🤝｜tech-support

[deep canyon]

just a larger featureset?

[ornate elk]

Comfyui can do the same as Auto1111 but is comepletely Node based. You can create very huge and specific Workflows in it

its good for the more advanced users, but can be very overwhelming

[deep canyon]

oof, do i gotta replace python 3.11 with 3.10?

im on 3.11.2

[ornate elk]

if you need python 3.11.2 you can install 3.10.11 too and then change the path in the auto1111 settings

but if you dont need it, best would be to uninstall it

yes you need to use the lora

[fair oxide]

ok, thanks, what value would you use as lora weight?

[ornate elk]

0.6 is recommended by the dev

[wraith bronze]

I've installed "3D Openpose", and generated the four images it spits out in the extension with the default pose, but when I press "Send to txt2img" there is nothing in the controlnet and the image generates w/e so the pose is clearly not being sent. There is no error message in the UI, and none in the terminal. Is the extension (ver: f2d5aac5) not working properly with 1.7.0 or am I missing something?

I've restarted SD as well

[ornate elk]

thats normal xD this 3d openpose extension is 8 months old and not updated

[fair oxide]

incidentally i used a weight of 1 with these settings.. and ...

[wraith bronze]

oh boo! when I try the normal open pose the pose isn't really getting the look I want 😛
Is there a working 3d poser?

[ornate elk]

no weight needs to be 1

lora weight 0.6

[ornate elk]

you can still save the images and put them into controlnet manually

[fair oxide]

here is the result of it ...

[ornate elk]

looks good

[deep canyon]

oh cool it just launches using the correct python

didnt even change path

[fair oxide]

great. so the settings are ideal?

[ornate elk]

was that with weight 0.6 and lora 1 ?

[wraith bronze]

I tried that, but that didn't work either. The image it saved was dimmer so I don't know if SD got confused.
I will try it again I guess xD

[fair oxide]

used 1 for that one, but yeah i could play around with lower values

[deep canyon]

why is this easier to set up than focus

fooocus was a PITA ngl

in comparison

[ornate elk]

because foocus added directml a few months ago, and auto1111 has this fork since a year, and my guide covers all needed steps

[deep canyon]

yeah the fooocus documentation is... lacking, to say the least

[mint pivot]

#👍

[deep canyon]

and upping my VRAM on there from 1GB to 20GB did literally nothing

for results

[fair oxide]

@ornate elk if you got few mins free now, can you show an example of face id from your end of the settings with this image, cause not all of the images are working out too well for me ..

[deep canyon]

it worked and properly allocated it but did not actually make it faster

[wraith bronze]

I've seen a lot of people mention blender for poses, is there a good resource for how to use blender to play aroudn with a pose ?

[deep canyon]

ooooopsy

[fair oxide]

also @ornate elk i think there is a clever way of using lineart with face id for pretty realistic effect

[ornate elk]

oh okay, why with lineart?

[ornate elk]

did you edit the webui-user.bat? like in guide step 5

[fair oxide]

the combination is for something as realistic as this

[deep canyon]

i did but i fucked it up

already figured it out lol

sometimes when im reading things my eyes have the habit of skipping lines just completely accidentally, or jumping from say, the amd guide to the nvidia guide without noticing

[fair oxide]

although i'd prefer to have better control over face id alone, so im curious what output iyou can generate with that image i shared so i can compare my settings

[ornate elk]

ahh yea, i know that

[ornate elk]

so you want it as realistic ?

[fair oxide]

the facial look yes but not entirely the pose

[deep canyon]

https://tenor.com/view/yayyj-gif-6510174559938943769

[fair oxide]

@ornate elk im also using face id plus

[ornate elk]

ill try

[fair oxide]

one thing i notice about face id is that the image we generate it may not follow the exact angle or positioning of the face, it just uses the facial resemblance of the source image .. which is perfectly ok for creating variations

