🤝｜tech-support

[upper linden]

yup

[ornate elk]

Ah that breaks the update function but can be fixed.
But this is a different error

[upper linden]

oh

[ornate elk]

Can you show me the full cmd log?

And did it worked before moving?

[upper linden]

yes

venv "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\Scripts\Python.exe"
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1-amd-18-ged0f9f3e
Commit hash: ed0f9f3eacf2884cec6d3e6150783fd4bb8e35d7
ROCm: agents=['gfx1101', 'gfx1036']
ROCm: version=6.1, using agent gfx1101
ZLUDA support: experimental
Using ZLUDA in C:\ZLUDA
Skipping onnxruntime installation.
You are up to date with the most recent release.
Traceback (most recent call last):
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\launch.py", line 48, in <module>
main()
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\launch.py", line 39, in main
prepare_environment()
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\modules\launch_utils.py", line 695, in prepare_environment
from modules import devices
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\modules\devices.py", line 6, in <module>
from modules import errors, shared, npu_specific
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\modules\shared.py", line 4, in <module>
import gradio as gr
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\gradio_init_.py", line 3, in <module>
import gradio.components as components
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\gradio\components_init_.py", line 3, in <module>
from gradio.components.bar_plot import BarPlot
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\gradio\components\bar_plot.py", line 7, in <module>
import altair as alt
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair_init_.py", line 649, in <module>
from altair.vegalite import *
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\vegalite_init_.py", line 2, in <module>
from .v5 import *
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\vegalite\v5_init_.py", line 2, in <module>
from altair.expr.core import datum
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\expr_init_.py", line 11, in <module>
from altair.expr.core import ConstExpression, FunctionExpression
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\expr\core.py", line 6, in <module>
from altair.utils import SchemaBase
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\utils_init_.py", line 14, in <module>
from .plugin_registry import PluginRegistry
File "F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\altair\utils\plugin_registry.py", line 13, in <module>
from typing_extensions import TypeIs
ImportError: cannot import name 'TypeIs' from 'typing_extensions' (F:\SD-Zuldav2\SD-Zluda\stable-diffusion-webui-amdgpu\venv\lib\site-packages\typing_extensions.py)
Press any key to continue . . .

[ornate elk]

Try delete the venv folder and relaunch

What I do to safe space is to put the models on an other drive and link them in the webui-user.bat

My guide covers this too

[upper linden]

i did that i linked my loras and models

[ornate elk]

Hmm okay

Let me know if deleting the venv and relaunching fixes it

[upper linden]

appreciate all the help

[ornate elk]

No problem

[midnight coyote]

Bump

[upper linden]

it worked, now i just need to add my extensions

[paper dagger]

Anyone ever try docker containerizing all of this so things retain their states? Seems a lot like early days of skyrim mods. One change and it breaks stuff?

Learned a lot here but still lost months later

[real fjord]

Hello, my subscription to StableAudio AI is expired on December. However, I still not downloaded all of my generated music in my dashboard. Is there any way I can differentiate the files I generated sing Free tier to those I generated using Creator License? Its just the dashboard for collection of generated music is just a scrollable window and there is no label or tags if that music is generated using free trier or pro tier

[balmy monolith]

Would anyone care to help me work out why SDXL isn't training in OneTrainer please. I've put my settings in #🏞｜general-with-images

[azure magnet]

Hi, i'm using comfy with python 3.11.8 what is the most recent version that i could install?

[daring ivy]

How to make animations?

[lucid yoke]

Is there no folder called upscale?

Its in the same model structure as loras, stable-diffusion, controlnet etc

[balmy monolith]

I think you should be using 3.10.11 https://www.python.org/ftp/python/3.10.11/python-3.10.11-amd64.exe

[balmy monolith]

You said that you are using 3.11, that will not be compatible, only 3.10.xxx

[copper turtle]

I just installed Ishqqytiger's amdgpu webui based on automatic1111 and i also installed zluda using this guide https://github.com/vladmandic/automatic/wiki/ZLUDA however i guessed i didnt need to install SD.Next since i'm using automatic1111, i can launch the webui but upon trying to generate my first image without touching any settings and also not having installed a specific model yet i get this error: "RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same"

anybody know what it is? i can give the full log if you need

[lucid yoke]

i see peoiple saying indeed to make a folder called ESRGAN

models/ESRGAN

might wanna refresh the UI to make it show up (refresh the page or reboot the ui)

yeah remove the real

its what i see on the a1111 github solved/closed

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/16593

hmmm

i mean its a guy with the same problem

but he fixed it by renaming

hmm

what version of A1111 are you using?

in the webUI somewhere in the bottom

hmm yeah your on the latest version. any reason your using A1111 instead of forge?

but stupid question did you reboot the web-ui

i recommend forge as its faster then A1111 for SDXL but yeah

in the pinned msg theres a guide

forge-webui

hmm it should show up after that

you could try closing the cmd and reopening it

just in case

and A1111 hasnt gotten a update since july

hmm yeah i would either recommend switching UI's to either Forge or Swarm (swarm has very usefull tooltips inside the UI)

but swarm can be a bit more confusing since it gives you more control

are you on a NVIDIA card or AMD?

ah then either forge and swarm are fine.

if you were on AMD i only know the guide for forge

there are some extentsions but forge and a1111 use the same backend iirc

ah Swarm UI uses "comfyui" as a backend. (noodles) so it has the benefit of an optional comfy ui use and speed of its back end

but without the noodles

i got it on darkmode but this is the UI you work with then

but theres more themes

[ornate elk]

Hey you should follow my AMD install Guide and dont use the sdnext webui.
Automatic1111 and Forge work much better with ZLUDA.
You find the guides here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

If you want upscalers to work create an ESRGAN folder in the models folder.
Then put the upscale model In there

The folder also will get created automatically when you use hires fix and choose an ESRGAN based upscaler

[copper turtle]

quick question, there appears to be no C:\Zluda directory

[ornate elk]

You have to create one

Its stated in the Guide

[copper turtle]

quick ctrl+F says that that's the first time C:\Zluda is mentioned, unless you mean the installation path of the webui?

[copper turtle]

oh

mb

[ornate elk]

No problem ^^

[copper turtle]

oh wait i'm looking at the ROCm HIP SDK 5.7 with ZLUDA Setup: section, in which the folder is named SD-Zluda

[ornate elk]

Oh thats the wrong one. Only use rocm 5.7 if you have an RX 580 gpu

[copper turtle]

i do got a 580

[ornate elk]

Oh then your on correct with the 5.7 guide

[copper turtle]

but i've installed SDK 6.x version because trying to install sdk 5.7.x gave me a general error 215

bruh

my god im just going to try everything from the beginning again

[ornate elk]

delete the .zluda folder and the venv then relaunch

[daring ivy]

how to make animations?

[copper turtle]

why is some 10gb pytorch clip model being installed

[ornate elk]

Its one time for sdxl models

[ornate elk]

Can you show me the full CMD log when you launched auto1111

[ornate elk]

Looks normal. Can you show the stable-diffusion-webui models folder?

And the ESRGAN folder

[exotic raven]

hello, im havin trouble getting an extension to work. can any1 assist?

im trying to get faceswaplab extension to work. its not appearing in my webui

[lucid yoke]

are you on AMD?

[exotic raven]

amd cpu, nvidia gpu

[lucid yoke]

automatic1111, reforge, swarm?

[exotic raven]

automatic1111

[midnight coyote]

Please help me with that 🙂

[exotic raven]

@lucid yoke

[lucid yoke]

ah i dont know, these were some details nessacary so someone could help

[lucid yoke]

the error shows a fix no?

[midnight coyote]

But i dont get it,all i want is to install flashattention 😦

[exotic raven]

hmm any1 around that may be able to assist? @lucid yoke

[lucid yoke]

open the file you use to launch it with notepad (probably its a .bat no ?) then change the line KMP_DUPLICATE_LIB_OK=FALSE
to
KMP_DUPLICATE_LIB_OK=TRUE

but that may cause crashes or silently produce incorrect results

[midnight coyote]

there is no text as that

thats the part what i dont understand

[lucid yoke]

hmm seeing how this isnt directly a SD Web ui problem but something else

https://github.com/Dao-AILab/flash-attention/issues

[midnight coyote]

its for comfyui

[lucid yoke]

id look into a more comfyUI dedicated space then. seeing how most of the help questions here are done by two people lately

[midnight coyote]

im in 9 groups and waiting for the rescue xD

[lucid yoke]

not that i dont want to help but its a very specific problem to a program that i personally dont work with

maybe create a issue on the github page

[midnight coyote]

i have no hope on those pages when i look at the question asked 3 weeks ago and they still didnt answered him lol

[lucid yoke]

i mean yeah its probably a small group of people doing this for a hobby

like im a K12 techie with some extra time on his hands but if i get bored of this i will probably move on to other communities/hobbies

its not like were getting paid for this

[ruby yoke]

your original error says you have multiple copies installed. you can't just rename a dynamic link library like that

the error also told you to go to http://www.intel.com/software/products/support/ - did you do that?

you are probably looking at uninstalling a bunch of stuff, then rebooting, cleaning up your system, rebooting again - and once you're sure all the duplicate copies are gone, and out of the registry, instaling fresh

[oblique walrus]

lol the intel site is just given as a generic "find support here" suggestion.

https://civitai.com/articles/9555/installing-triton-and-sage-attention-flash-attention-and-x-formers-win Step 2 there is about using a precompiled whl file. the generic error is something else entirely.

