 5 minutes for a 5 second video

[paper ridge]

oo does that support audio as well?

[ornate elk]

Nope Wan not, only ltx 2

[willow venture]

Please someone help me😭

[rose apex]

I am waiting for the sage attention installation guide :/

[ornate elk]

I tried to help someone installing it and it failed for him so idk.
Its so tricky with the visual studio build tools required that I can't guarantee that it will work. But I will make the guide soon

[rose apex]

Alright

Btw I dual booted Linux to test if there's any speed difference between Linux and Windows but there wasn't. It takes exactly the same time for Wan generations. For SDXL T2I, I think the VAE decode might have been slightly faster because it was taking ~10-12% less time compared to windows but it could be Placebo.

No difference for me because I think my 32 GB Ram could be the limiting factor. The benchmark I saw was using 64 GB.

[strange iron]

I want to download Stable Diffusion For video and 3d but I didnt get license tho i got mail. Can anyone help or update me?

[vocal burrow]

what do you mean exactly ? what did you do ?

[unreal talon]

why cant i see the batch option for controlnet? im using forge webui by neo

[strange iron]

I want to download it but it doesnt give me permission

as far as i know its open source isnt it?

[rose apex]

I did manage to install Sage attention 1.0.6 in Linux but idk how to verify it's working? The Patch Sage Attention KJ node requires sage attention 2 or above.

[ornate elk]

You have to use --use-sage-attention as launch arg

And you need to have triton installed too

[ornate elk]

It already shows "using sage attention"

So it should work

[rose apex]

Using patch sage attention KJ node gives this error.

[ornate elk]

Idk which version of sage attention gets used but my guess is V1

[rose apex]

Hovering over the Kj node says requires sage attention 2.2 or higher

Maybe I didn't install Sage correctly? I had triton automatically installed with the Rocm package from amd-script so I just activated venv and did pip install sageattention==1.0.6

This is my Triton version

[ornate elk]

Idk if it natively works in linux

On windows there we need to patch some files to make it work

[vocal burrow]

What is "it" ? What are you trying to download ? From where ?

[strange iron]

from stability.ai, AI model an open source code or software.

[vocal burrow]

ok I see, just wanted to make sure you didnt get it from some scam website. Well then it s probably best that you contact them directly by mail. This channel is mainly community driven and help people trying to generate locally.

[scenic gale]

Best 2026 method to get consistent character generation in stable diffusion / stable video please

[scenic gale]

Anyone?

[unreal talon]

okay igot reactor working on forge

finally

[ornate elk]

nice, what was missing?

[unreal talon]

nothing, i followed your advice and downgraded numpy then i had reinstalled cuda, cudnn and onnxruntime to newer version and it works
i was essentially missing requirements

[ornate elk]

ahh okay

[unreal talon]

thanks a lot for the help!

[ornate elk]

no problem, glad it works!

[unreal talon]

forge is so fast compared to A1111, the difference is night and day holy

[wicked monolith]

Trying to run forge on 5090 and encountering problems. It worked the first time and then every time after that it encounters various errors. Have tried looking it up online and there are problems noted but can't work out how to use certain solutions. Help would be welcome

[ornate elk]

Hey, you need to use Forge Neo, as Forge is outdated

[wicked monolith]

That seems to be working, thank you!

[ornate elk]

Perfect, no problem

[candid walrus]

How long do you think it will be till I can take image of my OC character I use for roleplay and say, change the thing she is holding to xyz and put her in a dress instead of a crop top or something like that even if I need to be a little more exact with my prompt and phrase it better and it will just do that as if I paid the artist more money for another few edits? And if it is possible on consumer hardware now how do I? I'm talking complete I ask an actual artist to make the edits as well and side by side people can't tell even if asked which one is AI?

I'm learning to draw but I'm wondering how practical that is because I'm doing it for use not because its fun aka I want good OC refs for my roleplays and I can't just pay people so sometimes I go no ref and not because drawing fun...

Maybe it is possible now but every time I have tried I never got good results and that was when I was just trying to make an already existing anime character not look creepy horror movie version and there are plenty of trained images on that already but I also don't just have 24gb of vram either.

[placid pier]

Hey guys, do you know what the extension is called that lets you describe an image?

[cobalt agate]

There are several. Florence2, Janus, QwenVL.

[candid walrus]

Like how GPT doesn't need you to be precise with the prompt and say no xyz and xyz and just black bla like you to describing things to real normal humans?

[vast crest]

Hi, I noticed thanks to a user that my UI/generator isn't loading LORA properly for some reason. I use SD Forge Neo from browsers.

[misty marten]

I can't install MagCache on ComfyUI.

[ornate elk]

Make sure Forge neo is updated

Then show a cmd log if you gen with a lora

[rose apex]

https://github.com/woct0rdho/triton-windows/pull/179#issuecomment-3707313213
This thread is talking about some Sage patches to make it work but they have a 6800xt. I dont know if this gonna work for RDNA 4

[vast crest]

and updated in your opinion?

[ornate elk]

Neo doesn't shows a version. You need to add
Git pull
To the webui-user.bat

In and empty row/line

Then save and relaunch

[ornate elk]

Yea that's what you have to use. A more detailed guide can be found here:
https://github.com/patientx/ComfyUI-Zluda/issues/170#issuecomment-2972793016
But you have to use other files for the gfx1100

[tribal oak]

is neo acting up for everyone else?

@ornate elk what am i missing?

i just fixed the extensions folder _unprompted thing and did git pull in webui-user.bat

[ornate elk]

Have you removed the unprompted extension, as it causes these errors

And add git pull to the webui-user.bat as the dev updates nearly daily

[candid walrus]

Trying to create ref sheets of some of the characters I have already created before I go and use AI to make new ones and GPT spat out a mostly fine ref sheet with some minor things wrong, from the original and was to censored to allow just the hair to cover the breasts it made me give them a sports bra for the cervitaur (deer centaur)

I asked in a different server and did not find a better online alternative. If I ask GPT to fix any of the mistakes it only makes the entire image worse.

TLDR if I setup something local can it get better then this so I don't have that looks like it was done by AI look? It did stuff like the bow string blended into the chest or the girl had one bandage on her leg and in some angels it had 2 some it had the one and minor stuff like that I'm sure in 4-6 months I can learn to fix in drawing myself but still...

[summer agate]

anyone can help me with making my amd gpu work? so far i think ive done most of the steps but now it says this : Warning: caught exception 'No CUDA GPUs are available', memory monitor disabled

[tribal oak]

it told me to rename it from unprompted to _unprompted

so i have to actually delete it?

its working now. thanks @ornate elk

[ornate elk]

alright, no problem 🙂

[ornate elk]

renaming it makes it useless, better delete it

[ornate elk]

hey, whats your gpu?

