🤝｜tech-support

[bitter pier]

Ohh

Then you should install the env first

[acoustic chasm]

env?

[bitter pier]

Btw do you know you have Gpu installed?

[vocal burrow]

ctrl+shift+escape -> performance -> GPU

[bitter pier]

Environment

[acoustic chasm]

I have radeon

[acoustic chasm]

how do I do that

idk what that even is

[vocal burrow]

screenshot the whole thing

[bitter pier]

SO which guide did you follow to run stable diffusion?

[acoustic chasm]

60 second guide

https://www.youtube.com/watch?v=i5hvZvzcxoo

[acoustic chasm]

aight

[bitter pier]

It does provide some installation guide on AMD?

[vocal burrow]

lol no of course

[acoustic chasm]

GRAHH

[bitter pier]

Perfect!

[vocal burrow]

i meant the gpu window

[acoustic chasm]

so I just cant have stable diffusion?

[bitter pier]

Not really

[bitter pier]

You can use that and should install the envrioenment first

[acoustic chasm]

yeah but how do I do that

also is there any other ai image software I could use as a replacement?

[vocal burrow]

Ok there we go, you have a RX 6600 XT

and you followed a tutorial for NVIDIA gpus

so of course it s not gonna work

[acoustic chasm]

sigh

this is what it says

for amd

and its a lil too complicated for me

I went into python

and python wont even wana open

[vocal burrow]

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

[acoustic chasm]

way too much work

I dont wana restart my pc either

[vocal burrow]

obligatory reminder to not get scammed in DM

[acoustic chasm]

I have something running

[vocal burrow]

Welp can t help you then. I can copy paste the same tutorial in smaller portions but the work will be the same.

[acoustic chasm]

alright thank you

im just wondering if there are any like

ai generators I can use instead

scam shazam

[ornate elk]

Localy only the ones from the guide.
Online are only paid services.

[acoustic chasm]

the one he sent?

[vocal burrow]

There are multiple ones in the page I ve sent you. But yes

[acoustic chasm]

I might just be better off without any of it then

[vocal burrow]

tldr zluda fast, directml slooooow

[acoustic chasm]

too much work and editing stuff

and I wana understand what I am doing

and not blindly follow tutorials

[ornate elk]

Then your out of luck because there aren't any YouTube videos about the correct or updated setup of stable diffusion on AMD GPUs.
Thats why I made these Guides to easily get it working.

The setup is not as easy as running an .exe file but you wouldn't understanding or control what an .exe does in the background too

[acoustic chasm]

thanks anyways

I might give it a shot later

[ornate elk]

Sure, no problem. Feel free to ask me if you have any questions about the setup.

[prisma moss]

RuntimeError: CUDA error: misaligned address 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. Unhandled exception caught in c10/util/AbortHandler.h 00007FF9D2FA805400007FF9D2F8C190 torch_python.dll!THPGenerator_initDefaultGenerator [<unknown file> @ <unknown line number>]

And what does this error mean?

[brave egret]

Hi everyone,
I got a Question about an issue with some of my presets in SD A1111

After writing a prompt I save it from time to time as a template within the 'Styles' as preset
But I´m having issues with about 1 of 3 presets

If I place the Prompt (positive and negative) in my SD the pics look fine
But when I clear the Prompt (positive and negative) and select the preset, which has the same positive and negative prompt saved within and use the 1:1 same settings, the pics have issues
And if there are issues with the preset and I run the generation with my Adetailer it makes it even worse, but when I place the prompt (positive and negative) directly in my SD instead of the preset, the pics look fine even with Adetailer

Pic 1 = Prompt directly in the SD not as preset
Pic 2 = Same prompt but as preset without Adetailer (only with one time face detailer)
Pic 3 = Same Prompt as preset but with Adetailer (only with one time face detailer)

(Using SD A1111 with 1.5 Models)

Wish you all a good weekend

[prisma moss]

how does one unload models on forgeui manually? when switching between them, i often get stuck on one of them

uh what?

and why not here

sure you do, you are not even in that server lol

fk u want bro?

@light elm

[ornate elk]

Banned the scammer

[ornate elk]

If you dont switch between them while an image is generating it shouldn't get stuck

[rose apex]

I saw this in SDnext ZLUDA. First, does this means FA2 works on AMD GPUs? (Chatgpt says otherwise cause). Secondly, How can I check if it's already enabled or not in SDnext?

[rose apex]

Works fine for me, you don't need to manually unload models I think. Happens automatically when you switch to a different model. There still might be an extension for it ig.

[dry jungle]

I've tried installing a LoRA into automatic1111 but after setting everything up and reloading the webui the lora is not appearing in the Additional Extensions, any idea why??

[prisma moss]

Uh, the loras dont go into extensions

[lusty edge]

Hey hey, I am kinda lost actually. Does anyone know how to train LoRA locally? I want to make one that captures my likeness. I tried the native Lora trainer, the advanced lora trainer, and the FL Kohya LoRa trainer in ComfyAI. What I could get out of the native lora trainer was not good enough. And the advanced & FL Kohya Lora trainer had me trouble shooting endlessly. I asked Co-Pilot to help me through it, since I don't know much about it yet. But after installing version after version of ComfyAI, requirements, custom node versions, creating virtual environments and python versions, I think ai doesn't cut it when it comes to helping me through this. I've been at it for nearly 11 hours today. It's not for lack of trying, I just lack the expertise. I believe help from an actual person is my best bet at getting the results I'm after. 🙂

[dreamy lodge]

any reason this isnt working for me?

[prisma moss]

@ornate elk sorry for bothering, since i randomly get a huge memorly leak which fills up my 24gb vram, i enabled tiling again, but i dont see why it needs to tile upscale after every damn inpaint, simply slowing the process, and its not as if with it disabled that it uses much vram anyway

ive set this to that, and ill report back i guess

[prisma moss]

Ok so it still happens on its own out of nowhere, i wasnt even underload, i was just masking an image..

and yes i tried fresh venv too, idk if i should try, and fully wipe forge

[ornate elk]

Can you screenshot the settings when it happens again?

[prisma moss]

Sure

[chrome tangle]

So as i said, i am not sure what to post this in. I am getting very bright skin and a weird red outline in every picture. It is very unflattering. I have tried doing ti with prompts, without prompts, with vae, without vae, with addons (like DT and adetailer), with refiners, without, starting from scratch with a new config, trying different checkpoints. they all give me the same issue. It can't be the prompt either because it happens without prompts. Its driving me mad. The images i will post have no Lora on them so they look bad. But yeah... https://postimg.cc/gallery/3KCf4Zt

[vocal burrow]

probably an "issue" with the upscaler

depends how you ve set it up

[chrome tangle]

it happens with and without an upscaler. but this is how i have my hires.

[vocal burrow]

try to change the upscaler model

also if the outlines are here without upscalers at all

[chrome tangle]

What upscaler would you suggest?

am not sure of any good ones other than remacri

[ornate elk]

Try lower the hires steps to 10 and set the denois to 0.4

Or try Fatal Anime 50000 as upscaler:
https://openmodeldb.info/models/4x-Fatal-Anime

[vocal burrow]

@mild flicker you need to start with the argument --listen to be able to access your sd instance on other computers from the same LAN.

[tiny peak]

I was wondering where would be a good place to ask for help with Comfyui I have a custom node with custom templates but they are not working in comfy. Like the templates show but they are greyed out and don't load and the nodes don't work either . I noticed that there are 3 custom node foldes the one under user the one in Local app data and there are 2 there which should I use?

[tiny peak]

well tried copying custom node folder to all of the other ones no luck

[tiny peak]

lmao

[vocal burrow]

Just in case there was a link to some external discord or whatever. Don't follow it, the only official help you'll get is from https://stability.ai/ or their mailing system. This very channel is community driven

also usually you can install custom nodes throught comfyui interface itself.

that way you're sure that you won't mess up the installation

if installed correctly then look at the logs for errors.

[prisma moss]

aight so

which settings

[quick roost]

I'm running an old script, but something's not working right
Flux

[gleaming lodge]

Base on this script it shows like you're using an old version of the training script

Which means the script won't actually train correctly unless you either run it with a supported gpu

[crystal nacelle]

I’ve been using SD1.5 but want to upgrade to another one. What do you recommend? I have AMD RX7600 (8GB), i5 9th gen, SSD, 32GB pc. My priority is prompt adherence. Was thinking of SDXL…

[ornate elk]

Yea that would work or if your into anime or semi realism then go for Illustrious based models

[crystal nacelle]

Thank you. Which UI do you recommend? I’ve always used A1111, would like to keep using it or something similar

[ornate elk]

Automatic1111 or Forge with Zluda should work fine

[ornate elk]

A screenshot of the whole page with to Settings you used, like resolution, steps etc.

[prisma moss]

I assure you nothing out of the ordinary there

[patent fjord]

Hello, I have a problem, I followed the entire tutorial for Forge with ZLUDA, AMD Ryzen 5 5600G Processor with Radeon Graphics, 16.0 GB RAM, years ago I had used automatic1111, and it worked although slowly, after removing it now I wanted to try the new ways, I have been trying all night between zluda, DirectML and it always gives me this error

[ornate elk]

Hey, can you share the whole cmd log?

[patent fjord]

The tutorial told me to use Python 3.11.9

[ornate elk]

Python 3.11.9 is okay

But rocm shows no agent was found

Have you downloaded the correct gfx files for your GPU?

[patent fjord]

I suppose, just in case, you can tell me which one it would be so I can check that I download the correct one?

[ornate elk]

Do you only have the igpu of the cpu and not a dedicated GPU?

I checked the the CPU and it seems that the igpu is not compatible with Zluda. Only DirectML would be an option.

[patent fjord]

I had tried it with DirectML and it gave me the problem, I will check and send what comes out

[ornate elk]

Full cmd log pls

[ornate elk]

For directml you need to change the launch args of the webui-user.bat to --use-directml instead of --use-zluda.
Then delete the venv folder and relaunch

[ornate elk]

also add --skip-python-version-check
then delete the .zluda folder and the venv folder and relaunch the webui-user.bat

[patent fjord]

I added and deleted both folders but it recreated them when opening the webui.bat

[ornate elk]

never launch the webui.bat

only the webui-user.bat

thats why it didn't worked

delete the venv folder and launch the webui-user.bat again

[patent fjord]

Oh, I don't know why in my mind I remembered it being like that.

[patent fjord]

Thanks! Now it works. It's a bit slow, I don't know if it's because it generated the first time, but at least it works.

[ornate elk]

no problem!, you can try remove --lowvram and instead add --medvram and make sure --attention-quad is also there

[patent fjord]

I will try it, thanks for your help

[zenith sail]

Hey guys doing a 2049x1024 image and would like to be able to make it a little more crisp, what do you guys reccomend? SD upscale looks like shit:

[storm knoll]

Hi I've just updated comfy and all my nodes, I'm getting an annoying 2-3 second freeze every single time I click on my text encode nodes to change prompts. It only happens after I've already ran the workflow. If anyone else has seen the same issue and found a fix plz let me know. Not much to show, problem is as I said and removing nodes isn't really an option

?

why r u trying to get me to join this discord server like your life depends on it

seems sus asf

rofl

literal braindead

@marble python is trying to fish people into his discord server for reasons unknown

not sure who to @ for mods

[vocal burrow]

took care of it.

