🤝｜tech-support

[covert veldt]

Large controlnets here https://huggingface.co/stabilityai/stable-diffusion-3.5-controlnets

Anyway maybe someone will see this and figure it out - it feels like something is off on their backend rather than my images and prompts.

[foggy gate]

Hi! Sorry a little out of the blue here.

I'm trying to install a model trainer like EveryDream 2 however I'm on a M2 chip on a Mac mini and I've heard best is to use other training models as EveryDream is not optimised for M2.

I hope if it's not too bothering if someone could help me out.

(The tutorial I'm following uses Automatic 1111 and EveryDream) I've instealled Automatic 1111 so far.

The idea is to run everydream 2 on Mac or somehow find an alternative workaround to achieve the same thing. I'm following the rock papers scissors tutorial on a Mac M2.

[ornate elk]

hey, does everydramer supports mac?

for me it seems that it only supports nvidia

But it seems like DrawThings has Mac Support and has an Lora Trainer build in:
https://drawthings.ai/

[foggy gate]

hey, does everydramer supports mac? No I don't think so. Ill try using drawthings.ai thank you so much!

I''l let you know how it goes >D

[ornate elk]

You need the openpose sdxl controlnet model:
The bottom one.
https://huggingface.co/lllyasviel/sd_control_collection/tree/main

Just one. But some work better than others

What also can work is the controlnet sdxl union model.
Its an all in one model

The promax model:
https://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/main

The bottom one

Has to be renamed
controlnet++_union_sdxl_promax

You need to select it

[barren gyro]

hi how i can download worked reActor face swap extenstion for webui stable diff

thanks!

https://github.com/Gourieff/sd-webui-reactor this removed

[ornate elk]

This is the new version:
https://github.com/Gourieff/sd-webui-reactor-sfw

[barren gyro]

put it here?

i dont see it here

the control net and faceswap

why?

[ornate elk]

Restart the webui

[barren gyro]

i do it

trying again

i cant open the app now

help :X

thanks

anyone?

[ornate elk]

Make sure your webui is updated

Then delete the venv folder and relaunch

[barren gyro]

i fix it but its use my cpu not gpu

--lowvram --precision full --no-half --skip-torch-cuda-test

i have to put this to make the web ui work

[ornate elk]

What's your GPU?

[barren gyro]

2080 super

8 gb

its was work before idk

all scrwed up

cause face swap extension

[ornate elk]

Then remove --skip-torch-cuda-test --no-half --lowvram --precision full
And then only add --xformers --medvram-sdxl --no-half-vae

Then delete the venv folder and relaunch

That should fix it

[barren gyro]

where the venv

located

same

@ornate elk fix it . thanks. now i just need help with reactor sfw if u can help

[ornate elk]

No problem.
Add a face to the reactor tab and then generate

And dont forget to enable it

[barren gyro]

i enable it i cant fint it dude

i dont see controlnet or anytrhing

[ornate elk]

Click on the extensions tab.
Then check for updates.
Then relaunch the webui-user.bat

[barren gyro]

its like that . i reset same

:X

[oak hearth]

Hello, I'm trying to install the dreambooth extension but I am not getting the tab for it. Can someone please help?

[ornate elk]

Hey, dreambooth is outdated.
It also breaks the webui.
To fix it delete dreambooth from the extensions folder.
Then delete the venv folder and relaunch the webui-user.bat.

For training use the standalone tools like OneTrainer or Kohya_ss

[ornate elk]

Can you show me the full cmd log?

When you relaunch

[oak hearth]

Are you able to direct me to anything that shows me how to install/use onetrainer?

[ornate elk]

For installation look here: https://github.com/Nerogar/OneTrainer?tab=readme-ov-file#installation

[barren gyro]

@ornate elk lol again the controlnet break it every time

[ornate elk]

Make sure your GPU driver is updated

[hallow latch]

I have ran into a issue.. When I generate my images they are not saved to my location of choose??

I tried to change to another location but will still not save it.

Is there a setting that somehow got turned off?

I need to download each image or zip them to get the images at the moment.

This happend today what I noticed, yesterday it was fine :/

Nvm, the "save all images" was unticked :/

Not sure why it got unticked

[knotty turret]

Yes both work

[whole crypt]

I have no idea how to use ai stuff, just downloaded web ui stable diffusion on my Mac after watching a YouTube video, then watched another on Lora’s, I used the same exact Lora with all the same settings and prompt and image isn’t done but I’ve don’t it a couple times and none have turned out good, lora was Eris from civitai, is it not turning out good cause it’s on Mac?

[tawny saddle]

Yaknow

i think

that sum1 in this server, must be in some server, that has a bot, that is using some kinda hook, to use ppls accs to send these fake help tickets, i think this bc, that last acc that sent that, isnt a "new" acc

must be ppl that are in a diffirent stable diff server

[noble idol]

lower the hires denoising to like 0.4, never put that on 7. up at the top it shows youre using the model that came with stable diffusion, you should probably download a different one

you can find other checkpoints on civit where you got the lora

[tawny saddle]

i personally like to merge the pruned defualt wit ones i get, cuz i dunno, just seems to make stuff look really good

[whole crypt]

Got it thx, do I have to download everything again, or just that new thing

[whole crypt]

Ya ur right,

[noble idol]

no, just download a new checkpoint

some popular models rn are

ntrmix
autismmix
noobai
WAI

make sure the lora youre using is compatible with your checkpoint as well

[whole crypt]

I’ll try it in the morning, and hopefully that fixes it

[noble idol]

gl

[tawny saddle]

once u get go to civiti to get a new model, and down load it, navigate to your model directory, in the root of your stable diffusion directory, for example "C:/stable-diffusion-webui/models/stable-diffusion/"
and drag and drop the models into that folder, "lora" is another form of low level model, and are used to add modifcations to an image formed by a model, these can range anywhere from like a few megabytes to about 200mbs, models or "checkpoints" can be like 1GB to around 20GB, lora goes in the Lora folder in the directory, i.e C:/stable-diffusion-webui/models/lora
:3
no need to reinstall stable diffusion for model placing uvu

[ornate elk]

You can't train loras in Auto1111.
Better use OneTrainer or Kohya_ss
Both are seperate tools made for lora Training

[solemn atlas]

Hi a quick question I have gpu 4070s but when I generate nothing works. Everything freezes. Is there a reason to that? It wasn't like that when I started. Then I switched to comfyui and it was really fast now its the same as automatic1111. It freezes and I can't do anything expect wating till it finishes generating.

[lucid yoke]

Got a error log?

[topaz tangle]

Can you recommend a YouTube tutorial so I can install it correctly? This is my first time with this type of technology.

[ornate elk]

Here is the install guide:
https://github.com/Nerogar/OneTrainer?tab=readme-ov-file#installation

[ornate elk]

Does it only freezes in Comfyui?

[topaz tangle]

I already have this downloaded

To get results like this, a parrot is trained with my own character, what is the best option? I find the parrot training process very confusing. Is there a more simplified option?

[ornate elk]

Idk much about training sry

[solemn atlas]

in both

[ornate elk]

How much RAM do you have?
And how much free space on C ?

[solemn atlas]

32GB

[ornate elk]

Ok good

[ornate elk]

So when does it freeze?

I remember you where the guy who switched your OS drive by accident

[solemn atlas]

yeah, that's me ..

when I start generating it happens. Actually it starts when i open comfui or automatic1111

[ornate elk]

Is your GPU driver updated?

[solemn atlas]

I'm doing it right know.

The studio driver I download right know

[lucid yoke]

i do remember switching OS drives (like having windows on D) is gonna cause a whole bunch of issues due bad programming hard coding a LOT of stuff to C

all of the issues and none of them documented lol

[solemn atlas]

what is that? Fetch?

[vocal burrow]

fetch <=> "download git related data"

[calm moth]

Is it possible to have forge/reforge remember past settings or change the default values specific settings? I'm annoying having to adjust all the Hi-Res and adetailer settings every time I start/restart the WebUI.

Found it, nevermind

[sick elk]

Hi, I don’t understand how use Lora ?
I just use different models ? 🤷🏻‍♂️

[wicked widget]

use lora with a checkpoint

[calm moth]

What are you using to generate images? ComfyUI/Forge/A1111?

[nimble elk]

hey everyone! hope someone has a brain powerful enough to problemshoot this for me.. I am very new to stable diffusion so let me know if you need more info.

The problem is i have downloaded 2 models that seem not to work: 1. "JJsExhibition_Flux.safetensors" + "HouseBlurprint_v1.safetensors". When i run those models the show whats on the screen shot and end up giving me a grey picture in the end.

When i run the model "wildcardxXLTURBO_wildcardxXLTURBOV10.safetensors", it works (it gives me a image, not just a grey solid color).

Any suggestions? the directory for the models are placed in: G:\Programs\stable-diffusion-portable-main\stable-diffusion-portable-main\models\Stable-diffusion
Image

BTW, the result is the same, with or without Controlnet enabled.

