🤝｜tech-support

[hybrid shore]

Why does the IPAdapter preprocessor I downloaded not show up in ForgeUI? Trying to use it with SDXL models and the 3 default preprocessors aren't working.

[ornate elk]

It goes into models/controlnet not preprocessors

Oh nice, your the first user with the new AMD GPU on Zluda.
Which if the guides did you followed?
Any other issues other than the need of --no-half
How's your speed?

[ornate elk]

Hey, I think the normal koboldcpp.exe should do it

[fair oxide]

lol i switched to LM studios sorryyyy

because i heard kob is better on cpu

while i have 24gb gpu

[ornate elk]

Ah I'm also on the XTX and using LM studio and ollama

[fair oxide]

should i consider ollama?

all of this is just to use it in sillytavern

lol

[ornate elk]

Its really fast but has no gui

[craggy orchid]

}

const initImageBase64 = 'data:image/png;base64,' + imageFile.buffer.toString('base64');

const payload = {
  init_image: initImageBase64,
  prompt: prompt,
  width: 512,
  height: 512,
  steps: 30,
  cfg_scale: 7.0,
  model: "flux", // Specify the model
  controlnet: {
    enabled: true, // Enable ControlNet
    control_type: "canny", // Example control type
  },
};

just paid for api to write an web app to make photo with controlnet, but why it keep saying wrong api

}
}
Stability AI API Error: 404 Not Found
Stability AI API Error: 404 Not Found

help me

[fair oxide]

doesnt matter if it has gui or no since i will just use its api and shove it in sillytavern

[craggy orchid]

https://api.stability.ai/v2beta/generation/flux/img2img
the endpoint im using

[fair oxide]

there is no way its better than SL

right?

[ornate elk]

Sillytavern is just a UI

It can use LM studio and ollama as backend

[fair oxide]

am stupid i thought wendy said backend

let me just say this

i have no clue what ai is i just jumped this up yesterday

got sl direct ml, git, python, wsl,

so many stuff

lm studio

kobold

oobabooga

my pc has a headache

[ornate elk]

Yea true it can be confusing

[fair oxide]

this is the first time i saw this pc lags maybe cuz of the llama model

running currently

[ornate elk]

Dont use larger models than your vram can fit

Mostly 14b is enough

[fair oxide]

but i have 24vram

needs more for bigger models?

also dude i saw a 400B llama model like WTF IS THAT i think it was 281gb

you need one of these nvidia 100 series to run that stuff

[vocal burrow]

yup you figured it out

[fair oxide]

i was wondering what are these for... now i know

[fair oxide]

can i chop its legs a little so it doesnt consume as much as vram so maybe it would stop stuttering

fuck i knew i shouldve got the 5090

wasnt avalible here

[vocal burrow]

it would not help, on the contrary

[fair oxide]

Interesting...

https://youtu.be/bHqJ6mo1P-4?si=buHgNeyDk09b9NKx

am trying to replicate the end result of this video

its like a whole Novel game you make yourself

[formal ridge]

That sounds interesting. Let us know how it works out if you get it working.

[fair oxide]

will do

youre using a nvidia should be easier for you

ngreedia

sry had to

i hate them

[formal ridge]

Only 8GB of VRAM, all I could afford. Q_Q

[fair oxide]

still they excel in AI stuff more than AMDs

[shadow path]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-automatic1111-with-zluda --use-zluda --update-check --skip-ort --no-half

Slow on the base model tbh, about 2 mins to run one default generation

[lofty cargo]

Anyone else has run into a "could not allocate memory" issue when trying to install insightface? For reference, I am doing it using conda install in a docker build

[lucid yoke]

With 8gb you can still run a llm but like a 7-8b model

It works for short rp but hmm its not ideal or if you dont mind complete amnesia from the model after a short while

[ornate elk]

2 minutes? With 1.5 or sdxl models?

[zenith loom]

*** Error loading script: openvino_accelerate.py
Traceback (most recent call last):
File "C:\Users\proje\stable-diffusion-webui\modules\scripts.py", line 382, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\Users\proje\stable-diffusion-webui\modules\script_loading.py", line 10, in load_module
module_spec.loader.exec_module(module)
File "<frozen importlib.bootstrap_external>", line 883, in exec_module
File "<frozen importlib.bootstrap>", line 241, in call_with_frames_removed
File "C:\Users\proje\stable-diffusion-webui\scripts\openvino_accelerate.py", line 47, in <module>
from diffusers import (
File "C:\Users\proje\stable-diffusion-webui\venv\lib\site-packages\diffusers_init.py", line 5, in <module>
from .utils import (
File "C:\Users\proje\stable-diffusion-webui\venv\lib\site-packages\diffusers\utils_init.py", line 38, in <module>
from .dynamic_modules_utils import get_class_from_dynamic_module
File "C:\Users\proje\stable-diffusion-webui\venv\lib\site-packages\diffusers\utils\dynamic_modules_utils.py", line 28, in <module>
from huggingface_hub import HfFolder, cached_download, hf_hub_download, model_info
ImportError: cannot import name 'cached_download' from 'huggingface_hub' (C:\Users\proje\stable-diffusion-webui\venv\lib\site-packages\huggingface_hub_init.py)

[zenith loom]

help me guys, cannot accelerate with openvino

[shadow path]

1.5

horrificly slow tbh hah

might have to try a linux install

the first generation took as long as these are taking now tho interestingly, no 15-20 min wait

[formal ridge]

Sounds like Re:Zero tbh.

[lofty cargo]

my dmesg output https://bpa.st/ARXQ

[formal ridge]

A1111? That's always going to be slow. If you can, try Forge/Reforge or ComfyUI.

[ornate elk]

Can you show me the full cmd log?

[ornate elk]

Which webui did you tried ?

[shadow path]

A1111

[ornate elk]

The question was not for you xD

[shadow path]

ah sry haha

[ornate elk]

But the one with the full cmd log

[zenith loom]

yea A1111

i kind of noob with the code, but i kind a need help on this haha

[ornate elk]

Basic Auto1111 won't work with openvino.
Your best bet is to try SD next:
https://vladmandic.github.io/sdnext-docs/OpenVINO/

[shadow path]

https://bpa.st/SUKQ

[zenith loom]

alright, i try this first, thanks ya

[ornate elk]

It says no rocm agent found. So its not using Zluda at all

Make sense because AMD HIP SDK 6.2.4 was released before the 9070 launch.
So its missing the GFX1201 library files.

You would have to use Linux or WSL2 on Windows to get native ROCm support

[shadow path]

ah I assumed because it listed the version in the next line that it had found it, alright!

WSL2 it is then

[ornate elk]

Thats the rocm Version that got detected

Not the GPU gfx version

[shadow path]

noted, thanks dude! is there a WSL2 guide pinned?

tbh I did notice that the CPU was spiking which was suspect

alright I'll give it a bash tonight

[vocal burrow]

there s no real point going for a wsl2 install tbh

[ornate elk]

You can also try Directml for the webui but its slower and uses more vram

[velvet spoke]

Righty retrying here

Is there a way to convert/merge a sd1.5 model to/with an sdXL? I have a great model i made with great lain but would like to upgrade it with the data of an sdXL model

[vocal burrow]

nope

different architecture you cant merge them.

[velvet spoke]

Awwww welp can't help it if its incompatible

[soft nest]

hey guys, is the captioning style important for SDXL for creating loras? should I keep the captions like:
"A man sitting on a bench with black hair and blue eyes wearing brown suit with black tie"
instead of: "A man sitting on a bench, black hair, blue eyes wearing brown suit with black tie"

Does the ai want captioning in a particular away?

[ornate elk]

Hey, you can use either.
It depends on which model base you train on.
Normal sdxl models like captions in sentences and anime based models like illustrious or pony like tag prompting more

[soft nest]

im using normal sdxl model

my prompts looks more like: "A man sitting on a bench, black hair, blue eyes wearing brown suit with black tie" so, would that affect my generation ?

mine is sorta mixed and I don't want my lora give bad generation because of it

[ornate elk]

Shouldn't make to much of a difference

But I'm not sure

[hazy pumice]

How do I fix SwarmUI interface when it says "Backends are still loading on the server" highlighted in yellow? None of my prompts will generate and it's stuck.

[ornate elk]

Try relaunch it.

[hazy pumice]

Yes I have 😦

[shadow path]

hmmm still not managing to get my 9070XT listed as the rocm agent in WSL2 following this guide https://rocm.docs.amd.com/projects/radeon/en/latest/docs/install/wsl/install-radeon.html#post-install-verification-check

Failing to understand how people managed to even get this running for benchmarks they did for reviews

Guess I'll be waiting for official rocm support

[ornate elk]

Yea WSL + docker is not that easy too

If someone ports the GFX rocm library to windows you will be able to use ZLUDA

[shadow path]

I'm not seeing the rocm library available on linux

might give DirectML a shot

[cold echo]

Hi! I asked about LoRA training for hairstyles the last time, using KoyhaSS

Can someone help with that?