I would suggest they just have to restart and try again

[ornate elk]

but i said you need to create an ESRGAN folder

not realesrgan

and in this folder you need to put the upscaler

rename is fine

[midnight coyote]

I did that so many times i don't want to count anymore

[oblique walrus]

@midnight coyote honestly, i think you can skip triton for windows. Triton has never really been important for end users

What is your goal with using that guide? are you doing development work? You might be better off creating a proper linux environment for that.

[fierce moat]

yep I give up on triton on windows long ago

the torch compile thing claims to make things faster

[oblique walrus]

i'm not even sure what the benefit would be of triton

[fierce moat]

u know what, I tried that triton stuff and torch compile

in linux

but unable to use it becoz out of memory 🤣

waste of time

[fierce moat]

to utilize torch compile to run generation faster

https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html

I think it only make sense when u have huge vram, and then you want to utilize them for faster speed

and then we also have deepspeed by microsoft that only works in linux, too, 🤣

but torch compile really speed things up, if u have the vram

[oblique walrus]

that link says torch.compile doesn't need triton as a dependency. doesn't even mention it

i honestly don't think there's any benefit to it at all. it's more of a data center tool to get 0.5% gains

[fierce moat]

I think it needs, at least there is the error I am getting when trying to run torch compile

I'll said its more than 30% increase at least

[oblique walrus]

i chalk it up to a massive waste of time from stock cuda

[fierce moat]

I think you are right, I think when I say torch.compile, i mean torch compile with inductor which claims to be the fastest

tensorrt still an options tho

[wicked widget]

yes it does. you spread allot of misinformation

[oblique walrus]

demonstrate how to get speed boosts from triton on windwos then please

i'd love to know better if my assumptions are wrong

[fierce moat]

he is not wrong, if you using torch.compile with other options like tensorrt, you don't need triton

but of coz folks are talking about torch.compile with inductor which is also the fastest and default options

[wicked widget]

gotcha

[wicked widget]

do you know if can get compile+ to work efficiently on 3060?

[oblique walrus]

i thought you were the expert. i guess you dont know after all. i'd appreciate if you'd come correct next time.

[wicked widget]

Who claimed I was an expert? Youre a fool.

[oblique walrus]

it implies you know better when you tell someone they're spreading misinformation.

i never made a direct claim. i will actually own not knowing. i've drawn assumptions based on what i've found, but i do not know

[fierce moat]

I think you can, on linux, for sd1.5 and sdxl, however for flux it require 24gb vram to compile

[wicked widget]

ah bio

boo

i do use linux but im capped at 12vram

[fierce moat]

and then you will have issue with lora something again

yea annoying

i gv up

[wicked widget]

i was able to get eugene backend to work

but that shit is so slow

not worth

[fierce hollow]

Why is OpenPose not working?

[wicked widget]

preprocessor is empty

use openpose

[fierce hollow]

No 🙂 If I select any Preprocessor, it doesn't even see the skeleton 🙂

[wicked widget]

very strange

everything looks good

[fierce hollow]

BUT, if I insert a non-black background... Just a second

[wicked widget]

I ditched a1111 and forge cos it would break like this for no reason. comfy hardly breaks on me

so im unsure how to proceed to fix

is there anything whacky in the terminal when you run it?

[fierce hollow]

@wicked widget 🙂

[fierce hollow]

3070ti 8GB

[wicked widget]

some reason its saying you dont have enough mem lol but 3070

wiat misread

[fierce hollow]

This started happening when I turned on Conrolnet

[wicked widget]

using any vram lines in webui ?

like --low-vram

etc

[wicked widget]

you got the 8gb vram card right?

[fierce hollow]

Yes

[wicked widget]

you may need to use vram lines args cos youre ooming. youre missing 500gb vram to run the process
turn off any wallpaper engine etc, run nvitop in console to monitor vram use and turn off anything not important

[fierce hollow]

This is a warning error, can it really affect the generation?

[wicked widget]

yes

also

[fierce hollow]

How do I do this?

[oblique walrus]

whats the openposexl2 model?

i tend to stick to the controlnet loras . usually the 128 rank version

[wicked widget]

add --xformers --medvram

youre missing xformers too

which you should alwags use as rtx user

[fierce hollow]

I was initially advised of it https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/blob/main/control-lora-openposeXL2-rank256.safetensors
But the result is the same

[wicked widget]

xformers alone will help prevent ooms

[oblique walrus]

https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank128 these are the ones i have success with in forge ui

[fierce hollow]

@REM set COMMANDLINE_ARGS=%--xformers --medvram% ^
Yes?

[wicked widget]

yep

think thars right format too, unsure cause your webui has allot more stuff in it than mine

[fierce hollow]

Now I'll try it with new arguments and then download yours.

@wicked widget Did it trigger the argument?

[oblique walrus]

i'm thinking that thibaud's model, which appears to be a lora style cnet as well, is not properly supported by forge.

it's a very hacky situation with the current version i find. unless you're using flux, i might recommend you stick to reforge fork of it instead

[wicked widget]

no

its still using normal vram

and no xformers

when adding it follow the format of the other ones

so take off that extra % at the end

[fierce hollow]

I started another one.bat, launched with arguments, let's wait...

[oblique walrus]

set COMMANDLINE_ARGS= --xformers --cuda-stream --cuda-malloc --enable-console-prompts

is my line. i'm not sure if you support cuda-stream or not on the 30 series. i also have console prompts on cause i think its neat

[wicked widget]

it does

use malloc and stream

should also help

biggest one tho would be xformers tbh

[fierce hollow]

The main thing is that my VPN wouldn't suddenly turn off right now, otherwise it would be fun 🙂

[oblique walrus]

interesting that it's reinstalling pytorch with cuda again, but hey. that can't hurt. maybe. lets see

[ornate elk]

Wrong Python Version

[oblique walrus]

ah yeah 3.12 is the iffy one isn't it. you want a 3.11 version

good spot @ornate elk kudos for you.

[fierce hollow]

I'm still waiting 🙂

It scares me a bit that the download is not continuing 2.1/2.4

[oblique walrus]

you're going to want to gut your python install anyways. isntall the proper 3.11, and then either manually reconfigure your venv to point at the right one, or just delete it and let the bat file rebuild it. meaning, the whole download thing again.

[fierce hollow]

)

@oblique walrus

[oblique walrus]

there's going to be a lot of problems until you repair your python install. for stable diffusion use, you'll want 3.11 , not 3.12. python, that's just how it is

[fierce hollow]

Which Preprocessor should I use?

[oblique walrus]

preprocessors just convert an image into what the controlnet model is trained on. so an openpose model wants an openpose input. since it seems that you've already got an openpose style image, you wouldn't preprocess it.
if you had a photo of a character, you'd use an openpose preprocessor to detect the pose and pass it to the model.

[fierce hollow]

If I choose Openpose, there are no models of yours, only mine, which does not work.

[oblique walrus]

did you load the ui before or after copying them into the folder? there's the little blue refresh button next to the drop down. that'll refresh the list.

you don't need to select the radio buttons. that'll just narrow it down. you can just manual it and select the one you want from the full list.

[fierce hollow]

I know about updating the list) But I even restarted the AI.

Excuse me

I'll try every process now.

@oblique walrus The situation is the same

[wicked widget]

does it still say the same thing in terminal

missing 500gb

?

[oblique walrus]

youre trying to preprocess an existing openpose image, so i'ts detecting no pose there and converting it to a black image, sending that to the model. Use no preprocessor instead. you only need to do the preprocess if you want to detect a pose from an existing image. but youve got your own pose image already

[lusty hedge]

for some reason my replacer extension isnt showing up, its saying im missing a infotext_utils.py file in modules. it was working fine a while ago and now it doesnt work anymore

[oblique walrus]

did you upgrade to forge? it uses a newer gradio. that seems like a gradio 4 thing. i might be wrong. i dont want to spread misinformation

[ornate elk]

Do you have a full cmd log?

