🤝｜tech-support

[vocal burrow]

it happens to all of us. or so I ve been told

[civic nova]

I refuse to accept it, I am going out like I came in Kicking and screaming

ok Seriously opened CMD line as admin

[vocal burrow]

you dont need to run it as admin but it s okay it should not mess everything

[civic nova]

tried cd f:/ to access drive but nadda what am I forgetting here?

[prisma moss]

you didnt open the cmd from the search bar in the folder

[civic nova]

Ahh ok try this again

[prisma moss]

nav bar whatever is called

[vocal burrow]

to change disk you need to type f:(or x: or whatever)

forgot about that my bad

[civic nova]

ok back step, lost line of thought here. Opened file explorer, opened SD Folder at search line top right what now?

[prisma moss]

go into webui folder

[vocal burrow]

type cmd in here

[civic nova]

there

[vocal burrow]

press enter

then run git remote show origin

what does it say ?

[prisma moss]

not there 😭

[vocal burrow]

here , not there

[civic nova]

ACKKK

Would screen share help? with voice?

[vocal burrow]

i m at work. I cant do voice chat and stuff like that

[civic nova]

ok back at local Web-ui folder

[prisma moss]

open cmd and run the command above

[civic nova]

ok that is where I am losing y<ou, open command? Where

[prisma moss]

here

in webui folder

not stable-diffusion-webui

just webui

which is found in stable-diffusion-webui

[civic nova]

ok I have several of them in that folder, user.bat, user.sh. webui.bat, and webui.py

[vocal burrow]

this folder

[civic nova]

I have no specific webui folder

yes that is where I am at

[prisma moss]

hes on a1111 then

in essence this

[civic nova]

timed out

ok I am back

[vocal burrow]

@civic nova don t send too many pictures at once. Otherwise our bot will time you out indeed

[civic nova]

ACKKK

this is where I am at

[prisma moss]

look at the red text on this image

[civic nova]

ok followed the vid, but you did it waaay too fast LOL..

CMD is open in the webui folder

[prisma moss]

git remote show origin

[civic nova]

Microsoft Windows [Version 10.0.26200.8328]
(c) Microsoft Corporation. All rights reserved.

F:\Stable Diffusion\stable-diffusion-webui>git remote show origin

• remote origin
  Fetch URL: https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
  Push URL: https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
  HEAD branch: master
  Remote branches:
  Authenticated-Sysinfo tracked
  Forbid-too-small-crop-region---compact tracked
  SD3-Lora-page-filter---detection-not-implemented tracked
  add-Usage-Syntax-to-Prompts-from-file-or-textbox tracked
  add-middleware-after-app-has-started tracked
  bump-Pillow-blendmodes-dependency tracked
  canvas-undo-hotkey,adjust-brush-size-by-area tracked

compact-Soft-Inpainting-infotext tracked
consistency-dec tracked
cu126-10-series-gpu new (next fetch will store in remotes/origin)
dev tracked
fix-sd_hijack-infotext-cased-by-conda-cache tracked
gradio-4.x-test tracked
gradio4 tracked
hires-button-opt-insert-gallery tracked
lora_emb_bundle tracked
master tracked
method-3-fix-sd_hijack-infotext-cased-by-conda-cache tracked
more-extra-transparency-fix-again tracked
option-reduce-connecting-to-huggingface tracked
reduce-server-request tracked
refiner_alt tracked
release_candidate tracked
revert-use-of-gitpython-hijack tracked
server-reload-API tracked
Local branch configured for 'git pull':
master merges with remote master
Local ref configured for 'git push':
master pushes to master (up to date)

F:\Stable Diffusion\stable-diffusion-webui>

[prisma moss]

yea

[civic nova]

Hit me enough times with a brick and I usually get it LOL

[prisma moss]

thats some outdated stuff by now

[vocal burrow]

ok so you re using automatic1111 which is indeed very outdated

[civic nova]

It was like 3 years ago. A lot has been happening here life wise.

[vocal burrow]

What gpu are you using ?

[vocal burrow]

it s okay. And AI stuff move pretty fast.

[prisma moss]

ait hes cooking

[vocal burrow]

Ok then I d recommend you to install forge neo by following this guide. https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-forge-neo

Keep your already existing Automatic1111 install folder for now to save you from having to redownload all your models.

[civic nova]

so any of that help?

[vocal burrow]

The file structure and UI between Automatic1111 and forge neo is near exactly the same. So you should not be lost.

[prisma moss]

yes but illd avoid using speccy

[civic nova]

it is old I know.

[prisma moss]

not only old, but inaccurate, and i believe it had secuiry issues

[civic nova]

So rename the current Stable diffusion folder?

[prisma moss]

no

make a new one

[civic nova]

so at install simply give it a new folder name no aut update ot the old folder?

[prisma moss]

do you have anydesk?

[civic nova]

anydesk?? Nadda, plain old windows 11

[prisma moss]

i mean its a program for remote control

[civic nova]

not to my knowledge.

I had a rmoete desktop app once but not sure if it is here anymore.

[prisma moss]

reinstall anydesk and lmk

that is if u want me to do it for u

[civic nova]

links?

[prisma moss]

https://anydesk.com/en

[vocal burrow]

Just know that anydesk will give full control of your machine to them. So only do it if you re comfortable with it and DO remove it afterwards. Otherwise try to do install forge neo yourself by following the guide I sent you.

[prisma moss]

👆

that sounds a bit threatening but ok xD

[civic nova]

but it is oh so true

[vocal burrow]

I mean I do have to warn users about that.

[prisma moss]

it is

[civic nova]

As mentioned I think it best we stay manual for now.

[prisma moss]

then follow the guide

[prisma moss]

👆

[civic nova]

That link sends me back to install Python 3.13.12

[prisma moss]

as you should if u havent already

https://www.python.org/ftp/python/3.13.13/python-3.13.13-amd64.exe

[civic nova]

ok so I have Git-2.54.0 64 bit installer and Python 3.13.12 installer downloaded

[prisma moss]

u already have git installed

so just install python, and during the installation (first step) check the ADD PYTHON TO PATH box.

[civic nova]

Welcome to the Python installation manager configuration helper.

---

Your app execution alias settings are configured to launch other commands besides 'py' and 'python'.

This can be fixed by opening the 'Manage app execution aliases' settings page
and enabling each item labelled 'Python (default)' and 'Python install
manager'.
If the items are already enabled, you may need to disable and re-enable them.
Open Settings now, so you can modify App execution aliases? [y/N]

[prisma moss]

Screenshot would help

[civic nova]

that is all that it says

[prisma moss]

Open the python installation

Check the add to path option

And go next next next till it finishes

[civic nova]

here is vid so far see no add path

[prisma moss]

thats an install manager

[prisma moss]

downlaod this

remove that from app store

[civic nova]

ok back step... uninstalled that instillation, now back to square 1

[prisma moss]

did u just remove the executable or actually uninstalled it fully from app store

[civic nova]

went into settings and uninstalled

[prisma moss]

ait then install what i sent

[civic nova]

not sure what remove form app store is.

it does nadda when I click it.

[prisma moss]

check if it went directly into ur downloads

if not then

https://www.python.org/downloads/release/python-31313/

[civic nova]

anyway I can share screen in here?

[prisma moss]

idk

go in voice channel and see there

[civic nova]

I now have in downloads an install manager and a 3.13.134 exe file. Run the exe?

[prisma moss]

ignore the manager

and yes run the 3.13.14

[prisma moss]

check both

[civic nova]

ok

customize instillation?

[prisma moss]

no just install now

etc

[civic nova]

ok installed

[prisma moss]

then make a new folder where u want the neo to be installed

open the folder

run the cmd like before inside the folder

[civic nova]

ok hold a tick

[prisma moss]

you can do it at ur own pace

when cmd is open: paste git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo

[civic nova]

cloning

done

[prisma moss]

right click the webui-user.bat file

and open it in notepad

add these arguments --cuda-stream --sage --fast-fp16

[civic nova]

@echo off

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

set COMMANDLINE_ARGS=

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

call webui.bat

[prisma moss]

--sage may or may not work depending if triton is installed more on that later

`@echo off

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

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

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

[vocal burrow]

extra ` at the beginning there

[prisma moss]

good catch

[civic nova]

ok lost again back step, note pad open in edit, add what where?

[prisma moss]

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

[civic nova]

@echo off

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

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

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

call webui.bat

??

[prisma moss]

yes

[civic nova]

next?

[prisma moss]

save and try running

same file

[civic nova]

Couldn't launch python

exit code: 9009

stderr:
Python was not found; run without arguments to install from the Microsoft Store, or disable this shortcut from Settings > Apps > Advanced app settings > App execution aliases.

Launch Unsuccessful! Exiting...
Press any key to continue . . .