[tawny saddle]

ahh

[tawny saddle]

i remember dis non sense xd

idk, for me, it worked to disable it

and then remove it

buh dunt do that, incase theres a way to fix it, i never learned how to use control net, so i jus removed it, i have a mid tier gaming pc anyway, so, meh '

i dint need it

[sick elk]

How can I see that’s…

[sick elk]

Checkpoint are model or not ?

[wicked widget]

yes

checkpoints are the base models youll run, loras are an additional guidance filter models you run on top

[nimble elk]

The thing is that it doesnt work even without controlnet enabled.

[sick elk]

It’s better to use diffusion bee ?

[wicked widget]

idk what that is, use forge, swarmui or comyui

[sick elk]

https://diffusionbee.com/

[wicked widget]

if youre a mac user ig, but i still stand by my previous statement, those uis are the best and works on mac aswell.

also check the pins

mac install guide

[calm moth]

Checkpoints are models. Imagine LORAs as highly specialized mini-models that are NOT loaded like full models. How they are loaded depends on the WebUI itself. Forge/A1111 have you type the name of the LORA in the form of "<Name of LORA:Strength of LORA>". ComfyUI has separate nodes for including LORAs in generations. I can't tell you how DiffiusionBee handles them since I'm not a Mac user, but LORAs are such a basic tool in t2i generation, I'm sure it's able to use them.

[sick elk]

Able to use them with Diffusion bee ?

[calm moth]

"I can't tell you how DiffiusionBee handles them since I'm not a Mac user, but LORAs are such a basic tool in t2i generation, I'm sure it's able to use them"
From what I just wrote

[sick elk]

Understand (sorry iam not English 🥲)

[sick elk]

I will try

[calm moth]

Supported models : - SD 1.x, SD 2.x, SD XL, Inpainting, ControlNet, LoRA
From the DiffusionBee Github Page.

https://github.com/divamgupta/diffusionbee-stable-diffusion-ui

[tawny saddle]

commentary node

[dark mural]

Does anyone know which checkpoint was used for this image? The Civitai post only mentions the LoRA, but not the checkpoint

[calm moth]

Dragging it into ComfyUI should show, since the workflow was attached.

[dark mural]

how pls

nvm ty

[analog hill]

is there a guide on what folders are put in where for stable diffusion?

cuz ive have no clue for tools and others

or upscalers

idk where to put the folders at

[flat yarrow]

In general with control nets in Comfy do they go after or before Loras? As I've read they go before and I pass an open pose to a KSampler (text2image) to get an image (pose is great), then image 2 image going through 2 Loras and a KSampler and it destroys the pose. The pose is a human figure at full length through open pose and 1 KSampler great images, then I sent to KSampler 2 with the denoise at 1 to maintain the first images, but returns headshots. Is this a setting I've not configure of a model issue?

[spiral tinsel]

Hi there, i wonder how to use 2 RTX 4060ti 16GB on stable diffusion large model? I run the program with comfyui.

[lucid yoke]

using both cards for one image generation is sadly not possible but in SwarmUI (also uses comfy) you can run a model on each GPU so you generate 2 images at once

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Using More GPUs.md but if you want to use GPU 1 and 2 to get more vram to run lets say a big model is not possible iirc

[spiral tinsel]

oo, thanks for provide tutorial too

[ornate elk]

Flux is not supported by auto1111

Houseblueprint could be a lora and not a model

[balmy shore]

whats the way to go about updating the various sd webui's and theyyre verious requisites? does it happen automatically when u launch or should --update-check be passed
im using webui-ux currently its just a fork of a111 so most of the backend should be the same im jw
cuz its been a good few months since ive used it and im guessing theres been some progress in the meantime but launching the bat didnt seem to even check

[balmy shore]

cortana says this but im not sure if itll break my shit but ill try it

[ornate elk]

Git pull is the command to update webuis

Nothing more

--update-check just tells you if an update is available

[heady zephyr]

why are my tokens limited in comfyui

[vast pulsar]

hey guys, can someone pls explain to me how to use remacri? i'm so confused. Using webui from a1111 btw

[rocky citrus]

I didn't see a specific help chat for what I am having trouble with. Is it ok to post the issue here and see if anyone can be of assistance?

[lean elm]

I have a image that I horionztally displaced an object using a depth map, i want to inpaint those pixels (without a prompt, just in context with the image) but they are sporatic and some models are having trouble with it. Is it possible to use the original image which i displaced to better these results? What would be the best approach.

[old maple]

hey i keep trying to make stable diffusion work but it keeps saying that i dont have importlib.metadata, but when i do the pip install it still doesnt work

[tawny saddle]

oi mate

why...anyone got stable diffusion 1.7.0 anywhere

?

not sure i like this new one u-u

dunno, seems slower, by a consirable amount, and i dunno, models dont seem to load properly, and it always throws error after error bc it cant "load model quickly"

jus wan the old one again, preferably 1.7

is there some kinda of old repo sum where

found it uvU

they should def make an old repo tho, no cap

like.. for experimental purposes or sum yk

or for like, if theres a bug on the new one, should be easy to revert

<v>

[ornate elk]

Put it into models/ESRGAN
Then you can select it in hires fix

[ornate elk]

Can you show the full cmd log?

[midnight kiln]

Anyone recognizing this error from controlnet? (used for upscaling in my case with the tile model):
RuntimeError: mat1 and mat2 shapes cannot be multiplied (154x2048 and 768x320)

im inouting a 512x512 or 1024x1024 - notthing special at all

The only place I can see these numbers in the error coming from is the yaml file that comes with the tile model - but im not supposed to directly edit that right? I've never seen a tutorial mention that at least

[ornate elk]

Something not compatible is used.
For example controlnet models for 1.5 but with an sdxl base model

Tile doesn't need a .yaml

[midnight kiln]

that makes sese - but what model version should I be using for SDXL tho? thats what I generally use most

[ornate elk]

Here are the 1.5 controlnet models:
https://civitai.com/models/38784/controlnet-11-models

Here are the sdxl ones:
https://civitai.com/models/136070?modelVersionId=267493

[midnight kiln]

controlnet11Models_tileE is current one

[ornate elk]

Thats for 1.5

[midnight kiln]

You sir - are a gentleman and a scholar.

that was the whole issue

[tawny saddle]

thank goodness i screenshotted my hash for 1.7.0 xdd

may not have ever been able to revert

im basically doing wut youd do with a computer that bugchecked hard, "last config that worked" setting on the bios

new one is jus meh atm >-<

[tawny saddle]

wut do i do bout it

; ^;

anyone got that dll by chance

[ornate elk]

What's your GPU?

[tawny saddle]

amd 5600G

[ornate elk]

Model doesn't loaded quickly is something that can be ignored

[tawny saddle]

this dll is there

[ornate elk]

So your using directml with the igpu?

[tawny saddle]

so why it tryna reinstall it

[tawny saddle]

ya

im installing the same version of Stable diff that i was using on my pc last year

and i dunno why it says no permission

bc i was able to last year <v>

[ornate elk]

Did you copied that version over, or did you reset the PC in the meantime?

[tawny saddle]

i downloaded 1.7.0 direct-ml, then did git reset --hard 03a88617fcd9442313ce2ed7facfecf6cdf72c36

it started, shows the correct version 1.7.0, downloadded a a bunch of stuff, then stopped at that dll, bc "no permissions"

by the OS for sum reason

[burnt salmon]

ive been using regional prompter in forge and the matrix mode works great but i have been trying to use the mask mode and it doesnt seem to work at all. ive been looking for guides on how to use it but every guide i find the UI is completely different than what i have. even the github page shows a different UI.

this is what mine looks like

does anyone have a link to a guide that tells me how to use it with that layout?

im sure im missing something simple but the guides arent any help because there are settings they tell me to turn on that are completely missing from my UI

strange. it seems to be working now although im not sure what i did to fix it lol

[tawny saddle]

my issue seems to be steming from Pip

not sure wut to do about that, maybe unisntall pip?

and then reinstall ovO?

[vocal sundial]

hey, could anyone help me with this? I follow the instructions to put these arguments but every way I try it says there's something wrong

this is kohya for training loras

[tawny saddle]

:^

i have pytorch for direct ml, wuts the issue u-u

i think i solved the onnxruntime thing

[lunar kite]

Hi! I keep getting this error with SD Deforum regarding xformers ~ does anyone know what to do to fix??

[tawny saddle]

screams

[ornate elk]

Wouldn't recommend fixing the old broken version and instead focusing on getting the latest to run normal

[tawny saddle]

nuuuuuuuuuuuuuufawufbafbaosbgf

i have new one, it works fine, its just slowww, the new one was gud, i used it for like 2 year

i only lost it bc i lost my ssd

i dunno why i cant install 1.7.0 fine, is the same version i was using

[lunar kite]

Any luck? @ornate elk ??