[lusty hedge]

already resolved it, had to update my stable diffusion

[fierce hollow]

Yes.
Tell me if I install ComfyUI, will you help me set it up so that I can do poses here?

[wicked widget]

yeah

install comfy, and comfy manager

and throw in my workflow, install missing nodes

[fierce hollow]

I'll install it now, can I send you a private message as soon as I install comyfyUI??

[wicked widget]

yea if u want

[fierce hollow]

I wrote to you)

[exotic river]

well, i am learning alot due to the number of crashes/issues installing new nodes causes =p

[fierce hollow]

Tell me, who here is good at working with OpenPose? Who can I write to?

[ornate elk]

Can you share an full cmd log + Screenshot how you try to get it working?

[fierce hollow]

Which program is better for you? ComfyUI or Forge?

[void quail]

basically this is the 2nd time i ran the cmd, this is the error that also appeared on the 1st run through

[fierce hollow]

@wicked widget helped me set up, Maybe he will add something, but it doesn't work or it works every other time with a drawing of a person, but I definitely need him to work with the skeleton with OpenPose edit...

[wicked widget]

do this @fierce hollow

[ocean scarab]

can i generate pics in discord?

[fierce hollow]

@ornate elk

[lucid yoke]

Brother ellie said to post full cmd log too😭

[lucid yoke]

Check artisan faq

I think its on you instead of the hardware, i get amazing results on a 3070ti

Show prompt?

[lucid yoke]

then its not struggling to get good results but your struggling in general

mood

what im not berating you. but if you are struggling to get good results its probably a prompting or setting issue but if you cant reinstall it because your GPU is dying its a whole different problem

but to awnser your question. 5080 is a decent AI card yes. its worth waiting for as you dont have to use all kinds of tricks like you do with AMD

i am getting excellent results on a 3070 so hence the comment struggling in general to get it working.

[pale salmon]

sup, dont tell me my 1,3k dollars worth of gpu having a hard times to upscale a picture by 3, is there any option to boost it idk

[ornate elk]

Which settings and model do you use?

[pale salmon]

WAINSFWILLUSTRIOUS, upscale was by 3

[ornate elk]

and whats your gpu?

and does upscale by 2 works?

[pale salmon]

RTX 4080 SUPER

[pale salmon]

yeah, 3 work too but it make me lags and generate slowly

[ornate elk]

make sure you dont run any games or wallpaper engine in the background while it generates

[pale salmon]

i guess i just use discord and youtube videos in 720p

[pale salmon]

tryed with 2.5 upscale

[ornate elk]

Whats in your webui-user.bat?

[elfin smelt]

Hello everyone, could someone help me achieve this kind of ultra-detailed result like in the second picture? I’d like to learn how to add more details to my photos and add a black outline to the lines for a more animated effect, like in the second picture.

This image is from MJ; I didn’t create it myself, but I would like to achieve the same level of detail as the one on the left. However, I don’t know how I can achieve this

The image was created using MidJourney, bro

Hmm, I don’t know enough about ComfyUI to answer that. I’m a complete beginner when it comes to Stable Diffusion, sorry

Haha, we’re in the same boat, bro

[lucid yoke]

just the eyes? or you mean an overall style

[lucid yoke]

oh depending on the lora you can get pretty decent results. tough i like to upscale the image first and then impaint the yes

eyes

but i use swarm ui so i can just <segment face>

tough it heavyly depends on the style and main model

oh wait are you using SD 1.5?

what is the model you are using?

well if you are using 512x512 pixels the ai has little space to work with but thats the downside of SD 1.5

whats your graphics card?

there are ways to utilize a AMD card for this

pinned messages show a tutorial on how to use stable diffusion with a AMD gpu. i personally use Nvidia so i cant help you on that one but again i dont know what graphics card you have so i cant say if its enough for it

[elfin smelt]

Up 👋🏼

[lucid yoke]

that is CPU

[lucid yoke]

i replied to you initially but conronron replied to me lol

brb but after i can check

open task manager

[elfin smelt]

Oh, I didn’t see it, sorry 🤣 How could I improve the details of my images like the second one above? I import an image and just want to enhance the overall details and add black outlines around the shapes in an anime style

[ornate elk]

with that gpu you should try out forge webui

[lucid yoke]

Cs1o is typing but he beat me to it

was helping ma with homework lol

[lucid yoke]

hmm just "enhancing" a image i do not know but you can generate a image like that and upscale it

i would impaint it

to edit some things

but not the entire image

i wouldent say its bad. its a good GPU for general use

[elfin smelt]

Enhancing the details, like making the hair, clothing, and outlines finer in a manga style, adding reflections, shadows, etc. I want the image to be ultra-detailed.

[lucid yoke]

both work fine

its the model

that needs to change

no clue from me homie

[lucid yoke]

yeah thats something you either do with photoshop or do in advance

it does, but you have to mention it

[elfin smelt]

What do you mean, bro?

[lucid yoke]

well if you want reflections, thick outlines etc you have to tell the ai to do it

or you gotta after edit it manually

would you believe me if its different for every model

if you got the model from civitai.green (sfw) or .com (nsfw)

[elfin smelt]

Do I need to install specific things? Can you guide me? Because I’m not sure exactly what to do.

[lucid yoke]

a checkpoint on civit ai for example

[lucid yoke]

whats your GPU

but CPU based generation is not gonna be really fast so i hope you got the patience

[ornate elk]

for your card forge

it will work better

[elfin smelt]

I have an RTX 4070 Ti Super

[lucid yoke]

ah yeah then a LOT is possible

i could hop onto a quick call

[ornate elk]

yep there is even an extension for better eyes

[elfin smelt]

I speak French, bro! If you do too, let’s go! Ahah

[lucid yoke]

english or dutch for me fam

[ornate elk]

you can try enablen the never oom button at the bottom but only for tiled vae.
Also try the fp16 vae sdxl fix as vae:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[lucid yoke]

but i will say all AI terms in english because i am NOT translating things into dutch

[elfin smelt]

Oh, my bad… Can you explain it to me in writing? Maybe in DMs?

[lucid yoke]

hmm i could condense it into a few lines
go on civitai.green
images tab
find a image you like (that looks good) lets say this one https://civitai.green/images/48136708

copy their resources used
copy their prompt
copy the negative prompt

edit the negative prompt and positive and edit as you go (so i would advice to keep the "masterpiece, best quality, newest, absurdres, highres," but change the rest if you want to keep that style etc)

there are also style loras in the "models' and "styles" but you need to use the filter button to match the model you are using (so like NightNoob in this one is a illustrious model)

[elfin smelt]

So I need to use img2img and add these prompts then?

[lucid yoke]

existing image: not possible as far as i know

new image: yes!

because you dont ADD detail to existing images

you can generate a image without a detail/style lora

and use the same parameters and to generate it anes

anwe

BRUH ANEW

but a random image without the parameters? tough luck as far as i know

[elfin smelt]

It’s an image I already created on MidJourney, and I just want to add more details to the overall image, kind of like Leonardo AI.

[lucid yoke]

☝️

[pale salmon]

ok thx

[lucid yoke]

midyourney made it

so you dont have:

1. the model
2. the original prompt (midjourney enhances it iirc)
3. the seed
   4.specific style they used
4. image parameters, karras, euler, euler a, dpm ++ 2m etc

thus impossible

[elfin smelt]

What Leonardo AI does, Stable Diffusion can do it even better, right?

[lucid yoke]

its a different model

they excel in different things

online image generators work vastly different to local ones

[elfin smelt]

So adding details to a base image using img2img is something Stable Diffusion can’t do?

[lucid yoke]

hmm let me reiterate

you can apply a detail lora to pimp a image before its made

you cannot say

hey ai enhance the image in this way and keep all the things i want

reflections etc is just a tad too advanced to add in post

[wicked widget]

you can throw an image in img2img and keep the denoise low

[lucid yoke]

its not gonna do what he wants

[lucid yoke]

like this

hes better off adding those details using a lora before making the image instead of getting one off midjourney and then editing it

[wicked widget]

and why cant it? that image looks like it was ran through photoshop tbh

im2img with low denoide to add stuff

yhen ps to change the post processing

which is what that image has

[lucid yoke]

if you dont have the parameters its gonna vastly change stuff

but he could try i suppose

tried adding cellshading and thick outlines (no same seed, parameters, prompt and sampler) (OG)

end result:

overal downgrade

but i said all that needed to be said. good luck though. i do recommend if you really want detail on a image without the parameters to use photoshop and edit away

[elfin smelt]

Well, I’ll give up then if it’s not possible… I thought Stable Diffusion was capable of doing as well as Leonardo AI. At least that’s what I was told.