[summer agate]

Rx 580

Yeah i know its kinda bad but there was something for it to make it work

[ornate elk]

it can work with zluda or directml

guides are in the pinned messages of this channel

[summer agate]

Yeah ive tried zluda but it says the no cuda gpu thing

[summer agate]

Hm ok

[ornate elk]

you need hip sdk 5.7

[summer agate]

I do

[ornate elk]

hmm okay

[summer agate]

I follosed that guide already

Thats ehy im asking for help lol , idk whats broken

[ornate elk]

can you show the complete cmd log?

[summer agate]

Sure

Altough its gonna take a while because im not home

[ornate elk]

ah alright

[summer agate]

venv "D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1-amd-53-g32f147b7
Commit hash: 32f147b7eb59be991db6619b9fa63874d59e1ed3
WARNING: No ROCm agent was found. Please make sure that graphics driver is installed and up to date.
ROCm: version=5.7
Setting HIP_VISIBLE_DEVICES and --device-id at the same time may be mistake.
ZLUDA support: experimental
ZLUDA load: path='D:\SD-Zluda\stable-diffusion-webui-amdgpu.zluda' nightly=False
Skipping onnxruntime installation.
You are up to date with the most recent release.
W0129 22:47:28.661111 23080 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\timm\models\layers_init_.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
warnings.warn(f"Importing from {name} is deprecated, please import via timm.layers", FutureWarning)
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_only has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from pytorch_lightning.utilities instead.
rank_zero_deprecation(
Launching Web UI with arguments: --device-id 1 --use-zluda --update-check --skip-ort --skip-python-version-check --no-half
Warning: caught exception 'No CUDA GPUs are available', memory monitor disabled
Loading weights [6ce0161689] from D:\SD-Zluda\stable-diffusion-webui-amdgpu\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors
Creating model from config: D:\SD-Zluda\stable-diffusion-webui-amdgpu\configs\v1-inference.yaml
Running on local URL:

here

[ornate elk]

it says no rocm agent was found

have you patched the gfx librarys?

[summer agate]

Yes

Altough i might have fked it up but im not sure

[summer agate]

and i put in the bin that other file

that needed to be in there (forgot the name)

[ornate elk]

i think they are in the wrong folder

[summer agate]

What.

[ornate elk]

the library folders need to be in:
C:\Program Files\AMD\ROCm\5.7\bin\rocblas

[summer agate]

Whaaat

In rocblas it told me to put just one thing

Well now ehat do i do

[ornate elk]

have you followed my guide?

[summer agate]

Yes lemme pull it out

[summer agate]

wait

THE WHOLE ZIP IN THERE?

[ornate elk]

no only the content

not as zip file

[summer agate]

[summer agate]

i put in only the rocblas

lmfao

ok now wtf do i do with the lib

oh wait nvm forgot ab the old lib thing

now i just dump rocmlib here

[ornate elk]

the files of the zip into this folder

[summer agate]

yes

now it SHOUD work right

also why is onnxrun not working

nope still same error

venv "D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1-amd-53-g32f147b7
Commit hash: 32f147b7eb59be991db6619b9fa63874d59e1ed3
WARNING: No ROCm agent was found. Please make sure that graphics driver is installed and up to date.
ROCm: version=5.7
Setting HIP_VISIBLE_DEVICES and --device-id at the same time may be mistake.
ZLUDA support: experimental
ZLUDA load: path='D:\SD-Zluda\stable-diffusion-webui-amdgpu.zluda' nightly=False
Skipping onnxruntime installation.
You are up to date with the most recent release.
W0129 23:00:48.252098 4488 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\timm\models\layers_init_.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
warnings.warn(f"Importing from {name} is deprecated, please import via timm.layers", FutureWarning)
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
D:\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\pytorch_lightning\utilities\distributed.py:258: LightningDeprecationWarning: pytorch_lightning.utilities.distributed.rank_zero_only has been deprecated in v1.8.1 and will be removed in v2.0.0. You can import it from pytorch_lightning.utilities instead.
rank_zero_deprecation(
Launching Web UI with arguments: --device-id 1 --use-zluda --update-check --skip-ort --skip-python-version-check --no-half
Warning: caught exception 'No CUDA GPUs are available', memory monitor disabled
Loading weights [6ce0161689] from D:\SD-Zluda\stable-diffusion-webui-amdgpu\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors
Creating model from config: D:\SD-Zluda\stable-diffusion-webui-amdgpu\configs\v1-inference.yaml
Running on local URL:

[ornate elk]

you have to restart the PC, then delete the venv and the .zluda folder and relaunch

[summer agate]

.zluda being where exactly

i forgot

[ornate elk]

same folder as the venv is

[summer agate]

ah nvm im blind

i just saw it

oke wish this time it will work

[summer agate]

relaunch as in run webui?

do i not change anything in it?

[ornate elk]

remove --device-id 1

[summer agate]

oh ok

well moment of truth

[summer agate]

[ornate elk]

hmm still:
No ROCm agent was found.

[summer agate]

HOOOOOOW

[ornate elk]

so its not recognisen the new library files

[summer agate]

or i di another dumbass moment

[ornate elk]

wrong folder...

they have to go into library

[summer agate]

IM GONNA FUCKING KILL MY SE-

[summer agate]

ok ty

wait in?

do i have to make a new one or?

[ornate elk]

select overwrite/replace

no new folder needed

[summer agate]

ok

do i sitll need to restart?

restert then delete or delete then restart

[ornate elk]

restart then delete the venv folder again and relaunch

[summer agate]

Im dumb as rocks with this god dam thing ngl

now the true moment of truth

if it still wont work il take another look twomorrow cuz ive been trying to make this work for 3 days already

[ornate elk]

oh okay 😄 sure let me know then

[summer agate]

[summer agate]

im gonna..... im gonna break my monitor

[ornate elk]

hmm no rocm agent was found

[summer agate]

As u said in the lib

In the amd thing

Then why the heeellllll

[ornate elk]

idk something is still wrong :/

maybe reinstall rocm hip sdk and redo the library steps correctly

[real mist]

hey, can i ask something about comfy here?

i have 2 "load lora" nodes in the workflow, and both have clip outputs. if i connect one of the loras clip nodes to text encode, i cant the other

is there a way to connect both of them?

[rose apex]

This is how

[ornate elk]

@rose apex my comfyui AMD manual guide is ready

With sage attention as optional setup

[rose apex]

Thank you. I'll give it a try right away

[ornate elk]

Nice, let me know how it goes. If you try the optional sage attention steps, can you Skip the install of visual studio build tools? I think it might be not needed

[rose apex]

Ok. I dont have it anyways

[rose apex]

After launching the bat file, it started but gives this warning

[ornate elk]

Strange, and you have run the install requirements before?