Do you think Textual inversion or dreambooth is worth a shot instead of it?

Also, I need help setting up comfy workflows to use LoRAs. Just starting out with comfy btw, is there a good tutorial or blog post for it
PS: keep ping on with reply

[leaden pivot]

Given how trash the NVidia 50XX launch was, I'm curious if it's a pain in the ass to run SD on an AMD card, and how the performance is. Anyone have insight?

[lucid yoke]

your gonna run into some jank but theres guides in the pinned messages

mostly zluda/direct ML eating up vram

so you have less useable iirc

[leaden pivot]

ah gotcha. thanks!

[whole prairie]

Hey guys I'm traveling right now for a couple of months so I don't really have access to a competent PC, are there any cloud GPU services you would recommend?

[tawny surge]

Hiyo, is anybody here experienced using Controlnet on SDXL in forgeUI because i cannot get the model to adhere to any of my prompts

Any help appriciated

[shadow path]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-comfyui-with-directml was able to get my 9070XT working with DirectML and ComfyUI

16 secs for default gen

[ornate elk]

try use the controlnet++ union promax sdxl model, that should work.
https://civitai.com/models/136070?modelVersionId=655836

[ornate elk]

ah nice!

[brittle lantern]

Has anyone got the new 9070 XT and used Zluda or ROCm on it if so how much faster is it from the XTX?

Or if there’s any tests already done on it

Please let me know

[tawny surge]

Sadly seem to be getting similar issues with the other, i've tried with different models, and get same issues? am i using it wrong?

[ornate elk]

try input a normal image

to test

[tawny surge]

as in using a high quality realistic image for my control net?

[violet swift]

Can someone explain to me why many people can run Wan 2.1 and Hunyuan with up to 4GB of VRAM, but I can't run any of them with an RTX 4060 with 8GB VRAM?

I've used workflows that are supposed to focus on the VRAM I have. I've even used the lightest GGUF programs like Q3, and nothing. I don't know what to do.

I get an out of memory error

[ornate elk]

just as test and check the cmd log

[tawny surge]

The CMD log is empty other than the progress bar being at a 100%

[ornate elk]

is your forge updated?

[tawny surge]

yep, i'll do a restart just for sure

[tawny surge]

As a last ditch effort, here is the pastebin of me rebooting, the UI, and running inference.

https://pastebin.com/J7gNgDCE

[ornate elk]

Ah your using python 3.12

But thats not supported in Forge

You need 3.10.16 on linux

[noble idol]

Wait wait wait

Why don't I have that installed

It says ++ its gotta be better right? Should I have been using that CS?

[ornate elk]

Yea its really good for sdxl stuff as it can be used for multiple modes

[noble idol]

Oh my goooooddd

[ornate elk]

Openpose, canny, depth in one model

[noble idol]

Does it run better

Like more optimized I guess?

Like how firge runs better than a1111

[ornate elk]

It should run normal or the same as other sdxl controlnet models

[hollow topaz]

Guys I have 2 5090s I can use for stable diffusion but idk much about prompting and stuff

Do you guys think this would be hard to do? (The comics in the video)

https://youtube.com/shorts/Bn4RO_wb57E?si=synQlzAB28IWHevr

[ornate elk]

Scam

[hollow topaz]

Got it, thank you

[ornate elk]

Dont join der scam server

[hollow topaz]

Oh wtf

[ornate elk]

Its a scam bot and fake support

[rapid flicker]

So my rtx5080 that i thought i was going to be waiting months on showed up this morning. I installed it but i was having issues with pytorch. I followed the link suggested by the error and ran the command it said to

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

but now i get
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.

[ornate elk]

Removed him

[hollow topaz]

Do u know where I can get some help/advice on stable diffusion stuff?

[ornate elk]

For installation and technically stuff your right here.
For generating images and prompting questions there is #📝｜prompting-help

Maybe also checkout some YouTube videos. No one will give you a full tour from 0 knowledge

[hollow topaz]

Got you, thanks. That’s fair

[ornate elk]

The 50 series card need some extra steps to get running, I think @lucid yoke can help you there

[lucid yoke]

I can! But im currently in bed

Will be on tomrrow 5pm due work

Lemme see if i have a link somewhere

[hollow topaz]

All good, I appreciate it! I actually haven’t built the PC yet lol. Trying to get more info on stable diffusion

[lucid yoke]

Ahh its gonna be soon though

[rapid flicker]

ah it might help me with my issue if you have that link though

[lucid yoke]

Since i boot i get like a dozen errors but still works, cant get text to speech working etc

[rapid flicker]

I need to restart my computer so I get the temperature readings. I only just installed the new 5080

But I have the best gif in my case

Priorities

[ornate elk]

Awesome haha!

[ornate elk]

If you have a ryzen CPU don't forget to select the right GPU in the NZXT cam software

Because it defaults to the Igpu

[crude trellis]

when setting up zleuda, does only amd hip sdk 5.7 version works or can I download most recent versions?

[ornate elk]

Depends on your GPU

[crude trellis]

rx 6750xt

[ornate elk]

If you follow my install guide it lists everything.
Its the first link from the pinned messages.

[ornate elk]

You can use 6.2 then

[ornate elk]

Make sure your on the latest nvidia driver.
And which webui did you tried to install?

[hollow topaz]

Anyone here using a Mac?

[ornate elk]

In the pinned messages there is an install script for MAC

For auto1111 and Forge

But here are mostly windows users

[rapid flicker]

Ah yes the drivers were installed not long ago which is why I said I needed a restart. I had the forge one for nvidia from the pinned message. I had the GTX 1080 before

[hollow topaz]

Got it, was just wondering

[rapid flicker]

I'm away from the computer now, it's late and I'm trying to fix my sleep pattern

I'll fail but I can try

[ornate elk]

I'm off now too gn

[azure magnet]

@ornate elk do you know someone that could help me with comfy?

[obsidian magnet]

Hello! I have been using Stable Diffusion for about 2 years now.
Recently i have messed with other AI stuff, and after updating lots of things, my Easy Diffusion image generation went even faster. The generation speed almost doubled, even though i already had CUDA enabled, and the wait between 2 pics went super fast.

Now i did the worst thing you could possibly do... I updated Windows... Now my GPU uses sometimes 13/12 GB vram, the generation is still quick as before, but the time between two pictures randomly becomes extremely long, and after a few pictures it just hangs. It also takes a long time and it can hang when i cancel a task. Why would it even take time to cancel?

I tried updating Easy Diffusion, but no matter what version i select in the settings, it doesn't really work well anymore. It feels like i am generating images that are too large, but even when i make the resolution tiny, it is still insane. I think when i made some smaller 786x786 pictures my VRAM usage used to be around 8-9 GB, now the vram is full with the exact same setting, AND using 6-7 GB RAM.

Has anyone experienced anything like this and has a solution or maybe even an idea how i could solve this?

I found a few things, like disabling the CUDA Memory thing in NVIDIA Control Panel, but that didn't seem to change anything, except that SD ran out of memory once and had to be restarted... The only thing i found is that many people complain about having this kind of problem after updating things, but none have a solution.

It also seems weird that in the task manager, after the image reached 100% or has been canceled, the python process has 12% CPU, but 0% GPU usage. Nothing has any GPU usage. And yet, in the other tab of the task manager, and in Afterburner, it shows around 99% GPU usage.

[next pilot]

@obsidian magnet

[obsidian magnet]

The ticket bot is rather useless. It basically says that it's a network issue, even though the exact same thing happens when my cable is pulled and it doesn't even tell me if a ticket has been created, or if i have to keep talking to it until it tells me that it's done... I am not posting logs from my PC. The console doesn't give a single error and it's basically a successful fresh installation.

[hybrid shore]

When I use the model it still doesn't copy anything from the control. I'm trying to get the clothing to transfer over an existing image that I managed to get into the right pose.

[broken goblet]

Soo I installed ComfyUI from the website instead of github, I need to use the "--normalvram --fp16-vae" commands but there is no "run_nvidia_gpu.bat" file, where can I put them?

[ruby yoke]

@pure tusk

[rapid flicker]

Bad bot

[ornate elk]

What do you mean from the website?
Did you download it as zip or did you git cloned it?

[ornate elk]

We have a #🧣｜comfy-ui channel where you can ask.
If its technical like setting it up, ask here.

[broken goblet]

I downloaded it as an .exe

[ornate elk]

From where?

[broken goblet]

https://www.comfy.org/download

[ornate elk]

Ah okay

I haven't used this exe installer so I idk if they added a launch bat or not

But you may need to use the normal github installation method for the args

[ornate elk]

What's your GPU?

[sacred ermine]

Does anyone have suggestions on how I can work on face swapping, specifically for swapping portraits? I tried looking for extensions and tools, but most options seem to be available only for Windows. I'm using a MacBook Pro with an M4 Pro.