[lucid yoke]

its capable pre processing, inpainting etc but its not fair to compare a detail editing feature if you have the manual and parts for leonardo but have nothing for SD

[ornate elk]

Its possible but it requires a lot of steps

[elfin smelt]

What are all these steps ?

[lucid yoke]

managed to change it into this in the meanwhile but getting a filter over it, ticker lines, cell shading, reflections while keeping the original image intact? nah

Maybe CS1o is a wizard

but at that point its a creative process instead of tech support

[pale salmon]

i test comfyui and i get no problem, the problem is stable sadly idk why @ornate elk

[lucid yoke]

i used like three prompts

first the girl:
masterpiece, best quality, good quality, newest, detailerlora,
<break>
1girl, thick outline, cell shading, solo, (calico ears), two-tone hair, gradient hair, (white hair), medium hair, (orange hair tips), (long sidelocks:1.2), green eyes, yellow eyes, multicolored eyes, gradient eyes, (red berry hairpin:1.1), red scarf, black coat, raised eyebrows, wide-eyed, happy, open mouth, tongue out, (tongue stuck frozen), frozen tongue tip,licking popsicle, (looking at viewer:1, side view, close-up, glint, steam, daylight, sunlight, winter, snow, vivid colors,

then i masked the background. like 5-7 tries till i got a nice result

best quality, good quality, newest, detailerlora,
<break>
space, galaxy background, stars, dark background

no i used image to image

but i had the seed, main prompt and sampler (and model) so i could change it in the same style

its not done at once

i did it in like three to eight steps

you can

you use image to image

mask the background

change the background that way

🎉

Neither, i use SwarmUI

its not CPU based

so yes

i would assume so

if theres a setting to run it on CPU i dont know as i havent ran into that barrier

[wicked widget]

anything and everthing related to images will always be processed better on a gpu than a cpu as gpus were made for that specefic task

which gpu?

and how much vram

[lucid yoke]

☝️

[wicked widget]

oh cool ty

oof 4

youre deff gonna need to use --low-vram and get xformers to help

[wicked widget]

yeep, conversation with him is always stimulating

yup

just slower

than what other ppl would do

theres a few 4gb vram users

[lucid yoke]

can a 1650 run sdxl?

ah i guess the current ui's have them very optimized nowadays

back when it was new i couldnt run it with my rig lmao

[oblique walrus]

So many wizards here. Truely a magical time in technology space

[wicked widget]

pretty sure yeah, esp on comfy

[lucid yoke]

though the newer the user the less they need to use comfy ngl

depending on their background ofc

[wicked widget]

yeah lol im just defaulted to comfy lol

habitual

[lucid yoke]

swarm habitual user here lol

sometimes i mess with the noodle but mostly i keep it simple

[wicked widget]

i went from a1111/forge to swarm to complete comfy

noodles are fun

[oblique walrus]

A lot of it is the libraries that are optimized. The UI's actually tend to mess up those optimizations fairly often. I would recommend swarmui as the king of the realm

[lucid yoke]

yep

[oblique walrus]

what i don't like about using non swarm workflows in comfy is that so many custom nodes are poorly coded and unoptimized

[lucid yoke]

i just get really tired after work so i dont have the concertation left for the noodles and nodes but i FINALLY got my gigabit connection working after drilling a hole thru my floor so im gonna spend some weekends looking into it more

[lucid yoke]

got a seperate comfy setup for the custom ones

[wicked widget]

giga internet is such a great experience

[lucid yoke]

keeping the swarm lean

right. my previous house had like 300/30

okay but not too great

now 1000/1000 stable

[wicked widget]

rthats like bare minimum for these days

great 10 yrs ago

[lucid yoke]

for 15 a month extra i can upgrade to 4000/4000 but thats a bit much for me ngl

[wicked widget]

if you do NAS and whatnot i would

but optherwise

good to stay on giga

[oblique walrus]

i've got 150/150. i've never had issue with it. I download to a 3TB hdd primarily which prbably makes the write speed fo the HDD the bottleneck. I could be wrong about that. Never really done the actual math

[lucid yoke]

im the onedrive manager at my work i get like 100tb per licence of cloud storage

so no need for a nas

[wicked widget]

lmao

fair

[oblique walrus]

❤️ volume licensing

[lucid yoke]

granted i dont trust my work enough to not randomly snoop here and there so i only store safetensors or files without a extension

[wicked widget]

upload ur nsfw folder u coward

[lucid yoke]

i work in K12 rather not get fired lmao

[wicked widget]

oh yeah

no

lol

hard no

[oblique walrus]

boss makes 10 dollars and i make a dime so thats why i use volume licenses all the time - updated rhyme for modern days

[lucid yoke]

but because of my work experience i can easyly translate the people skills to some of the users here

[wicked widget]

yeah some people need people skills xD

[lucid yoke]

i got the full adobe licence for free :3

[lucid yoke]

some people need reading comprehension skills

lmao

[oblique walrus]

my bro used to work school districts. when kids found ways to bypass his protocols and get minecraft running, he'd let them do it. cause it was impressive

[wicked widget]

hey in my defense, i know how to read, im just LAZY TO READ IT ENTIRELY lmao

[lucid yoke]

oh i got the network open for portable CS 1.6

if they manage to get it

[oblique walrus]

why not tfc? teach them the way

[lucid yoke]

blocking the fortnite online copy tho

i only reward certain games, i consider a lot of it brain rot and just block it

[oblique walrus]

hide a mounted image of brood war and see what happens

probably good to not get fired though

[lucid yoke]

blocking tiktok on a few accespoints but allowing it on some (all the good light spots are blocked)🤭

really like to see them get annoyed by the tiktok but hey its a school its not for that

[oblique walrus]

sooner than later the kids will realize their parents use tiktok now. and they will stop

[wicked widget]

yeah good thinkg tiktok is banned, shit is pure brainrot

[lucid yoke]

alas im not the final decision maker i would have blocked way more on the school web

[oblique walrus]

https://tenor.com/view/evil-laugh-gif-20076210

[lucid yoke]

got minecraft for education though, its dope for projects with the local gov of designing new neighborghoods etc

kids love it

[wicked widget]

oh get

ECO

for them tbh

[lucid yoke]

ECO?

[wicked widget]

its a minecraft like game but more depth, teaches cooporation amongst other things

https://play.eco/

[lucid yoke]

ah id need to see if theres a interest for it with the more tech oriented teachers

most teachers however are better then the kids with tech and thats not saying much

[wicked widget]

dang yeah

[oblique walrus]

version 6?

SDXL will fit into 6-8gb of vram. I'd suggest that as your bare minimum. Really depends on what your gpu is

[lucid yoke]

Its not weak, its just not supported for sdxl

[oblique walrus]

maybe if you used comfyui with high level of optimized workflows, like VAE tiling or whatever, maybe you could get away with SDXL models. Don't even think about SD3.5 or Flux models imo. Though i'd applaud you if you tried experimenting with GGUF versions of those.

If i were you, i'd focus most of your efforts on using SD1.5 ecosystem of models. There are tons of refines in that space

[lucid yoke]

SD1.5 based64 has a plethora of solid loras to enhance style, concepts and more, you just have to upscale for nice results

For good speeds yeah but the quality will be lower. If you dont mind waiting. Potential crashes etc sdxl pony is fine

[oblique walrus]

Pony is an SDXL refine so you've gotta have more vram to happily run it. You may be able to get away with it, but it'll be like 20min per image .. heh..

[wicked widget]

I prefer illustrious over pony for xl

[oblique walrus]

i dont even do anime gens and i concur with that. PonyXL poisoned the XL refine scene with bad practices

[wicked widget]

"good" is subjective

[lucid yoke]

Illustrious gets you nice images with low effort

Illustrious gets you great images with effort

[wicked widget]

go to civitai, to browse for models, turn on filters to only show illustrious or pony models

[oblique walrus]

SD1.5 the best anime model is "Anything" and the merges/refines based around that

[wicked widget]

and look for a mdoel you like

[oblique walrus]

so nice you had to say it twice

a lot of that is operator dependent

[lucid yoke]

I used a different seed to generate that sentence:^)

[oblique walrus]

sampler, steps, prompt, cfg, all the settings and inputs determin a lot of that. it's not just the model you use

[lucid yoke]

It's a numbers game and frankly personal styles etc. its a whole factor of things you gotta try n learn. the prompts i do post are not the ones i keep to my self (call it gatekeeping idc)

I can show you how to do simple stuff but i cant shove the best prompting practices into your head as it differs per model merge etc

[oblique walrus]

gatekeeping is like "you aren't a real ai user if you don't prompt the exact way i am" .. i think if you're keeping some secrets about your skills and approaches, that's something else entirely.