[tawny saddle]

it just, worked, like tom would say it

took me like 2 mins to install 1.7.0, i dunno wut the issues is

i dont believe i updated or changed anything regarding pytorch, or python, or even torchvision, it all should just work

[ornate elk]

Can't help with google collab sry

It says your GPU is to old

[tawny saddle]

yk wut i think

i think

that all the repos that 1.7 pulled from, have been updated to only support 1.10

so all the stuff it downloads, is not even compaitiable

u~u feck

my only hope is that sum1 has 1.7, or a verison earlier that 1.10, that also is for AMD, and dunt mind giving me that in the form of a zip u-u

thatbeing said, now that im friggin stuck wit this version crys, how do u stop it from tryna load models "fast"

[ornate elk]

Thats exactly the issue

[tawny saddle]

:((

so heart breaking

certainly wish they made it more "smart"

[ornate elk]

What issues do you have with the 1.10?

[tawny saddle]

well, apart from it like, tossing errors at me and lagging mypc bc it trys to loads models "fast"
then immdialty goes to load them "slowly" smh
some things i try to do, take longer in this verion, i.e, making a picture on 1.7.0 took like 1 min, whereas on here, it takes like 5 to 6 mins, and sometimes 15m

:V

is this bc python is still 3.10.6

do i need update it or nu

bc that appears to be the indirect/direct cause of this error

the one that keeps models from loading "fast"

i like to keep a neat terminal, the url errors arent keeping the ting from working, its just annoying >-<
theres gotta be a way to disable it, it seems to ve trying to find a model on hugging face, automatically, and by "none", it appears its not meant to, so, theres got to be like a setting or an arg that can be used to stop it from searching the repo for models, and just use local models

to stop the error

and, my intuition is telling me that, this error may be causing my local models to load more slowly

[ornate elk]

That error is caused by runwayml because they deletet the 1.5 base model from huggingface a few months ago.
Auto1111 tries to pull the model and the configs from the empty link.
The model can't load fast doesn't affect model loading times and can be ignored.
You can try forge webui for amd if you want a more updated version which should also run better with your GPU

Your model loading speeds depend on your RAM amount.
If you only have 16gb you have to increase the Windows Pagefile.

[ornate elk]

Yes please update to python 3.10.11 64bit

Much better

[tawny saddle]

got it :3

[tawny saddle]

for anyone that may have clicked those links

do not proceed

im afriad there is no store on steam called a "stora"

can barely understand why scammers still do that scam xD

like cmon, missplelled URLs is so old school, theyre def scraping the bottom of the barrol

[vast pulsar]

Thanks, I'll try it out

[hallow latch]

Is it possible to hide this text without removing it from the info panel?

[vocal burrow]

It s a way to filter people ready to fall for "easy scams". That way they have a list of "easy targets" to scam. They expect it to fail for 99% of people. But those 1%... They ll focus all their efforts on them instead.

[vast pulsar]

only have "RealESRGAN" folder

[tawny saddle]

tru

[ornate elk]

Then create an ESRGAN folder

[vast pulsar]

yep, found the answear on reddit, works now, thanks

[tidal thunder]

Hello all! I need some help please, i'm using flux in stable diffusion forge and getting bug in img2img (any tab, inpaint or just img2img) and im getting in inipaint ovesaturated selected area, in just img2img i'm getting blurred image
What i tried:
Changing: steps, method, type, cfg scale, denoising, vae, text encoder, change near the same flux model

[fair oxide]

why does my image look watered down

it looks weird

[ornate elk]

You need to use a custom VAE
Do you use 1.5 or sdxl/pony/illust?

Hey, what's your GPU?

[fair oxide]

i dont know what a VAE is
I use pony

[ornate elk]

Can you try using my setup guide?

It covers all steps needed for a correct installation, you mostly missing the needed args for your card (--xformers --medvram--no-half)

Its on the first link of the pinned messages

[ornate elk]

A VAE is sometimes needed for color correction.
You need one like this:
Direct Link to sdxl fp16 vae:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
Put that into the models/vae folder

Then go into Settings, User Interface, Quicksettings and there add sd_vae
then hit apply and reload ui.
Then you get a vae dropdown where you can select the vae

[indigo lark]

Hello, I need help with controlnet, if it's not the place for it please tel me where can I find help becouse I'm losing my damn mind.

After a long time of not using sd at all I made a fresh instal and wanted to use the openpose 2d, and I tried everything that I could find on the internet to fix it but The Tabs Are Not Showing, not openpose nor controlnet, I have right now a clean instal, nothing but controlnet instaled, still nothing,
I reinstaled it, updated it, tried 2 diffrent links for instaling, I reinstled the whole sd, I restarted my conputer, I have no idea what else I can do.
I am completly stomped
please help

[lucid yoke]

first off lets start from the beginning shall we.
what UI are you using? Automatic1111? Webforge-UI? Swarm?

[indigo lark]

Automatic

[lucid yoke]

what model of SD are you trying to run? 1.5? XL?

but generally:
Navigate to the Extensions tab and click on Available sub tab

Click Load from: button

In the Search box type in: controlnet. You will see an Extension named sd-webui-controlnet, click on Install in the Action column to the far right.

unsure how well it will work though as A1111 is quite dated and i recommend Forge webui from the pinned messages in all cases

[indigo lark]

already have it

also I am stupid and I forgot with it is, sorry,

[lucid yoke]

hmm it is there. how is your main screen looking like?

[lucid yoke]

and you clicked the apply and restart ui button?

[indigo lark]

yep

[ornate elk]

whats your gpu?

[lucid yoke]

gpu wont prevent it from showing up right? its an exstension tab

still worth exploring

[ornate elk]

then edit the webui-user.bat
at the line commandline_args= you add: --xformers
then save.
open up a cmd and run pip cache purge
then delete the venv folder and relaunch the webui-user.bat

[indigo lark]

also I followed this gaid https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers

still nothing

[ornate elk]

oh thats the wrong controlnet extension

[indigo lark]

oh?

[ornate elk]

thats the correct one:
https://github.com/Mikubill/sd-webui-controlnet

delete the other one before installing this

then relaunch

[indigo lark]

yea yea

whould it take this long?

[ornate elk]

Depends on the internet speed

[fair oxide]

the one I need is schedule type karras

[fair oxide]

where do I find that one

[indigo lark]

it says it's instaled but nope

no tab

[ornate elk]

Thats a sampler

[ornate elk]

You have to relaunch

[indigo lark]

I did

[ornate elk]

Also make sure you dont have opera VPN enabled and the webui whitelisted in any browser adblocker

[fair oxide]

im copying from an already existing image and when I do this thing is in the bottom but the image wont generate unless I get rid of it which I thin is why it looks so weird

[indigo lark]

nope

I dont have it

[ornate elk]

Its under scheduler

Just select it there

[ornate elk]

Can you try open it in an other browser?

[indigo lark]

sure

firefox

[fair oxide]

sorry but I dint see anything called that

[ornate elk]

Can you screenshot the webui?

[ornate elk]

There it is

[indigo lark]

edge

[ornate elk]

Also there

[ornate elk]

Ohhh your on an old webui version

[fair oxide]

do I need to update and if yes then how do I

[indigo lark]

? the tab is not there

[ornate elk]

Yep its pretty easy.
First update your python to 3.10.11 64bit
Press "upgrade" when asked.

Then edit the webui-user.bat
And in an empty row (like the second one) you add
Git pull

Then save.
Then delete the venv folder
Then relaunch the webui-user.bat

[ornate elk]

I see it in both images

At the bottom

Click on it to uncollapse

[indigo lark]

ohhhhhh, Im stupid

[fair oxide]

ok I will look how to update python

[ornate elk]

download it from here and install:
https://www.python.org/downloads/release/python-31011/

use the Windows installer (64-bit)

[fair oxide]

thanks its updating currently

[indigo lark]

so I have the extension and I'm not sure where it is, I do have the pth files

[ornate elk]

dont use these files, use the smaller ones from here:
each model is only 700mb:
https://civitai.com/models/38784/controlnet-11-models

then put them into models/controlnet

[indigo lark]

UU , okay, I have this, and im not sure how to use it , im sorry lost time I used sd was more then a year ago (or maybe more) a lot has changed

[ornate elk]

input an image and select the model you want to use in the right dropdown

[indigo lark]

why does it promp me to draw on it?

thank you for your help a lot ❤️

[ornate elk]

maybe you clicked something wrong, like sketch

have you downloaded all models from my link?

[indigo lark]

yes

[ornate elk]

and they are all in the models/controlnet folder? and you see them in the dropdown?

[indigo lark]

I saw only one, the 11-models

or do you mean this?