[quiet abyss]

Macbook Pro 16.2’’ M4 Pro - 14c - 20c - 48GB RAM - 512GB SSD
Is this good option for AI? I want to be able to handle Flux / Qwen or stable diffusion models locally. And maybe some smaller LLM

[vocal burrow]

should be fine for SD (and probably flux/qwen too)

cf pinned guides for tutorial about how to install things on your mac

[crystal nacelle]

are there any arguements you recommend I use like ---usemedvram or something? i only got 8GB vram on my puppy, 32GB ddr4 ram

[ornate elk]

For Auto1111:
--medvram-sdxl
For Forge:
--cuda-stream

[crystal nacelle]

okay thanks. the link you pinned also advised adding --use-zluda --attention-quad --skip-ort --skip-python-version-check

[ornate elk]

Yep they are required too

[quiet abyss]

Do you have one? How many seconds for image generation?

[vocal burrow]

No I don't, if I remember correctly I ve seen viking having around 2it/s on his mac (m1 ?)

I see people having from 6 to 10 it/s on sd1.5 models on a m3

src : https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html (search for darwin)

[storm knoll]

Spreading the word that if your UI freezes for multiple seconds everytime you click on text encode nodes downgrading the comfy frontend to 1.24.4 worked for me. Just add this to your launch .bat file "--front-end-version Comfy-Org/ComfyUI_frontend@v1.24.4"

[crystal nacelle]

im getting warning messages about setting GPU weights, should i do this? how do i do it?

[crisp apex]

Anyone here good at Comfyui? I am new and I am running into an issue where, I upload a new workflow to COmfy, but the missing nodes that I need to install are completely hidden. I can't even see what I need to install, there is just massive gaps of blank space in the workflow. Anyone know how to fix this?

Will do, thank you

I'm pretty sure your a scammer

But it was a good try

Bro... stop. Im not falling for it

I love how you delete it immediately

[safe fern]

!Help me please! I tried to exchange the wheel of the car using the flux kentext. But getting weird result. Could you give me any recommendation or any recommended workflow? Thanks

[tawny saddle]

dunt fall for the Wumplus nitro bot scam, dunt add it, spread the word 💜 +-+

[crystal nacelle]

got a problem with getting controlnet to work on my new forge zluda. getting a controlnet.py error on the log :

looks like the code on this link could help but i have no idea how to use it other than to make a notepad file, paste the code in the error directory and rename it to a .py file type lol
https://huggingface.co/pantat88/ui/blob/3004933886514ae448e66ec8e925b4aedf0fcf42/controlnet.py

[nova loom]

Hi, I'm getting NansException: A tensor with NaNs was produced in Unet. when I try img2img inpaint with illustrij_v8.

Tried "Upcast cross attention layer to float32" but the error keeps coming.
Tried --disable-nan-check but then I just get all black.

txt2img with illustrij works.
img2img with other models, even bigger size works (32GB cyberrealisticPony_v130 vs 6.9GB illustrij_v18).|

I'm using a Macbook M1 Pro.

Can anyone help?

SOLVED: Fixed after using an external fp16 sdxl-vae instead of the checkpoint's embedded vae. Thanks to this post.

[rain crater]

Hello, @nova loom ! Try reducing CFG scale to 3-5 (instead of the default 7). Higher CFG values can cause instability with some models on Apple Silicon.

[rain crater]

Hi @safe fern . follow my approch step by step.

1. Mask only the wheel you want to replace, not the entire car area
2. Use "Inpaint masked" not "Inpaint not masked"
3. Set Mask blur to 4-8 pixels for smoother blending
4. Use "Only masked" padding mode with 32-64 pixels padding

[alpine comet]

??? i downloaded from automattic 1111 github in the install and run on nvidia gpus bit

[vocal burrow]

rtx 5xxxx ?

[alpine comet]

i have rtx 2070

[vocal burrow]

laptop ?

[alpine comet]

desktop

[vocal burrow]

any iGpu on your cpu that might take priority over the rtx 2070 ?

[alpine comet]

idk what iGpu is, so im assuming no? idk sorry!

ohh intergrated

no dont think so

[vocal burrow]

run dxdiag and screenshot system and displays tabs

[vocal burrow]

2 display tabs

[alpine comet]

they were both the same soz

[vocal burrow]

how did you install a1111 ?

[ornate elk]

Please follow the install Guide from the pinned messages

[alpine comet]

oki

[balmy shore]

does anyone use sd-webui-ux? it doesnt have the normal env cmd prompt in the base folder and i dont know how to run the g it pull to update

or is there a run flag i can add to user.bat

[nova loom]

hi people, is anyone able to run controlnet XL on M1 (MacOS)?
SD1.5 controlnet does work
but whenever I try a XL checkpoint + XL controlnet models I get Placeholder storage has not been allocated on MPS device

btw, I'm using automatic1111. I just saw that diffusionbee mentions support for XL and controlnet but it has way less github stars than a1111. what do you guys use on mac?

SOLVED: Worked once I used a controlnet model more specifically compatible with my checkpoint. Turns out being "XL" is not specific enough.

[balmy shore]

i might go back to a111 from my current ui

[vast crest]

my sd forge is dead 🙁

[ornate elk]

Its installed the wrong way

So it can't be updated

[ornate elk]

Add git pull to an empty row/line of the webui-user.bat

Then relaunch

[vast crest]

How am I wrong? I did it right a long time ago. What should I do then? Do you have a good link?

[ornate elk]

You downloaded forge as zip but not the correct one.
A correct guide is on the first link of the pinned messages.

[vast crest]

Do you have the correct link you can send me?

[ornate elk]

Here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[vast crest]

ok, but is it still inside a winrar file? Is it good?

[vast crest]

is ok?

[woven walrus]

Am I taking crazy pills? Almost every time I load a comfyui workflow, there are missing custom nodes, even after "installing missing custom nodes" via the manager. This time I'm trying to load the kijai workflow for wan2.2 animate, and I'm missing WanVideoAnimateEmbeds and FaceMaskFromPoseKeypoints. I've already updated comfy and I've updated the wanvideowrapper, what am I missing? Feels like I'm the only person with this issue.

[alpine comet]

in the a1111 nvidia guide where do i add the '--medvram-sdxl --no-half-vae' in commandline_args because i already ahve --xformers in but when i add the vram command it crashes and doesnt open, says it doesnt recognise it

[worn oyster]

anyone is familiar with IP-Adapter from controlnet to mix images together ?

[ornate elk]

At the line commandline_args=

[rose apex]

Using forge (Zluda with HIP6.4) takes quite a bit of time to generate images. 4 mins for 832x1216 images (Later upscaled by 1.5 with Hires fix) using only 2 loras. My GPU memory usage also seems to be very low throughout

It spikes in between but mostly stay at around 3-4 gbs. Is this a problem?

[ornate elk]

Your system ram is a problem as its only 16gb.
You need to increase the Windows pagefile. That will fasten things up.

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.

[ornate elk]

Also 3-4 GB usage is not normal. Can you show me the full cmd log? It seems like its using CPU+RAM instead of the GPU

[rose apex]

You've told me this before and I have it set to 16 and 24 gbs yes.

[rose apex]

I'm outside rn sorry, I'll show you as soon as I get home

[ornate elk]

Alright

[rose apex]

It's the total opposite in ComfyUI. There, the vram usage goes above 15.5 Gbs and crashes my display

😑

Same 832x1216 latent upscaled by 1.5x

[ornate elk]

Enable Tiled VAE Decode, that should fix the crash

[rose apex]

I think it's already enabled or maybe it's missing somewhere. I'll verify it again

[ornate elk]

Should be the last Node before image output

Instead of VAE decode

[vast crest]

Unfortunately, nothing starts anymore. I followed the github guide (which I must say is a terrible guide), but nothing works anymore. I moved the loggers and checkpoints from the old forge to the new one, but it doesn't start anymore. The version of forge I downloaded doesn't have the COMMANDLINE_ARGS= line where you can add: --cuda-stream

[ornate elk]

The one I sent you? Why is it a terrible guide?

[ornate elk]

The commandline_args can still be added in the webui-user.bat which is in the webui folder

Also you showed me your webui-user.bat yesterday so what happened since then?

[vast crest]

It only started once. I left because it was too late. I just logged into my PC, and since then it hasn't started anymore. 🙁

I'll have to do it all again from scratch

[ornate elk]

Yea its not that difficult, download the forge zip then run the update.bat and then change the stuff in the webui-user.bat in the webui folder

[vast crest]

I did it, I edited this file: webui-user.bat

[ornate elk]

Okay

And what does happen when you launch the run.bat?

[vast crest]

when I start with that file, it opens, but in the run.bat file there is almost nothing (I mean the command lines) do I have to write them?

[ornate elk]

No the run.bat just points to the webui-user.bat of the webui folder. So its fine if you only edit the webui-user.bat but always start the run.bat to launch it

[vast crest]

ok, then I reinstalled everything and tried again

[ornate elk]

Alright, let me know if it works then 🙂

[vast crest]

in your opinion, it's better this way: @echo off
git pull

set PYTHON=python
set GIT=git
set VENV_DIR=venv
set COMMANDLINE_ARGS=--cuda-stream --xformers --opt-sdp-attention --opt-channelslast

call webui.bat

or like this: @echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat

editing webui user.bat?

[ornate elk]

remove --opt-sdp-attention --opt-channelslast as one doesnt work with xformers and the other one doesnt improve performance

[vast crest]

ok

[ornate elk]

Dont do this:
set PYTHON=python
set GIT=git
set VENV_DIR=venv

leave them empty after the =

[vast crest]

so is the second line I sent you better?

[ornate elk]

This one yes:
`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--cuda-stream --xformers

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat`

[vast crest]

ok,perfect,thx 🙂

so - git pull - I don't use lod anymore?

[ornate elk]

not needed, you can run the update.bat to update forge

but you can also add it to the webui-user.bat if you want

but Forge doesnt get much updates lately

[cobalt agate]

@woven walrus You need to update your Comfy to the latest.

@worn oyster I use animated masks to drive the IPAdapter to mix two image streams together to form a video.

[light depot]

hi bro, i try to install stable diffusiion, my laptop with rtx4060, but problem with the torch version, i try download & instal manul for the correct version - torch 2.5.1+cu121, still not working.