[prisma moss]

https://tenor.com/view/forehead-tired-sweaty-gif-8579944

[civic nova]

should you not have like a nurse swabbing your forehead???

If I am reading this correctly< there is something in default settings that has ot be turned off right?

[prisma moss]

sec

:: set PYTHON="C:\Users\YOURUSERNAME\AppData\Local\Programs\Python\Python313\python.exe"

try setting this

check if u have python 3.13 installed in: C:\Users\YOURUSERNAME\AppData\Local\Programs\Python

[civic nova]

we are in the web ui user bat right not the web bat?

[prisma moss]

yes

[civic nova]

just making sure, trying to run again

same message

[prisma moss]

did u check if the python 3.13 is in that folder?

[civic nova]

can you use voice or are you at work also

[prisma moss]

i can

u can call me

[vocal burrow]

run where python, python --version and py --list-paths say again please

[orchid patio]

realize kuli bangunan indonesia have wife and son daughter on stage

[prisma moss]

i got him up n running

[civic nova]

Yep it was a mind bender but he helped me get it running thanks to all agaion

@vocal burrow should get this guy on staff his patience is extrodinary...

[prisma moss]

😅

i believe Maxfield is the one responsible for that, but lets not bug him

[civic nova]

Later folks gotta go kill some plants...

[prisma moss]

Have fun

[vocal burrow]

it was a path problem, wasn it.

[prisma moss]

he didnt install the venv

the guide lacks that

it doesnt get automatically installed so i had to use the haomings steps with uv packages, and setting up the venv

[chrome forge]

can anyone help me?
Comfy ui with zluda
im using the cyberrealisticZimage_v40 model also it says model type = flow in cmd
Clip - qwen_3_4b_fp8_mixed
Vae - ae.safetensors
I have an image that i want to use like that face should be the face in generated image how can i do that? im having all type of issues with ipadapter is there any other way without ipadapter or can it work with it?

@ornate elk sry for the ping man ive been dealing with ts since yesterday

[lucid yoke]

Your trying to do a face swap basically?

Facetools reactor, custom comfyui nodes that can do that in a safe way

Since theres a nsfw filter on it

[chrome forge]

yeah kinda

[chrome forge]

wdym?

[lucid yoke]

Well the custom node checks for any "nsfw" material and if theres any detected in the image. It wont do the face swap

[chrome forge]

ahh okay

[chrome forge]

so what if the character is in something like a bikini? will that work?

[lucid yoke]

Probably not

I dont face swap but i know the node exists

[chrome forge]

okay imma try reactor, i generated a face on google flow now i just need that face to be on the images i generate with comfyui

[lucid yoke]

Reactor on its own is very lightweight but theres also a way of using something like flux klein 9b to swap the faces

It helps for making bigger datasets if your making a consistent character lora but theres simply not many images (such as a niche game character)

[chrome forge]

i dont think i can make lora cus im using an rx 6700 xt and i cant pay online to rent a gpu

[lucid yoke]

Well reactor worked on my 4gb gpu work laptop with no issues

[chrome forge]

nah im talking about creating a lora

[lucid yoke]

Fair enough!

[round sail]

Can I get some clarification about the forge-neo command arguments:

1. I keep getting conflicting ai advice about --sage and (--disable-sage)
   I AM USING WINDOWS 11 HOME VERSION 24H2 Build 26100.7171
   Graphic Card: Nvidia Geforce RTX 4070 Ti Super (16gb)
   Intel(r) Core(TM) i7-147000kf (32gb)
   GPU driver version -- 4/22/2026 =Version 595.79, Cuda compilation tools, release 12.4, V12.4.9, Build cuda_12.4.r12.4/compiler.33961263_0
   Video BIOS version -- 4/22/2026 Video BIOS Version 95.03.45.40.48
   BIOS Version/Date: American Megatrends International, LLC. F7c YY, 7/8/2025
   BIOS Mode: UEFI
   BaseBoard Manufacturer: Gigabyte Technology Co., Ltd.
   BaseBoard Product: B760 DS3H AC
   SMBIOS Version: 3.6
   Embedded Controller Version: 255.255
   NVIDIA Studio vs Game Ready drivers -- 4/22/2026 = NVIDIA Game Ready Driver (GRD), Version: 595.79, Driver Type: DCH
   CUDA version -- 4/22/2026 = CUDA 12.4 (Build: V12.4.99, compiler build: cuda_12.4.r12.4/compiler.33961263_0)
   Python environments -- 4/22/2026 = Python 3.13 (64‑bit) — Default environment, Python 3.10 (64‑bit)
   So, the dilemma is that when I try to gen wan videos, I get a lot of "Encountered NaN in Latent processing.py :: WARNING
   If you are using SageAttention, try --disable-sage"
   Not always but almost always

[prisma moss]

https://tenor.com/view/vae-gif-15357506855217945854

[round sail]

So after considerable trial and error (and a lot of hair pullng) I came up with this method for my system:
🚀 Forge-IMAGES (webui-user.bat)
set COMMANDLINE_ARGS=--theme=dark --sage --flash --cuda-malloc --pin-shared-memory --fp16-unet --fp32-vae --autotune --no-hashing
The "Sprint" Launcher: Optimized for high-speed, native 16-bit math. Use this for all standard image generation.
SDXL (XL)
ANIMA
KLEIN
ZIT (Z-Image / Z-Image Turbo)
SD (Standard 1.5 / Legacy)
🎬 Forge-VIDEO (webui-userVIDEO.bat)
set COMMANDLINE_ARGS=--theme=dark --sage --flash --cuda-malloc --pin-shared-memory --bf16-unet --fp32-vae --no-hashing
The "Safety" Launcher: Optimized for heavy 32-bit VAE stability and BF16 math. Use this for high-precision models and all video generation.
WAN (Both I2V and T2V)
FLUX
LUMINA
QWEN
I just want to say that the UI Preset System that was introduced in Forge-neo is AWESOME, besides the command Arg settings (shown above) I was able to update the presets in the settings menu and your implementation where the presets remember that last used Model, VAE, and Text Encoder is just genius. Thx Guys, really enjoying the fork!

[pearl timber]

yo guys im trying to install stable diffusion but i keep getting this error anyone help pls?

Command: "C:\Users\shybo\OneDrive\Desktop\stablediff\stable-diffusion-webui\venv\Scripts\python.exe" -m pip install https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip --prefer-binary
Error code: 1
stdout: Collecting https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip
Using cached d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip (4.3 MB)
Installing build dependencies: started
Installing build dependencies: finished with status 'done'
Getting requirements to build wheel: started
Getting requirements to build wheel: finished with status 'error'

stderr: error: subprocess-exited-with-error

Getting requirements to build wheel did not run successfully.
exit code: 1

[17 lines of output]

[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed to build 'https://github.com/ope`nai/CLIP/archive/d50d76daa670286dd6cacf3bcd`80b5e4823fc8e1.zip' when getting requirements to build wheel

[ornate elk]

Hey, what's your gpu?
Also Auto1111 is outdated and shouldn't be used anymore.
The updated setup guides are in the pinned messages of this channel.
Also don't install stuff on sour desktop location as it seems like yours is synced with OneDrive.

[pearl timber]

rtx 5050

also im using the forge one now
and i got it to run once by using cmd in the stable diffusion file but if i try to run it using the webui batch i still get an error

[ornate elk]

Okay, then you should use Forge Neo

And only launch it with the webui-user.bat

[livid phoenix]

Hi

I am starting to build parts for my new pc and wanted to know what new possibilities (or improvements) this new build

For now I only do image gen (mostly illustrious) and LLMs (LM studio) but want to slowly try other things like video/sound/music/agents.

What are the best things for each branch I could run with a setup like this ?

Actual build

• CPU : Ryzen 5600X
• GPU : 3060 12gb
• RAM : DDR4 32gb 3200MHz

Projected build

• CPU : Ryzen 7 7800X3D (already bought)
• GPU : 5070 Ti 16gb
• RAM : DDR5 32~64gb 5600~6000 MHz

Thanks.

[ornate elk]

5070ti is good, the more vram the better

[hidden night]

Guys can iPad pro M5 with 12gb of shared ram run Illustrious? (7gb)

[hidden night]

3090 for LLMs & Video
5070 ti for image

it's up to you but 5070 ti would be more reliable.

as a 5070 ti owner the only model I'd even consider using is Qwen 3 coder 30B

Efficiency will probably get better in the future but running LLMs locally is a stupid idea IMO. they need much better hardware compared to image & video

[ornate elk]

Yes it can. You should checkout the DrawThings app

[hidden night]

Does it support LoRAs? Can my ipad run it with LoRAs?