[fair oxide]

do I need to rename it to .txt to edit it

I cant do anything since its a .bat

[ornate elk]

right click and edit

on win11 its under more options at the bottom

[ornate elk]

there are mutliple you need

[indigo lark]

Ooooh

[ornate elk]

select each and then download

[lunar kite]

Does anybody know if SD Deforum Google collab notebook is glitching? Cannot get it to work recently

[fair oxide]

what is the venv folder

[ornate elk]

a folder inside the stable-diffusion-webui folder

[fair oxide]

this?

[ornate elk]

no

its in the same folder as the webui-user.bat

[fair oxide]

sorry 😭

[vast pulsar]

ok, now that i got remacri working... what does the "smoother" version do exactly? preview pics look the exact same to me

[fair oxide]

I think it updated

now what

[ornate elk]

now you can use the scheduler

drop the image back in and it shouldnt show the karras setting at the bottom

[fair oxide]

I dont see anything

[indigo lark]

huh?

[ornate elk]

What's on the right to sampler

[ornate elk]

Incompatible model used

The controlnet models you download work only for 1.5 based models

For sdxl/pony etc you need other controlnet models
https://civitai.com/models/136070?modelVersionId=255155

[indigo lark]

god damn it

[ornate elk]

your webui didnt updated

show me the cmd log when you relaunch

[fair oxide]

is this what you wanted or is there a logs folder like minecraft has

[ornate elk]

it cant update because of this file:

delete that and then relaunch

[fair oxide]

tf is that

I will try and find it ig

should I delete it

wait wrong one

its this one

I think that worked

there is new stuff now

[ornate elk]

Yep now your updated

What's your GPU?

[fair oxide]

rtx 4070 super

[ornate elk]

Okay, then your good with only --xformers as launch arg
In the webui-user.bat

You see now the sampler and scheduler are seperated

[fair oxide]

yea

what do I set it to

[ornate elk]

Best is to leave scheduler on automatic

It will take the best for the selected sampler

[fair oxide]

ok thanks

[ornate elk]

And sampler try DPM++ 2m or Euler a

They are fast and good

[fair oxide]

the generated image still looks weird

is it possible the lora is just not a good one?

[ornate elk]

How does it look?

[fair oxide]

my image

vs what I copied

I have the same model

and same lora

[onyx solar]

Looks like a different VAE.

[fair oxide]

https://tenor.com/view/cat-dance-gif-22187542

[onyx solar]

Is that a 1.5 model?

[fair oxide]

I dont know

what that means

[onyx solar]

I haven't looked at earlier posts...so my apologies if that was shared.

What model are you using?

[fair oxide]

sweetMix_v22Flat.safetensors [83326ee94a]

idk its what the image linked to on civitai so thats what im using

[onyx solar]

The HuggingFace site shows two different versions with baked-in VAEs that differ:
https://huggingface.co/digiplay/Sweet-mix_v2.2_flat/tree/main

The bottom 2 files are the 2 different bake-in VAE versions.

[fair oxide]

do I download them and put them in the VAE file?

[onyx solar]

That will likely be the difference in your colors at the end.

[ornate elk]

I already told you that this is a VAE issue and I linked you the VAE + the settings you need to do

[fair oxide]

sorry I thought that was just a random VAE and not the one I needed

[ornate elk]

Here

Its for sdxl and works with every sdxl based model

[onyx solar]

Sometimes we just need to hear/see things in a different way for it to be understood. It's all good.

[fair oxide]

I put it in here and the restarted stable diffusion and it still looks weird

[onyx solar]

You likely need to adjust the settings to use the VAE. But I'm a ComfyUI user so remembering where that's set in A1111 is a bit rusty for me. I'm sure @ornate elk knows it by heart.

[fair oxide]

[ornate elk]

Then go into Settings, User Interface, Quicksettings and there add sd_vae
then hit apply and reload ui.
Then you get a vae dropdown where you can select the vae

[fair oxide]

I already have something in here

should I remove it

[ornate elk]

No

Thats the checkpoint dropdown

Dont forget to hit apply

Then reload

[fair oxide]

did all that

it looks the same

[ornate elk]

Really? Can you check if the sd_vae is still set in the setting?

Also try to close everything and relaunch

[fair oxide]

yep its here

[fair oxide]

I will do that

that didnt fix it

[tawny saddle]

anyone here good at CSS style files

i wan me button to be a heart shape

:3

or is that not..like a possible thing to do with jus a css file, i bet its prolly like, sum u have to do within stable's code u-u

all i can do is make the stuff pink and make the bg a diff shade

[timid flicker]

hello, does forge support v-pred or do I have to install reforge instead?

I'm just wondering 😅

[ornate elk]

Is the webui whitelisted in any browser adblocker?

[ornate elk]

Forge should support v-pred

[timid flicker]

idk if I have to install something, because I couldn't make it work

[fair oxide]

the adblocker ignores automatic1111

[ornate elk]

Make sure its whitelisted in it

Can you screenshot the webui?

[ornate elk]

Hmm for me it worked. But the quality was really bad with v-pred

Haven't tried reforge

[fair oxide]

im just going to totally turn off the adblocker

nope still the same

[timid flicker]

I get this kind of results, and I'm using a v-pred checkpoint

[ornate elk]

Oh okay

Thats broken
Is your forge updated?

[ornate elk]

Can you take a screenshot of the webui?

Can't believe its not there xD

[ornate elk]

Its there at the top

Just select the sdxl vae

And then generate

And dont forget to enable the adblocker again

[fair oxide]

[ornate elk]

Ohh my fault you used an 1.5 based model

[fair oxide]

so now what

[ornate elk]

I'll send you the correct one

[timid flicker]

that was the issue, thx

[tawny saddle]

heartss

[ornate elk]

https://civitai.com/models/23906/kl-f8-anime2-vae

This one goes into models/VAE then relaunch and then select it

You can keep the other one for sdxl models if you have some

I'm off now

[fair oxide]

ok thanks for the help

I learned alot

[tidal thunder]

Hello all! I need some help please, i'm using flux in stable diffusion forge and getting bug in img2img (any tab, inpaint or just img2img) and im getting in inipaint ovesaturated selected area, in just img2img i'm getting blurred image
What i tried:
Changing: steps, method, type, cfg scale, denoising, vae, text encoder, change near the same flux model

[old maple]

venv "X:\Documents\stable-diffusion-webui-forge-main\stable-diffusion-webui-forge-main\venv\Scripts\Python.exe" Traceback (most recent call last): File "launch.py", line 4, in <module> from modules import launch_utils File "X:\Documents\stable-diffusion-webui-forge-main\stable-diffusion-webui-forge-main\modules\launch_utils.py", line 9, in <module> import importlib.metadata ModuleNotFoundError: No module named 'importlib.metadata' Press any key to continue . ..

[visual bane]

I'm looking for a pretrained, distilled VTON that can generate results faster than CatVTON (which takes around 30–40 seconds). I’m facing installation issues with DistillVTON (DM-VTON), so if anyone knows of a faster alternative or has successfully installed DistillVTON, I’d appreciate any guidance.

[brittle lantern]

Hi peeps just a quick question what seems to be the best UI to use on MacBook Pro with M chips a fren wants to start AI image gen

[ornate elk]

You installed it the wrong way by not downloading the correct version.
On the forge webui github scroll down to the install part there is the recommended version as download.
Then extract it in a folder you created.
Like on C:\Ai\

[neon orchid]

hello

i have a question

I have a friend who wants to install Stable Diffusion... but he has almost no Vram, is there any way he can install it? :/

128mb vram

[tranquil token]

So I installed String Essentials from the node menu in comfy ui and now I cant run anything with Hunyuan or any other video model.
It goes "clip missing: ['text_projection.weight']" and then pauses, i.e crashes.

Replicated on 2 installs one fresh, after restarts and deleting the nodes

[old maple]

didnt work same error

[ornate elk]

Can you show me the full cmd log?

[tawny saddle]

anyone remember how to install new fonts to the system terminal on NT based operating systems,i i.e windows 11 x64

I did it once back in XP days I believe

once more like at least 4 times

I think it used true type font

dunt remember

wanted to install this 7segment display font to it, I think it'd look baller

bru imagine if there was like a this display about the size of a space heater that was like a windows terminal relay that used vacuum Florencent displays, that'd be killer

[cinder shore]

hi guys, I'm a newbie to SD, and this is my laptop configuration. I wanna install SD on it, so which version is the most suitable for me? 🤔

[wicked widget]

nvm im stupid 8gb. you can do sdxl with some limitations

use forge or swarmui

check the pins for install guides

[leaden lintel]

did anyone know a good model and vae for anime character generation because i try to download xRikishi's Unstable Diffusion (Alpha version) and the rendering is bad

i'm using stability matrix with stable diffusion webui forge

[lucid yoke]

Any illustrious model with a sdXL vae would to the trick

Idk if the rikishis is a xl or 1.5 one

[regal mason]

Anyone get DreamBooth working on an AMD GPU? Looking to train SDXL LoRA

[lucid yoke]

dreambooth is outdated and probably broken. i recommend using another tool to make a lora. Kohya, onetrainer or civitai (paid)

oh and on AMD mb. well thats gonna be a chellange

but i managed to find this ||not so simple|| guide on civit ai that might help your case. https://civitai.com/articles/4332/offline-lora-training-on-amd-gpu-rocm-62-and-ubuntu-2402-noble-numbat-updated-august-2024

[tender barn]

need help

why does it keep saying error connection errored out

[ornate elk]

Only OneTrainer on windows for AMD

Works for lora training

[violet isle]

hi, idk if somenthing wrong.
when i use "realistic vision" it works very fast "1min or less" on generating images/inpaint, and when i use "pony realism" it takes more than 20/30 min

[ornate elk]

Hey, what's your gpu?