[vast crest]

I'm really unlucky today XD. keeps giving me trani errors: ModuleNotFoundError: No module named 'triton'
xformers version: 0.0.27
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4070: native
Hint: your device supports --cuda-malloc for potential speed improvements.
VAE dtype preferences: [torch.bfloat16, torch.float32] -> torch.bfloat16
Installing bitsandbytes==0.45.3
CUDA Using Stream: True
H:\webui_forge_cu121_torch231\system\python\lib\site-packages\transformers\utils\hub.py:128: FutureWarning: Using TRANSFORMERS_CACHE is deprecated and will be removed in v5 of Transformers. Use HF_HOME instead.
warnings.warn(
Using xformers cross attention
Using xformers attention for VAE

[ornate elk]

that triton warning can be ignored

[vast crest]

ok, but the program page doesn't open on the browser. I've been waiting for 5 minutes and nothing: Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: f2.0.1v1.10.1-previous-669-gdfdcbab6
Commit hash: dfdcbab685e57677014f05a3309b48cc87383167
Installing requirements
Launching Web UI with arguments: --cuda-stream --xformers
Total VRAM 12282 MB, total RAM 32694 MB
pytorch version: 2.3.1+cu121
WARNING:xformers:A matching Triton is not available, some optimizations will not be enabled
Traceback (most recent call last):
File "H:\webui_forge_cu121_torch231\system\python\lib\site-packages\xformers_init_.py", line 57, in _is_triton_available
import triton # noqa
ModuleNotFoundError: No module named 'triton'
xformers version: 0.0.27
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4070: native
Hint: your device supports --cuda-malloc for potential speed improvements.
VAE dtype preferences: [torch.bfloat16, torch.float32] -> torch.bfloat16
CUDA Using Stream: True
H:\webui_forge_cu121_torch231\system\python\lib\site-packages\transformers\utils\hub.py:128: FutureWarning: Using TRANSFORMERS_CACHE is deprecated and will be removed in v5 of Transformers. Use HF_HOME instead.
warnings.warn(
Using xformers cross attention
Using xformers attention for VAE

[ornate elk]

hey, follow the setup guide from the first link of the pinned messages. Then you shouldnt have any issues.

[vast crest]

very strange

[light depot]

thanks

[ornate elk]

can you show the full cmd log?

[vast crest]

do you mean the cmd window that appears when the program should start?

[ornate elk]

ah okay, i see, forge has an issue with xformers because triton is missing.
Close the cmd. remove --xformers from the webui-user.bat and delete the venv folder inside the webui folder.
then relaunch the Run.bat and it should start normaly

[vast crest]

Do I have to install this Triton? What is it for?

[ornate elk]

its a module needed for xformers. --cuda-stream should be enough as forge is more optimized than auto1111 which needed xformers

[vast crest]

I discovered something very strange. I was moving all my files to the new forge, but I noticed that it no longer starts when I move my checkpoints. How is this possible?

Could it be that I have a checkpoint file that is giving me an error?

He does it to me even when I marry my Loras

the program starts even without the vens folder which I don't have

[ornate elk]

Forge doesnt need a venv

[vast crest]

Ok, but the fact is that if I move my time and my checkpoints, then maybe it won't start, look.

if I remove them, it starts, why???

[ornate elk]

how much models and loras do you have?

can you try just move the loras into models/lora folder and then relaunch?

[vast crest]

I have 849 Lora

what: ControlNet preprocessor location: H:\webui_forge_cu121_torch231\webui\models\ControlNetPreprocessor

the folder is empty, I don't understand why it goes out

If I then remove the time, that error no longer appears

[worn oyster]

i'm trying to mix two image into another image but for some reasons the image given to the ip-adapter is not considered, it's generate colorful patch and nothing else, do you know if the IP-adapter is supposed to launch in the webui-user.bat ?

[ornate elk]

probably forge needs to calculate the lora hashes for all of them and that takes very long. Try adding only a few to see if it still starts up with them or not

[ornate elk]

controlnet models go into models/controlnet

[balmy shore]

ty will try this

[vast crest]

Hi, in the end, nothing to do, I no longer have the energy to check 850 lora and find out which one is defective. Maybe it's the version of Forge that doesn't work for me. Is there another type of version I can download?

[balmy shore]

it said this and then just didnt update

Updating d8cc7ad0..df7045bb error: Your local changes to the following files would be overwritten by merge: extensions-builtin/sd-webui-ux/javascript/anapnoe_sd_webui_ux.js Please commit your changes or stash them before you merge. Aborting
not sure how to p[roceed

is there a --force-update flag

[ornate elk]

You could also try setup ReForge

[ornate elk]

Okay to fix that:
Go into the **stable-diffusion-webui-ux **folder.
Click in the File Explorer bar. (not in the search bar).
then type cmd and hit enter.
after that type and run:
git reset --hard HEAD
hit enter and then
git pull
Then rerun the webui-user.bat

[balmy shore]

oh shoot its doing it thank u

[fair oxide]

Is ROCm available on Debian 13?

[vast crest]

Hi, I don't know what this ReForge is. I've tried over and over again, but my Forge never starts. Is there another version I can try?

[fleet pawn]

Hello folks, I've got a question and im not sure where to ask it but i figured here would be a good place to start. I've been using Automatic 1111 for the past few days, just got into it and started playing around. I installed the dreambooth extension from the webui's extension tab, and now when i try to run the sd.webui i get a runtime error : RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check. I am very new to all this and I'm not sure how to proceed.

I've got a 5070ti and its been running sd.webui beautifully so far.

[ornate elk]

There is Auto1111, Forge, ReForge, Forge Neo
But have you tried just not moving the loras over and just adding --lora-dir "PathToLoras" to the webui-user.bat?

[ornate elk]

Hey Dreambooth is outdated and breaks the Auto1111.
You need to delete it from the extensions folder and then delete the venv folder and relaunch the webui-user.bat.
Then you need to do the steps again to change torch for rtx 50 series.

[vast crest]

no. should I add --lora-dir / to the webui-user.bat file?

[grim totem]

my a1111 suddenly stopped booting with error "Cannot find empty port in range: 7860-7959". i retried multiple times. it was okay just yesterday and i don't seem to have changed anything in my system

[fleet pawn]

I see no folder called venv

[ornate elk]

Oh right the sd.webui has no venv

[fleet pawn]

I went to the extensions folder and removed the extension.

Would it be easiest to fix by just moving my Models out, deleting the entire folder, and re extracting the initial automatic 1111?

[vast crest]

where should I add it?

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat

[ornate elk]

To the commandline_args=

Where you also added --cuda-stream

[vast crest]

since strangely enough it's already there but it's useless

[ornate elk]

No thats for webui linking

And disabled

[vast crest]

ah,ok

[ornate elk]

Best would be to install Auto1111 the normal way or Forge by using the setup guides.
You find them on the first link of the pinned messages here

[vast crest]

now

this is what happens if I insert the lines like you told me

and yet I did everything right

[onyx solar]

You have "--lora-dir" as an argument, but you don't tell it what that directory is. If you're not going to specify that in COMMANDLINE_ARGS, don't put it there. Otherwise, give it the primary directory where you would store your LoRAs.

[foggy needle]

Is there anyone that has managed to get InstantID working within Fooocus? Without InstantID, it works just fine, but when enabling it, gets stuck on "Preparing task 1/1", then does this

[balmy shore]

tried doing a fresh install of auto111 but for some reason xformers wont install, i did the stanard winnividia directions on their github and then tried --reinstall-xformers flag

actually i just looked and it does look like it started with xformers

but when hit render theres an error and says xformers is not defined

sdp works though it was defaulting to doggetx

[ornate elk]

Remove --reinstall-xformers and --force-enable-xformers
Just add --xformers
Then delete the venv folder and relaunch the webui-user.bat

Also upgrade to python 3.10.9 64bit before deleting the venv folder or delete it afterwards again.

[balmy shore]

ah ty

[woven walrus]

Trying to get my friend up and running with comfyUI. We used the same workflow and settings, but his video gets blurry near the end for some reason. Any idea what is causing this?

[maiden flint]

I installed Stable Diffusion WebUI AMDGPU Forge on Stability Matrix

I'm getting rocBLAS error: Cannot read C:\Program Files\AMD\ROCm\6.2\bin/rocblas/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1032

[ornate elk]

Dont use stability matrix for AMD.
Follow the Forge with Zluda Setup guide from the first link of the pinned messages here

[vast crest]

hi

is ok now???

[vocal burrow]

looks good to me

[vast crest]

ok thx 🙂

[vast crest]

Are there any commands to make startup faster? I would also like to know if I need to enable any options so that I can see all my checkpoints immediately, since now I always have to press: refresh page

[ornate elk]

Is your H drive an HDD or SSD?
And how much RAM do you have?

[vast crest]

ssd 🙂 2tb

anyway

look,now is ok???

[ornate elk]

Yep looks good

[vast crest]

ohhhhhhhhhhh yes

[lapis raven]

https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu/issues/628
How do I fix this error?

[novel owl]

hello i have a problem with my comfyui all works but at the moment of the VAE decode it put me this error

the error log

someone can help me please ?

[novel owl]

@ornate elk my lord can you help me please?

[novel owl]

if you want you can again take control of my computer

[ornate elk]

hey, follow the updated AMD setup Guides from here and then let me know if which exact error you get if any:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[lapis raven]

Therefore, it is like this.

[ornate elk]

can you paste me the cmd log you get in here? Because on the Github Link Logs there are multiple issues.

[lapis raven]

Sorry
I restarted PyTorch? and launched it after reinstalling...

[ornate elk]

okay, can you just copy and paste the whole cmd log in here when you launch the webui-user.bat?

then i can check for any issues

[ornate elk]

looks correct so far

Let it compile, it can take a while

also whats inside your webui-user.bat ?

[lapis raven]

I made a mistake when opening it.
If it's webui-user.bat, it's this.

set COMMANDLINE_ARGS= --use-zluda --update-check --medvram --no-half --opt-sub-quad-attention

[ornate elk]

wait

remove --no-half

that just slows down everything

[lapis raven]

ok!

It won't progress past 0%.

[ornate elk]

just wait its still compiling

can take up to 1 hour max on some gpus

[lapis raven]

Hmm...

It threw an error...

Checking with GPT...

[ornate elk]

Can you show the complete cmd log again?

Please don't use chatgpt for zluda stuff

[lapis raven]

why

[ornate elk]

Thats not the complete cmd log

But can you try just generate 1 image at default settings without using a lora?

Does that work?

You may need to relaunch the webui-user.bat first

[ornate elk]

Because chatgpt doesn't know about custom webui amd forks with Zluda. And will output stuff that breaks it even more instead of fixing it.

[lapis raven]

I see...
A similar error was generated.

[ornate elk]

The issue comes from the dynamic prompts extension it seems

Also you added a lot of embeddings which all didnt load because they get blocked because they are .pt files (insecure)

There are several other issues involving extensions. For me it seems you copied them over from an old install?

[lapis raven]

I installed it from the official GitHub.

[ornate elk]

The extensions?

Did you installed them all today?

[lapis raven]

yes

[ornate elk]

Then try generate an image again but dont enable the dynamic prompt extension

If that outputs an error post the cmd log again

[lapis raven]

おｋ

[ornate elk]

Also Umi-ai extension is not compatible with AMD GPUs. I looked it up right now.

So that could be an issue too

[lapis raven]

thank you

[ornate elk]

To fix the webui, delete the umi-ai extension.
And then also delete the venv folder.
Then also add --skip-ort to the commandline_args of the webui-user.bat and then relaunch it.

[lapis raven]

ok

[ornate elk]

I'm off now, its late here. Let me know if it worked. If not we can fix it later

[lapis raven]

Thank you.
Please wish me success.

[maiden flint]

I'm getting the same error message

This isn't working either/

[maiden flint]

Nvm.. I believe it's because I'm using 6800s and went with 6800+ guide

[gloomy bolt]

Recently when i launch Auto1111, the startup (terminal) gets stuck loading on some lines for a really long time. Is there a way to soft reset this, such as deleting the venv file or something?

[maiden flint]

ZLUDA device failed to pass basic operation test: index=0, device_name=AMD Radeon RX 6800S [ZLUDA]
CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)
...
TypeError: '<' not supported between instances of 'NoneType' and 'int'

Now I'm getting this

[ornate elk]

Yes deleting the venv can help.