In my test it worked normal. Have you interrupted any command in cmd?
You should retry the steps St the activate venv and below

[rose apex]

Yes, during the install requirements my wifi disconnected, so i re ran it

[rose apex]

Ok

[rose apex]

If i set sage attention to qk int8 pv fp16 triton then it doesn't even pass through the sage attention node and gives the cannot import error.

If i set sage attention to auto, then it passes the sage node but gives a PassManager::run failed error in the Ksampler node. Here's the log for this one.

My rocm version says 7.2 instead of 7.1 even though thats what i installed rn right? Also, I think i have lot of cached stuff downloaded so i want to get rid of them to try a cleaner install

[ornate elk]

you dont need a specific sage attention node, sage attention already works if you can generate a normal image

some nodes try to use sageattention 2 or 3 but AMD only works wit hsage 1

[rose apex]

I did a clean reinstall of everything, didnt get the comfyui version warning. But sage attention is not working, it's falling back to pytorch attention.

[ornate elk]

any error, which workflow?

[rose apex]

The one you sent, about 1.5 upscale

The image generated but the logs says " using pytorch attention instead " during the K sampler node.

[ornate elk]

ah okay strange, then the vs build tools are maybe needed

[rose apex]

Does pip show triton not work?

And yeah, i'll go ahead and try with VS build tools now.

Which ones to install?

[ornate elk]

but idk if it works with the vs build tools from 2026

also it seems the curl patches are not needed if triton updates in the next days

[rose apex]

I installed 2022 only and yeah it still gives the same error

Maybe it doesn't work with RDNA 4 idk

Wait, I did the sage installation over my main Comfyui install and here it successfully generated an image.

[ornate elk]

hmm rdna4 should work too

[rose apex]

Yeah it seems to be working now on my main Comfyui. Maybe it didnt on the test one cause Pytorch and Rocm version mismatch?

[ornate elk]

hmm maybe

which ones do your main ones uses?

[rose apex]

This the test one

This is my main one

[ornate elk]

hmm okay the main onse uses the rocm 7.2 stable release

[rose apex]

Yeah. I tested 2 runs on both of them once again. The test one doesn't work, always falling to Pytorch attention but the main one works like normal.

[ornate elk]

okay, good then i change the pytorch version for the gfx12 cards

maybe you can the nretest

[rose apex]

Yeah sure

Also here it says AI folder but I think you mean the Comfyui folder

[ornate elk]

oh yes, thanks copy and paste error from my other guides ^^

[rose apex]

and here Start-ComfyUI-AMD.bat
I mean people won't do this wrong anyways but I guess it might confuse a few

[ornate elk]

okay changed it, can you restest?

now there are two pip commands needed for rocm 7.2

[rose apex]

Ok

[rose apex]

Do i have to install both?

[ornate elk]

try the second command first

if you get errors then also use the first for rocm

[rose apex]

Ok

[ornate elk]

if it works without the first command, i can remove it

[rose apex]

Okay

[rose apex]

I get this with the 2nd command

[ornate elk]

okay so the first is needed

[rose apex]

Running the bat file is giving this error and instantly crashing

I thought I broke something so did a clean reinstall, once tried installing both pip packages and once with just the first package but same thing happens when i run the start bat file. This shows up for a milisecond and crashes

Rocm and Pytorch did get installed though

[ornate elk]

can you show a pip list inside the venv ?

for me it looks like you installed torch+rocm on system level not inside comfyui venv

[rose apex]

how

[ornate elk]

⁨⁨venv\scripts\activate⁩⁩
⁨⁨pip list⁩⁩

and then open up a normal cmd, not inside comfyui or venv and run this only:
⁨pip uninstall torch torchvision torchaudio rocm rocm-sdk-core rocm-sdk-devel rocm-sdk-libraries-custom rocm-sdk-libraries-gfx110X-dgpu -y⁩

[ornate elk]

wrong torch is installed

run boath pip commands one by one and before that, activate the venv with
⁨⁨⁨venv\scripts\activate⁩ inside comfyui folder

[rose apex]

I did activate venv last time as well but ok let me do it again

[rose apex]

Ok it has (miraculously) started. I'll try sage now

[tall pier]

@ornate elk hey bro how do I fix this issue, my adetailer isnt working

[ornate elk]

hey, which webui do you use ?

[rose apex]

Yeah Sage attention is working now in this one Sadly I coudln't test if Build tools is needed or not since I already have it installed now

[tall pier]

forge

on macOS

[ornate elk]

okay great! yea me too would have to wait for someone without it to test xD

[ornate elk]

is adetailer updated?

which model did you use?

[tall pier]

Anzhc Face seg 640 v3 and full eyes detect v1

[tall pier]

not really, I havent touched anything after downloading it

[ornate elk]

any error when you use faceyolo only?

[tall pier]

very inconsistent

6/10 gens look alright

rest 4 sucks

the 2 models I use also same

[ornate elk]

i mean if you get an error on them in any way

like you did with the other one

[tall pier]

sometimes i need to restart to get it back and running

[ornate elk]

but i probably think your issue is related to this:
https://github.com/Bing-su/adetailer/issues/817#issuecomment-3462595521

can you run (inside forges main folder):
⁨⁨⁨venv\Scripts\activate.bat⁩⁩⁩
⁨⁨⁨pip list⁩⁩⁩
and then check for the ultralytics and mediapipe version?

would be interesting which ones you have

[tall pier]

lemme check it out and let you know

and also bro

does forge neo support mac

I heard its better than the regular forgeUI

[ornate elk]

if you manage to set it up yes, but there is not an install script for it

[tall pier]

how do I do it? is there a guide for mac?

[ornate elk]

nope

also someone on github said this:
⁨The problem was that I was using the following model for detection: Anzhcs woman face⁩
so i guess this Anzhc detail model is broken

but first let us take a look at the versions

maybe it will work after upgrading ultralytics or mediapipe

[night locust]

Hey, anyone know what's up when ComfyUI suddenly starts making only completely black images after working well? I'm not changing any settings at all, just changing the prompt a little and maybe the seed and then suddenly every picture is just black until I restart comfyui. No errors or anything.

[ornate elk]

whats your gpu?

which models do you use?

[night locust]

5090, illustrious (it happened on 2 different versions)

I've used comfyui for a long time, this just started happening recently (might have updated comfyui recently, could be why)

[ornate elk]

make sure its updated again and the nodes too

then also make sure gpu driver is updated

[night locust]

my gpu driver is a bit old by purpose (issues with vr in the latest nvidia driver)

[ornate elk]

ah okay

[night locust]

read something about black images being a NaN issue but it's a very technical article lol

[ornate elk]

yea it can happen due a lot of bugs

also clip skip issue etc

or if your torch is old

[night locust]

its just so odd that it persists between purging and changing models etc until i restart comfyui

[ornate elk]

does it happen with all models?