[ornate elk]

Yep they should work

[hidden night]

Have anyone tried it and is there a video or something? this is something extremely unpopular

[ornate elk]

Idk, but here is the link to the app:
https://drawthings.ai/

Its the best you can get for an iPad. Feel free to test it out

[hidden night]

I'll purchase an iPad so I'm thinking is it worth to buy a Pro M5 instead of Air M4

The M5 is insane, most powerful mobile CPU

but much more expensive too

[hidden night]

Running draw things on anything other than M5 is useless though

[ornate elk]

the newer the better yes, it works on macos too

[hidden night]

M5 ai accelerator stuff makes a huge time difference

Uhm

When is the next apple showcase thing?

Would it be worth waiting

[tawny saddle]

apple

pear phone

[lilac idol]

Hi,
today I installed openterminal with multiuser active. In the terminal, all the permissions work well. Problem is, multi user permissions can be fully bypassed by any unprivileged user via the file browser. This goes as far as having the ability to see and open every file except other user home direcories. Had anyone else had this issue?

Nice, was it resolved by a update?

Sorry I am not going to follow some random discord link, but appreciate the help! :D

[lucid yoke]

its not

and hes banned

[lilac idol]

True xD thanks

[lucid yoke]

yeah sometimes fake support URL spammers come in here, sorry for that

[lilac idol]

All good, no need to apologize ^-^

[lucid yoke]

personally i cant help with this issue since i dont use that but i really recommend popping into the github of open terminal instead

[celest delta]

how do i prevent forge neo to use so much memory even when idle? I didn't have that issue when i was using auto1111

[lucid yoke]

isnt it just the model being loaded in memory and the current models being much larger then in the A1111 sd1.5 era

iirc theres a unload option after X amount of minutes in idle, not sure

[celest delta]

no idea, but i'm using the same models i was using before, illustrious stuff

[lucid yoke]

how much memory does it use and what gpu?

@celest delta forgot to use the reply option oops

[lilac idol]

Gonna do that, thanks!! :D

[celest delta]

rtx 5080, using 80% of my memory in task manager (so it's not vram i think), and after a while the webui just slows down a ton, even if it's not generating, forced to close it completely and open it again

[lucid yoke]

probably a memory leak, did you update forge at all? and what browser do you use to acces forge neo?

[celest delta]

firefox, and it updates automatically, also it's been that way since the start i've been using forge neo, a few months ago

just i'm getting tired of it, lol

[lucid yoke]

did you install it thru matrix by any chance?

[celest delta]

i don't remember, i don't think i used anything named matrix

also not related, but should i ask here for help on how to use anima? despite following tutorial i can't get any decent image

[lucid yoke]

for anima prompts i actually do recommend hopping into #🍥｜anime , the guys there are using it so they know more then me :P

[kind dove]

Kohya SS

I’m looking for help creating a LoRA training Kohya_SS config for my dataset of 33 high-quality images. I’m using an RTX 4060 with 8GB of VRAM, and my goal is to get the best possible quality from the LoRA.

I used WD14 Captioning

I would appreciate it if the config was completely ready, Just that I just have to change the folder destinations.

I’ve asked ChatGPT and other AI tools for settings, but the configs I’ve gotten haven’t worked well. I’m still new to this, so I’d really appreciate it if someone experienced could take the time to help me make a proper config.

Thanks in advance.

[lucid yoke]

your trying to do illustrious training?

[kind dove]

realistic

sdxl

Is ai toolkit better??

i'm really new to this I'm sorry

[lucid yoke]

No worries!
i do recommend though checking the github page for documentation as they often provide examples https://github.com/bmaltais/kohya_ss#sdxl-training

i personally use OneTrainer but it shouldnt be too big of a difference.
https://github.com/kohya-ss/sd-scripts/blob/main/docs/sdxl_train_network.md

[kind dove]

:))

[thorny creek]

Hello! I want to see if I can use any closed source models in comfy UI for video generation. I heard runway is good. Any suggestions?

[lucid yoke]

the api provided workflows can be used yes.

[kind dove]

can we talk in dm?

:))

[lucid yoke]

why?

[kind dove]

i need help

we could talk here too

im not a scammer

lol

:))

[lucid yoke]

im not too familliar with koyha SS

[kind dove]

Okay no problem, if you are not comfortable enough to talk in dm

it's ok.

Do you recommend training LoRA's with 8GB VRAM?

btw is there a way i can train the LoRA for all models?

Do you know the game "Five Hearts Under One Roof"?
I have 33 high quality images of the character named "Gran"

I wanna train my models but theese apps are tooooooo complicated

I'm right now installing "OneTrainer"

https://klipy.com/gifs/fineignoremethen-fine-ignore-me-then--k01KR4B9ZMCPJ8T8KNR3FVYJ0TF

wow this is complicated

[lucid yoke]

im not iginoring, im also busy

[lucid yoke]

i do not know

with 8gb training anytthing beyond sdxl is gonna be very difficult

[kind dove]

I choose

[lucid yoke]

yes

[kind dove]

where do i put the images

[lucid yoke]

concepts

[kind dove]

add concept?

[lucid yoke]

but i really recommend watching a video as i cant really do a full hand held walkthough, but yes add concept

then select the folder

open the concept

[kind dove]

on path i choose the folder where my images and WD14 captions are

[lucid yoke]

yes

[kind dove]

prompt source

should it be from image file name?

1.png
1.txt

[lucid yoke]

yes

[kind dove]

theese text variations loss weight etc.

should i touch them?

[lucid yoke]

if you dont know what your doing i would recommend leaving them as default

[kind dove]

okii

i have

2348 woman images

idk what they were called

but something like

Re...

do i have to add them somewhere?

it was to like teach ai what "woman" are

[lucid yoke]

??

[kind dove]

oh

REGULARISATION

[lucid yoke]

i do recommend maybe asking this in the onetrainer discord as i mostly know about the models itself

[kind dove]

okay..

do i just press start training now?

[lucid yoke]

i dont see what your doing so i cant tell but if you did everything right and added a sample prompt sure

[kind dove]

nope i didnt

i just made the config #sdxl 1.0 loRA

and added a concept

with the images i have

and i added prompts

Do i need a hugging face token? isnt it all local?

[lucid yoke]

no its local

i really really recomend reading the documentation

[kind dove]

😭

[thorny creek]

@lucid yoke hey! Do you know if one could use a runway model (like aleph) on comfy UI?

[lucid yoke]

#🤝｜tech-support message

if theres a API workflow, you can. otherwise, you cant

[thorny creek]

There is!

[thorny creek]

They have API Do you have any YouTubers who have tutorials on this? I tried looking it up but I didn’t find anything. More precisely, I might not know what to look for :p

[lucid yoke]

hmm? i think the workflow should have documentation included. do you have a api key? (DO NOT POST THIS)

[prisma moss]

@ornate elk

@vocal burrow @lucid yoke

https://tenor.com/view/ping-mod-megumi-mahoraga-gif-12791569535980489290

[lucid yoke]

Cleaned, thanks 🫡

[mellow rock]

❤️

[round sail]

My method is to use the .json files. I load my base settings with a ,json, then add the path locations, pick a model to train on, run the WD14 caption utility, then combine that set of caption files with ones I made with Joy Junction captions (using a3rd party app called taggui.exe) then I use the Prepare Data Files button in Kohya_ss to set the folders up, then add the Repeats and Epoch settings based on the number of images. Here is a sdxl .json file (i create one for FLUX and One for Animie, etc) using the .json files and then loading them in kohya_ss (load config drop down) will save you tons of time, all of the settings in the .json files are alphabetical, you dont have to search all over the place in the GUI for the setting you want). Here is my sdxl ,json file: NOTE it is for a Graphic Card: Nvidia Geforce RTX 4070 Ti Super (16gb), Intel(r) Core(TM) i7-147000kf (32gb) system and will need to be adjusted. Dump the .json into an ai chatbot, tell it its for a 16gb gpu and ask it to give you update to make it work for an 8gb system. Dump it in Kohya_ss and begin trouble shooting. Dump the screen traces into the ai chat to get ideas about what to fix. Its a beast of an app, but it makes good LoRAs. GOOD LUCK, watch the videos - they help

[kind dove]

thank you so much 💞

thanks for the time

[ornate elk]

can you show the complete cmd log?

and when do you get the error?

[kind dove]

can we talk in dm

:))

[ornate elk]

sure

[kind dove]

I added my ESRGAN "Anime6B.pt" but i cant see it on hires.fix Upscaler

[prisma moss]

did u add it to esrgan folder?

[kind dove]

yes

[prisma moss]

did u restart?

like close-open

[kind dove]

not the cmd

only f5

[prisma moss]

lmao

[kind dove]

LOL

[prisma moss]

f5 does nothing

[kind dove]

this guy says

"Light Cheery Atmosphere"

how can i find this style

ive been searching for it on civit

[prisma moss]

idk mate

[kind dove]

oki

[kind dove]

on some models it says sd 1.5

Will it work on illustrulattisonstatia or SDXL?