[ornate elk]

Can you show the complete cmd log?

[lapis raven]

It didn't work out.

[ornate elk]

Okay, then please send the new cmd log

[lapis raven]

Here it is, though it's late.

[ornate elk]

Okay do these steps:
Edit the webui-user.bat and add git pull to the 2 line below echo off.
Then add --skip-ort to the commandline_args=
Then download and install Python 3.10.11 64bit.
Then delete the venv folder and the .zluda golder.
Boath are located in the stable-diffusion-webui-amdgpu folder.
Then relaunch the webui-user.bat

[lapis raven]

ok

error…

[ornate elk]

You still have python 3.10.10 and not upgraded to 3.10.11

[lapis raven]

oh…

[ornate elk]

Okay can you screenshot your webui browser with your settings like model, resolution etc?

[lapis raven]

Is this okay?

[ornate elk]

Dreamshaper v8 should work. Have you tried just using the default values to gen 1 image?

[lapis raven]

default？ This should be the original settings with no modifications.

Since it seems like a CUDA-related error, should I reinstall it?

[ornate elk]

AMD doesnt support cuda. Zluda emulates the cuda instructions.
Can you open up a cmd and run:
Zluda --Version

[lapis raven]

Unrecognized argument: --Version

Run zluda.exe --help for more information.

[ornate elk]

Oh try zluda --version

[lapis raven]

ZLUDA 3.8.4

[ornate elk]

Ahh okay thats old

3.9.5 is the latest

[junior sparrow]

I can help you

[ornate elk]

@lapis raven Then I would recommend the following steps to update Zluda:
Delete the content of C:\zluda but keep the folder.
Also make sure your AMD Adrenalin Driver is updated.
Download the latest zluda 3.9.5 rocm6 zip (not nightly) from here:
https://github.com/lshqqytiger/ZLUDA/releases/
And place the unpacked files of zluda in your C:\zluda folder

Open up a cmd and run:
Pip cache purge
Then in the webui delete the .zluda and the venv folder again then relaunch the webui-user.bat

[lapis raven]

no…

[novel owl]

hello i hear you ^^

[ornate elk]

Okay can you try remove all extensions from the extensions folder.
Then delete the venv folder and relaunch the webui-user.bat?

[ornate elk]

Zluda didnt found the GPU gfx version

Have you downloaded the custom gfx1032 files and put them in the correct folders?

[maiden flint]

I did this

Go into C:\Program Files\AMD\ROCm\6.2\bin\rocblas folder.
There copy and rename the library folder to old_library
Open the gfx .zip file and drag and drop all files of the library folder into the library folder. Not into the old_library folder.
If the zip contains a rocblas.dll, copy that into the C:\Program Files\AMD\ROCm\6.2\bin\rocblas folder.

Should I have kept the files from old_library in library?

I got the files from rocm.gfx1032.for.hip.sdk.6.2.4.navi21.logic

[ornate elk]

Yes you should have kept the files from old library in library and then add the new ones to library and overwrite

[maiden flint]

That worked! Thanks

[ornate elk]

Perfect, no problem 🙂

[maiden flint]

Getting this error + a crash. Dedicated 8.0 GPU. Tried lowering GPU weights

[vocal lichen]

I had to reinstall webui, and I chose a different folder without deleting the old one. Now the lora tab is acting weird, showing the loras from the old installation instead of the new one. Even weirder, new loras won't display even after adding them to both installs. Is there a cache of it outside the install folder or something?

[ornate elk]

how did you install the webui?

and which one?

[ornate elk]

the joblib error can be ignored or you install it inside the venv

you need to show your settings when you get the out of vram error

[maiden flint]

Show my settings?

[ornate elk]

like resolution, steps, model etc

[vocal lichen]

git cloned lshqqytiger's amd version of webui

[ornate elk]

okay, and now you see old loras? are any loras in the new installation?

did you linked something in the webui-user.bat?

[vocal lichen]

I see the loras I had yesterday before reinstalling

but not the loras I added today, to both folder

[ornate elk]

are the loras made for the base model you selected?

also whats in your webui-user.bat and can you show your cmd log when launching it?

[maiden flint]

512x512, 20 steps, ealismIllustriousBy_v50FP16.safetensors

[vocal lichen]

unsure, I tried a bunch and none worked... but I even deleted one and it's still not showing the change in the ui. It's like there's a cache somewhere

[ornate elk]

have you reloaded or restarted the webui after removing them?

[vocal lichen]

yes

[ornate elk]

okay then please cmd log and webui-user.bat content

[vocal lichen]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--use-zluda --update-check --skip-ort --skip-python-version-check --precision full --no-half

set HIP_VISIBLE_DEVICES=1


call webui.bat

[ornate elk]

remove --precision full --no-half, they just slow down and are not needed for AMD

[vocal lichen]

PS C:\ai\stable-diffusion> .\webui-user.bat
venv "C:\ai\stable-diffusion\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.10.1-amd-45-g3e1758c5
Commit hash: 3e1758c509354f44c1bc2aab06891ddfb433259f
ROCm: agents=['gfx1100']
ROCm: version=6.4, using agent gfx1100
ZLUDA support: experimental
ZLUDA load: path='C:\ai\stable-diffusion\.zluda' nightly=False
Skipping onnxruntime installation.
--------------------------------------------------------
| You are not up to date with the most recent release. |
| Consider running `git pull` to update.               |
--------------------------------------------------------
W0923 14:51:57.564000 18784 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\ai\stable-diffusion\venv\lib\site-packages\timm\models\layers\__init__.py:48: 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.
C:\ai\stable-diffusion\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: --use-zluda --update-check --skip-ort --skip-python-version-check --precision full --no-half
Loading weights [dd240b5bb5] from C:\ai\stable-diffusion\models\Stable-diffusion\unstableIllusionPRO_pro.safetensors
Creating model from config: C:\ai\stable-diffusion\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 16.4s (prepare environment: 25.5s, initialize shared: 0.8s, other imports: 0.2s, load scripts: 0.5s, create ui: 0.4s, gradio launch: 0.3s).
Applying attention optimization: Doggettx... done.
Model loaded in 3.6s (load weights from disk: 0.2s, create model: 0.9s, apply weights to model: 1.4s, apply float(): 0.2s, load textual inversion embeddings: 0.2s, calculate empty prompt: 0.6s).

[ornate elk]

can you try again and send the full cmd log if you get an error?

[ornate elk]

why is your webui outdated if you said you installed it today?

thats not possible

your sure thats the "new one" ?

[vocal lichen]

yes, 100%

[maiden flint]

This time it didn't crash\

[vocal lichen]

yeah, I made some changes in git, so I guess it's just using it to check for updates

[ornate elk]

did you installed the new version in the last few days?

[vocal lichen]

old version yesterday, new version an hour ago

[ornate elk]

then your log is from the old version

[vocal lichen]

in git, I made a new branch, removed some stuff from gitignore, and made some commits

but it's not an outdated version

[ornate elk]

can you try to setup the webui again the normal way and dont mess with the git stuff?
then it will probably work normal

[vocal lichen]

reverted back to master, ran it again... same problem

venv "C:\ai\stable-diffusion\venv\Scripts\Python.exe"
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.10.1-amd-43-g1ad6edf1
Commit hash: 1ad6edf170c2c4307e0d2400f760a149e621dc38
ROCm: agents=['gfx1036', 'gfx1100']
ROCm: version=6.4, using agent gfx1100
ZLUDA support: experimental
ZLUDA load: path='C:\ai\stable-diffusion\.zluda' nightly=False
W0923 15:02:50.668000 6476 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\ai\stable-diffusion\venv\lib\site-packages\timm\models\layers\__init__.py:48: 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.
C:\ai\stable-diffusion\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:
C:\ai\stable-diffusion\venv\lib\site-packages\torch\onnx\_internal\registration.py:159: OnnxExporterWarning: Symbolic function 'aten::scaled_dot_product_attention' already registered for opset 14. Replacing the existing function with new function. This is unexpected. Please report it on https://github.com/pytorch/pytorch/issues.
  warnings.warn(
ONNX failed to initialize: module 'optimum.onnxruntime.modeling_diffusion' has no attribute 'ORTPipelinePart'
Loading weights [dd240b5bb5] from C:\ai\stable-diffusion\models\Stable-diffusion\unstableIllusionPRO_pro.safetensors
Creating model from config: C:\ai\stable-diffusion\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 17.0s (prepare environment: 24.8s, initialize shared: 2.1s, load scripts: 0.5s, create ui: 0.5s, gradio launch: 0.1s).
Applying attention optimization: Doggettx... done.
Model loaded in 3.3s (create model: 0.9s, apply weights to model: 1.3s, load textual inversion embeddings: 0.2s, calculate empty prompt: 0.7s).
Compilation is in progress. Please wait...

[ornate elk]

this messages is very strange:
W0923 15:02:50.668000 6476 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
can you delete the venv folder and relaunch the webui-user.bat?
and now it doesnt show launch args, dont forget to add --use-zluda --update-check --skip-ort --skip-python-version-check

[vocal lichen]

stupid venv, I don't want to go through that again... the mirrors I get are like 20 kb/s, it's a pain

[ornate elk]

its cached dont worry

just let it rebuild clean

a corrupt venv can cause much trouble

[vocal lichen]

that's the reason I had to reinstall everything

I've always hated python for their terrible package manager

[ornate elk]

yea but something is clearly off with your installed version, so better get a fresh venv

[ornate elk]

looks okay, zluda just needed to compile fully.
But as you only have 16gb of RAM you need to increas your windows pagefile to speed up sdxl/illus gen time

[maiden flint]

Ah, I edited the bat a little before I ran it this time. Lemme run it again and see

Right now my pagefile is 40960

[ornate elk]

ah okay

should be fine then, but you can recheck it later. make sure its only enabled for C drive and disabled for any other drive and then set to 16000min and 24000max or 32000 if you want

[maiden flint]

Ok I'll change it to that and see if it fixes crashes

[ornate elk]

a PC restart is required after chaning the pagefile

[vocal lichen]

fresh venv, same issue

venv "C:\ai\stable-diffusion\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.10.1-amd-43-g1ad6edf1
Commit hash: 1ad6edf170c2c4307e0d2400f760a149e621dc38
ROCm: agents=['gfx1100']
ROCm: version=6.4, using agent gfx1100
ZLUDA support: experimental
ZLUDA load: path='C:\ai\stable-diffusion\.zluda' nightly=False
Skipping onnxruntime installation.
You are up to date with the most recent release.
W0923 15:17:54.633000 1852 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs.
C:\ai\stable-diffusion\venv\lib\site-packages\timm\models\layers\__init__.py:48: 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.
C:\ai\stable-diffusion\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: --use-zluda --update-check --skip-ort --skip-python-version-check
Loading weights [dd240b5bb5] from C:\ai\stable-diffusion\models\Stable-diffusion\unstableIllusionPRO_pro.safetensors
Creating model from config: C:\ai\stable-diffusion\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 38.2s (prepare environment: 65.4s, initialize shared: 1.1s, other imports: 1.1s, load scripts: 0.4s, create ui: 0.7s, gradio launch: 0.4s).
Applying attention optimization: Doggettx... done.
Model loaded in 3.3s (create model: 1.0s, apply weights to model: 1.3s, load textual inversion embeddings: 0.2s, calculate empty prompt: 0.7s).