[ornate elk]

Hey, the best tool for that is FaceFusion.
It also has a Mac installer:
https://github.com/facefusion/facefusion

[sacred ermine]

Thank you so much, that was a really good suggestion!

[ornate elk]

Np 🙂

[obsidian magnet]

Oh... I told the bot, but then didn't add the info to my post... GeForce RTX 4070. I could always easily make 100 pictures in a row, as long as there was enough VRAM for the first picture. Now it feels like every time 100% is reached, or i hit cancel, there is a random chance for it to hang. Also, before it took about 1 second between two pictures, now it constantly hangs for 10-30 seconds between 2 pictures of the same task.

From what i can tell by googling, it seems that something withing stable diffusion (all clients, Easy diffusion and 1111 at least) has been updated, and been causing trouble ever since.
Here are the versions of the things that are being used. I read about chaning around some of the versions, but the recommended versions always change, and sometimes are the version that i have installed anyways, so i have no idea what to even try changing.

git version 2.40.1.windows.1
conda 22.11.1
Python 3.8.5
torchruntime: 1.15.1
torch: 1.13.1+cu116
torchvision: 0.14.1+cu116
sdkit: 2.0.15.17
stable-diffusion-sdkit: 2.1.5
diffusers: 0.21.4
Easy Diffusion v3.0.9c

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/9082 seems to be about my problem. If it's not, it sounds very similar. Lots of people have the problem and no solutions.

[ornate elk]

Seems like a normal vram problem. The VAE step at the end requires a lot of resources.
You can lower that by using the tiled VAE extension for Auto1111 or by using the sdxl fp16 vae.
Also if your on Auto1111 try using xformers. It helps in lowering the vram usage.

[obsidian magnet]

i used to make 786x1024 with no problem, now 512x512 is impossible with the same settings. Also i am using Easy Diffusion

Also the VRAM problem *was extremely consistent. When i used too much, i would have to wait 30 seconds after every image. But THIS is completely random, until it hangs completely.

[ornate elk]

Do you use Wallpaper engine by any chance?

Its known to cause issues while generating

Also is your GPU driver updated?

[obsidian magnet]

No, not using anything besides Win 10, Afterburner, and nothing else open. Even Pulling the cable and shutting down firewalls makes zero difference.

Yeah, i reinstalled them after windows updated them even

thought windows might have been... windows, and screwed up

[ornate elk]

You can open up a cmd as admin and run
sfc /scannow
That scans and fixes bugs that occur after Windows updates

[obsidian magnet]

A bit clearer... Windows updated stuff, including GPU, then stable diffusion started doing this, and THEN i reinstalled GPU with newest version

I'll... maybe try that when i have some time later and got over my chkdsk flashbacks... 🙂

Thanks for the help already, i'll write what it did later.

[ornate elk]

No problem yea let me know!

[mighty knot]

Hi, I want to see old generations from old models, which channels can I see it ?

I think that they might delete all the old channels and old models.

[ornate elk]

Hey, I think they are gone

[mighty knot]

Yes

Thats sad

[hybrid magnet]

how do i fix this error?

[fair oxide]

So fellas i decided to instal sd once more since i updated gpu. Whats this though?

Git last version dowloaded,python too.

[fair oxide]

Reinstall python, enabling "add path"

[hexed quarry]

Niki

[ornate elk]

wrong python version

you need 3.10.11 64bit
and you have to git clone the repo

[ornate elk]

reinstall python and check "add python to path"

[vocal burrow]

And once you re done reinstalling python, delete the venv folder if there is any

[violet swift]

Does anyone know how to avoid those horizontal lines in images created by flux dev?

[ruby yoke]

i think @fair oxide might know what's causing that

[fair oxide]

Thanks! Finally worked

[zenith loom]

Guys, any help with the error on using openvino ? i use i5 chip with intel iRIS xe graphic

[slate totem]

guys ,anybody know why this question had happen? it always waiting, but no photo has produced

[ornate elk]

What does the cmd show?

[slate totem]

To create a public link, set share=True in launch().
IIB Database file has been successfully backed up to the backup folder.
Startup time: 26.7s (prepare environment: 8.2s, import torch: 4.5s, import gradio: 1.6s, setup paths: 1.1s, initialize shared: 0.3s, other imports: 0.4s, load scripts: 3.6s, create ui: 3.5s, gradio launch: 2.4s, add APIs: 0.9s, app_started_callback: 0.3s).
you say this?

I don not know what its meaning, maybe meanless about this question

[slate totem]

sorry，i think there is no feedback abou my question in the cmd

[ornate elk]

Seems like the browser and the cmd are not connected.
Make sure the webui is whitelisted in any browser adblocker and that your not using a VPN

[zenith loom]

there is an error at my sdnext, seems like no picture shown in the UI

[ornate elk]

Your webui version is old. Did you downloaded the outdated openvino fork of Auto1111?

[zenith loom]

any link for me to use beside SDnext by vladmandic ?

[ornate elk]

Nope sry. Openvino is not commonly used

How much vram does your iris XE GPU have?

You may be stuck to using CPU only mode in auto1111

[zenith loom]

really ? oh nooooo

btw, thanks for clarify it for me, otherwise i'll keep trying again

[crystal nacelle]

my stable diffusion SD Zula for Automatic111 stopped working. the cmd shows me this

[ornate elk]

Your on a old rocm hip SDK version

You should uninstall rocm hip SDK 5.7
And install rocm hip SDK 6.2
Then restart the PC.
After that delete the .zluda folder and relaunch the webui-user.bat

[crystal nacelle]

thank you my friend!

[crystal nacelle]

bro theres like a thousand of these things. am i uninstalling them all?

[ornate elk]

Yes all that starts with HIP SDK

[brave comet]

Im trying to install stable diffusion cpu only but its been stuck on this point for the past 10 mins

[ornate elk]

Your using an old Version + you have it in a OneDrive synced folder which causes issues.

[brave comet]

Oh

Is there a new version that works

With cpu only

I dont have nvidia

I have intel

[ornate elk]

Yep git clone the latest auto1111 webui.
Then add
--skip-torch-cuda-test --use-cpu all --no-half --precision full --opt-sub-quad-attention
In the webui-user.bat at the line commandline_args=

Thats it

[brave comet]

Ok send link and code seperate

Please

[fair oxide]

mostly try more steps and using a more noisy sampler e.g. SDE/ancestral samplers
not all of them have been adapted for Flux, they need the noise schedule changed so that the variance doesn't explode
in comfy DPM++ 2SA works
there used to be one in Forge that worked but it got removed

[brave comet]

Also does stable diffusion allow nsfw just cant post in here?

I do both sfw and nsfw

By here i mean this server

[fair oxide]

the model is just a grid of numbers
it can't do anything on its own
so the software decides what is allowed

[brave comet]

Ok

[fair oxide]

some software might phone home so you need to check

[brave comet]

Phone home?

[fair oxide]

phone home as in send data to the company

[brave comet]

Oh k

Im hosting my own

[fair oxide]

it could still phone home potentially

[brave comet]

Ok

[fair oxide]

I'm not saying they all do but you need to check if you care about that

[brave comet]

Its just anime and furry

[fair oxide]

dozens of comfyUI custom nodes will phone home sadly

[ornate elk]

[CPU Mode] Here is a quick Guide to install Automatic1111 Webui for CPU Mode (Stable Diffusion) on Windows

Preparation:
You need to install Git 64bit: https://git-scm.com/download/win
and Python 3.10.11, 64bit: (any python above or 3.10.6) Not 3.11, 3.12 https://www.python.org/downloads/release/python-31011/
And check "add python to path" when installing Python.

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
4. Press enter and after its done you can close the cmd.
5. 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
   6.After editing, save and launch the webui-user.bat inside the stable-diffusion-webui folder.
   After the Installation, that can take a while is done, the webui will autolaunch in your Default Browser.

[brave comet]

Ty

[fair oxide]

if you don't mind it sending data then its fine
some people are more bothered about it than others

[brave comet]

As long as it doesnt ban me thats ok

[fair oxide]

yeah I haven't heard of remote shutdowns happening

[brave comet]

Ok

[charred rover]

should I spend more money for value on a used RTX 40xx vs 30xx to go fp8? (converting SDXL)

will it be faster per dollar?

[brave comet]

How do i change anaconda to python 3.10.11

[ornate elk]

If you dont need anaconda, uninstall it

[brave comet]

I dont?

[ornate elk]

Normaly not

[brave comet]

Ok

[ornate elk]

A 40gen is better than a 30gen in any way for ai and gaming tasks.
Only if you get a 3090 it would be a better price/value

[ornate elk]

You installed the wrong python version

My guide linked the correct version

[brave comet]

Ik it came with anaconda

[ornate elk]

Have you uninstalled it now?