[lucid yoke]

No, sd1.5 is a different model

[kind dove]

😭

[round sail]

FYI, in the old forge-UI there was a massive number of styles you could invoke and edit in the gui. In Forge-neo they dumped it for a slimmed down version, but it is easy to restore, simply drop this in the forge-neo root folder. The one I uploaded is an edited copy of the original that I edited in notepad, and the format is pretty simple, you can sort thru it and pull the ones you like and dump them directly into your prompts.

[chrome forge]

can someone help me? i want to do img to img but i want to like exactly create the same image but w my character

[lucid yoke]

Hey what model are you using for this?

Or trying to use

i think flux klein (or another editing model) could do that if you have a reference image

[chrome forge]

hey im using Zimage turbo to generate images

like i want my character to doing exact same pose w same expression, and same backround

im using
model : z_image_turbo_fp8_e4m3fn
text encorder : qwen_3_4b_fp8_mixed
vae : ae.safetensors

[chrome forge]

this is just a random picture from pintrest btw

[lucid yoke]

well if you have a reference picture of your character i do recommend flux klein

[chrome forge]

well the problem is storage, what is the minimum req and can you share the workflow?

[lucid yoke]

Its the basic workflow
Im assuming you use comfyUI?

[chrome forge]

yes

[blissful gazelle]

Hey everyone, just wanted to share a custom training engine I’ve been building and see if anyone has some feedback or thoughts on the approach. I finally decided to ditch Kohya entirely for this run to avoid version conflicts and get full control over the low-level logic.
I wrote this custom script from scratch, specifically tailored for a Dual T4 Kaggle environment.
Here’s a quick breakdown of the architecture and pipeline:
Pure DoRA (Attention Only): I completely threw out the Feed-Forward Networks (FFN) from the training targets. The DoRA adapter only targets the attention blocks (to_k, to_q, to_v, to_out.0, add_k, add_q, add_v, add_out.0).
V-Prediction Base: It's training on top of an SDXL V-Prediction base.
Pre-computed Latent Caching: To save compute, it uses a pre-extracted latent cache dataset with a strictly enforced 1x aspect ratio (no padding/cropping during the run).
Hardware Optimizations: Running with expandable_segments:True to survive the T4 VRAM limits, PagedAdamW8bit, and custom batch sampling (ARBBatchSampler).
Safeguards: Implemented custom Gradient Clipping (Max Norm 1.0) alongside Min-SNR gamma (5.0) and 10% dropout to prevent loss spikes.
I'll attach the script below. Has anyone else experimented with purely attention-focused DoRA training like this, completely ignoring the FFNs? Would love to hear your critiques or if you spot any potential bottlenecks!

[velvet pelican]

nice ty

[heavy estuary]

anybody know whats causing this error? ModuleNotFoundError: No module named 'pkg_resources'

[lucid yoke]

Personally i can't dive into it right now but what webui are you trying to use?

[heavy estuary]

imma be fr w you

ion know what a webui is but im using stable diffusion

[lucid yoke]

And could you try to change your nickname

[heavy estuary]

what do you mean

oh right i forgot about that

[lucid yoke]

Hmm chances are your trying to install a1111, i recommend not using that (as it's broken rn) what gpu are you usingm

If its a 5000 series i recommend forge neo, if its older then a 4000 series, i recommend just forge

In the pinned messages theres a few guides

[heavy estuary]

think my shits like a 7800

[lucid yoke]

Ah amd

Hmmm

[heavy estuary]

lowkey i had like 5k saved up and i said fuck it ill make a pc

ion know how to use it lowkey

[lucid yoke]

Hmm if your using a amd card i recommend waiting on CS1o since im not too familiar with rocm / current amd drivers

[heavy estuary]

gotcha

[lucid yoke]

But there's a amd guide in the pinned messages aswell, i think he documented it pretty clearly on what to do

[heavy estuary]

i forgot about that ngl

[maiden pendant]

anyone can help i not able to install stable diffusion

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed to build 'https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip' when getting requirements to build wheel

[lucid yoke]

A1111 your trying to install is broken, i recommend trying to install a more up-to-date version such as forge/forge neo depending on your hardware see guide in the pinned messages

[maiden pendant]

where i can download ? i have rtx 4080 gpu

[lucid yoke]

The pinned messages

#🤝｜tech-support message

[maiden pendant]

which one have best forge neo or forge ?

[ornate elk]

With that GPU follow any of the AMD ROCm Guides from the pinned messages. For example Forge Neo with ROCm

[ornate elk]

Forge Neo

Its more up to date and supports the latest stuff

[maiden pendant]

thank you 😄

[maiden pendant]

Thank 😄

[signal pebble]

i followed the steps for neo forged but still getting launch unsuccessful, any help?

[maiden pendant]

showing this error exit code: 9009

stderr:
'C:\Users\vma\AppData\Local\Programs\Python\Python310' is not recognized as an internal or external command,
operable program or batch file.

Launch Unsuccessful! Exiting...
Press any key to continue . . .

[lucid yoke]

Im currently at work but it looks like its not recognising your python, you have it installed right?

[lucid yoke]

Please post the cmd log, I'm currently at work so i cant check but maybe someone else has insights

[ornate elk]

Post your cmd log here then we can check

[ornate elk]

Wrong python version installed.
Uninstall 3.10

[signal pebble]

I managed to figure it out, Im sure I'll have more questions tomorrow, thanks anyway

[maiden pendant]

how ?

[maiden pendant]

which version i download ?

[ornate elk]

Forge Neo wants 3.13.12

[signal pebble]

Listen to CS1o, Im the wrong person to ask

[maiden pendant]

still not working

i try all

[maiden pendant]

'C:\Users\vma\AppData\Local\Programs\Python\Python310' is not recognized as an internal or external command,
operable program or batch file.
Creating venv in directory C:\AI\sd-webui-forge-neo\venv using python
Unable to create venv in directory "C:\AI\sd-webui-forge-neo\venv"

exit code: 9009

stderr:
'-m' is not recognized as an internal or external command,
operable program or batch file.

Launch Unsuccessful! Exiting...
Press any key to continue . . .

[prisma moss]

You ve been told already

Uninstall python 3.10, install 3.13.12

[maiden pendant]

i already did

[maiden pendant]

still showing this error

[prisma moss]

I cba atm try with the -uv package installations check the neo guide

[maiden pendant]

ok

[ornate elk]

Delete the venv folder and relaunch

[prisma moss]

@vocal burrow

@lucid yoke

[strong dawn]

@lucid yoke

[lucid yoke]

Ipdadapter with sdxl? May i recommend you trying a simpler UI instead of comfy since your new?

Forge neo is recommended since it does a lot of stuff out of the box without the hassle

[strong dawn]

yes because i want this artstyle back

[lucid yoke]

Hmm well ipdadapter is most definitely the cause of the slowdown

[strong dawn]

i paid 560$ for this workflow so

its so slow

that i want to pump images in less than 60 seconds

[lucid yoke]

You pay for workflows???

What, its just style transfer

[strong dawn]

fivver

[lucid yoke]

Well anyways i do NOT recommend paying for workflows at all

[strong dawn]

now i know.

but i want to reproduce this style

im trying another simple workflow

and its horrible 🙁

[lucid yoke]

Slower gen times are to be expected with ipdadapter

[lucid yoke]

And frankly its not the best

[strong dawn]

could you help me get this style back?

[lucid yoke]

How many images do you have of that style?

[strong dawn]

like a lot

but that isnt the case

if you drag and drop the image into comfy you can see the entire workflow pretty easily.

[lucid yoke]

Your better off making a lora in that case

[strong dawn]

im so overwhelmed i dont want a lora im sorry if i seem frustrated and or annoyed im just trying to get on with my day.

i just want what i have just simpler and faster.

[lucid yoke]

Well lets say:
Lora = simple workflow, probably 10s gens

Ipadapter = minutes

[strong dawn]

could you help me rebuild that

[lucid yoke]

Its just how the technology works with the hardware, ipadapter has to do all the steps from the beginning every time. A lora is just a very tiny model trained that you tackle on the main model you yse

Roughly speaking

[strong dawn]

i dont need the ipadapter but everything else especially with wildcards would be great.

i just want something simpler Euro I'm sorry I just want to get directly to problem solving I dont want to remake a lora i have different characters that works with this style i just need to get to the source.

[lucid yoke]

well thats the direct problem solving and getting to the source so you can use that style for all characters / images
Well otherwise theres no solution for any speed increases

[strong dawn]

could you help me rebuild a workflow that will bring me back to the style you see above

something easy simple and clean just without all the junk and ipadapter nonsense

[lucid yoke]

You already have one with ip adapter. Theres no other way to do that without a lora is what I'm telling you

Or specifically knowing the style prompt that was used before ip adapter was applied

[strong dawn]

so i gotta stick to this shitty workflow then

[lucid yoke]

That or spend maybe a hour or two of your time to make a lora for free on your device and enjoy using the style way faster.