[ornate elk]

can you remove any lora of the webui and relaunch and then check if you see any?

if every lora is removed of models/lora you shouldnt see any of them after the relaunch

[maiden flint]

unfortunately did not fix crash this time

crashed after 100%

[ornate elk]

on 512x512 resolution?

[maiden flint]

Yep

[ornate elk]

strange

on which value is your gpu weight slider?

[maiden flint]

7152 but I’ve lowered it to 2000 and it also doesn’t work there

I’ll try updating drivers and going into overclock mode etc

[ornate elk]

do you havy anything open in the background like twitch, wallpaper engine or games?

[maiden flint]

Maybe that will fix

Nah I close it all

But will try everything next time and report back

Thank you!

[ornate elk]

okay, no problem, now its only a vram issue

you can also try to enable Never OOM for VAE at the bottom of forge

[vocal lichen]

that kinda helped, at least I'm seeing different loras now. Thanks a lot for the help, I appreciate it

[ornate elk]

No problem, good to hear, but your seeing loras now even if you remoed all of them from the new webui?

[vocal lichen]

I've already readded them

[ornate elk]

ah okay

[maiden flint]

Still crashing lol

Basically my laptop fan begins to get louder and louder and at 100% the entire PC blinks off

[paper dagger]

Set your cpu to 80-90%, its throttling it but might help? 🤷‍♂️

[ornate elk]

What's your pagefile set to? And how much space do you have left on your C drive?

[balmy shore]

anyone mess with openoutpaint, its pretty neat

wondering how important using an inpaint model is for the base gens

i got the controlnet part of the extension working with its inpaint model but i dont relaly have that many inpaint like sd1.5 models

[past tusk]

Can anyone please answer me my forge ui lora is being detected but not being used in images and there are not error massage

[ornate elk]

with sdxl/illus inpaint models are not needed anymore

[ornate elk]

is it compatible with the base model?

[past tusk]

@ornate elk yes it is compatible with base model i just changed some setting on my display after that it stopped working

Name of lora is manga illustration

[ornate elk]

for which model version?

[past tusk]

Sdxl

[ornate elk]

and you already restarted the webui ?

can you try generate an image with the lorad added, and then drop it back into the PNG-Info tab to check if the lora is mentioned in the meta data?

[past tusk]

No I didn't do that I will get back after I do that .

I restarted webui and did fresh installation also for what I did not do was checking png info

https://civitai.com/models/1255751/manga-mode-illustrious this is the lora

[ornate elk]

ah this one is for illustrious models not sdxl

[past tusk]

But when I used this before it worked perfectly

[ornate elk]

compare the images meta data in PNG-Info maybe it was with an other checkpoint?

because illustrious is based on sdxl but that doesnt mean it works with every sdxl models or not at all with them

[past tusk]

I used this lora with anylora checkpoint, and checkpoint named WAINSFW-illuatration-sdxl

For both of them it worked like wonder without any issues

[ornate elk]

wainsfw is an illustrious based model, so that works

but anylora isnt, or which did you used?

[past tusk]

Both and it worked best with wainsfw

[fluid maple]

Wan Animate native works with 640 x 640. But when I change dimensions, it bizarrely no longer uses the reference image - anyone hit that?

[ornate elk]

yes, because the lora and the model match the base version, can you share the link to your anylora one?

[past tusk]

https://civitai.com/models/23900/anylora-checkpoint here

[maiden zephyr]

i havent used a1111 in a while and tried to launch it... it just stops doing anything after Launching Web UI with arguments: --xformers --medvram... I havent touched the install since the last time i ran it, and it was working then.... is there any reason why it would stop working?

[ornate elk]

as i tought, thats an 1.5 model, it doesnt work at all with sdxl or illustrious loras

[balmy shore]

i cant or havent been able to run xl models so still on 1.5 but it seems to be sortof working

[ornate elk]

you mean in general? whats your gpu?

[ornate elk]

hey, can you share the full cmd log? then i find the issue

[maiden zephyr]

Python 3.10.8 (tags/v3.10.8:aaaf517, Oct 11 2022, 16:50:30) [MSC v.1933 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Launching Web UI with arguments: --xformers --medvram

[balmy shore]

gtx1060, 6gb. its like just on the edge i think of being able to run xl. but i just have been using 1.5 models since i started

it might just be loading sometimes it takes a beat

[ornate elk]

ah well ,yea it can run sdxl models with Forge webui but it will be much slower than 1.5

[balmy shore]

ya due to that ive avoided them and just try to improve my results with 1.5 , so i should try to find some inpaint models to get the most out of this?

[ornate elk]

hm okay looks normal but the webui version is not updated.
do this:
edit the webui-user.bat and in the 2 row/line below echo off you add git pull
then download and upgrad to Python 3.10.11 64bit
Then delete the venv folder and relaunch the webui-user.bat

[maiden zephyr]

i tried launching it without the arguments and it does the same thing... just stalls out after the launching command.... are there any logs for a1111

[ornate elk]

nope no logs :/

but if you do the steps from above it should launch again

[maiden zephyr]

webui version is at the most recent version

[ornate elk]

oh true lol

sry

yea then just upgrade python and delete the venv + relaunch webui-user.bat

python 3.10.7 and 3.10.8 were a bit buggy back then

[maiden zephyr]

will it recreate the venv ?

[ornate elk]

yes, with the latest torch and xformers version

[ornate elk]

hm if you want to use inpaint then yes, but i dont think there are that many
but most semi realism models dont need one

[balmy shore]

i used UNIGetUI to update some random things like torch and transformers but i dont know what directory its updating but it didnt break anything

[ornate elk]

unigetui doesnt update webui stuff because its inside a venv

[balmy shore]

ah i figured

i have no idea where all these packages are installed that its detecting lol but i had a feeling the gradio servers like a111 were basically compartmentalized

https://i.imgur.com/p8pLhmh.mp4

[wild tendon]

can someone help me

[balmy shore]

the openoutpaint thing is sort of working it gens like 4 essentially irrelevant sections and one that actually fits in context

[ornate elk]

probably denois to high, but if you want to improve your 1.5 images more then go for hires fix upscaling + adetailer

[balmy shore]

i do not think

[wild tendon]

yeah

[balmy shore]

because of birds?

[wild tendon]

no waffles

[balmy shore]

u know what they say

[wild tendon]

when the waffles hit orange man

orange man is allergy

[balmy shore]

it aind dumb if the ends justify the means

[wild tendon]

:((((((

[ornate elk]

@wild tendon with what do you need help now?

[fluid maple]

Wan Animate native works with 640 x 640. But when I change dimensions, it bizarrely no longer uses the reference image - anyone hit that?

I would advise walking outside the front door for a while

[ornate elk]

bye spammer

[maiden flint]

I tried a bunch of different permutations. Rn it's 40-64 GB

But I have 215 GB free on my C drive

[maiden flint]

It seems like this may be a solution for nvidia users but I have amd: https://nvidia.custhelp.com/app/answers/detail/a_id/5490/~/system-memory-fallback-for-stable-diffusion

[prisma moss]

Can sopmeone elaborate why dfq is this happening to me (at random)?

[ornate elk]

Could be an out of vram error maybe.

[rain crater]

I think your GPU is running out of memory when trying to allocate space for model operations.
This is my approach:
1.reduce batch size.
2.clear GPU cache regularly
3.use gradient checkpointing
4.kill other GPU processes

I hope this is helpful for you.

[prisma moss]

24gb is not enough for 1mpx hrfix generation??

[rain crater]

24gb should definitely be enough for 1 megapixel hires fix generation!. 1MP is only around 1000x1000 pixels, which is quite modest.
But even though , 24gb is enough, there are some issues can cause errors.

-Multiple generations without clearing cache
-Denoising strength too high → more VRAM usage
-Upscaler method → some use way more VRAM than others
-Hires steps too high → try 10-20 instead of 50+

[prisma moss]

im using 0.35 denoise

remacri model

10 hr steps

ALTHO, i set 0 in tiling size

(because of inpainting reasons)

[prisma moss]

a friend of mine is having this issue what is it?

[umbral grove]

For about 2 days comfy just will not work. I been using it for 2 years and havent had this issue before. Whenever i queued something it just crashes. no errors nothing. when i rarely do get errors it says i am out of memory. I even tried setting the res to 30 by 30 and nothing.

I feel like my account will get hacked if i join lol

I join but ima make an alt real quick lol

Yea i joined the server with an alt and yall asking for my wallet lol @shy falcon

[ornate elk]

Thanks for the hint, we banned him.

[ornate elk]

What's your GPU and what does the cmd shows when launching Comfyui?

[umbral grove]

I fixed it, it was a really dum issue that I would have no figured out without someone help

[ornate elk]

Ah alright xD

[umbral grove]

I had to mess with this to fix it. Tho I might have another question, anyway to lower vram usage? Am trying to use wan at a higher res but my 4070 sorta struggles

[ornate elk]

Ah the Pagefile.
There aren't many nodes that help with vram usage.
There is Tiled VAE Decode which can help with large resolutions.
And there are some vram cleaner nodes that can be added in between tasks to reset the vram.
Then there is stuff like sage attention and nunchaku which can help too.
But video gen on 12GB vram is not that great. Even with more its not that great xD

[umbral grove]

I am using sage attention but never heard of nunchaku before

[ornate elk]

It got a hype lately:
https://github.com/nunchaku-tech/ComfyUI-nunchaku

But idk if it works with video stuff

[grim totem]

I've figured it out, so in case somebody will be also looking into it:
If webui cannot find empty port in the default range 7860-7959, it means the ports are reserved/used by something else. on windows, it can be checked via powershell:
netstat -ano | findstr :7860
to check if a specific port is in use. if nothing pops up, it's vacant; if you see a PID and want to know which process it is:
tasklist /FI "PID eq xxxx" # replace xxxx with the PID
Then you can stop that process (but only if you know what you're doing) by:
taskkill /PID xxxx /F
But it might be launching on every boot anyway.
And you can see the reserved port ranges by:
netsh interface ipv4 show excludedportrange protocol=tcp
You can then safely set a different port by adding argument --port=xxxx with any bigger number in the webui.bat
The ports are often reserved by networking components like Hyper-V, WSL, Docker, NAT
In my case, everything from 1300 to 8018 was reserved. By looking into windows optional features, i figured that Windows Sandbox could be the culprit since I enabled it recently. So I disabled it, rebooted and the reserved range disappeared, and I was able to launch webui on default port again.

[vocal burrow]

scam, it's redirecting to something else most likely

[steel nest]

Hi there,

I would really appreciate some quick advice. I have no interest in making videos, so please could someone let me know which checkpoint to use for Wan txt2img, and which VAE, text encoder, Sampler/Scheduler etc.

I am running the new version of forgeNeo and have the following, but I get shit images:

Checkpoint:
wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors
VAE:
Wan2.1_VAE.safetensors
Textencoder:
umt5_xxl_fp8_e4m3fn_scaled.safetensors
Sampler/Scheduler:
DPM++ 2M with SGM Uniform.

Any help would be appreciated!

[cosmic cove]

can i ask? when you said it should be enough. You mean 1000x1000 before pressing hrfix? and ussually what do you mean hrfix? 2x resolution or 1.5? what should I assume when people talk about hrfix? thnx

[tribal anvil]

Webui has been really slow lately when I generate for the first time or switch to a different Lora any way to fix it?