[brave comet]

I am rn

Do i have to install all my python 11s

And my 13

[ornate elk]

You only need 3.10.11 64bit

After you installed it.
Delete the venv folder on the stable-diffusion-webui folder and relaunch the webui-user.bat

[brave comet]

I mean i have other newer python

How do i set it to use 10

3.10.10

[ornate elk]

You can edit the webui-user.bat and there at the PYTHON=
Set the path to the python 3.10.11

[brave comet]

Ok ty

[ornate elk]

You get the path by opening up a cmd and type
Where python

[brave comet]

Ok

How do i load a model?

And does this work with integrated intel

[ornate elk]

It won't use the Intel iris and only the CPU

[brave comet]

Ok ty

[ornate elk]

To load other models than the default you need to download them from Civitai.com and place them into models/stable-diffusion folder

[brave comet]

Ok

Ty

[ornate elk]

Np

[brave comet]

I only see lora and checkpoint

Will they work

[vocal burrow]

checkpoint ? screenshot plz

[vocal burrow]

models are checkpoints

[brave comet]

Oh ok

[brave comet]

Getting closer!

Does it matter that memory monitor is disabled

[vocal burrow]

no, that's because you're running it raw on cpu / without a gpu

good luck and be patient, this will take a while

[brave comet]

Ok ty

Like 5 mins

?

[vocal burrow]

judging by your laptop way more

[brave comet]

10?

Also is this for 8 images?

[vocal burrow]

for one

[brave comet]

Oh

Doesn’t stable diffusion generate 8?

[vocal burrow]

SD is meant to run on a gpu, you can run it on cpu but yeah.....

[vocal burrow]

it generates how many you ask for it. But the more the longer it will take.

[brave comet]

Ok

[vocal burrow]

Once you get something running I'd recommend you to take a look at model labelled "turbo" or "lighting" as they generate descent results in fewer steps => it goes faster.

[brave comet]

Does the default generate anime and furry

Also will it make my pc blue screen

[vocal burrow]

the default model ? probably but it won't be that good.

[brave comet]

Ok

[vocal burrow]

it should not

[brave comet]

Well vr chat (in vr) did

[midnight roost]

hi does anyone know why a ckpt checkpoint file takes up all my 32 gb ram when generating a img i get those "freezes" and stutter sounds during it.

[ornate elk]

Which Checkpoint?
What's your gpu?

[midnight roost]

i have this issue when i merge something as safetensors i get a error sometimes but i figured out i can make it a ckpt file instead

but when using it i get the 100% ram issue so what my tempoary solution is to make it a ckpt file and then turn that ckpt file into a safetensors by merging with a 0% checkpoint merge

but then it takes up double the amount of storage on my drive 😦

6950 xt

[ornate elk]

Do you try to merge sdxl or flux?

[midnight roost]

sdxl or illustrious

[ornate elk]

Because I heard for merging sdxl you need 64gb ram xD

What could work is to install the super merger extension.

https://github.com/hako-mikan/sd-webui-supermerger/

[midnight roost]

yea i have thought about trying to learn super merger it just looks complicated compared to the checkpoint merger

[ornate elk]

True, I haven't used it too

[midnight roost]

but the 64 gb ram might be when i try to use x y z plot

i like using that but it could be ckpt is bad with that

[ornate elk]

Nope XYZ works fine as it only switches between the checkpoints

You should always prefer .safetensor over .ckpt

[zinc comet]

Hi everyone, I got a 3080 10Gb, can I use sd 3.5 ? without offloading everything on the pc memory

[ornate elk]

Hey, Sd3.5 medium mostly

[zinc comet]

I managed the fp16 variant but only using enable_xformers_memory_efficient_attention() and enable_sequential_cpu_offload(), I end up with my pc memory being eaten alove (32Gb) and my 3080 using not even 2Gb, is there a trade off I can do to balance that ?

[ornate elk]

Dont use fp16 models

There are fp8 aswell as smaller gguf models

For the best performance you would need to use a model (checkpoint) that fits inside your whole GPUs vram

There are also sd3.5 large q8 or q4 gguf models that could fit

Here is the q8 sd3.5 large:
https://civitai.com/models/879251?modelVersionId=985027

[zinc comet]

oh fantastic

thanks a lot I'm gonna try

[regal mason]

my comfy-ui zluda setup suddenly stopped working, and can't seem to fix it (radeon rx 6800 XT)

Are there any updated guides/github repos for setting up comfyui with Zluda?

[cosmic furnace]

@ornate elk Do you know any ways to improve inpaint? I mean to use it with more speed & comfort. Espec to do more accurate mask or have less noticable mask edges >_>

[ornate elk]

Hey, what's the issue? And did you used my Comfyui guide or the patient x one?

[regal mason]

It's been a while, I can't remember, but there was kind of an "All in one" github repo that did most of the work. It required HIP 5.7, python 3.10

I'm not sure what changed, yesterday I tried running comfyui and it updated some libraries and no longer works. I have since tried a lot of "fixes" and nothing helps. Tried some scripts to test the bare minimum functionality, and get output like:

Very small FP32 linear failed: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling `cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)`
Very small FP16 linear failed: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling `cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, &fbeta, c, CUDA_R_16F, ldc, CUDA_R_32F, CUBLAS_GEMM_DEFAULT_TENSOR_OP)`
Very small BF16 linear failed: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling `cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16BF, lda, b, CUDA_R_16BF, ldb, &fbeta, c, CUDA_R_16BF, ldc, compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)`

[ornate elk]

Ah yep thats the patient x guide.
What's your GPU?

[regal mason]

radeon rx 6800 xt

[ornate elk]

Then I would recommend settinf comfyui up clean again with my guide which uses rocm hip SDK 6.2 and latest zluda

[regal mason]

will do, where can I find it?

[ornate elk]

In the pinned messages of this channel or here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[regal mason]

ah gotcha. I looked through the pinned messages for comfyui and AMD but missed that one

[charred rover]

better, but also 2-3x more expensive than a used 30xx. and 3090ti adds no value vs 3090, only price

[charred rover]

sorry I meant 3090ti is 10% faster in AI but 30-40% pricier

[formal ridge]

I can merge SDXL on ComfyUI using only 32GB of RAM, but I have to shut down almost everything (including explorer.exe) before I can do so.

[zinc comet]

Hi ! I wonder if i got a 3090 24Gb will i be able to do most of the things i want, like lora training sdxl (3.5 medium maybe ?), run flux/sd3.5, wan or other video model ? I think my question is what people use for the cool video or image i see, high end consumer gpu, or they go for colab/sagemaker, or they got access to those entreprise high end gpu ?

[sullen raft]

please tell me why SD does not perceive negative promts when running the Prompts from file or textbox script?
As I understand it, the promt database should be designed like this:
--prompt --negative_prompt
but it doesn't really work

[cosmic furnace]

Erm? Cs1o?) About inpaint? ;o

[ornate elk]

oh sry, in auto1111? forge? comfyui?

[cosmic furnace]

https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P also will this work on forge and where theres 7 models?

forge) like it pretty much ;o

[ornate elk]

forge doesnt support video stuff

[cosmic furnace]

fast-convenient and frequently updated. it seems. although weird it still didnt updated control net for flux =\

[cosmic furnace]

Dont tell me this just like hunyan only for Comfy -__-

https://www.reddit.com/r/StableDiffusion/comments/1j8w6go/wan_i2v_720p_can_do_anime_motion_fairly_well/

awww ? 😦

[ornate elk]

yep

[cosmic furnace]

mmmkay

and something to improve inpaint in forge?)

[ornate elk]

always set inpaint area to "masked only"

that should make inpaint really fast

[compact salmon]

any updates on the 9070xt?

[ornate elk]

Yes forge is still being updated

[hollow pulsar]

module = self._get_module(self._class_to_module[name])
File "F:\AIDiffusion\stable-diffusion-webui\venv\lib\site-packages\transformers\utils\import_utils.py", line 1088, in _get_module
raise RuntimeError(
RuntimeError: Failed to import transformers.modeling_utils because of the following error (look up to see its traceback):
module 'numpy' has no attribute 'dtypes'
Press any key to continue . . .

has anyone seen errors like this before? i updated numpy, transformers, torch, and python to the required version but nothing seems to get me back in

[rotund idol]

tengo problemas con stable difusion me ejecuta con amd pero siento come muchos recursos para la tarjeta grafica que tengo que es buena, y aparte no me toma bien los modelos y loras alguien puede ayudarme?

[brave comet]

How do i export my tensor.art ai model?

So i can add to civit ai

[cosmic furnace]

Lol) 😄 That one i know... I mean some good extensions or secret techniques 🙂

Espec it annoys that u can see where inpaint was made if not very acurate... -_-

[ornate elk]

Can you show the full cmd log?

[ornate elk]

I can help, what's your GPU?

[odd zodiac]

all my images have some inperfections , like in her eyes . why?