Or even use civitAI and dont even do the hassle yourself, but if your not interested in that im afraid we're going in loops

[strong dawn]

the reason why the lora wont work

is because

i have different angles and poses

the AI will get confused.

[lucid yoke]

For a style lora that doesn't matter AT ALL

What

[strong dawn]

and i also have NSFW in it too

So

[lucid yoke]

Ive made two before, your misinformed my guy

[strong dawn]

i dont want it to be confused

if you think this is going to work then

i have 200+ images ready to go.

[lucid yoke]

If you want the real hassle free route:
civitAI is the way to go but its paid.

If you wanna learn a little and in the future wanna make your own for free. I really recommend learning onetrainer

I'm not able to train it for you

[strong dawn]

yeah so im fucked.

560$ down the grave.

[lucid yoke]

Well, its an expensive lesson for sure. It's not entirely useless because you can use it to make synthetic datasets of other styles

[strong dawn]

@lucid yoke Hey I want to say sorry

I got upset at you and thats not fair + shitty.

I removed a lot of junk that I didn't need anymore.

[lucid yoke]

No worries! It happens to the best of us

[strong dawn]

So now I got it to go from 100 seconds to 60 seconds.

No its something i have to worry about

youre kind. so i dont want you to feel like a POS cause of my shit attitude

[lucid yoke]

That definitely does help but the main bottle neck is still the ipadapter

I'm honestly thick skinned so don't worry about it haha it's fine

[prisma moss]

whats the problem

[strong dawn]

my current workflow is trash

i want a simpler

cleaner

faster workflow

instead of waiting 100 seconds i want it to be 60 seconds

[prisma moss]

well, the speed itself is gonna be dependant on ur specs, arguments, preview type

[strong dawn]

i have a 4070 super

[prisma moss]

that doesnt change what i said

[strong dawn]

gotcha

im new to this all the vagueness and random terms i dont compute it

[ornate elk]

These images remind me of the ilustmix model. You could test that out

[strong dawn]

thats the thing i dont want to keep testing

i just want to get to the answer

[ornate elk]

Thats difficult without at least trying stuff out

[strong dawn]

gotcha

[prisma moss]

while at it lemme see ur workflow

[strong dawn]

json ?

[prisma moss]

just screenshot it for now if its not big enough

[strong dawn]

its too big

[prisma moss]

i mean, its definetly a bit stacked with stuff i wouldnt bother adding for a simple style

[dusty hawk]

I've installed SD for my windows laptop as a new user of this software. Can someone recommend a model for me?

[lucid yoke]

what GPU does your laptop have? i assume you installed A1111 or forge as software?

[dusty hawk]

A1111 installed

Installed RAM is 16GB (if that matters at all)

GPU seems to be 4GB

[lucid yoke]

im afraid you wont be able to run the most used models. it will probably offload to ram and take a few minutes per image.
my old work laptop had similar specs and took 2-3 min for a SD 1.5 model at 512x512 resolution

maybe if you optimize it with --lowvram args in the launch file it maybe could improve performance but its gonna be pretty limited what you can do

[dusty hawk]

Could you still tell me about the most used models for this software? Maybe I can find something that I can work with

[lucid yoke]

due how it works you could probably could run any of the SD 1.5 models.
most used ones are Illustrious or Z-image/Anima

but the other models are simply too large

[dusty hawk]

Are you able to provide download links for these?

[lucid yoke]

https://civitai.com/models is one of the most used sites for these.

be sure to filter for checkpoints and sd1.5

[ornate elk]

for the best performance oyu should setup Forge classic.
and then you need to increase your windows page file to 32gb so that you dont get a slow down wehn using sdlx/illustrious models

[lucid yoke]

you can run sdxl with 4gb vram nowadays?

dang i missed out on the latest news, well yes listen to Cs1o in that case my bad!

[dusty hawk]

The thing is I got A1111's download link through an install tutorial, so I don't know what Forge's looks like

[ornate elk]

you can check the pinned messages of this channel, there are all the guides listed

on the first link

[dusty hawk]

How do I uninstall A1111 in favor of Forge Classic?

[ornate elk]

uninstall python 3.10 if installed, then just deleting the folder of auto1111 is enough

[dusty hawk]

python has to be deleted?

[ornate elk]

uninstalled yes, if its 3.10 or 3.14

[dusty hawk]

"increase your windows page file to 32gb" what does this even mean?

Wait a minute, I have 2 GPUs

[ornate elk]

which ones?

[dusty hawk]

Intel(R) UHD Graphics
NVIDIA GeForce RTX 2050

So which one of these 2 commands is recommended to input?

[ornate elk]

--flash --xformers

[dusty hawk]

That's not 2 separate commands? Or is it?

[ornate elk]

seperated

[dusty hawk]

I have to edit the top line of the specified section to match the command, yes?

oh wait wtf

where do I put it????

[dusty hawk]

What's the difference between '--flash' and '--xformers' anyway?

[ornate elk]

different performance optimisations

[dusty hawk]

what are the tradeoffs of each tho?

[ornate elk]

nothing

[dusty hawk]

then why are there different options for this???

[ornate elk]

one is used for vae and the other for models

[dusty hawk]

vae? what does it stand for?

also is --flash the vae one?

[ornate elk]

xformer is for the vae in forge neo and flash for the models

[dusty hawk]

so --xformers is obsolete for this version?

can both commands be put together or just one?

[ornate elk]

boath together

[ornate elk]

hmm seems like it maybe dont work with python 3.11 anymore

try uninstall python 3.11.9 and install 3.12.10 from here:
https://www.python.org/downloads/release/python-31210/
then delete the venv folder inside forge classic and relaunch the wbeui-user.bat

[dusty hawk]

Bruh, the microwave fucking slowed down my setup while it was cooking

[dusty hawk]

What kinda PyTorch version would be suitable for this Stable Diffusion version?

[lucid yoke]

Update your drivers

[dusty hawk]

They're all up to date

[prisma moss]

r they studio drivers tho?

[dusty hawk]

Studio drivers?

[prisma moss]

if nvidia yes

[dusty hawk]

What's that?

[prisma moss]

ill take that as a no

[dusty hawk]

No, I mean like what are studio drivers?

[prisma moss]

studio drivers are studio drivers

[dusty hawk]

Maybe I have one

Lemme see

[dusty hawk]

I don't have one, so lemme download a studio driver real quick

[ornate elk]

you dont need studio drivers

make sure to use the normal nvidia driver from the nvidia app

[dusty hawk]

well what is that normal driver called?

[ornate elk]

game ready drivers

[dusty hawk]

Are studio drivers the incorrect type, or is game ready driver simply enough?

[ornate elk]

game ready is enough

[dusty hawk]

All I need is a model now

[lucid yoke]

well since you can run SDXL from what i learned today.
you can find models on Civitai SDXL/Illustrious is doable

[dusty hawk]

Can you provide me the download links for them??

I can't seem to find them

(also, are there any models suitable for me which would be usable for any style?)

[lucid yoke]

https://civitai.com/models
illustrious = anime
sdxl has some realism finetunes and some anime ones but its better for realism imo

sites slow today though idk whats up with that

[dusty hawk]

how does one even find out what version of SD they have?

[lucid yoke]

the filter button

[dusty hawk]

No, I mean like how do I know what version of SD I have

I downloaded that Forge Classic one

webui thingy btw

[lucid yoke]

oh that doesnt matter in this case, you should be on the latest

if you were on version 1 or versio 1.15 it would work the same

but settings i think

[dusty hawk]

What's the latest version??

[lucid yoke]

2.23?

its on the github

[dusty hawk]

So many version tags for SD

[lucid yoke]

SDXL 1.0, and " Illustrious"

the models often show the settings with their example images

so you know how to prompt them

[dusty hawk]

"most used" while it only has like 2k downloads

how far down did stable diffusion fall?

illustrious has like 3.2k downloads

and that's with the SDXL 1.0 tag that I selected for checkpoints

Are y'all sure SD uses checkpoints as the model type?

Or does it also have other model types?

[dusty hawk]

this does seem to be the case

but then why can't I find the popular recommendations that y'all gave me?

[supple osprey]

What do I do here?

[lucid yoke]

your probably sorting by top week

sort by top yeasr

[dusty hawk]

I sorted by all-time

nvm i found one

https://civitai.com/models/795765/illustrious-xl

Is this the Illustrious you were talking about?

[lucid yoke]

illustrious is a model type

[dusty hawk]

and checkpoints are models, right?

[lucid yoke]

yes
if you sort by this

youll find the anime based ones

[dusty hawk]

https://civitai.com/models/1277670/janku-trained-chenkin-and-noobai-rouwei-illustrious-xl?modelVersionId=2786084

is this model suitable for me?