Old laptop but it's never been this slow

[acoustic saddle]

I got my local install of forge working on Nobara (Linux!) Here's how to do that:

running stable diffusion webui forge on nobara linux (or fedora) if you already have a pre-existing installation created in windows:

(be warned, this process will modify parts of the environment and do things beyond my comprehension, you will not be able to generate images identical to those you had before. back up your stable diffusion instance before proceeding if you want to do that.)

starting within linux, run sudo dnf install python3.10

go to the folder your forge install is in, find webui-user.sh, open it in a text editor (like kate), change python version to 3.10,

(optionally, you can also add --xformers to the launch parameters, and any other parameters you may already be using)

for instance, mine has these two lines:

export COMMANDLINE_ARGS="--xformers"

# python3 executable
python_cmd="python3.10"```

save it

open a terminal (like konsole or gnome terminal) in the same folder as webui.sh

run `./webui.sh`

it will rebuild itself to work with linux, and that's it!

run at any time by changing directory to the /webui/ folder and launching `./webui.sh` from a terminal.

[ornate elk]

Whats your GPU and what's in your webui-user.bat?

[cosmic cove]

spent over an hour arguing with chatgtp on how to enable xformers. new old python... this that... bat file changing... nothing worked.. gave up...
4090 32gb ram, ssd, 7.5it/s at 1080p render speed with sdxl 1.0
I tried using cuda malloc command and no mem attention but same speed...
fallback disabled in nvidia
hardware sheduling disabled in windows
ive heard others have much higher speeds..
any tips?

[ornate elk]

which webui do you want to use?

normaly adding --xformers into the webui-user.bat is enough

[cosmic cove]

i tried all, and as a beginner i found a1111 forge never crashes, no problems.. that command line does nothing for me, i think even chatgtp gave up, told about a1111 having its own python and not using mine which would support gpu, im not sure what it told but it was over hour it giving me advices that did nothing to fix it lol

[ornate elk]

yea thats normal for chatgpt xD

so your using auto1111 or forge now?

and which python version?

[cosmic cove]

forge. not forge used to give oom messages, forge is just super trustworthy.
how do i copy paste text? discord prevents me to write what a1111 shows

3.106 2.3.1+cu121 gradio 4.4

[ornate elk]

mark the text and press ctrl+c and then in the chat ctrl+v

it wil maybe create a text file if the text is to long

[cosmic cove]

i think i had higher python version but gtp downgraded it in hopes of making xformers work, maybe xformers would not boost my peformance? im not sure.

ctrl c ctrl v doesnt work even for a single letter on discord ( i have web browser version) it works if i turn the "legacy" version somewhere, but legacy version bricks other discord channel commands

[ornate elk]

oh okay then paste it inside a txt file and send that

[cosmic cove]

version: f2.0.1v1.10.1-previous-669-gdfdcbab6  •  python: 3.10.6  •  torch: 2.3.1+cu121  •  xformers: N/A  •  gradio: 4.40.0

[ornate elk]

if you installed forge by downloading its recommended .zip file it does come with its own python version 3.10.6 no matter what you have installed on your PC

[cosmic cove]

dont remember how i have forge, been some days ago

[ornate elk]

the cmd log would help to figure it out xD

[cosmic cove]

when i launch forge?

[cosmic cove]

https://imgur.com/a/mHn26dG

[ornate elk]

looks normal but you didnt have --xformers --cuda-stream added to the webui-user.bat
anyway as forge doesnt get much updates anymore and you have a 4090 you should go with Forge NEO.
Its more updated and supports newer pytorch and xfrormers versions which makes it faster on your gpu than normal forge.
It can be found here:
https://github.com/Haoming02/sd-webui-forge-classic/tree/neo
For this you should have Python 3.11.9 64bit installed only

And dont download forge neo as zip, install with git

[cosmic cove]

https://imgur.com/a/mHn26dG now with added both lines like you mentioned.

second picture bottom

cool i will try force neo!

[ornate elk]

it doesnt list them

but yea forge had always issues with xformers... forge neo should work much better

[cosmic cove]

i got neo i can choose where to install, i dont want it to hide in some obscure folder

[ornate elk]

make sure you used its recommended git install command to get it:
git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo

if not you wont be able to update it

[cosmic cove]

how do i install in c:/a1111neo folder? it would be far better, or any other?

[ornate elk]

yea make a folder on C for it

then while in that folder, click in the explorer path bar (not searchbar), type cmd and hit enter

then paste the command from above and hit enter

[cosmic cove]

https://imgur.com/a/KIFMUd9 need help sir

[ornate elk]

what have you done exactly?

why is there powershell mentioned xD

you dont need to follow the uv install guide from its site, just git clone the repo, edit the wbeui-user.bat and launch it

[cosmic cove]

okay

but recommended way tells me to do it

to install uv and update it

[ornate elk]

yea but its not needed

the "normal way" without uv will work fine

[cosmic cove]

broo...

bro

25it/s from 7.5

[ornate elk]

Woah sounds good!

[cosmic cove]

maybe i could somehow figure out how to make videos now.. what speed i can hardly believe, it renders too fast 😄

nevermind,, default settings! not 1080p

[ornate elk]

Forge neo does support wan2.2, qwen image and flux kontext

Wan 2.2 can do txt2vid

[cosmic cove]

~8.5it/s so its a boost so far, no idea how to install wan or why people use qwen or flux kontext but maybe i will figure it out.. any website blog talks about these things?

[ornate elk]

I don't know any :/ as they are all pretty new

[cosmic cove]

what pc you have?

[near rampart]

RuntimeError: CUDA error: no kernel image is available for execution on the device 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.

What the problem?

[ornate elk]

I have a RX 7900xtx with an Ryzen 7800X3D

[ornate elk]

What's your GPU and please copy and paste the full cmd log when launching the webui-user.bat

[near rampart]

RT 5060

[ornate elk]

Full cmd log please and not python console. Nvm its all installed the wrong way.

Automatic1111 doesnt support the rtx 50 series nativly so you need to follow the setup guides from the first link of the pinned messages to get it working

[cosmic cove]

wanna thank you... without i would have stuck with sdxl, now i just trying flux cos with neo it just somehow got easier for me and i aint going back to sdxl.. flux soo nice.. watched long tutorial while it was downloading flux about presets and stuff and its super exciting

[ornate elk]

no problem 🙂 yea flux is great but needs a lot of vram, there are also smaller models of it out there like a Q8 gguf version.

[cosmic cove]

whats weird it doesnt crash! older a1111 used to crash a lot, a1111 forge almost never crashed but it was sdxl, now flux and i can even use discord while its running! its trully epic

[ornate elk]

thats awesome! yea auto1111 hasnt been updated since a year and Forge also haven't got much updates or support for newer models lately

will update my guides today to include the forge neo setup

[tepid crypt]

Need help to find some way to remove a character and keep the background, already tried many tools online and inpaint, but it akways generates random shit, no matter how much weight i put into background and empty, it keeps generating random stuff.

[dense tapir]

Where are Adetailer's default values located at within the "ui-config.json" file?
@ornate elk

[ornate elk]

Not in ui-config.json because its an extension

[vast crest]

Hi, sorry to bother you, but I had to format my PC, and now I have to redownload everything. I wanted to ask you which CUDAs should I download for Forge?

this is my sd forge: webui_forge_cu121_torch231

[ornate elk]

What's your GPU?

[fickle heart]

Hello guys, could someone please help me with the Krita Stable Diffusion Plugin?

I am trying to install it but getting this error. I have installed it before on my PC before formating

[rose apex]

https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html

Does this have any significance?

[ornate elk]

yea and nope, AMD is currently working on their project called TheRock which this driver is probably based on now.

Would wait for the official support before installing this driver.
Also TheRock works too without it. Im currently using it

[rose apex]

I see

[ornate elk]

But good that they are going forward

[vast crest]

rtx4070oc

[ornate elk]

would recommend trying forge neo then, as its more updated and supports the latest pytorch 2.8.0+xformers 0.0.32

or you install 2.8.0 manually into forge

[vast crest]

where do I find it?

[ornate elk]

Its on my Guide list:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[vast crest]

ok,thx

[fair oxide]

Brother this is the same issue I am having right now. Did you solve this, or can we try to solve this together?

[vast crest]

So this version of Forge is better in every way? Should I keep this one then?

[ornate elk]

Test it out or compare to the old forge and decide.
But as you have a 40series GPU and that supports sage attention and fp16, neo should be faster

[vast crest]

In fact, I generated 10 images in less than half the time it took me with the normal forge 0_0

[ornate elk]

Sounds good then 😄

[cosmic cove]

I see no "out painting" where i could drag drop image, cover it with tool and generate image around it to blend it in.. i tried many combinations with the buttons here and nothing worked so far. i probably need some extension

[tribal anvil]

GTX 970 and it's slow no matter what I use haven't tried it without a Lora tho

Should be done by now

[ornate elk]

Maybe you should try using Forge webui as its better for older gpus

[tribal anvil]

Mb that's what I meant when I said webui

[ornate elk]

Oh okay, and your using an 1.5 based model and lora?

[tribal anvil]

Idk I've just see a Lora I like and download it. How do I check?

[ornate elk]

You have to check the info box on the loras page on civitai

It shows "Base model:" and then it can be 1.5, sdxl, Illustrious etc

Same goes for checkpoints

As sdxl based stuff will take much longer than 1.5 on your GPU

[tribal anvil]

It just says illustrious

[ornate elk]

Ah okay and your also using an Illustrious checkpoint (model) with it?

[tribal anvil]

I believe so lemme check

Yep

[ornate elk]

Illustrious models are around 6GB on size but your 970 has only 4GB.
So it can't fully fit inside nor an lora too.
So it will get loaded partially and stored in RAM

So what you can do to improve the speed is to increase the windows pagefile.
Thats needed if you have 16GB or less of RAM.

[vast crest]

Hi, I had to format my PC. I was in a hurry and didn't save the upscalers. Can you tell me where I can download the good ones?

[ornate elk]

From here:
https://openmodeldb.info/

[vast crest]

thx a lot

[hoary surge]

Hello! I have suddenly developed a strange problem... when I try to generate anything I get this error:

RuntimeError: mat1 and mat2 shapes cannot be multiplied (1024x64 and 1x98304)

Any way to fix it?

[ornate elk]

hey, thats caused by incompatible model+lora

[prisma moss]

Can u get a mod to get this guy out, he's the scammer type, since he hasn't posted in this server yet, but did in another

[fair oxide]

how many scammers are in this damn server?

[vast crest]

Hi, just out of curiosity. When I start Forge Neo, it still takes a long time. I then noticed this strange message: Legacy Preprocessor init warning: Unable to install insightface automatically. Please try run pip install insightface manually.

I need to install something, right?

[ornate elk]

Insightface is made for controlnet IP-Adapter face ID or reactor faceswap.
If you dont use them or any other face swap stuff then you can ignore that.
Or if you want to install it then 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 2 commands one by one:
venv\scripts\activate
pip install insightface
Then relaunch forge neo.

[vast crest]

mmm,ok,thx

[peak sigil]

hellooo im trying to use local stable diffusion with the amd gpu fork but it seems that i am unable to do that i have searched alot but it just refuses to open i tried installing it thru github and stability matrix still not working mind you i have an rx 580 and i know it needs some tinkering to work but can anyone atleast point me to a documentation?