[ornate elk]

Hey, because you used an sdxl model which is trained on 1024x1024, at a resolution of 512x768

Better use a resolution that is nearer the trained one.
For example 768x1024

Also your model is a lightening model but you used 30 steps.

[odd zodiac]

thanks how many steps i should use?

[ornate elk]

5-6

Cfg needs to bet lowered then too

[odd zodiac]

kk ill try again

[ornate elk]

I think to 1

[half patrol]

does anyone know what I'm doing wrong? Every image I generate is a grey box

[ornate elk]

Yes, your trying to use flux in Auto1111 which is not supported

[half patrol]

what do I need to use?

[ornate elk]

Forge Webui or Comfyui

[half patrol]

i assume I need to reinstall the whole thing again?

[ornate elk]

Whats your GPU btw?

Flux needs a lot of vram

[half patrol]

4080

[ornate elk]

That will work.
Yea you can delete Auto1111 and then just install Forge.
You'll find all install guides in the first link of the pinned messages

Also don't use the 23gb flux model

There are smaller variants that perform much better on your gpu

[half patrol]

i was only one other model on the same download page for flux, I assume just switch to that one?

[ornate elk]

Nope there are different models based on flux and also so called gguf models. Theser are smaller.
For example here is a good flux model:
https://civitai.com/models/141592?modelVersionId=992642
Check the Files tab for the different sizes.

[half patrol]

how do i delete auto111

just nuke the folder?

if there are so many flux models how do i know which one the pixel art model i use actually needs?

[ornate elk]

The ones that fit in your vram will work the best.
Mostly fp8 safetensors or q8 gguf

For flux you also need some additional files too.
Like the clip t5 encoder and clip l aswell as the ae.safetensors vae

[half patrol]

so if i install the one you linked what else do i need?

[ornate elk]

checkout this page for the infos:
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050

It links the files I mentioned

[half patrol]

ive installed everything you said and im getting this error

[ornate elk]

You need the ae.saftensor

In the models/VAE folder

And then select the 3 files in the second dropdown

[half patrol]

i have done that and I still get the same error

[ornate elk]

Can you screenshot the webui interface

[ornate elk]

Remove the t5 fp16 clip

You only should use either t5 fp8 or fp16 but not both

And make sure the ae. Is selected

[half patrol]

i did that and tried to generate an image and my entire computer froze

no idea what happened

[ornate elk]

How much RAM do you have?

[half patrol]

32 gb

ddr5

[ornate elk]

When using flux you need to close all other programms and make sure you dont have Wallpaper engine running in the background

[half patrol]

i did have alot of chrome tabs open lmao

i'll try again

[ornate elk]

And make sure the model you used was fp8 or q8 (around 11gb)

[half patrol]

looks like its working now, thank you so much for the help!

[ornate elk]

No problem 🙂

[honest needle]

Hello,
I am billed on a monthly basis but cannot access stable assistant

it has been month and I have had no answer so far

The single sign on does not work and loops on the access page.

Can anybody help please ?

[vocal burrow]

Sorry, this is a community driven support channel. Your best shot to get this resolved would be to contact stability directly through their mail.

[honest needle]

Ah ok I understand. I've tried to reach them for months and never get an answer :/

[humble mica]

Is it even possible to use ngrok remote server with A1111 UI still? The only tutorials I was able to find are outdated and trying to figure this out on my own isn't working

I've already generated the auth token, ran the config file and edited user-webui.bat file properly, but it still says "'failed to connect session', 'The authtoken you specified does not look like a proper ngrok tunnel authtoken"

Also, I tried using --share and gradio, but it always times out after an hour or two

All I really need is a remote access to my PC running local, so I can mess around and generate images on my phone when I'm away from home

I think I've found the culprit, but no idea how to fix it

nevermind, all I had to do was match the port of my local URL with ngrok's http command - basically, I opened the wrong port

[brittle lantern]

I wish I knew bro if anyone has info on its performance please let me know

[regal mason]

Excellent guide - worked the first time, and seems much faster than before too. Appreciate all the work you do

[compact salmon]

https://www.reddit.com/r/comfyui/comments/1jaome4/working_comfyui_with_rocm_on_9070xt_a_quick/ this guys working on it last few days

@brittle lantern

if u get any info on how its running lmk bro lol @brittle lantern . going to tryt and get it setup myself today but cant go full linux i need windows app simultenously

[brittle lantern]

Alr thank u bro im considering getting a 9070 xt once prices comes down if not ill fall back on normal 9070 and if i cant get any for some reason ill have 7800 XT for £450 as last resort

[ruby yoke]

??

[flat nymph]

Guys, who can share the workflow of image imitating video action, I used mimicmotion, animatediff, and reactor, but it didn't work very well, mimicmotion mimicked the action to generate a bad image

[gray cloud]

do someone know why my stable diff (reforge) dont want to work with my lora's ?

[mystic zodiac]

Can someone please tell me what ControlNet models should i use for my Pony anime model in webui forge**?** i have tried a lot of versions, engine+model combinations,
I was getting all sorts of errors, most of which occur when using incompatible model
My console is constantly displaying error about webuitorch itself
I had this webui since early 2024 and constantly updating everything

[mystic zodiac]

loras need to be compatible with your model, pony-pony, sd-sd, flux-flux

[ornate elk]

have you tried controlnet union++ promax ?

[gray cloud]

thats the point

i installed the right one (if i did it well at least)

[mystic zodiac]

pretty sure i have tried something promax, im confused with the installation, something seems to only work with controlnet engine, something works strongly without it, all youtube videos on that topic seems to be out of date, hasn't got any positive result trying out 7+ controlnet setups

feels like my forge cant handle controlnet at all, only works with Canny on my custom small SD model, my goal is Openpose on pony, im gonna delete and reinstall all i have tomorrow just to try to get it to work

i want to know one thing, what is SDXL? is it a model or webui workflow

[ornate elk]

its a model

[mystic zodiac]

thank you, i were misleaded earlier

[ornate elk]

for me it works with forge and pony

[ornate elk]

you would need to share mor information, like which lora, model and what does the cmd log shows

[rotund idol]

Podrias hablarme por privado? xD

[hazy wigeon]

tried updating pip for comfyui portable, now i get this error on startup

[plain pond]

mr @ornate elk can u help me please.. I haven't used stable diffusion for a long time, when I wanted to use it in an emergency... this happened, even though I just wanted to change the model.

[gray cloud]

ok im gonna check that

unfortunatly those lora have NSFW in their preview so im not allowed to send them here, but i can send them to you in PM ?

[mystic zodiac]

why is it installing some random checkpoint by its own while reinstalling webui?

[ornate elk]

Because it needs at least one to work

[ornate elk]

Ah sry the 9070 cards are so nre there is no Rocm or zluda support right now.
The only guide you can try is the Automatic1111 with DirectML or Forge Webui with DirectML.
You find the Guides in the pinned messages of this channel.

[ornate elk]

Try delete the venv folder and relaunch the webui-user.bat.
Also send the fill cmd log if it doesnt work.

[grave fable]

guys my SD fucking broke its quality on itself

i have no idea what's going on, i only tried to enable xformers

quality went from this

to this

same prompt, same sample, same everything

tried to disable xformers, nothing changed

makes no sense

[vocal burrow]

different resolution, probably different model, etc

[grave fable]

same model same resolution

oh wait i was making at 500 for speed

hol up

oh yeah it was resolution

crazy

ty

[gloomy kraken]

hello..

[rotund idol]

but the 9mil series is current and it is assumed that there are already guides to make amd work with Rocm, it is only to install the library, in fact I have already made stable diffusion work using the guides, the problem is that there are some loras that do not take them well, I need to be able to download a repertoire but a window does not appear when it is executed, which if I am not mistaken was the ONNX window

[upbeat atlas]

So the webui did an auto update and ever since then I'm getting this error, how do I fix this?

[plain pond]

now it works.. thank youu

[ornate elk]

Delete the venv folder and relaunch the webui-user.bat

[mystic zodiac]

i finally got it to not spam errors, but it isnt working correctly with promax model

this is with openpose off

this is with openpose on

what am i doing wrong?

[flat olive]

hello, i have 1660 super 6gb, here are my args and the error i get, idk if i have to really do that of skip torch cuda test, or if i will have to install torch, im following the instructions in that installation method

[mystic zodiac]

i think its out of date method, my girlfriend has 1660Super, she just installed 1-click-webui-install and its working perfectly

[flat olive]

yeah i tried that one but i keep recieveng the 128 error code so i tried that one

[flat olive]

does she had the same issue?

[mystic zodiac]

no, no issues or errors were displayed throughout 3 days

[flat olive]

well, i will just wait for someone to help me with that

[ornate elk]

Hey update your GPU driver and also change --no-half-vae to --no-half

[flat olive]

that was it, thanks for the last

[ornate elk]

No problem

[ornate elk]

You have to rename the model also you have to select XL at the top

Also make sure the webui is whitelisted in any browser adblocker

Also make sure forge is updated

[flat olive]

fp16 for what is? i downloaded a model that uses it, should i change somenthing in argument line?