And in the age of youtubers stealing all the good tips they find online and acting like they came up wiht it all, i entirely recommend people do this. Break the meta and keep it to yourself .

You'll note that i rarely post any images publicly

[lucid yoke]

Ive been accused of knowledge gatekeeping before lmao

[oblique walrus]

Let them accuse you. You know your worth

Damn knowledge commies

[lucid yoke]

For this sorta stuff i wouldnt worry about it unless you get a better pc to make your own model

But at that point youd have the knowledge

[oblique walrus]

i often feel like 90% of the effort training a new checkpoint, or a lora even, is building the dataset and it''s captions.

The actual training is just flicking a switch and waiting.

[lucid yoke]

Once you start into making loras etc you learn lot of terms and tags really fast

Yeah

And honestly its sometimes dang hard to think of a lora to make

[oblique walrus]

For practice, i train instagram girls lol. I never publish the models i make though.

[wicked widget]

thats very shady

[lucid yoke]

O actual good idea for practice, but it feels hella immoral somehow

[oblique walrus]

is what it is. its all for private use and educational reasons

[lucid yoke]

I would use random anime screenshots of characters

Fan arts etc

[wicked widget]

one thing i truely loathe of ai and SD in general is how easy it is to enable the bad actors

[oblique walrus]

i draw the line at creating nude images, but in the tests i do a lot of the models will just bias that way on their own. I never publish any of that though. I dont even think any of it is saved

instagram just happens to be a very good source of datasets that i don't mind looking at

[lucid yoke]

Yep, i unfortunately had to learn how easy it is to fake stuff with it. (I give workshops to teachers for awareness)

[oblique walrus]

they're not all i train too. i find other sources and approahc other styles where i can. None of it is published either

I got ethics but i'm not going to lie to myself like i don't like looking at cute girls

[lucid yoke]

Oh i get it and for training i get it. Easy datasets etc but alsong they stay sfw private practice

It wouldn't be the first time hearing from a colleague a girl ending it all bc of a fake

[oblique walrus]

often when a nude image comes out of one button prompt situation (random prompt generator), i'll feel guilty. But then i remember i put (nsfw, explicit, nude, naked:1.5) into the negative and i'm doing my best. "No nuu, it's the models that are wrong"

I've actually stopped using cyberrealistic so much due to those strong biases

[lucid yoke]

Yeah and like you said you don't keep them and delete them so you have done in this case the best possible steps

[oblique walrus]

happens to guys too. "we'll post all these pics online" has caused many , not just girls, to end it all. sad and tragic.

[lucid yoke]

Oh yeah definitely but you don't hear a lot about the guys. Its just the first time i heard it of a 14yo was so tragic i started the ai awareness workshop at my work

[ornate elk]

Hey, just want to let you know this is the tech-support channel

[oblique walrus]

Yeah advocating awareness and safety is a big deal. Strong effort

noted

[lucid yoke]

Fair, conversation just took off

Hopping to offtopic

[ornate elk]

Thx

[neon orchid]

hey guys

a question

I was talking to a colleague and he said "Why are you still using Pony Diffusion? Illostrius is much better"
I wanted to ask... is that true?

[oblique walrus]

yes. For that realm of work it is.

Pony is poorly trained and disaligns the model from the base substantially

[neon orchid]

But if that's the case... why are there more pony models than Illostrius? of characters

[oblique walrus]

Pony has been out longer

[onyx solar]

"start"

lul

It's official name is the "Windows key".

If you're on a different OS, you'll see it called meta or super.

[loud pond]

Whats the difference between illustrious and pony? will i still be able to use illustrious on A111 with all my lora's for pony? or do i need to get new loras?

[wicked widget]

beep

[lucid yoke]

Yes + prompt

[shy charm]

the one time i want to use comfyui

i tried to manually pull it and it does this

so i guess something is wrong with the repo

[signal jasper]

Hi all, I am following this guide to learn how to train my own Loras in ComfyUI but I am stuck on this error message. I've followed every step. Installed the custom nodes, installed the requirements for each of them but I get this error when. I have searched and come up with nothing. Could anyone help please? Ty https://www.reddit.com/r/StableDiffusion/comments/193hqkz/lora_training_directly_in_comfyui/

[wicked widget]

use onetrainer or kohya ss to train loras

[signal jasper]

Thanks, ill pivot to that

[final horizon]

hello everyone, got a question. Can I use stable diffusion 1.5 loras in Flux comyui? or does the Lora needs to be matched with the base model for it to work?

[lucid yoke]

No 1.5 loras dont work for flux
Yes the lora needs to be matched with the base model
1.5 with 1.5
Xl with xl
Pony with pony
Flux with flux

In some rare cases a xl/illustrious/pony lora works on another of the previously mentioned but rarely

[ionic spire]

Hello there, so I installed this a little over a year ago and used it just fine constantly throughout the last year with no problems. I then tried to update it and everything seems to have broken. I even tried a new install and I just get the same error.

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

[lucid yoke]

Updated swarm? Comfy? Forge? A1111?

[ionic spire]

A1111

I think, thats the only one that looks familiar

It seems to be because of the models I have. I guess maybe there old and dont work with this newer version.

The weird part is, I did this new install and waited for it to finish compiling and did a test run and it made a picture using the same model. I then restarted my PC and now its giving this error.

Yea I have no clue, no matter which one I try to load it gives me the same error

Doesnt matter how old or new it is

[ornate elk]

Can you show a full cmd log?

[ornate elk]

Zluda couldn't be found.
Delete the .zluda folder and relaunch the webui-user.bat

[onyx elk]

Hi, I tried to run the openvino accelerate script and this error pop up, everything is still working normally but i can't use the script.

I'm on intel i7 1360p, intel iris xe, 16g ram

[vocal burrow]

You re probably best trying to use sd.next if you want to use openvino.
It s a fork from A1111, it should be easier to install. https://vladmandic.github.io/sdnext-docs/OpenVINO/

[onyx elk]

I have use SD.Next, it works well for 1.5 but I couldn’t load XL model with it due to low ram

[vocal burrow]

Sadly, I doubt it will work any better with A1111

[onyx elk]

I know, I'm just going for false hope

[vocal burrow]

intel iris are iGpu (understand gpu that does not have its proper memory)

[onyx elk]

A1111 actually loaded XL and pony though I don't know why

[vocal burrow]

therefore you only have 16gb of ram for windows + SD

for sdxl models it can be tricky.

ComfyUI with some OpenVINO stuff might be more optimized but I still doubt it will work https://nolowiz.com/how-to-use-comfyui-with-fastsdcpu-and-openvino/

(haven't tried it)

[onyx elk]

I thought ComfyUI only support ARC and not Igpu ?

Also Node based stuff is scary

[vocal burrow]

looks like it does support igpus

https://github.com/BetaDoggo/ComfyUI-FastSDCPU

[vocal burrow]

I get that ^^

[onyx elk]

Oh I have seen it before, it is very fast for cpu genning

[vocal burrow]

you ll have to convert models tho.

It happens a lot for those hardware specific system.

[onyx elk]

That is why I'm a bit put off

[vocal burrow]

same happened for using tensorRT on my nvidia

[onyx elk]

Thank you for your help, I'll check it out or if it doesn't work I'll just continue using A1111

[ionic spire]

That seemed to magically fix it and im honestly amazed at how fast it works now. I was lucky if I could get a pic per minute and now its shooting them out in 5 seconds.

[loud pond]

does anyone know if i will be able to use pony lora's with illustrious?

[vocal burrow]

unlikely, illustrious seems to be based on "Kohaku XL Beta 5" => not a pony model

[loud pond]

oh alright thank you

[vocal burrow]

to be more precise, it will probably technicaly work but results won't be good

[loud pond]

whats the difference between pony and illustrious? so far ive only seen illustrious make better images?

[vocal burrow]

Both are ultimately based on sdxl. But pony's training has been... let s say pushed to the point where it is very different from vanilla sdxl. So lora made for it are unlikely to cooperate with other sdxl based models.

regardless it s worth a try, but I would nt hold my breath

[loud pond]

yeah so like putting aside loras, is there any benifit switching to illustrious or is pony still great to use?

[ornate elk]

Yep Zluda is really fast

And you can also generate larger images and use sdxl

Zluda has a better vram management

[ionic spire]

I should have updated it long ago, thank you so much for your help.