[ornate elk]

Hey, checkout the setup guides on the first link of the pinned messages.
There is an explicit guide for the rx580 with Zluda

[peak sigil]

Thank youuuuuuuuu

[ornate elk]

NP, let me know if it works.

[peak sigil]

Alright will do

[tribal anvil]

How do I use img2img to generate while keeping the same artsyle as the image and how can I get a prompt from it?

Tried integrate clip but the prompt it gave me was awful

[tepid crypt]

found out this tool works great https://tensor.art/template/890082907731814649

[crystal nacelle]

can someone help me instal control net: open pose to Forge UI SDXL? I tried the other one but got mat1 and mat2 errors (they are for SD1.5 not sdxl thats why). i tried this link but it instals to the wrong folder, and even moving it doesnt help because no models appear in the ControlNetIntergrated tab...

[ornate elk]

Hey, try these:
https://civitai.com/models/136070?modelVersionId=267493

[ornate elk]

Controlnet union+ promax should be really good for sdxl.
For noobai or Illustrious you need other ones.

[crystal nacelle]

i almost exclusively use illustrious models: i use stable D as an illustrator and 2D animator.

its working thank you! any advice for a dw open pose model? i love how it maps complex hands and facial expressions

[ornate elk]

Here is the Illustrious controlnet openpose model:
https://civitai.com/models/1359846/illustrious-xl-controlnet-openpose

[woeful knot]

I update comfy and dependencies and my vram requirements shot up for some reason. I could regularly make 800x1120 videos previously and now it can't even generate 768x1024 videos without spilling over... 64gb of ram and 3090, this is how it looks like mid gen

[ornate elk]

Hm thats strange. I would recommend updating your GPU driver first and then test again

[worn patrol]

What text to video platform is the best

[violet swift]

Does anyone else experience ComfyUI's venv crashing and not recognizing dependencies when trying to open a new pod in Runpod?
I've been working with storage (1 TB) for a month now, without any issues, but since Runpod crashed 2 days ago, this garbage is useless.

[fluid maple]

Anyone find Wan Animate gives slow motion even with the ComfyUI native defaults?

[rustic nova]

anyone have any experience with comfy ui / stable diffusion? I'm running into a bunch of missing nodes when I'm trying out new workflows and I'm not exactly sure how to solve the problem without breaking everything lol

[solid vector]

hi guys good morning, afternoon, or evening. im been trying to install stable diffusion and i go this error "AttributeError: module 'torch' has no attribute 'dml' "
does anyone know how to fix it?

[ornate elk]

hey, it sounds like your on AMD, whats your GPU?

[rocky ridge]

Hi im on Rocm with an AMD card and im having an issue with Hidream where the preview in Kasmapler will show that it is worker correctly but when it goes through VAE decoder in comes out compeltely black in all cases. Are there any alternative nodes or fixes? Sorry if this has been asked before, done a search cant find an exact fit as it doesnt error or anything

When i use the first inmage generation model, it works fine, the difference being theres no decoding stage

[solid vector]

hey sorry i didn't reply. a black out happen
here my gpu

[iron lantern]

with control net do i prompt to the image or just randomize it ?

[ornate elk]

Thats your CPUs igpu. Are you on a laptop? Any other dedicated GPU installed?

[solid vector]

no im on pc.
wait let me check how do find my gpu

is this it?

It said its on my task manager

[ornate elk]

any other GPU?
because thats not a dedicated gpu, which is preferable needed for AI stuff

it could look like this:

[solid vector]

nope.
so i cannot use stable diffusion?

[ornate elk]

whats your cpu model?

[ornate elk]

ah okay that explains why you dont have a gpu.
But with this you can only use the slow cpu mode

[solid vector]

ok

how?

how do i use the mode

[ornate elk]

Automatic1111 with CPU Mode:
You need to have Git 64bit and Python 3.10.11 64bit installed.

1. Make a new Folder on your Drive (not on Desktop, Downloads, Documents, Programms, Onedrive) and name it Ai for example:
   C:\Ai\
2. You go into the folder you created in this case Ai, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
3. Then you copy and paste this command:
   git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
   Press enter and after its done you can close the cmd.
4. Edit the webui-user.bat (right click, edit),
   At the line COMMANDLINE_ARGS= You add: --skip-torch-cuda-test --use-cpu all --no-half --precision full --opt-sub-quad-attention
5. After editing, save and close the bat. Then launch the webui-user.bat
   After the Installation, that can take a while is done, the Webui will autolaunch in your Default Browser.

[solid vector]

alr i'll try it
thx

[ornate elk]

Np, There is also one thing that might work:
As you have 32gb of RAM you could allocate more RAM to your iGPU as VRAM. Thats done in BIOS.
If you give it like 8gb you could also use the DirectML Version of the setup instead of CPU.
But the best would be getting a dedicated gpu in the future.

[solid vector]

oh ok thx fo the info

[forest aurora]

Good day, people. Looking for some help with "cg-use-everywhere" in regional prompting as I am trying to have a different set of lora applied to a different inpainted region (e.g. Region 1 and Region 2 groups). Currently, most parts should be connected correctly, yet I am having an issue with finding a way to connect the MODEL output from " CR Apply LoRA Stack" to have it received by the KSampler model input as it's currently using the model output from "Attention couple" in Combine regions group to the KSampler model input. I have a feeling that I am trying to over-complicate it and there's a more simple means to have this archived, yet my brain isn't there rn looking for any suggestions or workarounds.

[solid vector]

hi guys good morning, afternoon, or evening, i got this error anyone know what does it mean?

nope i followed what CS1o said

and i got this error

alr

[solid vector]

anyone know how to fix it?

the support sever i joined earlier suddenly disappeared after i made a ticket and explain the issue

[ornate elk]

it seems like it somehow broke as it tried to get some stuff from XL.
you need to probably git clone the normal repo again and only adjust the webui-user.bat then relaunch

[ornate elk]

the "suport server" are scammers that want money and personal data, report them to us and discord if you see them and block

[solid vector]

git clone the normal repo?
how

[solid vector]

oh ok

[ornate elk]

1. Make a new Folder on an internal Drive (not on Desktop, Downloads, Documents, Programms, Onedrive) and name it AI for example: C:\AI\
2. You go into the folder you created in this case AI, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
3. Then you copy and paste this command:
   git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
   Press enter and after its done you can close the cmd.

but delete the other webui first

[solid vector]

do i have to delete the existing

oh

ok

[solid vector]

do i have to put the same commandline_args =?

[solid vector]

omg it worked i came back and im on the site thx for helping!!!

[ornate elk]

Perfect np

[crystal nacelle]

Hello, my OpenPose was working for like a day but now i get an error when i try run it on my ZLuda Forge SDXL or Illustrious models

[timber lodge]

Please, help me. WHY???

[crystal nacelle]

that looks like the same issue im having.

[ornate elk]

Can you show a complete cmd log?

[ornate elk]

Which preprocessor did you have selected?

[dusky osprey]

Hi! I'm kinda new to SD and read that Zluda was good with AMD gpus (I've a 9060XT) but when I tried installing it I got this error after trying to run the .bat

I was following this steps: https://github.com/patientx/ComfyUI-Zluda?tab=readme-ov-file

[ornate elk]

hey, probably the antivirus flagges the zluda files

its a common issue with that repo

[dusky osprey]

Thanks. I'll try reinstalling everything with the antivirus turned off

[ornate elk]

it will probably flag the files as soon as you enable it again, so better set the .zluda folder inside comfyui to the whitelist or ignore list but not the whole comfyui folder

[dusky osprey]

Gotcha. Should I set an exception for the zluda folder and then reinstall it?

Good and bad news. I got a new error

looks like torch is having some troubles, not sure why 🤔 , maybe I've a wrong version of something.
I just formated my pc today btw, I installed everything from 0 so I'm not sure that I've an "older version" or something like that, but I'm no experto

[ornate elk]

you have the wrong python version

python 3.11.9 is required

[dusky osprey]

Oh, ok

I'll replace the versions realquick

should I reinstall everything or just restart pc

[ornate elk]

only delete the venv folder

no pc restart required

[dusky osprey]

Gotcha

[dusky osprey]

I'm having troubles with the HIP. I tried installing HIP 6.4.2 but im getting this message

I'll try reinstalling it rn. Lmk if im doing something wrong

[ornate elk]

have you added hip to path?

[dusky osprey]

no I didn't

I'll check rn how to do it

[dusky osprey]

I'm having troubles with this step

https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/tag/v0.6.2.4

If i've a 9060xt what should I use¿

I managed to add the path btw!

[ornate elk]

you dont need custo mfiles

hip sdk 6.4 has native rx 90xx support

[dusky osprey]

Oh. Gotcha

What should I add to path then? I was doing the 6.2 way but looks like that's for older gpus. I'll try removing that and going with the 6.4 again

[ornate elk]

yea uninstall everything of 6.2

[dusky osprey]

on it

What should I add to path for the 6.4? I added C:\Program Files\AMD\ROCm\6.4\bin but idk if that's correct

[ornate elk]

thats correct

[dusky osprey]

Ok

Im getting this error now

[ornate elk]

is your driver updated?

Adrenalin 25.9.1 ?

[dusky osprey]

yessir

[ornate elk]

oh an you have to run the patchzluda2.bat , and dont run the patchzluda.bat

inside the comfyui-zluda folder

thats needed when switching to newer hip sdk versions

[dusky osprey]

You're 100% right. That was mb for changing the version

is asking me for a URL. Post says i gotta get them from here: https://github.com/lshqqytiger/ZLUDA/releases
Should i use the newer? 3.9.5

[ornate elk]

if you used the -n .bat files you need the nightly version

if not then the normal ones

[dusky osprey]

Im using this url: https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d23391f0fad66c7459cf/ZLUDA-windows-rocm6-amd64.zip
And is giving me this error:

I tried using the N version too but no luck

Should I re-install from scratch? Takes me like 6 minutes

Looks like I didn't fully uninstalled the hip 6.2

I'll see how to do it properly

[ornate elk]

Uninstall every module of hip SDK 6.2 then reboot the PC

[dusky osprey]

On it

Done. Trying again

It gave me the same error. Im reinstalling the whole thing

I removed every module of hip 6.2 and rebooted. I'll try to stick with the normal bat file (not the N one)

It worked!

Thank you so much mate. You're the goat

[ornate elk]

Perfect! No problem 🙂

[rose apex]

Comfyui now has native support for AMD GPUs and patientx says the performance is similar to his comfyui zluda fork. Would you try it or wait for a later version?

[rose apex]

Using PYTORCH_TUNABLEOP_ENABLED=1, the first 512x512 took 37s for the default prompt but the subsequent ones only took 3s avg. Then I upped the resolution to 1024x768 (16 words prompt). First one took like 11 mins and the subsequent ones took only 12s.

[ornate elk]

I already tried it yesterday,
Speed test:
SDXL 1024x1024 on a 7900xtx OC/UV on Adrenalin 25.9.1 on Windows:
Comfyui without changes with rocm 6.4.4 and torch-2.8.0 gets 3.6 it/s (using sub quadratic attention as pytorch attention doesnt affect it)
TheRock and ROCm Nightly with rocm 7.9.0 and torch 2.10.0 gets 4.6 it/s (using pytorch cross attention)
Haven't tested Tuneableop but set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 is needed for nightly and pytorch cross attention.