[night locust]

not sure yet, only tried with illustrious recently

wait i think i know when it happens now. it's after I run FaceDetailer

after doing that, going back to an earlier node to make another image with another seed always makes black images after that

[ornate elk]

ah so make sure these nodes are updated or show the setting you used wit hfacedetailer

[lime pivot]

FB_Qwen3TTSVoiceClonea
'Qwen3TTSTalkerConfig' object has no attribute 'pad_token_id' #anyone know what the problem is? i already downgraded transformers. im using flybird FB_Qwen3TTSVoiceClone

[summer agate]

yo, im back at it and ready to reinstall everything , altough i have a quick question , when i look at the guide theres 2 amd fork things, do i install the one for rx580?

[ornate elk]

either Auto1111 or Forge

[lament coral]

Hello, can someone help me with automatic 1111 install ?
I downloaded the sd.webui, extracted, i installed python 3.10.6 i added to path then git.
I run update then run. When i press on run i get an error to install clip
fresh windows 11 install
Gpu 4070

keeps asking me to lto github

[ornate elk]

hey, dont use auto1111, its outdated.
Check the pinned messages of this channel for the updated guides and go with Forge Neo

[lament coral]

Is it same as Automatic 1111 to just download a file from civitai and it just works ?

[ornate elk]

yes, even same ui

just more updated and more model support + performance tweaks

[lament coral]

@ornate elk THX A LOT IT WORKED !!!

[ornate elk]

perfect, np 🙂

[knotty turret]

What u guys recommend me to use with a RX6600

i heard something about TheRock but i dont if my GPU can use that

@ornate elk what i can use to get the best performance saw u saying auto1111 its outdated so what should i use now

[lyric vector]

Comfyui save image node suddenly stops working on my after I update all

%date:yyyy-MM-dd%/image_%KSampler.seed%```

I've been using since before but right now suddenly im getting error on the file name

[ornate elk]

Forge with Zluda would be good

[tawny saddle]

ba-baboop

[rose apex]

is it possible to install Flash Attention 2 on RDNA4? if yes then where is it worth using?

[ornate elk]

Yes its possible but I didn't noticed any major speed improvement

And it uses more vram than sage

[rose apex]

I see. I heard something similar as well

[dusk pebble]

Hello everyone.

[ornate elk]

With the latest ROCm nightly release today this changed.
Flash and Sage attention are now much faster then pytorch cross attention!

Currently sage is the best option as its as fast as flash but also uses less vram

[rose apex]

Damn. How do I install the nightly release?

[lyric vector]

I have 2 comfyui and I want to link my #2 to the models folder of my #1 but I'm having errors when I run it

[ornate elk]

You need to change the \ to / in the top path

[marsh forge]

i cant run stable diffusion

hello i need help i have this tab pooping out when i connect my account to stable deffusion and it sey "the string binding is invalid"

@pure tusk maybe you can help me with that obe

ouch

can someone pls help me

[pure tusk]

looking at this, what are you doing to connect your account?

[marsh forge]

just connecting my google account

[pure tusk]

where?

ah

[marsh forge]

i am begginer

[pure tusk]

hmm I don't think this is something on us, looks like a git issue

[marsh forge]

ouch

you earn from ai creating?

[vocal burrow]

Definitely a github issue. Make sure you have the latest git client and if you're from some "well known country with internet privacy issue that shall not be named" you might need to use a vpn.

[marsh forge]

im from poland

haha

i mess with it

and i think i f00k it up

[vocal burrow]

ah! misstep number two.
Automatic1111 main branch is broken. (and will probably stay like that as it is pretty much abandonned software at this point)

checked the pinned message for guides on alternative ones such as forge / forge neo / comfyui / etc

[marsh forge]

but it maybe bcs of me

i done this

1.open file

2find line with

[vocal burrow]

nah I'm pretty sure

[marsh forge]

ouch

but before this showed up

venv "D:\a1111\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Cloning Stable Diffusion into D:\a1111\stable-diffusion-webui\repositories\stable-diffusion-stability-ai...
Cloning into 'D:\a1111\stable-diffusion-webui\repositories\stable-diffusion-stability-ai'...
info: please complete authentication in your browser...
remote: Repository not found.
fatal: repository 'https://github.com/Stability-AI/stablediffusion.git/' not found
Traceback (most recent call last):
File "D:\a1111\stable-diffusion-webui\launch.py", line 48, in <module>
main()
File "D:\a1111\stable-diffusion-webui\launch.py", line 39, in main
prepare_environment()
File "D:\a1111\stable-diffusion-webui\modules\launch_utils.py", line 412, in prepare_environment
git_clone(stable_diffusion_repo, repo_dir('stable-diffusion-stability-ai'), "Stable Diffusion", stable_diffusion_commit_hash)
File "D:\a1111\stable-diffusion-webui\modules\launch_utils.py", line 192, in git_clone
run(f'"{git}" clone --config core.filemode=false "{url}" "{dir}"', f"Cloning {name} into {dir}...", f"Couldn't clone {name}", live=True)
File "D:\a1111\stable-diffusion-webui\modules\launch_utils.py", line 116, in run
raise RuntimeError("\n".join(error_bits))
RuntimeError: Couldn't clone Stable Diffusion.
Command: "git" clone --config core.filemode=false "https://github.com/Stability-AI/stablediffusion.git" "D:\a1111\stable-diffusion-webui\repositories\stable-diffusion-stability-ai"
Error code: 128

and this lines

and after complete authentication this showed up

[vocal burrow]

I'm sure

[marsh forge]

about?

[vocal burrow]

you can even see the error at the very top of your log

it can't read commit cf167yapyapyap because it does not exist anymore on github's servers

[marsh forge]

this log doesnt show any of my private date or something?

[vocal burrow]

automatic1111 is referencing some broken repo

[marsh forge]

so should i go with other thing then automatic1111

[vocal burrow]

no.

[vocal burrow]

yup

check pinned guide

[marsh forge]

you can help me?

do you earn from ai creating?

[vocal burrow]

if you have some problems following the guide feel free to ask and I (or someone else) might help you out.

[vocal burrow]

No I'm not paid by stability.ai nor do I earn / nor publish anything that I create with AI.