[violet isle]

rtx 3070

[ornate elk]

Ah do you use auto1111 webui?

[violet isle]

yes sir

is better other?

[ornate elk]

OK, then edit the webui-user.bat and at the line commandline_args=
You have to add: --xformers --medvram-sdxl
Then save and relaunch
That speeds everything up

[violet isle]

alright 🫡

[ornate elk]

Its needed for GPUs with 8gb vram.
Pony is sdxl based that's why it takes longer but with these settings it should take much less longer

[violet isle]

i see

ill try, thank you very much!

[ornate elk]

No problem 🙂

[violet isle]

as you said, works like a charm, 40s

again thank you very much!

[heady zephyr]

why are my results so different when i use forge ui instead of a1111

the results on a1111 are way better

[normal wedge]

How do you guys generate video with stable diffusion? Do we need an extension for that?

[neon orchid]

can you use stable diffusion with 128 vram guys?

Sorry for the unexpected question xP

[tranquil token]

Everytime I try to que a Hunyuan video gen now (after NO update, NO changes to layout) it kills itself. Just from one day to another, why is ComfyUI so unstable? It constantly crashes itself even when nothing changes from one day to another, yesterday atleast I was able to get it running once after re-creating the DualCLIPLoader node for some reason, but not anymore now it 100% bricked it self.

[midnight kiln]

Anyone good with the custom scripts for automatic1111? For the life of me I can not get the thing to show up in the drop-down list - but various extension scripts do.
I even made a minimal test just to rule out weird issues in the code - so I must be registering things wrongly in the GUI somehow
Weirdly enough my script can generate a new UI element, so it seems to run (no errors also) - but it's just not in the scripts drop-down (in any tabs)
Here is my minimal code that fails also:
`from modules import scripts

class TestScript(scripts.Script):
def title(self):
return "Test Script"

def show(self, is_img2img):
    return scripts.AlwaysVisible

def ui(self, is_img2img):
    # No custom UI elements are needed for this test script.
    return []

def run(self, p):
    print("Test script executed!")
    return []

`

[ornate elk]

What's your GPU?

[little bough]

Hi, installing SD on a new system, what is the most recent version of python that I can install for it? Thanks

[ornate elk]

3.10.11 64bit

[little bough]

Thanks

also, I'm wondering, I have a new OS sytem but still have my old HD connected with all my old SD installs, I am wondering, if I install python on my new OS will i be able to just load up my old instances of A1111/Forge that are on the old HD? or can I even just copy across the entire folder to my new HD?

Or should I just reinstall again? I have like 4 instances(a1111, reforge, forge x2) and a lot of extensions so i am trying to avoid having to reinstall everything again if not necessary

running same OS and it will be the same version of Python installed (I think I had 3.10.11 on the old one)

[ornate elk]

Forge should work if you copy it over

Auto1111s update function won't work.

Reforge should work

[little bough]

hmm ok

tried to actually load it up on the old drive and then tried copying it to new drive and got this:

despite having just installed python 3.10.11 and adding the path function

I guess it's looking for that user folder

is there anyway to divert the python folder its looking for to the one on my new drive?

[ornate elk]

You have to delete the venv folder

Before launching

Or you can add the path of the new python into the webui-user.bat

[little bough]

aha! nice, thanks, its reinstalling it now 🙂

btw, do you have any recommendations for sharing say, Controlnet files and models that belong to say faceswappers etc that you want to 'share' between different instances so you dont need to reinstall the same files for every seperate version of a1111/forge, I usually use symbolic links for controlnet models but I've found it has gotten messy, it works but I'm wondering if there's a better way or do you just keep your install extensions seperate?

[ornate elk]

You can easily do this by editing the webui-user.bat

There add --models-path "D:\examplePath\toModels"

It should point to the models folder of auto1111 or Forge/reforge

[little bough]

would it be like this? btw i tried setting up the my old models folder in the new folder with symbolic links and then with redirecting in the .bat in the picture but with both i get an error 13 no permissions, any idea why this might be? could it be something to do with fixing user permissions for the old ssd or something else? any idea? thanks

[ornate elk]

Because you installed the OS new.
All folders you created on the old system will get you permission errors

Because the user account changed which which these got created

A reinstall of the webuis would be fixing all issues

And yep the --models dir and lora dir will work too

[sleek depot]

I'm having an issue with Adetailer, where it is improving hands, but I've been getting a lighter square around the hand afterwards.. I tried checking the apply color correction after img2img that I saw someone suggested, is there another fix for this?

or is the VAE i am using just incompatible with it

[tranquil token]

A 3090. I wrote my own node that just "consumes" a model input in a new graph so it forces it to load and that seems to fix it? Once the model is loaded everything works fine, but getting it to load up the first time causes the issue for some reason.

[little bough]

uh ok, will just go with the fresh install, thanks

[dim sentinel]

hello, I'm searching for someone that can help to generate videos like these https://civitai.com/user/bhopping
He's using HunyuanVideo
I will pay in bitcoin for the help if required
I have installed stable diffusion but not sure how to make videos I tried to follow videos on youtube but without success

[midnight kiln]

It can be several things. Mixing of different styles or samplers (adetailer can be set to use it's own, separate from the rest of the picture) can sometimes make things blend poorly.
A VAE can absolutely do this, but it can't be incompatible with adetailer I think - only the model. However note that you could set adetailer to use its own VAE - and if you are using 2 different one in the same image there's no wonder you get poor blending
manually selecting resolution for adetailerto use (much better to use the option in settings if you want to maximize the possible resoution)
not enough steps to blend well.
improper clip-skip, especially if it was different in the original image. Most SDXL and Pony models use clipskip2, but there are exceptions.
I'm sure certain settings like a very small or non-existing mask padding can also cause seams
I have rarely had much issue with poor blending in adetailer, so if you have messed around with it a lot, try to leave it to it's defaults and see if it helps - and move from there. Maybe enable a whole new adetailer tab (settings) with default settnings in case you're not sure what those were.

Not sure any of these are actually your issue - but it's a few things to check at lrast

[chrome rampart]

Hi all. Question can you use the flux controlnet function like reference in Forge in combination with Flux Dev?

[sleek depot]

When i disabled the vae.. or switched to another one it was fixed.. both VAEs are for that model but one makes more pastely colors.. whic adetailer didn't like even though it was set to use the same settings. so I'll takt the great reply to try to troubleshoot a bit though

had pretty much standard adetailer settings.. it just doesnt like that one VAE

but.. adetailer for face wasn't giving boxes.. only the out of the box hand model running in tab 2

[midnight kiln]

Great to hear. It's always a relief when you can find the root of pesky issues like that. Hopefully all models moving forwrd will just have their VAEs baked in, because it's often not that clear what it's intended to be used with

see that you didn't click the checkbox for manually setting a vae in tab2 then

[sleek depot]

haven't changed any settings. didn't wanna go messing things up even more before i could userstand... so started with removing things to try to locate it..

seems everything is set to use same as checkpoint.. well except the Adetailer sampler.. which is DPM++ 2M.. no clue why? is that default?

yeah seems that way.. fond some other post with the same settings

[midnight kiln]

thats default yes. I use state manager extension to set one of a few personal presets where all the tabs are loaded with my most used settings

otherwise things are going to reset on you over and over - it can gte frustrating - especially whne it ruins a whole batch because of the ONE setting you forgot to change

[sleek depot]

state manager.. will write that down for later.. just kinda getting into this and dont wanna add huge complexity

yeah true

does state manager safe stuff if you ahve to F5 the tab?

[midnight kiln]

its just a button to save the UI as-is

[sleek depot]

if so.. then def need that

[midnight kiln]

and it also can autosaave the settings for each generation you did - so you can go back and continue tweaking or rerun something

[sleek depot]

so this post says for SDXL to change adetailer to 1024x1024.. I'm using a pony model at the moment.. uh.. where would i find the best size settings to go for?