[feral dust]

hi, friends. nice to meet you
I have a new idea to collaborate with you that could be profitable.
how about it?

[ornate elk]

Wrong channel

[feral dust]

what do u mean?

[ornate elk]

Read rule 3 and 5 of #✍🏼｜rules-and-tos

[crystal nacelle]

okay. first i used the explosion and i t read the control net and showed me the result:

but when i tried to generate an image with it i got the same error:

[crystal nacelle]

i managed to get this model working as an alternative tho https://civitai.com/models/135552/openpose-sdxl

takes fucker long though, maybe thats just the price i have to pay

[ornate elk]

You should use controlnet union++ promax

That should work and fast

[crystal nacelle]

yes, i have been to this link before but im not sure how to download/install it. i think i pasted the link in webui extensions url installer and it downloaded the files to E:\SD-Zluda\stable-diffusion-webui-amdgpu-forge\extensions\controlnet-union-sdxl-1.0 but the models don#'t appear on preprocessor>models

[ornate elk]

Dont install it as extension

Just download the model and put it into models/controlnet

https://civitai.com/models/136070?modelVersionId=655836

[royal helm]

Hello good fellows!
Wonderful to be here.
I have managed to not only install automatic1111 on fedora with al lthe broken instructions out there but I also got rocm working and everything's great.

I just can't generate anything at 1024x1024 with sdxl -.-
Running into torch.OutOfMemoryError: HIP out of memory. Tried to allocate 9.00 GiB. GPU 0 has a total capacity of 11.98 GiB of which 8.04 GiB is free. Of the allocated memory 2.60 GiB is allocated by PyTorch, and 894.80 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

Now I am on an older RX6700XT witth 12gb vram so it is pushing the limit but surely I can find that 1gb it's SO CLOSE or hand off some of the processing somewhere to get this to work?

already using --medvram and --medvram-sdxl

Processor is a 5700X3D not really a workhorse for this kind of thing.

Much love out there guys

[vocal burrow]

Which version did you install ? What command line args are you using ?

[royal helm]

standard automatic 1111 webui via python install
sdxl1.0
export COMMANDLINE_ARGS="--no-half-vae --skip-torch-cuda-test --medvram --medvram-sdxl"

smaller models work fine

[vocal burrow]

Is it even using your gpu ? Can you you check its usage during a generation with rocm-smi ?

[royal helm]

Any setting that would reduce that vram use by like 1gb or let it spill into the 64gb normal ram (ik that's not how it works and it's wildyly frustrating 🤣 )would be great

[royal helm]

yes it is defionitely using my gpu I did all of that earlier

Went from 45s/it on my processor to 4.5its/s on my gpu it's definitely working

[vocal burrow]

probably is indeed

[royal helm]

the py script calls the HSA_OVERRIDE now for my gpu class which is quite nice

I know I need a new gpu but I thought I could spit at least an image or two out of sdxl with 12gb vram.
With the 0.04GiB and the 894.8MiB that's so damn close to the 1GiB it needs
It's like smelling the finish line

[vocal burrow]

You not gonna like it... But I would not recommend using A1111 nowadays. It hasn't seen any updates for over a year. Forge is more updated but also on its way out because the dev is quite busy. But still it would probably do better regarding memory management than A1111.
Also Forge Classic or Forge Neo are probably the best ones out there for now but I haven't fully tried them yet.
ComfyUI too but that s a whole another beast.

Check the pinned comment for install guides on those.

[royal helm]

bags

[vocal burrow]

They re windows based guides but there s still some knowledge to apply there for linux stuff too.

Don t bother with zluda if you re using linux

run things using rocm directly.

[royal helm]

no way to force this mfer to use less vram?

by like a factor of 10%

[vocal burrow]

not without running much much slower

[royal helm]

bags

[vocal burrow]

you can try using --opt-sub-quad-attention but I would nt hold my hopes

[royal helm]

i need to free up a total of 0.1GiB

[vocal burrow]

until you add a few words or lora and then it will be 0,5gb or whatever :p

not sure what resolution you re using too but you could try lowering it.

[royal helm]

1024*1024

the kind of basic sdxl size I thought?

[vocal burrow]

yes, you can lower/increase it by 30% without too many issues

[nova loom]

Hi people. Has anyone experienced prompts being completely ignored when running Qwen-Image-Edit locally?

I'm using the GGUF models in ComfyUI and trying img2img. E.g., with a prompt like "add glasses to her face" I only get unrelated, slight changes to the image: a bit more saturation, eyes pointing elsewhere, etc.

I'm using these models:
Qwen-Image-Edit-2509-Q3_K_S.gguf on "Unet Loader (GGUF)" node
Qwen_Image-VAE.safetensors on regular "Load VAE" node
And these models on "CLIPLoader (GGUF)" node
mmproj seem to be correctly associated:

Using mmproj 'Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf' for text encoder 'Qwen2.5-VL-7B-Instruct-Q3_K_S.gguf'.

Tried with and without Qwen-Image-Lightning-4steps LoRA.

Btw, txt2img does align with the prompt, the problem is only img2img 🤔

Appreciate if anyone can shed a light 🙏

SOLVED:
I was using the wrong "Lightning-4steps". There are specific ones for Image vs Image-Edit vs Image-Edit-2509.
Now my output image is a bit messy but that's a different problem.

SOLVED:
Upgrading models from Q3_K_S to Q4_K_M fixed the messy output.

[royal helm]

took your advice. ComfyUI would work if it accepted the HSA_OVERRIDE argument as easily as A1111 because it uses neag 2GiB less VRAM. :/

Different problems are not the same problems it's all progress

VAE load device: cuda:0, offload device: cpu

would be nice if Comfy just used the gpu but that would be too easy

[vocal burrow]

comfyUI should be using your gpu by default

And you can just set HSA_OVERRIDE before running comfyUI, it s a system variable, not a a1111 argument.

So something like HSA_OVERRIDE=xxxx python main.py should work.

[dawn swallow]

automatic1111 anyone know how to get all old / new smaplers i wanna have them all to run test anyone know how to restore them

[vocal burrow]

easiest way is probably to just revert to an old version if you just want to run some tests

[dawn swallow]

dang nabbit thanks well if anyone knows how to get all versions that would be very nice. Thanks for the recommendation tho will help me out a bunch

[vast crest]

Hi, in webui-user for sd neo, are there any strings to make it faster at startup?

[ornate elk]

Is it installed on an SSD?

[vast crest]

yes,Corsair Mp600 Elite 1tb M.2

but I noticed that it was always slow to start up honestly

Even photo generation doesn't seem very fast. The PC was formatted only 2 weeks ago. I also updated to the new Windows 25h2 update.

[ornate elk]

Latest drivers installed?

Correct launch args used?

Also add git pull to the webui-user.bat to auto update it

[vast crest]

where should I write: git pull?

@echo off

:: set PYTHON=
:: set GIT=
:: set VENV_DIR=

set COMMANDLINE_ARGS= --cuda-stream --sage --fast-fp16 --xformers

:: --xformers --sage --uv
:: --pin-shared-memory --cuda-malloc --cuda-stream
:: --skip-python-version-check --skip-torch-cuda-test --skip-version-check --skip-prepare-environment --skip-install

call webui.bat

[ornate elk]

Add it below the @echo off line

The rest looks good

[vast crest]

like this?

@echo off

git pull

[ornate elk]

Yes

[vast crest]

ok,thx

What about faster image generation? Maybe it's just my impression that it's a bit slow XD

[ornate elk]

What was your GPU and how long does a illustrious 832x1216 30 step Euler a image takes ?

[vast crest]

7,9 sec

is ok?

[narrow sable]

hi everyone,
i'm looking for help using my character LoRA to create t2v animations with SD 1.5 and AnimateDiff,
the results i got so far are mediocre at best. t2i works fine using my LoRA(s)
i feel like its a training issue and i can share examples
hope this is the right place to ask

[dusky osprey]

Hi! I'm trying to use zluda with an AMD card (9060XT 16vram) and 16RAM. I can generate images with checkpoints from civitai and the standar SD ones but flux seems to hate me or something.
I tried reducing the ram to --lowvram but when trying to create any image with flux (txt2img or inpaint) it just freezes my pc and I gotta restart. I don't care if it takes longer to generate but idk why it crashes? The SD with 512x512 generates images in just like 10 secs. Is it that bad with flux?
I took the pic with my phone for obvious reasons

[hard cedar]

Hello I am experiencing difficulties with my Stable Diffusion installation but my english is very bad. Does someone here speak german?

[sharp shore]

Hello. I'm pretty new to this. I followed a tutorial only instead of the 3.5L I used 3.5M from here

When I try to load in the workflow, SD3.5M_example_workflow.json, I get this error.

How do I resolve this? Where can I get this SkipLayerGuidanceSD3 from?

I believe I solved my own issue. I'll leave the question for anyone else, and I will update as to whether this fixes it or not, but I'm updating ComfyUI and python dependencies now...

Well, running that made it so pytorch isn't working, because it "wasn't compiled with CUDA enabled". Probably worth pointing out haha

The update did resolve this issue.

[ornate elk]

Hey, make sure you dont use the 23GB large flux model. It doesnt fit in your vram and will freeze everything.
Use a smaller gguf variant like Q8 gguf

[ornate elk]

Hey, yes I do, feel free to DM me ^^

[upbeat trail]

Can anyone recommend me any easy to install all ai platform?

The only one I know is Pinokio

[ornate elk]

i dont recommend them but there is also stability matrix, but its only for image gen and traning webuis

[timber oak]

Heya my discord got hacked but im back now

[timber oak]

hey CS1o i helped him install comfy and stability matrix set up api key for civitai and redirect the extrafolderpaths.yaml

but then he got a cuda error

he got a 1070 8gb which should be fine

like his cuda cores couldnt be accessed

the installation went good

i mean we used comfy ui portable straight from the github

[timber oak]

yes i love stability matrix but for some reason the search function is still bugged as hell

i feel like only very few models get displayed on search

the previews get bugged and need a long time to get redirected

and then sometimes when i reach the end of the search results and load more models it loads a lot more and mixes them with the models i previously looked at

its extremely comfortable for browsing models and loras

[ornate elk]

he should check if he has the latest drivers from nvidia

[timber oak]

thx

ill tell him to update his drivers when i see him

[ornate elk]

but with a 1070 comfyui is no fun, i would suggest using forge or forge neo instead

[timber oak]

is forge like auto1111?

i see

[ornate elk]

more updated and some more features, better optimised on low vram or older gpus

[timber oak]

back then i started with my 1060

and sd1

damn it was slow and the images were bad

but when i swapped to comfy and the gen time got reduced to half the time

[timber oak]

if he installs forge

does he need the --lowvram stuff?

i forgot how to do it

thats long gone in my memory

[ornate elk]

nope not needed

[timber oak]

epic

also idk if this is the correct chat for it

but atm i am stuck at pony models

and maybe some flux

is there any big upgrade? like sd1 -> sdxl, sdxl -> pony?

[ornate elk]

whats your gpu currently?
Illustrious based models are currently the best for Anime stuff, newer training data on booru tags and no score tags needed.
And there are also semi and realism models of it

[timber oak]

i have a 4070