[marsh forge]

can you ping me to the guide

i can find it

[vocal burrow]

this place is pretty much community driven

[marsh forge]

cant

[marsh forge]

i mean do you earn bcs of your arts, chat gpt said that i should try to make ads for small
companies

and that is all i will try to do

[marsh forge]

should i uninstall something from this a1111 folder

[vocal burrow]

delete it completely you won't need it. (save models, loras, etc if you downloaded some for it)

[marsh forge]

do i have any advantage if i have rtx 5070 and ryzen 7 9800x3d? or its same for everyone

[vocal burrow]

I would NOT trust any kind of LLM / AI for pretty much anything. They can be great tools but they should not be trusted.
Even more when it comes to legal problems given that they're probably not especially trained on your country's laws.

So yeah, ask a lawyer or triple read every fine print in every software + model you're gonna use if you intend to commercialy use AI creations.

[marsh forge]

what does llm mean

[vocal burrow]

llm <=> text AI stuff

(simplifying here)

[marsh forge]

i still dont get it

am new

im

[vocal burrow]

cpu does not matter that much, however the more vram (gpu memory) you have the more AI stuff you'll be able to run.

[vocal burrow]

llm <=> text AI stuff <=> think chatgpt, claude, etc

[marsh forge]

so if i try to make ai ads with someone model they can
sue me

[vocal burrow]

depends of the model's license + the license of the client (eg : automatic1111, comfyui, etc) you're using the model with + your country's laws.

[marsh forge]

so how to make it lawful

good with the laws

[vocal burrow]

you select the good combo. Some models that allow commercial use + client that allow commercial use + check that your country allow AI content commercialisation altogether.

And I can't help you much with that.

[marsh forge]

what does the commercial acually mean

[vocal burrow]

basically meaning you can use it in a professional activity / make money out of it

[marsh forge]

ouch okauy

so maybe you can give me diffrent idea that i can start with to make money

i work in full time job and after job i have like 5 hours to do whatever i want and i like to play games like fortnite etc and i would like to start earning from doing something on pc not just playing 5 hours straight

[vocal burrow]

I'm afraid that's outside of this discord's scope. Start working on IT stuff as a freelance, video editing, etc...

[marsh forge]

okay thanks mate

[restive turret]

Yo tryna dm you cuz I need some help with nsfw generation as well

[hot thorn]

im on a 2week break from ai, so u will need to ask someone else

[prisma moss]

make anime gooner content >:D

[paper ridge]

I was wondering if anyone could please point me toward some good models + workflows that work well on consumer hardware

I’m building a free web app that wraps around ComfyUI, and I want to ship it with a solid set of workflow templates, but I’m still figuring out the best approach there

Any advice or pointers would be hugely appreciated 🙂

[iron lantern]

would this make faster gen times then a 3060 12gb ? https://www.pccasegear.com/products/70836/gigabyte-geforce-rtx-5060-ti-eagle-max-oc-gddr7-16gb

[marsh forge]

On tiktok?

[prisma moss]

Nah, pixiv/patreon/x/ig/fb

[marsh forge]

And you are earning from that?

[prisma moss]

A bit ye

[marsh forge]

How much per month

[prisma moss]

Round 500

[marsh forge]

I need this as a Side work

[marsh forge]

Dolars/euro?

[prisma moss]

Dollars

[marsh forge]

There is a video that helped you with your advanture

?

[prisma moss]

No

[marsh forge]

Do you use automatic1111?

Or something else

[prisma moss]

Forge (basically a1111)

[marsh forge]

Can you give me video that can help me with installation?

[prisma moss]

Follow the github guide

Im off to bed now

[marsh forge]

Okay

I had a error

While login

And i cant use the automatic1111

[marsh forge]

Here

[prisma moss]

You dont need to log in

[marsh forge]

Okay

I will try later to go with forge neo

I have rtx 5070 12gb

[oblique ether]

Hi everyone! 👋
I’m looking for a specialist who can help me set up a local Stable Diffusion workflow.

What I need:
– Stable Diffusion WebUI (AUTOMATIC1111)
– InstantID (or IP-Adapter for face identity)
– Proper pipeline for realistic portraits

Everything is already installed and running locally, I just need correct setup and guidance.
This is a paid task.

Where can I find such a person?
Thanks in advance 🙏

[rose apex]

You should use Forge classic/neo instead of A1111 since it's long outdated and people often have problems with it.

[oblique ether]

Thanks!
Got it.
So Forge Classic is the better choice for InstantID setup, right?

[rose apex]

Can't say specifically about InstantID but in general Classic is better and stable for SDXL workflows and Neo is better if you want to use Flux or Wan2.2

Maybe @ornate elk can help. I don't use forge

[oblique ether]

Thanks a lot for the clarification! 🙏
That makes sense. I’m focusing on SDXL workflows, so I’ll stick with Forge Classic.
I really appreciate you sharing the contact as well! Thanks again!

[fringe kraken]

Hello, I'm really new to stable diffusion and image generation. I ended up here because after setting it up with my sillytavern and koboldcpp it produce only image made of blended pixels. I test it "alone" and it still do the same.

[onyx solar]

Nowhere near enough information to help. You need to share things like what software you're using, what model, what hardware you're on, what the workflow is, etc.

[fringe kraken]

So as I say I'm new so most thing would be default. I just lunch koboldcpp with the model [stable-diffusion-v3-5-large-pure-Q4_0.gguf] without touching any other parameter. I have a AMD Radeon RX 7700 XT and AMD Ryzen 5 5600X 6-Core Processor.

[fringe kraken]

Else is there other thing that could help solve the case? I didn't find this error online.

[onyx solar]

On the koboldcpp github (https://github.com/LostRuins/koboldcpp), you'll find a section about AMD Users. Did you follow those instructions?

[fringe kraken]

yes it says to use the vulken nocuda option I think I did everything correctly. I will check again.

Yep, that was good

[onyx solar]

Show a screenshot of your Image Gen settings page.

[fringe kraken]

I use it via sillytavern so I got this:

else this:

[onyx solar]

First thing to try is to bump the resolution up. While SD3.5 can do some small resolution stuff, you should test 1024x1024.
I'm not a koboldcpp user, but 1 megapixel is the generally recommended minimum.
Also, because I don't use sillytavern for image generation, I don't know if the Scheduler set to default means it's using "Normal" or not, but if available, that'd be another adjustment I'd make.

All that said, it's definitely not the generation software of choice for images.

[fringe kraken]

Here are the potential scheduler.
I tried to use on Silly to agrement my chat, but I also want to create image for worldbuilding, what would you recommend as software?

[fringe kraken]

Preferably something with documention/tutorial for beginner.

[vocal burrow]

looks like you're trying to generate 512x512 stuff with models trained for 1024x1024

oh someone already answered. 👍

[fringe kraken]

Hello, so me again...
I try to follow this: https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-automatic1111-with-zluda
installation guide but get this error:

Zluda doesn't work/isn't detected but I don't get how?

the zluda file is directly under C: and i did the PATH thing.