[midnight kiln]

I also highly recommend PNG viewer for similar reasons. you can have settings encode metadata (all the gen settings) into the actual image - and PNG viewer (a new tab) will read this back to you when you drag a picture there. its a big help to help you remember how exactly you made something previously. This includes the exact seed, so you can reproduce it very closely

[midnight kiln]

you can treat SDXL and Pony mostly interchangably. Pony is just a further customized version of SDXL

[sleek depot]

ok

last time i tried these things was whent he first automatic111 came out so I'm trying to get back into how things work.

[midnight kiln]

As for resolution in adetailer, I would not recommend setting resolution manually - BUT - I would recommend letting it auto-maximize the framing for you. Essentially letting it do "superresolution" in that area that is higher than the output image - but this greatly helps the image develop more cleanly

you can enable it this way

[crystal cobalt]

I'm messing around with creating some stylistic LoRAs for the first time, and for whatever reason all of my outputs seems to have this problem with x4 creating duplicates of the intended subject, and most of the time they're all sort of merged together...

[sleek depot]

oh.. another question.. I ran a batch job while i was away from the computer.. and it seems like the generation slows down a lot when left unattended (like windows going to the screensaver) or even if I change tabs for too long.. i thought the SD engine was running the tasks.. not the browser

[sleek depot]

Are you using multidiffusion?

cause I had that kinda weirdness when i was playing around with that

and the upscaler

[crystal cobalt]

No, no upscaling or anything at the moment, very basic setup with minimal nodes in Comfy

[midnight kiln]

Note: strict is best if you rarely use other than SDXL and Pony (like me) - but "free" should generally work for all models. its just slightly more prone to bugs because weird resolutions can happen

[sleek depot]

cool might swap to strict then as I'm currently trying to limit myself to those 2

[sleek depot]

you try lowering the strength of the lora? loras seem to go all over the place when they go too "hard"

[midnight kiln]

that sounded a lot like what you'd get with a tiled upscaler (without controlnet) with too liberal denoiser yea.... but Comfy spesifically I am not experienced enough in

[sleek depot]

Side note.. I'm just a noob trying to guess things based off of limited experience

[midnight kiln]

to be clear - you are training your own loras, not using existing ones?

[crystal cobalt]

One of the few things I've yet to try and control for, no. Have just been using a strength of 1.00 so far, and have been getting good results other than the x4 image/merging problem. Guess I'll start testing LoRA Strength next...

[sleek depot]

I had a look at comfy, I do love node based stuff.. but the examples seemed... yeah... like something I'm putting on the todo list for when I've gotten a better grip on what things do in webui

[crystal cobalt]

If this is directed at me... yes, training my own

[sleek depot]

check the description for the lora... some state to use lower values at times

oh

nvm then 🙂

[crystal cobalt]

I'm creating from scratch, yeah

[midnight kiln]

I think once you understand what happens in the webUI - moving to Comfy is not actually that hard. It's just a different way to look at much the same stuff from a more programmer-ish perspective rather than menus

[sleek depot]

Also on my todo list... last time around i did some dreambooth training for my own face, but seems there are cleaner methods now

[midnight kiln]

I have to assume something weird happened in training then - or your training data

but maybe just send it here and I can test if its the lora itself that is the issue?

[crystal cobalt]

Sure, good idea...

[sleek depot]

yeah. I'm not unfamiliar with node-red or davinci resolve so nodes are fine.. just want to know more about what I*m controlling ^^

[midnight kiln]

i assume a lora should work fine betwen comfy and automatic - they use the same models after all

[crystal cobalt]

One thing to note, is this is specifically trained on SD1.5 ...

[sleek depot]

btw.. did you have any idea about my question on why generation seemed to slow down when computer had no attention?

[midnight kiln]

rough instructions for this lora? what is it of?

[sleek depot]

like... now.. I've been away from the tab for the generation on purpose to see what it deos and the console is saying 47s/it but as soon as I tab back to it it'll jump forward

[midnight kiln]

and for what model?

[crystal cobalt]

It's for SD1.5. It's a hand drawn illustration style, old vintage logos, that sort of thing

[midnight kiln]

Process priority is the only thing I can really think of... but unless you are doing heavy stuff in other windows it shouldnt have that large of an impact

check if the GPU/CPU usage also actually goes down

because it may just be some buggy time display and not something real

[sleek depot]

yeah,. fans on gpu spin down.. then back up again when i return to the browser tab

[crystal cobalt]

Example:

[sleek depot]

its a bit annoying since starting a 20 min batch needs me to be there to babysit the tab instead of beign able to let to computer sit idle

[midnight kiln]

The lora seems to work nicely - and designs come out real nice too

a little heavy, so I found it best to run 0.7 or lower to not get too many puppies

I think your problem is settings in generation

furries not intended lol - but I don't hate it? 😛

model = Dreamshaper 8 (SD1.5)

[sleek depot]

hm?

[midnight kiln]

ping - test results

[crystal cobalt]

Awesome. So, none of the weird duplication-merging I'm getting...

What sort of params did you use?

[midnight kiln]

no, just too many heads sometimes. the models often equate "stronger" with "more" - so that's not unusual

[crystal cobalt]

Besides 0.7 strength?

[midnight kiln]

nothing special. just some simple prompts about shields and signs with wold crest. 1024x1024. DPS++ 3M sampler and 50 steps

worked fine at low res too

lora:LoRA_Test:0.5
(carved wooden tavern sign:1.2), (embossed wolf crest,:1.2) BREAK
medieval fantasy street at night with rain in the background, focus on sign

[crystal cobalt]

Alright... good to know it's not an inherent flaw in the LoRA itself anyway... thanks very much for taking the time to test! Back to testing params for me...

[midnight kiln]

no problem. now you have eliminated half the potential problem size at least 🙂

[tawny saddle]

2_

[sleek depot]

gah, the annoyance when you gen a picture that looks so promising.. but 6 fingers

[primal pendant]

hey before i even try can a 6700 xt use xl illustrious ?

[ornate elk]

Yes

[primal pendant]

ok nice what it work best on ?

forge ? automatic ?

is this still fine

[ornate elk]

Both will work. Maybe forge will work better as it also supports v-pred models

[primal pendant]

it's been hell of a long time since i last tried anything yeah i went straight for forge

[ornate elk]

Looks okay but rocm 5.7 is old now.
Your GPU can work with hip SDK 6.2.4

[primal pendant]

will it break anything ?

if i jsut update

[ornate elk]

But if everything works you dont need to upgrade. But it could have a better performance

[primal pendant]

well i'm up for better

[ornate elk]

Depends on how you do it

[primal pendant]

tell me safest way lol

and that will be way

[ornate elk]

Okay make sure forge is updated.
Check if you have git pull in the webui-user.bat

If not then add it into an empty line

And relaunch

[ornate elk]

Okay so the cmd shows "already up to date" at the start?

[primal pendant]

it does now yeah

so how can i update rocm without breakign anything?

surely if i just remove manually and then update it will break things no ?

[ornate elk]

You have to uninstall hip SDK 5.7
Then Download and install hip SDK 6.2.1
Then you would also need to download the correct library files for your GPU.
Then you have to replace them in the rocm library folder.
Its all explained in my Forge AMD Zluda guide.
Then you need to reboot and then
Delete the .zluda folder in forge
Launch Forge and it should compile and work.

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

[tranquil token]

For hunyuan videos where in the chain should you place lora nodes?

[crystal cobalt]

I finally figured out the image duplication/merging issue, and am now generating a single clear image! I had the Latent Image resolution set to 1024x1024, and apparently SD1.5 is native 512x512, so I guess the noise generation was happening x4 in each corner quadrant 🤦‍♂️

[ornate elk]

Have you deleted the .zluda folder before launching?

[primal pendant]

yeah

it created

[ornate elk]

Okay thats good but its strange that is shows failed to load zluda at the top

[primal pendant]

how long should i be compiling for ?

[ornate elk]

That could be an issue later

[ornate elk]

15-25mins
And the first image can take between 15-40 mins.

[primal pendant]

ok even though not new install ?

[ornate elk]

But the one error is a problem.
Something isnt correct

[primal pendant]

should i stop it then ?

[ornate elk]

Yep because changing zluda and hip version does need the one time recompiling everything

[ornate elk]

Yep

Try delete the venv folder and the .zluda folder again and relaunch

[primal pendant]

ok

[ornate elk]

Also have you downloaded the latest zluda files and placed them on C:\Zluda?

[primal pendant]

yeah

i just dread each time i attempt something with this do you remember last year the hell with permissions and things not working and the compiling every launch for 1 hour lol

i will give up it it does that again

so far so good

[ornate elk]

Ohhh you had these f upd permissions xD

[primal pendant]

yeah lol i used a 3rd party software to fix it after some googling lol

[ornate elk]

Yep I remember

[primal pendant]

that was hell

[ornate elk]

True

[primal pendant]

i was losing my shit cause i was watching videos and following guides and it wasn't doing what theres did

😭

but got there in end

you helped me alot

[ornate elk]

Yea it was a ride ^^

Hopefully its easier to fix now.
Have you replaced the Library files correctly?
And dropped the library.dll into the /bin folder?