[ornate elk]

Which model?

[flat olive]

i think it is pony?

[ornate elk]

Is it 6gb in size?

[flat olive]

yep

[ornate elk]

Then its okay

Should work

[flat olive]

are there any guide that explain the whole concept of the usage of models? like whats xl, pony, loras, for?

btw is there anything i can do bc i still can creating images but this shows me that a1111 is using my cpu instead of my gpu

[ornate elk]

Check the vram usage of the gpu

Also are you downloading something rn?
Your network is maxed out

[flat olive]

yes but isn't somenthing related to a1111

[ornate elk]

But something related to CPU usage and RAM maybe

[flat olive]

this?

[ornate elk]

Looks good

You won't see a high GPU usage in % when generating

Try to generate a 768x768 image

[flat olive]

add 1 minute +- , i should use any parameters in the argument line to optimize this?btw this is in 512x512

25 steps 768x768 DPM++2M Karras

[ornate elk]

can you try an 1.5 based model to check the speeds ?

[turbid willow]

Anyone know why I might be experiencing a sudden and random drop in image quality? I found a prompt I like, sticking to it, made like 50+ pictures with it, getting good stuff, and then all of a sudden out of now where, with no change to settings or prompt, the images are poopy, all washed out and bleh.

I turned off SD and restarted it, no change.

[ornate elk]

wahsed out? seems like a vae issue

[turbid willow]

My model has a built in vae, I don't even have to mess with it

The image on the left is broke, I didn't do anything

It only makes images like that on the left now

[flat olive]

idk how good enough is(1.5 based model as you asked)

768x768

[ornate elk]

which webui do you use?

[ornate elk]

looks okay speed wise

you could try forge webui if you want faster pony speeds

[turbid willow]

Using forge

[ornate elk]

try drop your old image back into the prositive prompt. then press the blue button to auto apply the settings and press generate

[turbid willow]

Oh, I didn't know you could do that, I'll try. Sec.

generating, looking good so far, so weird. So what happened? Did I tweak my prompt slightly and some how that broke it all?

Yeah, image turned out normal.

Thanks! lol, I don't understand, but thanks XD

[flat olive]

will try later

[mystic zodiac]

which model should i rename? also no adblock, forge is freshly installed

[ornate elk]

np, you can compare the meta data of the two images in PNG-Info tab if you want to see whats caused it

[flat olive]

any installation guide you recommend?

[turbid willow]

Must have just been some weird prompt tweak?

[ornate elk]

maybe, you should see it in the meta data

if you send both originals, i can also check them

[ornate elk]

the first link in the pinned messages covers them all

[turbid willow]

How do you look at the meta data?

[ornate elk]

if you drop the image into the PNG-Info tab in the webui

[turbid willow]

Ha, thanks, there's so many things I don't know. Awesome.

[ornate elk]

the union model should be named:
controlnet++_union_sdxl_promax

[turbid willow]

Oooh, I think I found the issue. somehow it got droped from 30 steps down to 10. I see, I see. Thanks CS1o! You're the best ❤️

[ornate elk]

no problem 🙂

[turbid willow]

I think I tried to drop the batch count down from 10 to 1 and I accidently dropped the step count instead.

Or somethign liek that

Cool!

[ornate elk]

awesome image btw

was that with an illustrious based model?

[turbid willow]

Yes. I am using Illustrious

[ornate elk]

ahh nice, i have that but didnt used it so much

[mystic zodiac]

same bad result

[turbid willow]

I don't know if it's always this hard, but I spend hours tweaking my prompts till I find one that generates good images. Thanks!

[ornate elk]

can you set the seed to -1 and the resolution to 1024x1024 ?

[ornate elk]

do you have the booru tag autocomplete extension installed?

it helps with prompting with illus models

[turbid willow]

Idk, never heard of it. Tell me more! ^_^

I just learned about lora's last week, I keep learning and learning. Getting stronger all the time, but I havn't reached my final form yet. 😉

[mystic zodiac]

it got even worse)

[turbid willow]

Does this discord have a channel for posting art people make? I can post a few pictures.

[ornate elk]

Illustrious and pony models are trained on booru image board tags, like 1girl, masterpiece, long hair, blue eyes, etc.
The boorutag extension can search these tag librarys and auto complete words like an keyboard autocorrect feautre

[ornate elk]

can you disable the dark reader extension on the webui and test again?

[turbid willow]

Ooh, that's cool. I've heard about it being trained on that website you mentioned, booru, I just forgot the name. Having an extrension that auto fills and corrects though sounds helpful!

Is it easy to download and install?

[ornate elk]

yea you also see how often the tags is used so you can guess how likely the model will use it

[turbid willow]

Oh, It's in extentions. Sec, I just learned about how to do extentions yestday, I think I know how

[ornate elk]

Go into the extension tab,
click on available,
click on laod from
then in the list search booru tag autocompletion

[turbid willow]

Yeah, I just learned that yestday. Nod nod Just took me a moment to have my brain register what you said and what I had to do. Awesome, thanks!

Is there a place in this server to post pictures of work?

[ornate elk]

yep in #🏞｜general-with-images or #🍥｜anime or #✨｜sdxl

[mystic zodiac]

no such extension, only ones i got on chrome is SpeedDeal and youtube ShowDislikes

maybe the picture i put is from a wrong source?

[ornate elk]

can you show the last picutre here:

is it pure black ?

[mystic zodiac]

yep

[ornate elk]

thats the issue

wait im dumb xD

Thats the issue

[mystic zodiac]

heres the options

[ornate elk]

the input image you have is Already a Procesed POSE
so you have to select NONE

as Preproceesor

if you would drop an image from a normal person into controlnet then it would have worked

because then it needs to process the image to a pose

[mystic zodiac]

where can i write you a check?

thank you so much, never have i had working controlnet, even from youtube guides

and the issue was in such plain sight 😁

[ornate elk]

yep we even had this issue a lot here when controlnet launched

but its so obivous (for me) i forgot to check xD

[mystic zodiac]

does this model work with multiple maniquens?

[ornate elk]

multiple input images at the same time?

for pose?

there is a batch function that could do it

[mystic zodiac]

no, like 1 image with 2-3 human poses intercepting

like a group photo

[ornate elk]

ahh yep

thats possible

[mystic zodiac]

thank you for your time, ill try to research further possible troubles myself

[ornate elk]

no problem there are also site with poses:
https://openposes.com/

[austere charm]

is it possible to have both the A1111 and ComfyUI of SD just so i can kind of decide which one I want to try out and keep?

[ornate elk]

yes totally

[austere charm]

does openpose work on sdxl yet or is there another way to gain a little bit more control over composition/posing and that type of stuff? Tried looking up openpose xl models but those don't seem to work out

[ornate elk]

Controlnet union++ promax works very good

Its for sdxl only

[austere charm]

okay ill try that out, thanks! 🙏

[paper crater]

Hi! just to confirm, I can't install Stable Diffusion with python 3.12.7?

[ruby yoke]

you can. where'd you see that you can't?

[paper crater]

Thanks for responding, I guess I must have come accross old guilds cause It specifically stated to have python 3.10 some version

I'm pretty new to this so sorry if its an obvious question but I don't have to git clone etc if I just download comfy ui?

[ruby yoke]

most of the issues are going to be more related to the custom nodes or other code you want to use, or to comfyUI or forge or auto1111 than to stable diffusion itself

[ruby yoke]

you ust git clone in order to do an install of stuff from github. when you are reading the install instructions on a specific repo, it'll tell you if you need to use that command or not. most people do that, a few (like swarmUI) have installiers

[paper crater]

got it thanks ^^

[faint widget]

Does anyone know where I can find a working version of the fixed i2v hunyuan workflow?

[idle bolt]

Can you use sdxl/pony with ltx video or hunyuan? If not, what is the best way to create videos with sdxl/pony at the moment? Thanks.

[formal ridge]

You could easily set up a longer workflow (with a switch from one of the popular packs) that takes your created image and pipes it into i2v using hunyuan or wan on ComfyUI.

[upbeat atlas]

Now I'm getting this error

[dapper rampart]

oh its ai imn design,im an architectural student lol

im getting this error may i ask why

i dont understand anything

[craggy root]

hey y’all how do i cancel my free subscription?

[ornate elk]

Hex, what's your GPU?

AMD GPUs dont work with xformers

[dapper rampart]

its geforce 1660 ti

nvidia

[ornate elk]

What's in your webui-user.bat?

[dapper rampart]

where am i gonna check it

[ornate elk]

Right click and edit the webui-user.bat

Its in the webui folder

[dapper rampart]

im using the web version tho

[lucid yoke]

? Yes but you run it locally. Its the file you use to launch it

[untold mauve]

hi, with forge should i use "--use-zluda --update-check --skip-ort --medvram-sdxl" like in a1111?