[ornate elk]

hey, you installed the wrong python version

3.14 is not supported

uninstall it, then install 3.11.9 64bit
then delete the venv folder and relaunch the webui-user.bat

[prisma moss]

now when i start it it undos my stuff to install 2.10

[prisma moss]

@ornate elk Imma need u on this one..

I installed triton>sage, both versions are compatible with pytorch version etc.

But when genning i get: (wan gen)

Error running sage attention: Command '['F:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\triton\runtime\tcc\tcc.exe', 'C:\Users\Kipac\AppData\Local\Temp\tmpzd3z9r6s\cuda_utils.c', '-O3', '-shared', '-Wno-psabi', '-o', 'C:\Users\Kipac\AppData\Local\Temp\tmpzd3z9r6s\cuda_utils.cp313-win_amd64.pyd', '-fPIC', '-D_Py_USE_GCC_BUILTIN_ATOMICS', '-lcuda', '-lpython313', '-LF:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\triton\backends\nvidia\lib', '-LF:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\triton\backends\nvidia\lib\x64', '-IF:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\triton\backends\nvidia\include', '-IF:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\triton\backends\nvidia\include', '-IC:\Users\Kipac\AppData\Local\Temp\tmpzd3z9r6s', '-IF:\ComfyUI_windows_portable\python_embeded\Include']' returned non-zero exit status 1., using pytorch attention instead.
for every step

no wonder my gens are slow af.

[prisma moss]

Im on cu130, idk how compatible those are with 4090

[fringe kraken]

So...
now I got everything installed and ready but when i try generating an image I get RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same
Does anyone know what it means?

[ornate elk]

Its using the wrong torch version

[fringe kraken]

So I have multiple torch on my computer?

[ornate elk]

No

It just installed the wrong one

Restart and send the complete cmd log

[fringe kraken]

here is the complete cmd:

[ornate elk]

Something is wrong with the hip path setting

Can you show how it looks in the system path settings

[fringe kraken]

Here HIP_PATH set to %HIP_PATH%bin

[ornate elk]

Ah that's the wrong path

You added it at the wrong spot or changed the good one

I'll show you how it should look

[ornate elk]

these two should point to 6.4 like here:

[fringe kraken]

yeah, I added instead of altering PATH. my bad

[ornate elk]

after that click on here;

and select edit

there add a new entry with the %HIP_PATH%bin

[fringe kraken]

and I add C:\Zluda too right?

[ornate elk]

after that restart the PC,

then delete the venv folder and relaunch the webui-user.bat

[fringe kraken]

ok so now it launch but the cmd log says: Compilation is in progress. Please wait... indefinitely and I can't generate image

[ornate elk]

yea thats normal

compilation will take 15-40 minutes

only one time, then you can generate

[fringe kraken]

oh ok, thx

[ornate elk]

no problem

[prisma moss]

Now mee 😭

[chilly knot]

NVIDIA GeForce RTX 5070 Ti with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90 compute_37.
If you want to use the NVIDIA GeForce RTX 5070 Ti GPU with PyTorch, please check the instructions

what version PyTorch need for install?

or what i need to do

for fix

[prisma moss]

r u on comfy portable

[chilly knot]

its stable diffusion

[prisma moss]

type nvidia -smi in cmd and check the cuda version

[chilly knot]

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:41:10_Pacific_Daylight_Time_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0

[prisma moss]

so nvcc --version works on ur end?

[chilly knot]

yes

[prisma moss]

well, it shouldnt be 11.8 thats for sure

[chilly knot]

12.8 i need to install?

[prisma moss]

at least

[chilly knot]

i try to install torch but fail

[prisma moss]

whats ur python ver

[chilly knot]

i try this command

python -m pip install --default-timeout=100 --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

[prisma moss]

do python --version

[fringe kraken]

finally it doesn't end, after a couple hour...

after relaunching multiple time I finally got a progress bar

[chilly knot]

Python 3.10.6

[prisma moss]

do u have multiple python vers installed?

[chilly knot]

no i think

firts i need install last cuda and after install toch?

[prisma moss]

i wouldnt be surprised if its some 50xx shenanigans, since they are a bit different compared to 40xx

that command should pull cuda as well

[chilly knot]

my comman?

[prisma moss]

what u did here

and i wouldnt know why it throws the error

[chilly knot]

after upgrade pip

[chilly knot]

now someting block file

python hell days🙂

every time the simplest procedure turns into an investigation

[ornate elk]

uninstall 3.10.6

and follow the guides from the pinned messages, install Forge Neo, not auto111

[chilly knot]

its not work with this card?

[ornate elk]

3.11.9 is needed

and dont install auto11111, its outdated and not made for rtx50 gpus

Forge Neo is updated

[chilly knot]

thx i need use wan video too

[ornate elk]

forge neo works with wan video too and comfyui also

[chilly knot]

now i isntall cuda 12.8

[ornate elk]

Setup Guide can be found here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[chilly knot]

for forge?

[ornate elk]

for Forge Neo, comfyui, etc

all guides are there

[ornate elk]

Guide for comfyui portable + sage is here:
#1243610878314807337 message

[prisma moss]

Im working with someone ill let u know if this hell ends

[chilly knot]

what i dont need in cuda install for forge or someting else?

too big

i cant change folder(

only system drive

sad

[ornate elk]

nope

only your normal nvidia driver updated

[ornate elk]

you can install it anywhere

[chilly knot]

how

it dont show me different folder

[ornate elk]

what?

which programm?

[chilly knot]

cuda

i dont have space in system drive

[ornate elk]

you dont neeed cuda

and make sure your system drive always has 15-20gb free,
a full system drive can cause a lot of trouble

[chilly knot]

facefusion need cuda

[ornate elk]

get a tool like WizTree and check whats taking up so much space

[chilly knot]

it dont show space

[mossy sierra]

Could someone help me with Stableaudio?

please?

[mossy sierra]

when I try to visit the generate page it gives me an error

[nova tide]

Hello

[mossy sierra]

Could you help me or no?

[prisma moss]

@ornate elk ait the dsa cuda error still here

there isnt anything i havent tried ffs at this point

fresh installs, correct toolkit, cuda/torch/python evrsions everthing

Prompt executed in 45.58 seconds Exception in thread Thread-13 (prompt_worker): Traceback (most recent call last): File "threading.py", line 1044, in _bootstrap_inner File "threading.py", line 995, in run File "F:\ComfyUI_windows_portable\ComfyUI\main.py", line 302, in prompt_worker comfy.model_management.soft_empty_cache() ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^ File "F:\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 1727, in soft_empty_cache torch.cuda.synchronize() ~~~~~~~~~~~~~~~~~~~~~~^^ File "F:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\cuda\__init__.py", line 1085, in synchronize return torch._C._cuda_synchronize() ~~~~~~~~~~~~~~~~~~~~~~~~~~^^ torch.AcceleratorError: CUDA error: unspecified launch failure CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

[idle cairn]

He's restarting

[prisma moss]

also this link is unknown

[mossy sierra]

could someone help me with my issue with stable audio?