[ornate elk]

No problem 🙂

[vocal burrow]

it s totally up to your taste

[loud pond]

OH alright, i thought illustrious was just a better version of pony that came out. And i thought i would have to use it to get the best generations possible

[loud pond]

@ornate elk What is an easy why to upscale your images on A111 stable diffusion? right now all the images i try to upscale through the image to image latent upscale and resize tool are very distorted

[vocal burrow]

what parameters are you using ?

[regal mason]

is there a good solution to multiple frames generating in the same picture? It keeps dividing it into 2 or 4 frames and generating the same subject with slightly different angles

[vocal burrow]

lower the denoiser strength

[loud pond]

oh alright? to 0.7?

[regal mason]

what's more important the aspect ratio or number of pixels?

[vocal burrow]

no more like 0.4 or even lower

also most people don t bother with the refiner

[regal mason]

interestingly I was using 1024x1024...
I just changed it to 768x1024 and it seems to have resolved it

[vocal burrow]

both don t strand to far away from your model's training resolution
sd1 : 512x512
sd2 : 768x768
others : 1024x1024

you can tweak them by 30% or so in any direction usually without getting too much "duplicate artifacts"

[loud pond]

ok!

[loud pond]

why not? my images always come out really strong? wont a refiner fix that?

[vocal burrow]

it takes more time and people dim that the results are not necessarily worth it. You try it out and determine what works best for you.

[onyx elk]

Oh...I've been genning at 512 and upscale them is that bad ?

[loud pond]

ok thank you!

[vocal burrow]

For sdxl, sd3 and others. Yes. Outputs will be garbage at 512x512

straight forward

[onyx elk]

Yeah they look like smudge at 512 until i scale them up

[vocal burrow]

Otherwise yes it s very easy to end up with stuff like that https://cdn.discordapp.com/attachments/1002602742667280404/1112151998586961920/twoheads.webp?ex=67879403&is=67864283&hm=cfddae3f954a679c2f1277c97f13d44d3de3ac23aacd0718a46a875ef415de85&

so set the initial output resolution to some resolution close from the training resolution and with the desired ratio.
And upscale it afterwards if you want higher resolution.

[daring ivy]

How to make animations?

[lucid yoke]

i forgot did you have a 4070?

[daring ivy]

4060

[lucid yoke]

theres a Text to image model thats really good but img2vid isnt there yet. with your card youd take like 3-7 minutes for a 6-8second video

[daring ivy]

kk but how to make one?

[lucid yoke]

https://www.youtube.com/watch?v=COJCleld_H8

[daring ivy]

is there diffrent ui?

[lucid yoke]

Swarm supports it natively

but youd need to know a bit more out of the box

its beginner friendly but theres less 24/7 chat support

[daring ivy]

i have no idea how to use this ui on video

[lucid yoke]

i dont know what to tell you its well documented on the github on swarm

and i just sent you a whole video explaining it on how to do it on comfy

[formal trout]

With cogvideo

[ornate elk]

Hires fix in txt2img and SD upscale in img2img

[ornate elk]

For anime illustrous is better than pony, also you dont need the score tags anymore

[loud pond]

oh sick

[loud pond]

thanks ill install the extension

[ornate elk]

SD upscale is already included but you could also install SD ultimate upscale

[loud pond]

alright

thank you

[signal basin]

hey everyone, doing launch.py and it wants to download a model from hugging face. but the download always fails. what do you guys do?

it always just freezes

[ornate elk]

Are you on Linux or windows?

Which webui do you use?

And if your from Asia you need a vpn for the setup

[signal basin]

linux

usa

i can't even get to the ui

because it wants to download a new model and it won't complete

https://imgur.com/XZJRKSt.png

and it hurts that there is no resume on the server

got 300mb of it last try

[ornate elk]

Is it stuck on 17%?

[oblique walrus]

it's python on windows that is the problematic download client. not the server.

why so many tabs open?

also, just download it manually and copy it into the folder then.

oh you're not on windows. probably still the cause

[ornate elk]

Use the Adetailer extension

[signal basin]

that's linux, i can't get it with wget or browser either

just won't complete or resume

[ornate elk]

Try press space

To resume

[signal basin]

will do

[oblique walrus]

have you confirmed that it has in fact stopped downloading? Like if you go look at the file folder and the size of the file in it, does it stop going up? or is the network traffic over your NIC dropping off?

[signal basin]

yes

[ornate elk]

If pressing space doesnt resume then do this:
download a model from Civitai.com
Like Dreamshaper and put it into the models/stable-diffusion folder.
There delete the 1.5 model if its in there.
Open up a terminal and run
Pip cache purge
Then relaunch the webui.
It will skip the download.

[oblique walrus]

heh dns server. atm machine. til learned

DNS probably isnt' the problem though. That'll only connect the domain name to an IP address, then it's purpose is done.

[signal basin]

dns is fine

[oblique walrus]

TIL LEARNED! always thought it was server

[signal basin]

working on it now

[wicked widget]

which distro

[signal basin]

debian 12

[wicked widget]

hmm ok i dont habe much knowledge with that one

[signal basin]

basically ubuntu except more conservative

[wicked widget]

oh use sudo apt

[signal basin]

forcing new line didn't work

civit ai also craps out on me

im wondering if it's my isp

that's the one

that i'm trying to launch

[wicked widget]

thats old

look at the pins here

for install guides

more up to date

[loud pond]

does it make any difference if I used the source_pony and source_3d prompts in stable diffusion?

[wicked widget]

CS1o wrote allot of good ones

[signal basin]

ok

[wicked widget]

altenativsly i also suggest installing Stability Matrix for those who have nvidia gpus

[signal basin]

will try

[wicked widget]

sure one moment

[signal basin]

lol, i'm using 8.8.8.8

connection drops

not a dns issue

[wicked widget]

check ur dms hat

[signal basin]

feels like days of dialup again

[wicked widget]

lmao

what isp are you even with

[signal basin]

i'm loving Xfinity!

lol

the ubuntu iso 5.8gb is going strong....

are there torrents for checkppoint files?

[regal mason]

Does anyone know the best way to generate 2 distinct characters on each side of the image? Not looking to load in specific faces or ip-adapters, just have them not use the same character prompts for both subjects (using comfyui preferably)

[signal basin]

ubuntu iso failed oto

too

[wicked widget]

oh god xfinity

if u wanna try my distro

Pop Os

ubuntu

not "mine" but i use it lol

[signal basin]

no

i was using as a test of downloading

should be pretty reliable

[wicked widget]

ic

[signal basin]

i think my isp, my modem/router, my computer

all suspect

[wicked widget]

ya

[signal basin]

all ols

old

really fresh insall though

so i suspect computer less

i'm on ethernet

sadness on my end too

any torrent sites for models?\

[wicked widget]

not that i know of but i wouldnt be surprised if theres a rentry

@ornate elk know anything like that?

i know in days of old there was rentry with all the 1.5 model torretns

[ornate elk]

your trying to use a style lora as model, nut its not a model

[wicked widget]

I also dm'd them about better installation guide ytou posted

[ornate elk]

have you changed the dns of your router or PC ?

[wicked widget]

well a1111 is not gonna question the user input

[ornate elk]

your car doesnt warn you either if you put the wrong fuel in it :/

[wicked widget]

lmao

i mean it kinda does...by not starting

put disel in a non disel car

xD

yeah lol 2 yrs ago when was things started

you can download and use that model with your gpu 🙂

[ornate elk]

models (checkpoints) are the the core file you need to generate images.
there are different versions of it.
1.5 based ones are 2-4gb
sdxl/pony/illus are 6gb
flux can be 4-23gb

Loras are smaller and got into the models/lora folder
they are used to alter an image on the way it got trained on

[wicked widget]

they got my card, 3060

[ornate elk]

thats a lora

[wicked widget]

I NEED TO LOOK FOR MORE THAN 2 SECONDS

i wear

swear

thats a cute lora

[ornate elk]

you need an illustrous based model for this lora

[ornate elk]

hires is the hires fix button

hires fix is for upscaling

with the correct settings it enhances the image

[azure magnet]

What is the last version of python that i can install for comfy

[regal mason]

I use 3.10.11

[safe inlet]

3.12.8 generally speaking, some older developed things might have issues, i personally don't.

[regal mason]

though my setup is specifically tailored for an AMD GPU with ZLUDA

[safe inlet]

Comfy Desktop uses 3.12.8 sooo...

[azure magnet]

How can i fix this?

[ornate elk]

Auto and forge have some extension for that

[regal mason]

no clue. I basically ran any errors through chatGPT o1 until I got it working

[azure magnet]

No i don't understand how to do that

There isn't a folder called cache_setting. json.