[primal pendant]

yeah i followed all exactly

far quicker

all looks correct right

[ornate elk]

One error

It can't find a needed DLL from zluda

Means it won't work

[primal pendant]

oh

[ornate elk]

That can be caused by different things

[primal pendant]

i will check path

[ornate elk]

• Wrong environment path settings
• python installed from the windows store instead from the .exe

[primal pendant]

python was hell to sort to back then it's proer old one not through windows

[ornate elk]

If you find paths to hip_sdk 5.7
Delete them

[primal pendant]

oh i left 5.7 in the rocm folder

[ornate elk]

Why

It should be uninstalled

[primal pendant]

i unistalled it from pc the folder it left

[ornate elk]

Ah okay

Then delete that folder

[primal pendant]

yeah deleted it

[ornate elk]

Check the "Path" variable too

Where you should have %HIP_PATH%bin
And the C:\ZLUDA folder

[ornate elk]

Okay both are there

And a lot of paths to python and conda xD

[primal pendant]

yeah i added everything cause it was griefing me

so everything i could i added lol

bit over the top but worked

[ornate elk]

What you can try is to reboot the PC again.
Then delete the .zluda folder again and relaunch

[primal pendant]

ok#

[ornate elk]

Also did you downloaded the correct files for your GPU?
It has to be gfx1031 for Rocm 6.2.4

And zluda files are rocm 6 version 3.8.7

[primal pendant]

yep 1031

[knotty turret]

@ornate elk think u forgot about me😫

[knotty turret]

can we try to continue to find fix now?

[ornate elk]

Need to go off now as its late here.
Maybe we can take a look at it tomorrow again

But I'm a little bit clueless now

[primal pendant]

ok no worries tomorrow sounds good

yeah coming onto 12am here

[ornate elk]

If you have the time.
Try a complete fresh install of forge or Auto1111

[primal pendant]

oh god

you think update corrupted it then ?

i guess can redo forge tomorrow

[ornate elk]

Normaly it can't but I don't know if your system is an issue xD

[primal pendant]

my permisiions are good i'm certain lol

system all upto date and checked

[ornate elk]

Run an SFC /Scannow
As admin in cmd to be sure

[primal pendant]

ok

[ornate elk]

Not now. And sry I forgot you.
Maybe you can update me on the issue again

[ornate elk]

The python is something that also can break it

As you have a lot of versions and env paths

[primal pendant]

hmmmm

well i only did that as python was griefing me alot

[ornate elk]

Yea but its a mess now

[primal pendant]

it's only a mess if you look for it

[ornate elk]

Rocm hip SDK is correctly installed.
Also you gfx files are correct.
The only problems is ZLUDA and that error can be caused by python

[primal pendant]

i never had that issue on older folder

[ornate elk]

What I suggest is to try install Automatic1111 with zluda to test If that works

[knotty turret]

having trouble getting ComfyUI to start, even though Forge and Auto1111 are working fine.

In our last conversation, you suggested running pip cache purge and then starting comfyui-start again, but that didn’t fix the issue. After that, I also renamed the folder to remove special characters and reinstalled everything from scratch, but the problem persists

and u checked my python version and where but its finne too

[ornate elk]

What was your GPU again and which rocm hip SDK version?

[knotty turret]

RX6600 Hip sdk 6.1.2

[ornate elk]

Alright. I'm off now. Can help you troubleshooting tomorrow

[knotty turret]

Ok

[fair oxide]

The interrupt and skip buttons on automatic111 are so insanely slow that it's faster to just close CMD and close the tab and just reboot stable diffusion than to wait for it to cancel the image

Is there a fix for this or will I just have to keep doing what I am doing

[ruby yoke]

I think you found the only real fix. i have to do that with comfy too.

[noble idol]

There's a fix

Let me find it

Just a sec

go to settings and type 'live'

click this one

restart UI

should interrupt way faster

[fair oxide]

thanks I will try that

[gilded merlin]

I am using ComfyUI with a PulidFluxII workflow I downloaded and it keeps stopping on insightface loader. I have antelope and buffalo in the models folder, but for some reason it keeps sticking there. Any thoughts on what might be happening?

[ruby yoke]

are you by chance also trying to use diffusers?

[civic bramble]

Question....is flux models meant to have no negatives. I changed the cfg scale and it shows up. Can anybody give me tips on negatives. when I scale it up...its look really bad

[ashen gate]

Do you know why my Automatic1111 might take forever using controlnet openpose, but just fine without?

[ruby yoke]

i don't use auto1111 - but @ornate elk should. however you'll have to wait for him to log back on as he's asleep right now

[austere charm]

does controlnet happen to work with sdxl a1111 yet? it didn't work for the longest time so i just lost track of it and I'm not finding anything that says it works now

[whole skiff]

Hello, I have a problem, I use Stable Diffusion and I can't recreate any of the images from civitai even when I match every single setting 😭 For example here is image I was trying to recreate (image 1) and what I get in result (image 2) and PNG info of mine image info and original image info (image 3) and they are exact the same but results ALWAYS different. I'm just desperate, please help 🙏

[ruby yoke]

you could try this https://github.com/Mikubill/sd-webui-controlnet

[midnight kiln]

Sure, it works. used it to upscale real nice. learning how to use it for syle-transference now. Theres a whole lot you can do with it

*style

[lucid yoke]

If the author used tools/extensions they didn't mention it makes sense, my images i upload usually miss some info bc it's not fully recognized and its a hassle to complete it every time

Maybe a control net

Or a seed variation setting its "seed.3" or something else that doesn't show

[whole skiff]

I mean it's like that with EVERY civitai image for me, I even contacted 1 person and he told me all info of his image are displayed I matched 1 in 1 and again disappointment, I asked him to recreate some images that I of course can't for unknow reason, waiting for result

[lucid yoke]

Hmm do you happen to be on a amd?

[whole skiff]

No, rtx 3060

Also reinstalled Stable Diffusion and when trying to download extensions get errors

Maybe that's linked somehow

[lucid yoke]

Are you using --low vram or xformers?

[whole skiff]

I tried with xformers and without

Same result

I'm already thinking about reinstalling Windows

[lucid yoke]

Hmm xformers will cause you to get a different result every time

Shouldn't be that tbh

[whole skiff]

Wait a sec

[lucid yoke]

Your even using the same ui as them?

[whole skiff]

Yes

With xformers and without

Minimal difference, pose not even close to original

Original

But all png info identical

And same ui

A1111

I need to sleep for now, please someone help 😭

[ornate elk]

If the image creator used AMD, Intel or CPU to create that image you won't get the same with Nvidia

Hardware makes a difference

[ornate elk]

Make sure to select flux as type at the top left

[lucid yoke]

Also question of my own, why do you need to recreate the exact same image? If you want to inpaint/change it why not their image as a initial?

Just feeling theres more to your question why you want the exact same one. As the images you made are still good?

[sleek depot]

Is there a better way to import a prompt from civitai into webui? I do the copy all generation data and then click the blue checkmark.. but it never picks the loras which I have to manually add the weights for

[hexed torrent]

Why do I get this error when install an extenstion?

[sleek depot]

Seems like a problem with your revocation server

[hexed torrent]

What is that?

[sleek depot]

essentially.. it could be that or something cascading to that.. if you look at the error you got after fatal: it says it can't access the repo.

possibly since there seems to be a trailing slash after .git which seems odd.

I'm guessing you are just clickling an install button? or have you pasted the url for the repo?

tied visiting the repo.. which is there.. so might be the revocation server.. which could mean something with your git settings for the installation.. have you been able to install other extensions?

[hexed torrent]

Yeah, I've got other extensions installed

[sleek depot]

Ok.. so.. you have a virtualenv for your install?

[hexed torrent]

Nope, it's installed on my server pc with only 1 OS

[sleek depot]

Ok.. so I have a possible quick fix, but it includes setting a setting that I would suggest not keeping after you install if you are interested

If the issue is in fact the revocation check, you can disable that globally in git settings with:
git config --global http.schannel.checkRevoke false
I'd make sure to restore it once you are done though, since that isn't something suggested to run in the long run

(restore by setting it back to true)

your milage may vary ^^

[woven elk]

I'm sorry if I misunderstood some things, but for like a year now, I am not able to fix stable Diffusion / SDXL generating horrible scuffed images like this. I'm running on a 8gig 3060ti with 32gigs of Ram, currently using A1111, but ComfyUI didn't work either. If it's a common bug let me know, request anything if I have to provide further information

[jovial mist]

Hello Guys, do you know how to change the default generation resolution from my multifotm api request ?

Can i do that with sd3-5 medium ?