[idle cairn]

What is it ?

[mossy sierra]

for some reason when I visit the generate page it says "Application error: a client-side exception has occurred (see the browser console for more information)." I don't know why though

[idle cairn]

What's the code ?

[mossy sierra]

what do you mean?

[idle cairn]

(see the browser console for more information).

You have a code like 403, 502, 503, 404... ?

[mossy sierra]

No I think here is the console code I got from the page

[idle cairn]

Using a VPN ?

[mossy sierra]

No I did not use one

[ornate elk]

It shows what taking up space

[prisma moss]

https://tenor.com/view/h2di-cat-dead-cat-depressed-cat-laying-gif-11402833644907079653

[prisma moss]

so uavailable rn

[chilly knot]

no )but i install cuda ,facefusion now work on cpu -and errors in start

after install i need only add to PATH 2 folders with cuda?

[prisma moss]

they just keep on ocming

[chilly knot]

What's happened?

[prisma moss]

everything

sageatt isnt fucking working

[chilly knot]

hardrive?

seagate

[prisma moss]

uh no?

[chilly knot]

sageattention?

[prisma moss]

yes

[chilly knot]

I just hadn't heard this name and thought you made a typo.🙂

[ornate elk]

Yea sry its in the comfyui discord

[woeful raven]

yo

should i use the comfy ui rocm

[ornate elk]

Hey, what's your GPU?

[woeful raven]

rx 760

7600

[ornate elk]

With 8GB vram?

Would suggest going with Forge Neo with ROCm

[woeful raven]

yes

[woeful raven]

its better?

[ornate elk]

Easier to use

And better to handle with 8GB vram

[woeful raven]

k

[woeful raven]

@ornate elk

[chilly knot]

cs10 i install neo forge all works fine for my new gpu thx

[prisma moss]

do u have comfy discord

i just did a manual comfy installation, followed every step properly, and it is still cooked, idk where to fin look at this point

[ornate elk]

Nice, no problem.

[ornate elk]

Rerun the commands again to change the torch version

[chilly knot]

wan support by default
its ok work win wan on this ui?
I read that it needs to be very precisely configured using various lora so that it doesn't take a lot of time.

[ornate elk]

Wan 2.2 14B works.but you have to enable the Refiner option in the settings

[woeful raven]

which commanda

commands

i ran the .bat file again and same thing

[ornate elk]

From step 8 setup rocm

[woeful raven]

i was in step 7

if i do 8 then go back to 7

would be same thing?

A wait

step 7 talks about that

"or get cuda error"

im blind my bad

[ornate elk]

Ah okay then yea, do the next steps xD

[woeful raven]

i got this when i tried generating image

[ornate elk]

Can you show the complete cmd log?

[woeful raven]

sure

[ornate elk]

Can you relaunch and just try to generate a normal image and don't use ponyv6, its very broken

[woeful raven]

ok

got same error

with v1.5 pruned

[chilly knot]

👋

cs10 what i do wrong -wan dont work

[ornate elk]

Lownoise goes into refiner

[chilly knot]

can you show how it

like this?

[ornate elk]

and high noise as checkpopint, low noise for refiner

[chilly knot]

and + vae + encoder?

[ornate elk]

check this:
https://github.com/Haoming02/sd-webui-forge-classic/wiki/Inference-References#wan-i2v

there you see how it should look

[chilly knot]

i have 32 gb memory it enough?

or cant work

16 gb vram and 32 ram

[ornate elk]

16gb vram could be to less

[chilly knot]

They wrote on the forums that they even launched on 8 gigabytes. Maybe I need to change the model for my system?

14 b gguf its heavy?

again error (

my settings

all like you said

maybe vae broken

[chilly knot]

i redownload vae +guf 3 models and it start works but take all ram and on 50 % its off

and cmd closed

and dont take gpu memory

[paper dagger]

Security breachhhhh lol

[prisma moss]

Bruh u should use comfy for wan

[chilly knot]

that not take all memory?

[prisma moss]

No, but its more functional than forge for i2v

[chilly knot]

Why does RAM start to be used up when AI should always use video memory first?

lora ltx very help?

[prisma moss]

Idk what ltx is

Because it cant put all the models in only gpu

So its putting it in ram, than to gpu, and back and forth

Unless u mean lightx loras

[chilly knot]

yes

[prisma moss]

They are necessary , unless the model has them integrated

[chilly knot]

and cmd stop

take all 32 ram

strange i set low res

[chilly knot]

i try low res 320 x 320

[ornate elk]

Your probably running out of vram

Maybe you should try comfyui for video. It supports Wan 2.2 5B, a much smaller model

[chilly knot]

its very strange people on reddit run it on 12 vram

when wan running

work only ram

after 1 min gpu 100% and after 0 % and still take ram

[river pier]

Hello friend, which build do you recommend?

[ornate elk]

Hey, go with Forge classic from the pinned messages

[river pier]

Thank you

[river pier]

Do you know how to make a video based on pictures?

I have several images, and I need to make a video based on them.

[hidden night]

Someone help

I was trying to use forge web ui

When I click generate it instantly cancels and gives me this error

[ornate elk]

With 4GB vram you can't do video stuff

[ornate elk]

Can you send the complete cmd log?

[meager prawn]

i am currently trying to install Neoforge for my rx 9060xt 16 gb but it is using python 3.11.9 all the time also in my amd software it is telling me i have 3.11 interpret but in the install it is showing it should be 3.12

could somebody please help me fix this

[ornate elk]

Hey, uninstall python 3.11 and install python 3.12.10

[meager prawn]

from the python website or should i do the install from my amd software again?

[ornate elk]

From python website

[meager prawn]

then add to path right?

[ornate elk]

And for a good setup guide check the pinned messages

No AMD app needed for that

[meager prawn]

i followed it but it gave me every time an error after i installed the rocm staff through the commands

[ornate elk]

Ok, yea then follow the Forge Neo with ROCm setup guide from the pinned messages

[meager prawn]

i will try now

thank you for your help

do i need to restart my pc before i start?

[ornate elk]

Doesn't hurt but not needed

[meager prawn]

it did not work again

and this is what it showed me before i reopened the webui-user.bat

[ornate elk]

can you delete the venv folder and redo the steps?

so launching it once, then running the commands, then relaunch

[meager prawn]

in my neoforge folder?

ok i will do it

the tmp folder also?