[lucid yoke]

sites broken for me but why not

[ornate elk]

forge cant use anima models, you need to use forge neo

[prisma moss]

It really isnt

Theres a big difference in inference speed the two

[supple osprey]

does it need python 3.13 exactly?

[formal forge]

Hola, soy Hugo. Estoy en Windows y quiero instalar Stable Diffusion, pero me da errores con Python y pkg_resources. ¿Alguien puede guiarme paso a paso en la instalación?

[lucid yoke]

If you joined that server, it was a scam!
Second:
I don't know Spanish sorry.

Cleaned up the spam bot, if you missed it. Good!

[formal forge]

El error que me aparece es: "ModuleNotFoundError: No module named 'pkg_resources'".
Estoy usando Python 3.10 en Windows. ¿Qué pasos debo seguir para solucionarlo?

[lucid yoke]

Wow another scam

Cleaned again

[formal forge]

Hi everyone! I’m currently installing Stable Diffusion on Windows with Python and Git, and I wanted to double-check the recommended versions for 2026. I’ve seen that Python 3.10.6 is often suggested instead of newer Python versions. Also, I was told to install Git and AUTOMATIC1111 WebUI. I’d like to confirm:

Are these still the recommended tools and versions to use?

Is there any other software or a newer version of AUTOMATIC1111 that I should install instead?

Thanks a lot for your help!

[lucid yoke]

No, i recommend
Forge classic/neo depending on your hardware
#🤝｜tech-support message

A1111 is outdated

[formal forge]

Thanks for the reply! So in 2026 the recommended installation is Forge instead of AUTOMATIC1111?
Do we still need Python 3.10.6 and Git, or is there a newer alternative?

[lucid yoke]

I recommend checking the guide. It tells you there.
Depending on your hardware a different version is recommended

[lucid yoke]

Te recomiendo consultar la guía que enlacé. Ahí encontrarás la información.

Según tu hardware, se recomienda una versión diferente.

Google translate

[formal forge]

Could you please share the official Forge installation guide link?
I want to make sure I follow the correct steps depending on my hardware.

[lucid yoke]

Its here @formal forge

[dusty hawk]

Just so I'm sure, I'd also need to have a VAE selected in order for this thing to understand my prompt?

Where tf do I get that?

[lucid yoke]

Should be automatic or embedded, vae is just color

Your prompt is very very bare bones however

@dusty hawk hey i saw a message from automod and we DO NOT have a "support discord", those are scammers who come in here, we only offer it here. In this channel

[dusty hawk]

What?

Then WHY IS IT RECOMMENDING THE SERVER???

[lucid yoke]

We are the discord

We wont forward to any other

The ones that post invites etc are scammers who join

I delete them when i see them

[lucid yoke]

So in short, any referrals away from this discord with like "support" shouldnt be trusted. Any dm/support links shouldn't be trusted.

But yeah your vae question. When you generate something does it come out grey or full of colors?

[dusty hawk]

I dunno, maybe?

Prompt: Spongebob

Prompt: Amy rose

That one came out a bit grey

wtf

Prompt: super mario, nintendo

prompt: winnie the pooh

such an inconsistent pattern of coloration

what's up with this???

[lucid yoke]

well your prompts are very very very short but looks like its working as intented 👍

[dusty hawk]

well why doesn't it color some of these sometimes then?

[lucid yoke]

well you didnt give it more details or instructions. if you were to say "greyscale" in the negative prompt. it will avoid that.

[dusty hawk]

So how do you suggest that I should figure out the prompting from now on then?

[lucid yoke]

well if you look at other model image examples, the negative tag is usually something like:
bad quality, worst quality, lowres, jpeg artifacts, bad anatomy, bad hands, multiple views, signature, watermark, censored, ugly, messy, bad anatomy

and the positie prompts along the lines:

 black hair, male focus, short hair, blurry, jacket, 1boy, depth of field, outdoors, looking at viewer, solo, gloomy, expressionless, dismal, masterpiece, best quality, highly detailed, high-quality, color theory, shadows

[dusty hawk]

hmm

I see

👍

[lucid yoke]

all i can recommend is look at the provided image examples from the model and maybe get a autocomplete function running but i use a different UI so i cant help with forge

[dusty hawk]

ok

thx

[lucid yoke]

oh found it:
the Adetailer extension and boorutag autocomplete extension from the extensions tab is highly recommended

[prisma moss]

also what are ur negatives?

[dusty hawk]

There were no negatives in all these prompts

[dusty hawk]

Wait, how do I get extensions??

[lucid yoke]

The extensions tab

[prisma moss]

Well duh, thats why ur results r shit

[dusty hawk]

Gimme some slack here. I'm in my "no common sense" period

[prisma moss]

This ur first time generating?

[dusty hawk]

yeah, kinda

I ain't no expert, but I think there's more to prompting than just finding the correct words and tags for waht you want to create

[lucid yoke]

well its mostly that though

[fickle herald]

anyone can tell me how to create such "beauty portraits"?

just imagine its sharp ^^

[celest delta]

@ornate elk remember when there was an error with the forge neoupdate? I didn't put back the automatic updates since then, i put back git pull in the .bat so that i update automatically again but i get a bunch of errors and it's not updating

[celest delta]

great i broke everything when i tried to get git to bypass the change i made backthen... it reseted everything to default and now i can't load my custom lora directory anymore and the whole thing fails, what the hell...

[ornate elk]

Any cmd log to check what's going wrong?

[celest delta]

i managed to fix it, now it's up to date, thanks anyway

i still have errors though

oh great my dynamic prompt extension disappeared too...

maaaaaan why did i try to update...

venv "H:\AI_ART\sd-webui-forge-neo\venv\Scripts\Python.exe"
Python 3.13.12 (tags/v3.13.12:1cbe481, Feb  3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]
Version: neo
Installing sd-webui-infinite-image-browsing requirement: hnswlib>=0.0.0,<1.0.0
Error details: Couldn't install sd-webui-infinite-image-browsing requirement: hnswlib>=0.0.0,<1.0.0.
Command: "H:\AI_ART\sd-webui-forge-neo\venv\Scripts\python.exe" -m pip install "hnswlib>=0.0.0,<1.0.0" --prefer-binary
Error code: 1
stdout: Collecting hnswlib<1.0.0,>=0.0.0
  Using cached hnswlib-0.8.0.tar.gz (36 kB)
  Installing build dependencies: started
  Installing build dependencies: finished with status 'done'
  Getting requirements to build wheel: started
  Getting requirements to build wheel: finished with status 'done'
  Preparing metadata (pyproject.toml): started
  Preparing metadata (pyproject.toml): finished with status 'done'
Requirement already satisfied: numpy in .\venv\Lib\site-packages (from hnswlib<1.0.0,>=0.0.0) (2.3.5)
Building wheels for collected packages: hnswlib
  Building wheel for hnswlib (pyproject.toml): started
  Building wheel for hnswlib (pyproject.toml): finished with status 'error'
Failed to build hnswlib

stderr:   error: subprocess-exited-with-error

  Building wheel for hnswlib (pyproject.toml) did not run successfully.
  exit code: 1

  [5 lines of output]
  running bdist_wheel
  running build
  running build_ext
  building 'hnswlib' extension
  error: Microsoft Visual C++ 14.0 or greater is required. Get it with "Microsoft C++ Build Tools": https://visualstudio.microsoft.com/visual-cpp-build-tools/
  [end of output]```

  ERROR: Failed building wheel for hnswlib

[notice] A new release of pip is available: 26.0.1 -> 26.1.1
[notice] To update, run: python.exe -m pip install --upgrade pip
error: failed-wheel-build-for-install

Failed to build installable wheels for some pyproject.toml based projects

hnswlib


Note: hnswlib is optional and only needed for experimental topic clustering/search.
Core image browsing features will still work without hnswlib in some cases.
To skip optional dependencies during installation, set: IIB_SKIP_OPTIONAL_DEPS=1
If you want to install hnswlib, please ensure you have:
  - A C++ compiler (gcc/g++ on Linux, clang on macOS, MSVC on Windows)
  - Python development headers (python3-dev on Ubuntu)
  - Or try: pip install hnswlib --only-binary=:all: for pre-built wheels
See: https://github.com/nmslib/hnswlib/issues/403 and #504 for more details.
Launching Web UI with arguments: --cuda-malloc --xformer --lora-dir 'G:\LORA```

the rest is the usual stuff where it works normally

i already have visual C++ 14 so i don't get it, as well as the others, i didn't have all these errors before updating right now

[ornate elk]

These errors come from an extension probably

Like an image browser extension or something else

Also its not --xformer its --xformers

And --cuda-stream should be added too

And I think its --loras-dir not lora-dir but I need to recheck this later

[celest delta]

nah this one is correct, as i said, i fixed that

but my dynamic prompt extension refuse to come back... should i just go back to my previous settings for it to work?

like this

[ornate elk]

Why is it not working?