[tawny saddle]

tbh

das neat image

looks like wut it'd maybe look like to starew into a diamond wit all the light refracting into different wavelengths

[tawny saddle]

dis

with the 21 meme in the center

xd

[woven elk]

a bit, but yknow, I asked for a volleyballer photorealistic

or a castle, im not sure

[tawny saddle]

xd

does the model u using have that as a potential trained asset '

maybe it messed up cuz it dunt, like asking for a boeing 787 dream liner from a model traind on fish

[vocal burrow]

what resoloution are you generating for ?

[woven elk]

It's 512 by 512 (I know SDXL is meant for larger things, but it won't look good with usual SD either)

[vocal burrow]

it's worse than "is meant for", SDXL will break and output nonsense results if you don't output for something close to 1024x1024.

[woven elk]

I see

[vocal burrow]

Regardless of the trained resolution you can usually diverge 25% above or below the trained resolution in one or both dimensions.

(so for SDXL it's ok to go 756x1024 for instance)

[woven elk]

mhm ok

and

Is there a way for me to actually run it in that quality because at least in A1111 those resolutions only caused more trouble...

[vocal burrow]

And from what I see

photorealistic, highly detailed, real, image of a Volleyball player.Negative prompt: drawn, scuffed, comic, 2D, disconnected limbs, more than 6 Fingers, less than 4 Fingers, cruel, extreme Kontrast.Steps: 50, Sampler: DPM++ 2M SDE Karras, CFG scale: 7.5, Seed: 1112568318, Size: 512x512, Model hash: e6bb9ea85b, Model: sdXL_v10VAEFix, Clip skip: 2, Version: v1.5.1

You were using an SDXL model here

[woven elk]

yes

wow is that in the metadata?

[vocal burrow]

yup

[woven elk]

teach me how

nvm got it

[vocal burrow]

"run it in that quality" what do you mean by that ? With an 3060ti you have 8gb of vram if I remember so you'll have to use --medvram-sdxl command line argument

Cf pinned comment for a detailed guide on how to install and set A1111, forge, etc

[woven elk]

ok, I'm currently trying to install forge, I'll tell you if it works when im done, thank you for the help and infos

[gilded merlin]

For some reason in my reactor node I have a slot for swap_model and inswapper is not showing up (ComfUI). Where should I drop inswapper in my folder structure for that to show up for that particular node?

[tawny saddle]

das not necessarily tru

I believe it largely depends on the model and wut those models was trained on, regarding image size, at least in my experience, I've made many a good image with 512z512, those now days I use something like 620x520, or around that

as chu can see

that being said, I believe that if a model was trained on images with like 4k res, any low or too low, at least that id which is significantly lower than the trained res, may out out gibberish, and models trained on primaraily smol images, may kinda get messed up at higher res

jus gotta kinda experiment

[vocal burrow]

That's what I'm saying...
SDXL, SD3 are trained on 1024x1024
SD2.1 on 728x728 768x768 (fingers slipped :p)
SD1.5 on 512x512

Sure you can try to generate for half resolutions but most likely it will break, it won't be able to apply the patterns it learned properly.

just like it will break if you attempt to generate at twice the resolution (witout any kind of guidance, tiling, hresfix, etc)

[whole skiff]

I contacted creator of images on civitai he has same ui and also nvidia and png info matches 1 to 1, but still no match in result

[oak pond]

Hi, How do people with RTX 50s train loras in flux right now ?

[chrome rampart]

Hey all. Anybody know if/how you can use flux controlnet options like pose/reference in Forge?

[proud terrace]

Hello,

I have an Image with 7000 x 5000 resolution and I want to use "img to img" on it.
My PC cant handle that, so I always seperate the image into tiles manually in Photostop and use IMGtoIMG on each tile, then manually reassemble them in PS.

Is there a way to automate this process in Stable Diffusion? I only find tiling tutorials for upscaling. But I dont want to use any Upscaler, only IMGtoIMG with the actual model. Thank you.

[ruby yoke]

@fair oxide might know

[fair oxide]

multi diffusion could help with this
it splits into tiles and then merges the trajectories at the edges so there are no seams

[tawny saddle]

ahh got u ;o

7kx5k? ;0
Mama mia that's some resolution

[ashen gate]

Is it normal for automatic1111 to take 10x as long to generate something when using controlnet? I've selected Low Vram and I get an error:

"Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)"

[tawny saddle]

now, im nu stable diffusion engineer, but sounds to me like its saying the models u used, one is for cpu and one is for gpu, or, that one is using cpu, and one is using gpu

[feral ether]

\photorealistic, highly detailed, real, image of a Volleyball player.Negative prompt: drawn, scuffed, comic, 2D, disconnected limbs, more than 6 Fingers, less than 4 Fingers, cruel, extreme Kontrast.Steps: 50, Sampler: DPM++ 2M SDE Karras, CFG scale: 7.5, Seed: 1112568318, Size: 512x512, Model hash: e6bb9ea85b, Model: sdXL_v10VAEFix, Clip skip: 2, Version: /

[tawny saddle]

i agree

[formal field]

can someone help me with a launch issue?

(after this it normally works ok, just occasioanally hanging at 50%)

when i launch my first queue of a session sd always gives me an error at 20% and hangs for ages, slowing or stopping the pc to a crawl and occasionally crashing it or giving a black screen or other fun things, if i manage to hit the cancel button it still takes minutes for it to act on it. Using 3.5 large but seems to be the same with turbo and medium. Using a 3080ti on Win11 and this is what it says when it takes a nap:

VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
clip missing: ['text_projection.weight']

[lucid yoke]

its trying to offload to the cpu. do you have this error with SDXL too?

[formal field]

no idea what that is, just installed it a few days ago 😮

but guess that explains why the pc pretyt much dies sometimes, no idea why it tries to use it over gpu, i mean 3080ti is no slouch

[lucid yoke]

try using a SDXL model from CivitAI.
Sd3.5 large turbo or the normal one can't both fit in 16gb vram, they require 24gb at the least. IIRC

so its failing to offload to the CPU

what UI are you using to generate?

[formal field]

i got 32gb ram

ddr5 even

[lucid yoke]

VRAM***

[formal field]

comfy

[lucid yoke]

videoram (graphicscard. yours has 12GB)

hmm im not familliar with comfy to edit its launching parameters like xformers or lowvram etc to let you run it better

[formal field]

ah ok, i tried the 8gb tenesor but had the same issue

[lucid yoke]

hmm since your new its safe to asume your fully new to stable diffusion?

[formal field]

the odd thing is that after the initial crash attempt it usually works kind of ok without slowing the pc that much, mostly

yep all new to it

[lucid yoke]

yeah thats the offloading

hmm well if you find comfyUI intimidating theres some other UI's you could use to make the image making process easier on the eyes

yes its the best of the best but its quite complicated for new people

[formal field]

yeah t looks like a school project

but looks are deceiving i think

[lucid yoke]

i recommend using https://github.com/mcmonkeyprojects/SwarmUI
It has a comfy UI back end (having the benefits and optional to mess around it) while still having a easy to use GUI with a lot of (?) to explain what it does

[formal field]

thanks i'll have a look, but that or sdxl won't fix my offload issue right? though i am confused why i still get it when i switch the 16 for the 8

[ashen gate]

Is it normal for automatic1111 to take 10x as long to generate something when using controlnet? I've selected Low Vram and I get an error:

"Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)"

[lucid yoke]

hmmm well im on a 3070TI and its running sdxl just fine. you could run 3.5 medium maybe but i think flux has more features youd like

[lucid yoke]

what control net are you using? but generally yes it can increase generate times. especially lineart, openpose and such. canny isnt that laggy

what graphics card do you have?

[ashen gate]

RTX 3090

[lucid yoke]

well you dont have low vram to be sure but okay. what checkpoint?

[formal field]

is flux online only?

[lucid yoke]

no

[formal field]

hmm interesting, flux seems to generate better realism

[ashen gate]

Extension is "sd-webui-controlnet version 56cec5b2". Checkpoint doesn't seem to matter as it happens with all of them, but "https://civitai.com/models/439889?modelVersionId=1199750"

[lucid yoke]

remove the --lowvram in your ui launcher and restart it

[ashen gate]

low vram is a checkbox i ticked not something i inserted in the launcher. ill uncheck it

[ornate elk]

Which Controlnet model did you used?

[lucid yoke]

ah cs1o is here. correct me if im wrong but with 24gb of vram. the lowvram tag will slow it down right

[ashen gate]

idk how to check. it was just in extensions available tab

[lucid yoke]

what were you trying to do with control net?

[ornate elk]

Do you use Auto1111?

[ashen gate]

pose a figure. they always tend to look lanky and distorted when i use the stick figure

[formal field]

probably not, nothing installed besides comfyui, the sd clips and a workflow

[ornate elk]

Ah okay, do you run any stuff in the background like games or WallpaperEngine?

[ornate elk]

You need controlnet models to use the controlnet extension

The models need to be downloaded manually