No, there is

Idk

[loud pond]

just so others may be able to help you better could you please paste the entire log in chat?

[azure magnet]

yes

Here

[safe inlet]

maybe a conflicting node, did any say import failed or anything at the end cuz that's still not the full log

[loud pond]

copy the entire text in the command log and paste the entire text in discord

[pale salmon]

step bro im stuck in the loading

bruv

btw ur off topic, Community support channel for questions related to local/cloud/colab installations.

yeah maybe but rules is rules people need help (like me) and u make message and helper cant see my message

[regal mason]

ugh.. anyone know how to access the comfyui navbar on mobile when its not showing? I need to save my workflow

figured out a solution.. open a new browser tab with the comfyui address, and it will load your current session without resetting it. The menu will display momentarily as long as you dont click anywhere else, but will let you save the workflow

[ornate elk]

go into settings, user interface, quick settings, there add sd_vae
then hit apply and reload ui

then you have a dropdown to select the vae

if the file is in models/vae

yep but if you dot it as i mentioned then you dont have to go in the settings everytime

you will have it in txt2img at the top

there add sd_vae

illustrous needs an sdxl vae

not the 1.5 one you downloaded

vae, aswell as loras always are only compatible with models that have the same base version

no problem

[wicked widget]

that is cursed

[uncut jewel]

Hey! Anyone experienced with Trellis node extension up here ? 👀

i guess not

Then my question will be, is it possible/recommended to change python version from Comfyui from 3.10 to 3.12 since it seems i installed all the xformers/triton/wheels etc... for Trellis in 3.12 version of python or something like that

[wicked widget]

you dont want to use 3.12

or 3.11

3.10.11 is the one because the others break

ah thats a vpred model

youll need to run it in forge for that one

forge has the extra stuff to run it

download the 3.1

see if its not vpred

if its not vpred youll be able to run it as normal

Forge is A1111 but with extra tweaks to it pretty much

a whole other webui installation

[cyan mountain]

im downloading comfyui with hunyuan and im wondering what hunyuan model should i install? to do videos in 720p at least constistently for like tiktok and insta reels. I can choose from the bf16 model to like q3 even to q8 (q7 skipped). i have a rtx 4060 ti with 16 gb vram and 16 gb ram
all the models are from the hugging face website

[deep mist]

hi what make this picture? I want to %100 same picture

[wheat flame]

why is this happening?

[lucid yoke]

Good question, if you got their model, seed, prompt and workflow its possible

Otherwise its not and it's gonna deviate

Hmm its possible they didn't fully list all tools used like clipskip? Or a comfy node changing stuff, euler a + karras etc.

Or even a variation seed that they didnt list

My images i need to manually add all tags/meta data on civit

Cant tell from the image but are you using the exact same aspect ratio?

[wheat flame]

it worked perfectly before :/

[lucid yoke]

Can you show the full report?

[wheat flame]

absolutely

File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 174, in _map_node_over_list
process_inputs(input_dict, i)
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 163, in process_inputs
results.append(getattr(obj, func)(**inputs))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\nodes.py", line 568, in load_checkpoint
out = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, embedding_directory=folder_paths.get_folder_paths("embeddings"))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 825, in load_checkpoint_guess_config
sd = comfy.utils.load_torch_file(ckpt_path)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 36, in load_torch_file
sd = safetensors.torch.load_file(ckpt, device=device.type)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\safetensors\torch.py", line 313, in load_file
with safe_open(filename, framework="pt", device=device) as f:
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: The paging file is too small for this operation to complete. (os error 1455)

Prompt executed in 0.02 seconds

there is more, just a second

here that is everything

[wheat flame]

here!!

:3

[loud pond]

i mean if you just want the same image, just use image to image? or is that not what you mean?

[lucid yoke]

Paging file is too small for this operation to complete. I am no comfy expert but the error should be the same ish

[wheat flame]

thanks, i will check!!

IT WORKED!!! THANK YOU SO MUCH!!

another error showed up

D:

C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Total VRAM 4096 MB, total RAM 16294 MB
pytorch version: 2.5.1+cu124
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce GTX 1050 : cudaMallocAsync
Using pytorch attention
[Prompt Server] web root: C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\web

Import times for custom nodes:
0.0 seconds: C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py

Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt

C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>pause
Press any key to continue . . .

[lucid yoke]

Ah you are using a 1050

Amd vram state to normal,

Set to low vram?

Its like that on other ui's

Scheduler on yours is automatic, hes using sgm uniform. Ensd?, cfgrescale. Isnt that a extension within a1111/forge nowadays

[wheat flame]

i know i change that in Main.py right? but there is a specific line?

[lucid yoke]

In other ui's its in the launch script iirc but comfy users correct me if im wrong

Or uhh google i suppose

[wheat flame]

i tried to google it, i found the commands but not where to put it

[lucid yoke]

add the --lowvram argument when running ComfyUI: python main.py --lowvram

[wheat flame]

which line?

[lucid yoke]

Not a checkbox, you missed the scheduler setting (euler a + sgm uniform) a ensd setting but that might be comfy specific, and a extension for cfgrescale

[wheat flame]

C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Traceback (most recent call last):
File "C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\main.py", line 2, in <module>
comfy.options.enable_args_parsing(--lowvram)
^^^^^^^
NameError: name 'lowvram' is not defined

C:\Users\igorh\Documents\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable>pause
Press any key to continue . . .

i put this lowvram everywhere possible in everyline i could try nd the result is the same

[lucid yoke]

did you add lowvram or --lowvram.

i googled the thing but i see you also revived a year old reddit thread bruh

[lucid yoke]

even this?

https://www.reddit.com/r/comfyui/comments/15jxydu/comfyui_command_line_arguments_informational/ (see top comment for readble ver)

[wheat flame]

yes, but i might have fixed it

by changing for (["--lowvram"])

yes, i fixed it, it is now working!

[midnight roost]

hi i have this issue when i swap between multiple ckpt checkpoints i get 100% ram usage and then it starts using the c drive up too

[lucid yoke]

Why are you using the old format?

The old pickle format needs to be uncompressed first when loading into memory. Hence explaining the ram issue

Then again how much ram & vram do you have?

[midnight roost]

i use ckpt because if i try to merge checkpoints then i get the error pystring/float etc. But if i change it to ckpt format i get no error and the merge works

[midnight roost]

i have 32 gb ram and 16gb vram

[lucid yoke]

Hmm shouldent be causing such an issue unless its the ckpt causing it

What ui

[midnight roost]

automatic

i can see it loads sd_vae thingy a lot

and something called optimizing dogtex?

[versed spruce]

Hello! I have a question: when I use the "img2img" function, the colors always end up being dull and grayish. How can I fix this issue? Is there a way to adjust the colors or download something to correct it?

I'm using Illustrious.
Also, where can I set the "Clip Skip: 2" parameter in Forge?

[pale salmon]

dang nvm

[lucid yoke]

Clipskip for sdxl is a meme, you dont need it. Also are you using a vae?

Also clipskip was hidden in a settings submenu cant recall where

[versed spruce]

I couldn’t find a VAE for Illustrious, so no, I’m not using one.

[lucid yoke]

Use sdxl vae

As its based on sdxl

[versed spruce]

th

[lucid yoke]

No vae = no colours basically or really dull

[versed spruce]

All of them will do, right?

[lucid yoke]

Just sdxl yeah, i use the basic one

[ornate elk]

How much RAM do you have?

[lucid yoke]

☝️

Asking in advance even if i cant solve it :^)

[ornate elk]

What's in your webui-user.bat?
Also how large are the model files?

[midnight roost]

what do u mean model files they are the 6-7 gb type of checkpoint merges. And i have this webui-user.bat

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --use-zluda --skip-ort --theme dark

call webui.bat

[paper dagger]

Guessing but dont you need to set

PYTHON=<YOUR PATH OR VERSION> GIT=<YOUR PATH OR VERSION> VENV_DIR=<YOUR PATH TO VENV DIR>

? Shot in the dark from my end

[shy charm]

this won't show in the controlnet dropdown

there's no models folder in the controlnet extension folder itself

where do controlnet models go if not the models/controlnet folder

[shy charm]

ok for some reason i had to manually add the path to the models/controlnet folder in the settings

[paper ridge]

Hello, I'm trying to setup Stable Diffusion w/ an RX 7900 XTX GPU (preferrably on Windows), and I'm not quite sure how to go about it

I've heard conflicting view points as whether to use Microsoft Olive, ZLUDA, or DirectML - and I was wondering if someone would be able to give me some advice on which approach I should go with?

Thanks in advance, I'd really appreciate it 🙂

[wicked widget]

Follow the Zluda guide for AMD gpus in the pins

[paper ridge]

Does it perform better than the microsoft olive approach? I keep reading that using olive is 10x faster?