[ornate elk]

--medvram-sdxl won't work.

The rest is okay

[untold mauve]

oh ok thanks

and should i continue to use fixfp16 vae?

i notice that forge is way slower than a1111 with rx6600, idk if i need some options to activate

[graceful oar]

hey idk why but when im using inpainting in a1111 the result is distorted and blue-ish on the edges? looks cartoony. when it is generating i can clearly see that it generates smth good but the result is bad. does anyone know a possible fix?

wanted to fix the face this is the result

also is rx6600 (what i currently have) decent for stable diffusion? from what ive read 8gb vram may be too little

[ornate elk]

Yes

And when upscaling always enable never oom for tiled vae

[ornate elk]

Your mostly using directml which is slow and vram hungry.
A much better option is to use ZLUDA with Auto1111

You'll find a guide for that in the pinned messages of this channel. (First link)

[short tulip]

Anyone done WAN LoRA training with images? I tried with diffusion-pipe and it runs through all the training but the checkpoint files are small, like 80mb. Also they dont seem to affect outputs when generating. Which leads me to believe its not working at all. Anyone who has successfully trained a WAN LoRA mind sharing your training config files? Or suggest a different way to train. NOTE: I have trained Hunyuan and it worked fine.

[graceful oar]

alright, thanks, will try to install now.

[ornate elk]

np, let me know if any step isnt clear

[untold mauve]

thank u but i don't understand why it's so slow compared to a111 (70/it vs 9/it)

[ornate elk]

70 ?

[graceful oar]

is there a significant boost when using zluda compared to direct ml?

[ornate elk]

yes about 2-4x and much less vram usage, allowing you to use sdxl models and even upscaling to larger resolutions with 8gb vram

[graceful oar]

alr, i see that its telling me that it works better with sd.next, do i need to install that?

[ornate elk]

nope

sdnext is a pain to use xD

[graceful oar]

ok, installed everything

ill let you know if i got any error or if its way slower

seems alr rn

it consumes NOTHING 💀

[ornate elk]

its mostly compiling

check the cmd

[graceful oar]

no its generating

rn

[ornate elk]

then something is wrong

[graceful oar]

maybe im js checking the wrong graph :3

no its alr

6s/it

[untold mauve]

yep 😦

[ornate elk]

can you share me the cmd log?

[ornate elk]

the full one pls

[ornate elk]

which settings have you tried in forge?

resolution and model etc

[graceful oar]

thats all

??

i mean it works

so

args

[ornate elk]

remove --no-half lol

thats slows down massively

[graceful oar]

it broke without it

[ornate elk]

it should not

[graceful oar]

hm alr

will try

[ornate elk]

pls try again

also your python is old. you should use 3.10.11 64bit

[graceful oar]

yeah no

[ornate elk]

after upgrading that delete the venv folder

[graceful oar]

doesnt work

[ornate elk]

whats the model you try to use

[graceful oar]

uhh wdym model

like this?

text2img

[ornate elk]

delete that model and use an other one

[graceful oar]

where can i get a diff one

[ornate elk]

also do this:
also your python is old. you should use 3.10.11 64bit
https://www.python.org/downloads/release/python-31011/
after upgrading that, delete the venv folder

[ornate elk]

here:
https://civitai.com/models/4384/dreamshaper

[graceful oar]

omfg torch reinstalling

download speed capped at 8mb/s for some reason

[untold mauve]

832x1216 and biglove2, in a1111 it takes 9it

[ornate elk]

And forge 70 s/it ?

With the same settings?

[untold mauve]

sorry 70 with 2 images, but with 1 image 35it

slow anyway

the gpu limit is like 85%

7000something gb

[ornate elk]

Hmm strange

I have nearly the same performance

On both

Maybe yours is not using Zluda correctly?

Feel free to share the full cmd log

[graceful oar]

getting this error rn

now it launches?? idk

nvm getting the same error as before

btw i was downloading stable diffusion to use with direct ml from a different tutorial so could that cause instability? that im trying to use zluda now?

[graceful oar]

its solved by adding --no-half, but as you said it slows everything down

any workaround?

[ornate elk]

Yes if you not follow the linked guide then thats the problem

[graceful oar]

i did follow it

[ornate elk]

Directml in the folder name is outdated stuff

[ornate elk]

But you said you downloaded it from a different guide

Tutorial

The other guide is not just a commandline_args change

You have to setup a new stable-diffusion instance

[graceful oar]

am i that dumb

brb

is this normal?

btw, it launched, but i cant generate anything (pauses)

[ornate elk]

Yes thats good

Compiling is needed for zluda

It has to do this only two times

And then its ready to use.
It doesnt do it every time. Only when torch or the AMD Driver updates

[graceful oar]

alr ig ill js wait

[ornate elk]

Yep, wait until the webui launches.
Then type cute cat at default settings and hit generate.
Then it will compile the second time until you get an image

And then it should take seconds for the images

[graceful oar]

ye i typed in a prompt and hit generate

ill wait now

how long did it take you to compile this?

ok i js got model loaded in 779s still compiling

[ornate elk]

It takes around 15-40 mins max

It depends on the CPU to compile

[graceful oar]

alr

[graceful oar]

it works 🙏

tysm bro

[ornate elk]

Perfect, no problem 🙂

[random night]

okay, i don't know what's going on anymore

[neon orchid]

Hi guys, a somewhat technical question.
To generate drawings with AI, do you need a good processor or a good graphics card/GPU?

[random night]

gpu my friend

[neon orchid]

So if I want the images to be generated faster, do I need more GPU?

👀

[random night]

the answer is, you need more vram, not more gpu

[neon orchid]

ohhh, thanks 😄

[random night]

get the gpu with the most vram you can get your hand of, then it's should be good

prioritize nvidia gpu

[neon orchid]

Wow, thanks man : D

I'll keep that in mind 🙂

[broken epoch]

Title: How to Keep Stable Diffusion API Images Consistent for the Same Product?

Post:
Hey everyone,

I'm using the Stable Diffusion API in my e-commerce website to generate product images dynamically. However, I have an issue: the images generated for the same product are not consistent across different requests.

For example, when generating images for an ASUS ROG laptop, each request creates completely different designs, instead of maintaining a similar style or details. This makes it difficult to use AI-generated images for real products.

Is there a way to ensure visual consistency when generating images for the same product? Would image-to-image (img2img) or any other method help?

[regal mason]

For some reason my "Canny" node is not working, just loads forever

Which is odd because it says it's from "Comfy Core".. shouldn't it be part of the base models?

Heh. I spoke too soon... it eventually rendered. I guess it's one of those things that takes a while the first time you use it

[tribal anvil]

Hey everyone I just got webui forge but I have no idea what I'm doing can anyone help or know any good guides?

[formal ridge]

Speaking of A1111 / Forge... Does anyone know of any ComfyUI nodes that perform the "Resize Seed from Width/Height" option that's on both WebUIs?

That's one amazing option I've not been able to find a replacement for.

[formal ridge]

Give this a try. https://stablediffusion.cdcruz.com/

[tribal anvil]

thanks ill have a look

[verbal mortar]

Does anyone have a link to really good guide for setting up SD on Ubuntu, including Nvidia drivers, CUDA, and A1111 (or any decent webui)? I feel like I'm 90% of the way there but can't shake this "ModuleNotFoundError: No module named '_ctypes'".
So, short of asking for Linux and Python help in probably the wrong forum, I'm willing to just scrap what I have and start from scratch with a good guide. Thanks!

[tribal anvil]

what folder do i put checkpoints in?

[lucid yoke]

flux in diffusion-models and SD in stable-diffusion

[tribal anvil]

i put the checkpoint in stable diffusion but its not working do i have to do anything else for it to work

it shows up but doesnt work

nvm it works now

[regal mason]

Is there a custom node for frame interpolation that works with AMD GPUs?

[formal ridge]

Just use a specialized tool like Video2x

[regal mason]

this one? https://github.com/k4yt3x/video2x

Does it work with AMD?

[formal ridge]

According to the issues, yes. Just make sure you use "h264_amf, h264_vulkan, or h264_vaapi"

[brittle lantern]

For £450 what’s a good GPU for image gen most taxing I’m doing is flux workflows with loras and custom models

[austere charm]

whenever i restart webui on a1111 i have to pull up my last image generation to bring back all my settings, is there a good way to set webui up to where it will let me just pick up where i left off whenver i close/restart it?

[regal mason]

Aight... "frame interpolation" might be too vague for what im looking for. I'm hoping for something that would use AI to generate a frame "between 2 images", that is natural and not just an average between the 2 images. For example, if I have a front picture of a face, and a left picture of the same face, I want to generate one or more images of the transition between those 2 poses (gradually turning head to the side)... does something like this exist that works well?

[formal ridge]

That's still frame interpolation, you just have to do it multiple times and blend to get what you want.