[ornate elk]

nope

[meager prawn]

ok so now i launched it once and i get the cuda error where it tells me to update my drivers

[ornate elk]

ok then now redo the steps to change torch and rocm

[meager prawn]

is this normal or is something wrong?

[ornate elk]

normal

[meager prawn]

i got this then i closed the cmd and relaunched the webui-user.bat as in the guide

but then i get

[ornate elk]

ok let me check something

@meager prawn ok, it seems that the Forge Neo Dev updated the webui to need Python 3.13, 2 hours ago xD

so, uninstall python 3.12, install python 3.13.12 from here: https://www.python.org/downloads/release/python-31312/
Then delete the venv and redo the steps, then it should work

I have to update the Guide real quick

[meager prawn]

ok will try it thank you

wait i have more then one python i have 3.13 then i have 3.14.2; 3.13.5

do i need to unistall all?

[ornate elk]

3.14 is not compatible at all and the 3.13.5 is older

[meager prawn]

when all the ai stuff installs rocm directly for them why do i need to have the rocm thing from amd in the software like is it giving me anything? i also read that lm studio which i use because of rocm also installs rocm for itself but idk if it is true that it is using rocm because i see no difference between rocm and vulcan there
also tried koboltcpp with rocm but when i try to launch it from the gui it closes itself
so is there any reason i should keep the stuff the amd software installed for ai?

[river pier]

And then what to do?

[ornate elk]

you dont need the stuff from AMD, the point of lmstudio and kobold using their own is that they can easily upgrade or switch to an newer rocm version much easier.

so if they need a different version they can use it without messing with each others rocm

rocm gets updated very often so its better to use the ones you install into the venv. so you can easily uninstall and reinstall a newer version

[river pier]

Damn, what can be done to generate a video based on an image?

[ornate elk]

you would need to use online services for that a friend with a good gpu

[meager prawn]

so i can unistall the my_venv from amd because it takes like 10 gb and my c ssd only has like 50 gb left

[meager prawn]

is there an benefit from using rocm in these because i saw almost zero difference and in kobold it even crashed

[ornate elk]

you cant use vulcan for image, vidoe gen so yea there is only rocm, zluda or directml

zluda is for rdna2 gpus, and directml is just slow (with onnx a bit faster)

[meager prawn]

Neo forge works now thank you really much

[ornate elk]

perfect np, feel free to dm me if you want to chat about ai stuff on amd in german ^^

[meager prawn]

Now the next step is for me to look into why kobold and silly tavern wont work

[meager prawn]

i will do thank you

[ornate elk]

havent tried them yet, so cant help much on that

[meager prawn]

nah thats ok you already helped me out so much

how many it/s is normal for an rx 9060xt 16 gb with an xl model?

at 1024x 1024

[ornate elk]

around 4-5

more like 5

[meager prawn]

i get like 2.5

[ornate elk]

which settings?

[ornate elk]

select the XL preset at the top left

also whats in your webui-user.bat?

[ornate elk]

hmm thats strange

[ornate elk]

are you on the latest AMD Adrenalin driver?

[meager prawn]

yeah i am on 26.1.1

[ornate elk]

okay, ill check if there is a better rocm version, send you a link later

[sand lichen]

hi could someone help me pls i have pyhton 3.10.6 cuda 12.6 with torch idont understand

[ornate elk]

hey, whats your gpu?

[sand lichen]

hi 3070 ti

latest driver

[ornate elk]

okay, then you should install forge classic or neo instead

auto1111 is outdated

all setup guides can be found in the pinned messages of this channel

[sand lichen]

ok thanks a lot

[sand lichen]

and will i need stability matrix or no ?

[ornate elk]

nope

[sand lichen]

ok thanks

[ornate elk]

no problem

[barren stratus]

where can i find an install guide

[ornate elk]

in the pinned messages of this channel

first link

[barren stratus]

ty

[ornate elk]

np, let me know if you have any questions

[ornate elk]

probably the latest stable nightly performas faster:;
you have to got into the forge neo folder then oipen up a cmd there then run these commands one by one:
venv\Scripts\activate.bat

pip uninstall torch torchvision torchaudio rocm rocm-sdk-core rocm-sdk-devel rocm-sdk-libraries-custom rocm-sdk-libraries-gfx120X-all -y

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ torch torchaudio torchvision
then relaunch the webui-user.bat

[meager prawn]

I will try it when I am on my PC again thanks for looking into it

[west socket]

Hello nice to meet you all, I'm gonna be honest, I'm a real noob at all this and I'm frustrated for how hard is being all this set up of a UI to use stable diffusion. The thing is that I installed the webui-directml version of automatic1111 for the first time and in fact it worked, but because it was using my cpu instead of my gpu (RX 580 8 VRAM) I decided to reinstall somethings or try others UI like comfyUI, Amuse, SwarmUI and not a single one worked (for some reason that I still don't know).

Then I tried to re-install A1111 to atleast use SD with my cpu and then turns out that it drop a really annoying error trying to install CLIP, and at this point I don't even know what to do, I tried other version of python, i tried intall CLIP manually, I tried guides from youtube, reddit, github, even before write this I tried the guides in the pinned post and nothing works, I know that probably I'm missing something but I don't know what tbh, so i will be glad if someone could give me a hand 🙂

[unreal linden]

I'm not sure whether I'm doing something wrong or it's simply not a thing anymore but is there any way to get the old cross attention optimization options? It seems to defer to 'Automatic' and the option to change it is crossed out.

[unreal linden]

would this be okay to do for forge classic?

[ornate elk]

It should use pytorch cross by default if your not using sage or flash

[ornate elk]

Hey, checkout the setup guides from the pinned messages of this channel. First link.
There try the Automatic1111 with DirectML guide.
If you want, you can also try the Zluda guide for the rx580, but its harder to setup.

[plain crypt]

Hello, I've been trying to install it for several days now, but I keep getting this message.

Can someone help me? I've been stuck for days.

[ornate elk]

Hey, what's your GPU?

[plain crypt]

NVIDIA GEFORCE GTX 1650

[ornate elk]

OK, then checkout the pinned messages of this channel for the updated setup guides.
Automatic 1111 is outdated.
With your GPU go with the Nvidia Forge guide

[plain crypt]

OK thanks

Is this the one?

[plain crypt]

Can you help me? I'm stuck, I don't understand.

[ornate elk]

Sure at which step?

[plain crypt]

For better performance, modify the webui-user.bat file (right-click, edit), located in the webui folder. On the line COMMANDLINE_ARGS=, add: --cuda-stream
Then, save and close the file. Relaunch run.bat

I can't find the place to edit

[ornate elk]

Do you see the webui-user.bat file?

Inside the forge folder

You can right click it and then select "more options" then "edit"

[plain crypt]

Okay, and then what?