[paper ridge]

thanks, in that case I'll prob go w/ zluda since this is gonna be my first time setting up stable diffusion locally

is automatic111 the recommended ui to go with, or comfy?

[wicked widget]

i dont even know what this microsoft olive is either tbf

id have to look into it

i personally use comfy, its all a matter of preference, each ui ahs their pros and cons

the general concensus is that forge and swarm are the main easy to use and comfy is for advanced users

[paper ridge]

cool thank you, if forge/swarm are easier maybe I'll start off with one of those haha

and yeah I think microsoft olive is able to convert cuda models to run in a different format optimized for amd cards etc

[paper ridge]

would u recommend one over the other, between forge and swarm? ie. does one have more support

[lucid yoke]

Pinned messages has a amd guide for forge. Havent checked/seen one for swarm but since its a comfy ui back end its basically the same steps as a zdula comfy ui build i think

But theres more support for forge in this server

[fickle panther]

My previously trained lora's that worked fine are now producing black screens and this error

NansException: A tensor with NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[paper ridge]

thanks, im gonna actually end up just doing comfy ui I found it in the guide for zluda, but I noticed here they recommend installing python directly, is that neccesary or can i use miniconda?

[ornate elk]

Okay looks normal.
With 32gb ram this issue is strange.
Does the ram fill up completely when switching models?

[lucid yoke]

Hm sorry i used python directly, no idea about miniconda

[ornate elk]

Miniconda is not recommended by the guides as you can install python 3.10.11 and use it for every Webui without a problem.
But if you know what you do you can install python 3.10.11 with miniconda and shouldn't get a problem.

Also for beginners I would recommend using Auto1111 or Forge.

[paper ridge]

gotcha thank you! I'll be sure to install that version then ❤️

[ornate elk]

NP, feel free to ask me if you get any questions.
Also zluda will be a much better experience than directml with onnx

[paper ridge]

Thank you, although the setup had quite a few hiccups along the way, the end result is significantly faster than DirectML ❤️

[versed spruce]

And one more question: does it turn out that all LoRAs and embeddings made for SDXL work with Illustrious? Where can I read about compatibility?

[chilly citrus]

i click checked box Upcast cross attention layer to float32 but still error

[ornate elk]

What's your GPU and what's inside your webui-user.bat?

[vital beacon]

Hi I'm getting confused trying to install Stable Diffusion and google isn't really helping

A lot of the information is from about a year ago

I have a Rog Strix 3080 and I'm following the step by step guide in the pinned information

But I'm a bit lost now

I'm trying to follow these steps

Is this correct?

[chilly citrus]

RTX 4070, inside idk, just run

[ornate elk]

Yep thats correct

[vital beacon]

Thanks! I've since managed to figure it out and it's now working. Just trying to understand how ComfyUI works

[ornate elk]

Np, If your new Auto1111 or Forge is easier to start with than comfyui

[vital beacon]

That's okay I'm excited to learn how it works. I'm completely new to this. This is Day 1 for me in terms of learing Stable Diffusion

[lucid yoke]

Sdxl loras only work on sdxl, pony on pony, and illustrious on illustrious

Idk about embeddings

[wicked widget]

clip

and text encoders are same thing pretty much

[midnight roost]

yea when i used X Y Z plot/test each checkpoint it does it if i just add 1 or multiple of CKPT checkpoint files. but i will try to disable add vae

@ornate elk when i choose a merged ckpt file
Loading weights [d0c6b4a29f] from D:\Automatic1111 Stable diffusion\stable-diffusion-webui-amdgpu\models\Stable-diffusion\OmniRillusmNoobAICyberRealistic-0.3.ckpt
Loading VAE weights specified in settings: D:\Automatic1111 Stable diffusion\stable-diffusion-webui-amdgpu\models\VAE\sdxl_vae.safetensors
Applying attention optimization: Doggettx... done.
Weights loaded in 11.7s (send model to cpu: 2.2s, load weights from disk: 7.6s, apply weights to model: 0.5s, load VAE: 0.3s, move model to device: 1.0s).

[midnight roost]

if i set "VAE Checkpoints to cache in RAM" to 1 and disable "Selected VAE overrides per-model preferences" would that work?

[iron lantern]

If I have a image of my wife and I want to run her into forge ai how hard would this be

Took a photo of her and want to change it up a bit

[chilly citrus]

i cant move this popup (properties panel) even i zoom in zoom out.

[wicked widget]

img2img tab

[iron lantern]

I wanna copy the pose and stuff ... and make her anime style

[wicked widget]

train a lora on her images, roughly 15/30 of them on different angles etc. then you'll be able to do a generation of her with controlnet and other model styles

if hardware allows, sdxl

[ornate elk]

How many do u select to test?

[midnight roost]

quite a few like 6 or something but today even just 2 ckpt files takes up all ram dunno why it takes Ram it uses ram and then continues to use up my C drive virtual memory since thats where python are

[iron lantern]

The control net ui folder is not showing

@wicked widget

[versed spruce]

Is there a way to set Clip Skip: 2 in Forge, and where can I find it in the settings? I think the absence of this setting is why I can’t achieve the same results as in A1111.

[onyx sonnet]

@halcyon zodiac @dreamy token @jagged idol @ashen plume @light elm Hey guys, I have the following script for in-painting when I upload the mask (as showen in attachment) on the the-man.png image, most of the times it doesn't perform well, I changed the seed number to different numbers, without seed (randomized seed) most of the time, it doesn't perform well with the prompt in the script, I wonder if there is something wrong with the code or it's the stability's bug that cannot perform well in in-paint.png feature.

I also have uploaded a sample result.png generated by the Stability.

Any help/guidance or intuition is really appreciated, we put some great hope on using the Stability in our production.

@quartz ferry FYI

[ornate elk]

Can you check the pagefile settings?
Also open up a cmd as admin and run
SFC /scannow

[warped canyon]

any way to put "script" next to "adetailer"?

[loud pond]

Is anyone here experienced with prompts? I am trying to get an image generated but everytime i generate the image it looks worse than all the other images on civitai even though i use the same lora's and strenght

[hollow portal]

maybe, check their seed

I struggled with that last night.

Drove me mad.

[iron lantern]

does anyone know how to fix this or whats going on control ui

[loud pond]

nah generation is still horrible

[loud pond]

Please paste the whole error so someone can better help you

just copy and paste the text in the cmd prompt into chat

[loud pond]

did you find any good youtube videos or sources of information to help with prompting which i could use? (if you dont mind)

[noble idol]

@ornate elk hey i did it before but i cant find out how to do it again but i figured you might know but do you know how to manually change the defaults for forge/a1111

[ornate elk]

Hey, you have to edit the ui-config.json

[noble idol]

oh hella, thanks man 🙏 youre the best ong

[paper ridge]

@ornate elk Sorry to bug you, just had a quick question about ComfyUI 🙂

I notice some models run fast and state:

CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16

Where as other models state current as "cpu", and run slower:

CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16

Is this an issue where the GPU isn't being properly utilized, and if so can it be fixed, or is it just a downside of using ZLUDA with some models not being compatible?

[fickle panther]

Can anyone tell me why all the loras i train in kohya produce black images?

[wicked widget]

whats ur settings

havw you tried another trainer such as OneTrainer

[fickle panther]

I have not tried onetrainer but I will give it a shot.

Ill send my settings rn

[fickle panther]

this is the settings

I've tried bf16 and fp16 both only give black images.

[wicked widget]

prolly cos ur using a vae model as the training source lol

check ur training model

try a different one

[fickle panther]

What confuses me is i've used this model before and the models worked, but now they don't.

[wicked widget]

strange indeed

[fickle panther]

Well regardless of it being strange, I just tried onetrainer and it works great the lora isn't just producing a black image. I used the same base model as in the settings as well.

AI is weird.

[wicked widget]

onetrainer is my prefered

change the learning rate warmup to 200 and this should serve as a good starter settings for training xl loras on a 3060 12gb vram

[fickle panther]

I have a 4090 should I crank it higher?

these were my settings and i got pretty good results

[wicked widget]

if you want speed nah leave it as is, one thing you could mess with is the lora dim and alpha. high dim will eat more vram, i pushed my besties 4090 up to 256-128 just for lolz

well actually

the batch count will affect the quality too

while yes with the 4090 you could do 4, make sure the dataset is.well balanced for it

[fickle panther]

I don't really know how to balance it, is there a method?

[wicked widget]

i have an epoch helper calc i wrote you could use if need help

[fickle panther]

YES!

pls

[wicked widget]

the link is in my bio