[formal ridge]

modify the settings file and change your defaults, that's the best thing you can do.

[regal mason]

Video2x did not seem great at this, at least in regards to finding a logical frame between 2 frams that wasnt just kind of smudging them together

it would probably be good for other use cases, like increasing the FPS from 30 to 60, but it definitely didnt seem to do any "thinking" about what might occur between the 2 frames, just kind of averaged between them

[alpine patio]

is anyone having problems running stable diffusion on google collab?

[tribal anvil]

How demanding is webui reforge Everytime I try to generate an image my laptop gets really slow

I have a GTX 970m which ik is old but I thought my laptop could handle it

[ornate elk]

Make sure you use 1.5 based models. They are around 2gb

Dreamshaper v8 for example

[tribal anvil]

Sorry but how can I tell I'm new?

I think the ones I have are around 2gb

They're all 200 mb

Ig I'm doing something wrong because that's way under

[formal ridge]

Is it personal preference, is that why you keep recommending DS8? There are much better models

[formal ridge]

If you're downloading from CivitAI, the models are called "Checkpoints" and it sounds like you're downloading "Loras"

[tribal anvil]

I can't use both?

[formal ridge]

You can if they're compatible and know how. Loras are used like an embedding.

[tribal anvil]

To try and learn how to use this I'm copying the models I want to use

So I use all the resources that were used to make them

[formal ridge]

I'd recommend clicking the link I sent 5.5 hours ago in this channel. the cdcruz one.

That should help get you started

[tribal anvil]

I looked it over again but I still don't know how to fix the lag

It happens when I enter prompts or try to generate

It was fine earlier

[formal ridge]

lag will always happen if using a laptop, heat builds up and most laptops aren't exactly equipped to handle lots of heat. Plus you're using a 9 series. I'm amazed it even works.

[tribal anvil]

Damn so there's nothing I can do about it?

[formal ridge]

If you're using A1111, one thing you can do is switch to Forge or ComfyUI (if they'll work with a 970m)

[tribal anvil]

I'm using forge

That's what it says at least

[formal ridge]

Scroll down and enable both Never OOM options

[tribal anvil]

Both?

[formal ridge]

Should say "Enabled for U-Net" and "Enabled for VAE"

You want checkmarks in both boxes

[tribal anvil]

Aight

[formal ridge]

Now run a couple images and see if that helps

[tribal anvil]

I'll try thanks

[noble idol]

@ornate elk forge broke again, help

[formal ridge]

Just because there's a button that says "Update" doesn't mean you should push it.

[noble idol]

Okay first off, fuck you koi, how'd you know I hit update 😅

Secondly how do I fix it

[formal ridge]

1. We talk enough that I understand your mindset. and 2) I have no idea, set everything back up again?

[noble idol]

Fuck you (affectionate) and fuck, setting stuff up again is crazzzyyy 😭

I'll look into it again

[formal ridge]

I'd say "maybe next time you'll remember not to push that button but..." 🤣

[noble idol]

See if I can't find the fix through pure luck

[noble idol]

😅

I'll uhh TRY to remember to not update it next time

The big green button though....

[deft hemlock]

i switched to comfy and wana make the models and output images use the same folder as a1111

how do i do that?

because ill use a1111 to gather previews of my loras aswell as the triggers and other things

from my research its looks like i have to mess with something in the comfy nodes folder

[formal ridge]

There's a file in ComfyUI folder called "extrapaths.yaml.something" (forgot the whole filename), but you can edit that.

[deft hemlock]

this ?

i changed it but i still dont see any of my models

[formal ridge]

Yes that. See the "#"? Those should NOT be there on the lines you've modified.

[deft hemlock]

i see

but its formatted correctly right ?

the \ and / arnt flipped

the red lines are what i changed

[formal ridge]

You need whichever slash your OS uses I believe. Windows uses "/"

[deft hemlock]

file explorer is \

so i guess \

[formal ridge]

If that slash doesn't work, just use the other one.

[deft hemlock]

alright

im rebooting

everything checks out ?

[formal ridge]

Now as far as images outputting to the same folder, I'm not sure how to do that.

[deft hemlock]

ill figure out the output later

the models are the main thing

[formal ridge]

I'd recommend using hardpaths (drive letters and all) for each line you want

[deft hemlock]

F:\ai\sd.webui\webui
F:\ai\comfy\custom_nodes

they are already like this

[ruby yoke]

c:\comfyUI - not f:\ai\comfy

[deft hemlock]

does it have to be like that? all the other webuis are in a similar structure to f:\ai\forge

i got it

but its still laggy

hmmmm

maybe my pc is the issue

probably

[flat olive]

how i'm able to use a lora that has a .safetensor? which folder it has to go? lora or Stabledifussion?

[formal ridge]

Loras always go into Lora folder

[deft hemlock]

lora is lora

[formal ridge]

".safetensors" is just a file format

[deft hemlock]

to know if its a lora or a checkpoint just look at the file size

if its more than 1 gb its a cehckpoint

if smaller its a lora

boys i need to know what can i add here

to maximize performance

i got a laptop with an rtx 2050 (4gb vram) and 12 gb of normal ram

the issue is that its too laggy

i tried comfy its les laggy but still laggy

[flat olive]

since i got 32gb of ram a1111 works better tbh

[deft hemlock]

lmao

flex

[flat olive]

mb but i'm not kidding

[ruby yoke]

if you don't put comfy where it wants to be by default, manager can have problems when you ask it to install stuff. and either not isntall or install incorrectly. if you're not going to put comfy on your system drive off root, then plan to do your installs by hand and avoid manager

[flat olive]

this thing takes so much ram

[ruby yoke]

vram, not ram. Video ram with your GPU, not system ram with your cpu

[deft hemlock]

hmmm i see

i tested comfy by manually moving pony and a lora to test it out

it still lagged my pc to hell so i guess im still sticking with reforge

since it lags almost asmuch

[formal ridge]

Forge and ComfyUI will still work much better than A1111 if you're using nVidia.

[formal ridge]

Depends if you use aggressive offloading like some of us do.

[flat olive]

lmao i just ended up downloading it

[flat olive]

i see that a better gpu obviously u get a better perfomance but it is also important the ram

[deft hemlock]

can confirm

[deft hemlock]

yeah

someohow it ran better on my old rig

my old rig had 16gb of ram

but 512 mb of vram

it took 30 mins per 512x512 image

the thing never lagged

now i have a proper rig and it lags

[ornate elk]

Nope, its to make sure it works.
The model is 2gb and and not a hyper, lightning, whatever model.
For the start its okay. But yea there are better ones.

[gilded laurel]

Could someone help with setting up models to N00mkrads Stable Diffusion GUI? When i paste the models safetensors files, it downloads on the gui, but when trying to generate image, it says error and that isnt in diffusion format? So what do i need to do?

[ornate elk]

Remove --skip-torch-cuda-test --always-low-vram --disable-nan-check

[zinc comet]

Hi everyone, can i use a lora trained with sdxl 1.0 with another sdxl based model ?

[keen falcon]

I've tried to install Forge-Webui with ZLUDA as a complement to my working Automatic1111 with ZLUDA so I can use Flux, but I get the attached error and it doesn't open a web browser window instead it closes.

Anyone have any advice?

[terse hull]

how do i install a1111 on amd and linux?

[ornate elk]

Hey, yes

[next nebula]

hey, does flux. 1.D work with sd automatic1111?

[vocal burrow]

no you ll want to use forge for flux stuff

[zinc comet]

Quick question, i want to buy on ebay a 3090 24gb, this model PNY GeForce RTX 3090 XLR8 Gaming EPIC-X RGB Triple Fan Edition M 24GB GDDR6X, i'm wining the bid with 500£ and 3 days left but no one apart 1 guy at the begining, tried to bid higher so far, am i making a mistake ? I haven't found much information online but it seems a decent card, any thoughts ?

[formal ridge]

I have a thought, did you do your research on the 3090 before you started bidding? Some specific models (not sure if that PNY is one of them) have bad capacitors on the back of the boards.

So when those capacitors blow, there goes your entire card.

[zinc comet]

Thanks will check

Mmmmhhh from what i can read the main problem would be fan and temperature potentially, otherwise decent perfs and havent read anything about faulty component

Will see, i would have prefer asus or a founder version, i might take the risk with this one 🙂

[ornate elk]

Use an other webui. Because its outdated

[safe inlet]

i dont use steanme

cormmunity lol

[tired lintel]

Yes but the results are nit as good.

[zinc comet]

Like ? How bad is the dégradation from your own experiments?

[sonic pine]

can I ask for reforge help in here?

So Im trying to transfer my forge settings to reforge

I dont think that link was legit

anyways

Its being a bit glitchy

First off, I get a black screen when I use some models

that might be because of vpred

[vocal burrow]

don't click any shady external discord link

[sonic pine]

yeah lol, I was confused why some new dude was sending a discord link for it