Have you updated the extensions with the extensions tab?

[celest delta]

it's an old extension that's not updated since 2024, but it was still working with forge neo, until i updated today

it's not showing up anymore, and my prompt aren't divided (used it for that)

like {glasses|freckles}, it would randomly select one of the two, but now it doesn't do anymore

[ornate elk]

Ah okay, yea nothing much you can do about that

Wildcards would be a way to to the same thing

[celest delta]

that sucks though, with the dynamic prompt i can just edit it on the fly in my prompts however i want

like between two loras, between two prompts, and many more... maaaan

guess i'm just gonna go back to my previous settings hoping it still works

[lucid yoke]

on swarmUI, thats kinda built in if that function is that important

[celest delta]

all that to try to make anima work correctly, and it doesn't, maaaan, i hate these stuff always breaking your settings with every updates

anyway to go back to a previous update of forge neo?

please i really need this, i have a fuckton of settings saved with dynamic prompt

ok... i found on reddit a fork that supposedly works with the latest update... BUT, cause nothing ever works with me

i deleted the old dynamic prompt extension, but when i try to install this new fork from the webui with the URL, it says the folder already exists... what do i do?

[lucid yoke]

Back up the existing folder of the extension and then delete it? Also i do hope you didn't lose the prompts by deleting it this fast

[prisma moss]

U just need to downgrade neo

[celest delta]

as i said, i deleted it completely

[lucid yoke]

From the ui or Also the left over folders in the root?

[celest delta]

also my prompts themselves are savec elswhere so it's fine

deleted directly in the root

[lucid yoke]

Silly question, did you reboot the ui?

[celest delta]

of course, relaunched it entirely, i'm gonna just reboot my computer see if it changes anything

STILL DOESN'T WORK WHXY

[prisma moss]

because u didnt downgrade neo

to previous version

right click the main webui folder, open git bash, paste git checkout cd42575. Delete venv, back up extensions or whatever

[celest delta]

that's not about downgrading here, just installing an extension that should work with the latest update

and now i have another issue, guh

i don't remmeber what i tried, tried to check previous versions or something, and now when i launch webui it says this :

You are not currently on a branch.
Please specify which branch you want to merge with.
See git-pull(1) for details.

git pull <remote> <branch>

have no idea how to go back

[prisma moss]

yes its exactly about downgrading

[celest delta]

i didn't need to, i managed to make the fork work now with the new update, i just had to give it a new name... but i still have that issue of not being on any branch

[prisma moss]

branch --neo obviously

[celest delta]

like i add --neo in the .bat?

[prisma moss]

did u remove the venv

[celest delta]

yes i still get this error

it says i'm not on any branch now

it still works mind you, but i dunno what ti means and if it can cause issues later

[prisma moss]

make a new folder and do a git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo

run to get the venv, close if opens, then in root folder git checkout cd42575

and copy whatever extensions/models etc u had in ur previous forge to the new one

[celest delta]

@ornate elk btw i have a 5080, is it worth adding --sage? no downside? in addition to xformers and cuda malloc

[celest delta]

thanks i'll try later, i need a break, i lost 2 hours on this shit, just to get back to how i was, without any improvement as i had before, so frustrating

lol

[ornate elk]

you should add --cuda-stream and --flash they improve the performance with no downside

--sage is even better but can degrade images a bit

[celest delta]

ah i rather it being a bit longer than degrading the image

is cuda stream still necessary for a 5080 though? and thanks

gonna add them and remove sage

[celest delta]

i tried with and without sage, there wasn't a singel speed difference, lol

for like 100 gen

[bold nacelle]

Hey guys I’m trying to get GNEISS Coin to show up on MetaMask. Who do I talk to about making that happen?

[lucid yoke]

well not here since this is about image generation and not Coins

[bold nacelle]

Yeah sorry clicked join MetaMask and brought me here somehow

[celest delta]

welp, i dunno what this update did, but my gen are slower than before even with the stuff you told me to add @ornate elk

i was doing 60-63 minutes for 300 gen, and now it takes 72-75 minutes, it's not a big deal, but it's weird

[prisma moss]

this is the third time you are blaming the update and not doing anything to revert it

[celest delta]

i'm just trying to figure out why, if it's something on my end or not

is it something happening to many people that they have slower gen with that new update? is that why you keep telling me to revert? (keep in my mind that i didn't update since march)

[prisma moss]

one more guy had the dynamic prompt issue till he reverted

[celest delta]

the dynamic prompt issue is solved thanks to that fork i've said earlier

now i'm just wondering about generation speed

[lucid yoke]

did you happen to add a lora?

more loras = slower speed

[celest delta]

i'm always using loras

and i don't use more than usual

[lucid yoke]

yes i mean usually load 2, this time load 3-4

[celest delta]

i'm always using the same numbers or lora so that shouldn't change anything

[lucid yoke]

im not sure what your previous version was or what changed after x amount of versions because his changelog is frankly horrendous but this is why i dont like forge neo. PROBABLY because they changed something in the backend. i recommend opening a issue mentioning your slowdown so they might be able to tell you more.
https://github.com/Haoming02/sd-webui-forge-classic/issues

[celest delta]

seems like i'm 2.23 already

at least i can't see 2.24 when checking with git log, only 2.23 and 2.22, and when even selecting one of the branches, it still says i'm not on any branches, it's so freaking weird

[celest delta]

okay it's defintiely not an update issue.. went back to a march update (the last one i did), and the speed is still not the same as before, i just don't get it, i always get the weirdest problem with this (CS1o can testify to that, lol)

[knotty turret]

Should i worry "triton not installed"

i used install-n.bat but still not have it install

[lucid yoke]

Hey there, I'm not familiar with amd installations but Cs1o is, i recommend waiting for him and i absolutely would be careful of random dm's

[ornate elk]

You can ignore that. Its only needed if you want to get sage attention working

[knotty turret]

This will make things faster?

[ornate elk]

yes whats your gpu?

[knotty turret]

RX6600

anything that helps i accept

[knotty turret]

already know CS1o he the goat

[ornate elk]

you could try setup ComfyUI with ROCm instead of zluda or use forge neo with rocm

[knotty turret]

ROCm is faster?

well ok there any guide for it

[ornate elk]

for rdna3 and 4 for sure, but idk if its the same for rdna2 gpus

the guide is in the pinned messages of this channel

[knotty turret]

tried this one

will try with rocm now

[ornate elk]

let me know if you see an improvement

[knotty turret]

dont see 1032

i believe 1032 is for rx6600

[ornate elk]

the x stands for every rdna2 here

[knotty turret]

oh mb

[ornate elk]

np

[ornate elk]

oh i need to say that rdna2 gpus need the manual comfyui setup

the portable one is not supported, i have to add that to the guides

[knotty turret]

this part?

ok

[ornate elk]

oh right i added the steps for portable too, i forgot xD

[knotty turret]

fine?

[celest delta]

my generation speed got lower again today... what the hell is going on

this is ridiculous, what could cause this

[knotty turret]

@ornate elk not used to this so think need to ask😂 fine too?

[prisma moss]

fresh install, and check if u have studio drivers installed

[ornate elk]

yep can be ignored

[knotty turret]

believe should do this too

[celest delta]

i never needed this studio drivers... and installing everything again? I feel like i'm doing this every month, it takes too long... 😔

CS1o you got any idea?

[prisma moss]

bro no one is going to fkin chase perfomance issues for you

ive already told you studio drivers offer better inference speed but you decide to ignore it 3 times

so what do you want?

[knotty turret]

good to know i installed this studio drivers

[celest delta]

you're confusing me with someone else, it's the first time you told me about studio drivers

sorry i know i complained a lot yesterday, but this time it's not me, lol

[prisma moss]

true that was someone else

[celest delta]

will that affect gaming? Apparently i can't have both nvidia drivers

[prisma moss]

no

[celest delta]

thanks, i'll try that

[knotty turret]

@ornate elk Not sure if triton its enable this time it is?

[prisma moss]

why do u need triton

[knotty turret]

to make thing run faster better generation time i guess

[prisma moss]

well u wont be able to tell unless u do --use-sage-attention

And you have pytorch argument

[knotty turret]

here?

[prisma moss]

yes

[knotty turret]

let me see what CS1o will say

[ornate elk]

no, you have to install triton manually. and sage attention needs some extra steps too for rdna2 gpus

--use-sage-attention then only works after sage attention was installed

[knotty turret]

already did this

[ornate elk]

the steps can be found here:
https://github.com/patientx/ComfyUI-Zluda/issues/170#issuecomment-2972793016

[knotty turret]

not gonna lie kind lost now alot steps now

[ornate elk]

i didnt include them in my guide for rdna2 and it also needs a manual comfy setup

it wont work in the portable mode of comfyui, manual setup is needed

[knotty turret]

shoud i remove everything

start again from here

[ornate elk]

patientx made a repo for comfyui with rocm all in one installer, you could try that

https://github.com/patientx-cfz/comfyui-rocm

[knotty turret]

only that?

ok let me delete what i did and start again

[ornate elk]

all of that yes, read their github or just wait until i add the steps to my mnaual comfyui setup guide

[knotty turret]

will try the guide above

and believe should put --use-sage-attention on bat too

[ornate elk]

yea its needed after the triton and sage install

the commands are on the other link i posted above, if the setup doesnt to it automatically

[knotty turret]

ok

after this

do it later

nvm idk if should keep doing this

😔

[ornate elk]

ohh

[knotty turret]

sad now that delete the zluda one

dam this taking long time

[ornate elk]

but rocm could be still faster even without sage for 6600

[knotty turret]

good to know

its it normal to take so long?

[ornate elk]

yep its downloading stuff in the background

around 2gb i guess

[knotty turret]

the cmd just closed

...

didnt do anything

[ornate elk]

hmm

[knotty turret]

guess its fine

alot new files

normal start too

the only difference didnt open my browser

find the install log

this one comfyui-rocm.bat or comfyui-user.bat

[ornate elk]

use the comfyui-user.bat

[knotty turret]

my AMD folder is taking alot space now after the install and from the time using comfyui 21 gb there anything i can delete?

Bin folder 10 gb
Lib 10,4 gb

[celest delta]

even with studio drivers, my generations are still slower than before... i dunno what that recent forge neo update did, but even reverting back to a previous branch doesn't fix anything at all... it's so frustrating going from 60 minutes to almost 100 minutes for the exact same gen, i just don't get it, feels like my new rig doesn't even matter anymore
guess i have no choice but to do a fresh install... hope this will fix it...

[lucid yoke]

Did you post this on the forge neo GitHub as an issue?

[celest delta]

not yet since i didnt have an account, but i will do it later today

[ornate elk]

AMD folder in c programms?

Because there are only amd driver stuff stored.

[knotty turret]

yes

[ornate elk]

Uninstall everything of HIP SDK 5.7, delete the folder after that

What's your gpu again?

Its strange that your 6.2 is that large

[knotty turret]

RX6600

[knotty turret]

ok

[ornate elk]

whats so large ind your lib and in your bin folder?

[knotty turret]

lib:

[ornate elk]

if you dont use any webui with zluda you could also uninstall hip sdk 6.2 completely and delete the folder afterwards

[knotty turret]

bin: (alot files)

[knotty turret]

i only using the rocm right now

[ornate elk]

ok, yea then you dont need hip sdk

[knotty turret]

so i can delete everything inside hip sdk?

and then delete the folder

[ornate elk]

uninstall hip sdk via system control and then delete the folder

[knotty turret]

this?

ok

god dam thx this saved alot space

[ornate elk]

np you can then also open up a cmd and run
pip cache purge
that will remove temp files from the install of the webuis.

[prisma moss]

thank you for reminding me lmao

[autumn wedge]

Hello everyone,

I made my first LoRA consistent character using ZIT with AI-Toolkit.

I followed the settings and workflow of this tutorial:

https://youtu.be/uHqOdGsR9cw?si=JH8TbrMmDGEQv3cs

What I did:

• I generated 30 images using FLUX.2 9B

• Then, I used Nano Banana to change the lighting/environment and make the images more realistic

• I used basically the same video settings

• I trained LoRA with 2500 steps

The problem is consistency.

At first, the character seems "correct" (sometimes), but after several generations, the face slowly changes more and more, and eventually the model begins to generate an Asian-looking face instead of the original character.

I'm using the default ZIT workflow within ComfyUI.

I also noticed that when I use styles like:

• "Phone photo"

• "High Quality Photo"

The consistency of the face gets even worse.

Sometimes, generations are very "accurate", and sometimes the character barely looks the same.

I have attached some example images so you can see the problem.

I've seen people recommend:

• 100-150 images

• About 4000 training steps

• And approximately 100 to 200 steps per image

For a much better consistency.

So I wanted to ask:

• Are 30 images too small for a solid ZIT character LoRA?

• Is 2500 steps too low?

• Are the "Phone Photo" / "High Quality Photo" styles causing identity deviation?

I use them because I can't use realism LoRAS (RealisticSnapshot/Instapic) because they degrade EVEN MORE!

• What would you personally change to improve consistency?

Any advice would really help, since this is my first serious LoRA training attempt

[misty marten]

I'm having this error; it says "Press any key to continue..." When I press it, it closes the command prompt and doesn't open the web UI. I'm using Windows 11 with Zluda, ROCM 6.2, and an RX 9060XT.

[light mist]

How long @misty marten

[neon pecan]

I'm trying to run stable diffusion on stability matrix. I got my pc upgraded, everything downloaded fine, had to download missing files for the script folder and I keep getting AssertionError: Torch not compiled with CUDA enabled
i'm using the amdgpu version. why does it keep bugging me about CUDA? It's an AMD version

nope

does anyone know why i'm getting a AssertionError: Torch not compiled with CUDA enabled message when trying to use an AMDGPU version of stable diffusion?

The last guy tried to scam me

[lucid yoke]

Did you see his @? The message is deleted

[neon pecan]

So I dodged a scam, good

Tried to get help with stable diffusion and he's asking for my wallet information

[lucid yoke]

Hmm with stability matrix i sadly can't help because thats a ragtag of random install scripts but i do recommend in the pinned messages the guides by Cs1o

[neon pecan]

so is that agent guy a scammer that the admin and moderators wont kick?

[lucid yoke]

And the "stable diffusion" ui is every outdated, i do recommend using a newer one

[lucid yoke]

We do ban them a lot

[neon pecan]

he took me to a create ticket server and the first thing im asked is did I assign my wallet information which was nonsense cause it never asked for anything so I closed the ticket, left the server and blocked him

it was obviously an attempt at getting information

[torpid frost]

Open a ticket

[neon pecan]

No

[torpid frost]

Who's that?

[neon pecan]

I just did and they tried to get my bank information

they asked me about wallet information

[lucid yoke]

Ahh yeah those guys we ban them every time but its a automated bot

[torpid frost]

Bank lmao 😂

[neon pecan]

they're dumb if they think it'll work.

[lucid yoke]

Youd be surprised

[neon pecan]

as soon as they asked if I added wallet information I knew right away something was wrong

I ditched the server and blocked the guy

and now i'm back here....waiting for help

[lucid yoke]

^^

[neon pecan]

where?

[lucid yoke]

The pinned messages. Ill link it
#🤝｜tech-support message

Rule 6 anyone who tries to lure you away from the server should be treated as a scammer

[neon pecan]

they werent being smart about it at all. straight up asking for wallet information? dead giveaway

not exactly discreet

[lucid yoke]

Eh they cast a wide net

[neon pecan]

well they dont seem to be catching any fish.

a free ai software doesnt ask for wallet information

[lucid yoke]

But yeah, i dont recommend using stability matrix

[lucid yoke]

Depending on your gpu there's a guide made by cs1o and i think he's online later if you get stuck

[neon pecan]

I guess i'll use zluda

[lucid yoke]

What gpu do you have?

[neon pecan]

I just installed a radeon rx 9060 XT. It was recommended to me to run stable diffusion

I'm stuck because of torch AGAIN

[lucid yoke]

I recommend waiting for Cs1o then, im not familiar with amd.
Your not trying to install a1111 are you?

[neon pecan]

no

ffs this is super annoying

[lucid yoke]

If you dont wanna wait try searching for the error in this channel, it's likely someone else had it too before

[neon pecan]

no results

ERROR: Could not find a version that satisfies the requirement torch==2.9.0 (from versions: none)
ERROR: No matching distribution found for torch==2.9.0

isnt torch a nvidia thing?

[lucid yoke]

Yes/no, the drivers get emulated iirc.

[neon pecan]

then why is it constantly bugging me about torch????

i'm getting impatient and frustrated cause its the same errors no matter what im using

[lucid yoke]

Are you sure you followed the guide for zluda?

[neon pecan]

torchi this torch that

[neon pecan]

yeah, there's nothing about torch

[lucid yoke]

Like in its entirety, hmm well i recommend taking a breather and then waiting for Cs1o

[neon pecan]

im following his AMD instructions for amd zluda and nothing mentions torch. so I dont get why the AMD version is bothering me about torch a NVIDIA THING

ive been trying to get this crap working for nearly two hours

and it's "Torch" this "error cant install torch" that WHEN IM DOING AMD

torch torch torch torch torch

[lucid yoke]

Yes torch is a requirement, but again something is probably set wrong and let him take a look at it

[neon pecan]

I dont have the patience to wait for this guy to come on.

the demands ai has is outrageous