[ornate elk]

yea thats what ip-adapter always does

you can fix a position with a openpose with face mask

[deep canyon]

did i mess up by putting a space here

[fair oxide]

yes or lineart for details

[ornate elk]

normally that doesnt matter

[ornate elk]

but you may need to delete the venv folder if you still get the skip-torch error

[deep canyon]

i deleted it all and am retrying now lol

i assume setting it up for nvidia first may have borked it

[ornate elk]

yes

then do my fix up step from the amd guide

run pip cache purge in a cmd

[deep canyon]

yep i did we are gucci

so did you make this?

or just the guide

[ornate elk]

i made the guide, the fork is from lshqtiger on github

[deep canyon]

gotcha

fooocus is cool but kind of ass

its terrible at making anything even remotely specific

amazing if you type "burger" but add more than 2 words and it shits itself

[upper narwhal]

iim have trouble with my art, idk why it make it like this. it does this last minute before completing and idk why its doing this. i could give more imformation if could

[deep canyon]

the prompt was "fortnite burger"

[ornate elk]

yea its made for easy prompting and getting crazy artistic results

[ornate elk]

hey, whats your model and vae ?

[deep canyon]

i find they rarely have much to do with the prompt

[upper narwhal]

where would i find that at?

[ornate elk]

what webui are you using?

[deep canyon]

woah auto1111 is way faster

[upper narwhal]

uh,,i actually dont know where would i find this imformation

[ornate elk]

do you use a local programm?

then in the browser, scroll to the bottom and check the version

[upper narwhal]

i think so, yeah its on my local fine

[ornate elk]

yea it depends on the models and its generaly better optimized as it supports more backend optimisations

[deep canyon]

looks considerably worse tho

sampling method i assume

[ornate elk]

your using the basic 1.5 model, thats over a year old so thats why

[deep canyon]

oh

should I be using a newer version?

[upper narwhal]

says version: v1.7.0

it was workign earlier but idk why its doing this

[ornate elk]

yea, you should get community made models, these are much better in quality, and have specific artstyles,
the largest databse is Civitai.com, there you can download them for free.

For example get the Dreamshaper v8 model from here:
https://civitai.com/models/4384/dreamshaper

Put that file into models/stable-diffusion folder

[deep canyon]

oh yeah i was using those on fooocus

thought you meant i should be using a newer version of auto

[ornate elk]

nope, auto1111 is up to date

[upper narwhal]

so any possible reason why its doing this last minute

[deep canyon]

i disabled filtration on this

and people are way too horny

[ornate elk]

whats the model name, thats in the top left corner?

[deep canyon]

ok this is sick as fuck

[upper narwhal]

you mean the check points?

[ornate elk]

yes

[upper narwhal]

currently using MFCG Doll Mix v2-No VAE.safetensors, tired other checkpoint it keeps messing up last minute

like its working give the right colors until very very last second it messes up

[ornate elk]

okay, i checked your image meta data,
your sampler + the upscaler settings doesnt work together

dont use 4 hires steps, use 10, and set the denois to 0.5

then use DPM++ SDE Karras for example

also lower the lora weight to 0.8

[upper narwhal]

can you explain why? just so i can learn

[ornate elk]

You used a latent upscaler, these upscalers can generate additional stuff into the image by denoising.
but 4 steps are to less for them. 10-15 is a good value.
The denois defines how much the image should get changed by the upscaler, going for a value like 0.7 is to high and can make deformations to the image.
Samplers are all different, some work better with Latent Upscalers than others.
Would recommend, Euler a, DPM++ 2M SDE Karras, DPM++ SDE, but not with Heun together.

There are also Upscalers like Esrgan4x+ that work better with any samplers, and they dont add stuff to the image, so if you just want a little enhancement then they are good too

@fair oxide here is what i get with id-plus

tried realistic vision model

[fair oxide]

ahh ok, i think some of the models i tried have different outputs depending on how they are trained

i checked with same settings too

[ornate elk]

ill try now with face full as comparrison

[fair oxide]

sure, would love to see that

[fair oxide]

did you use default values?

[ornate elk]

yes

nothing changed

[fair oxide]

hmm, face looks bit too different from the source, you suppose?

could be the realistic vision model

[ornate elk]

maybe a bit, but for me face full is still better than face id, because it can actually move over the skin and face accents better, every extension that uses insightface (like Reactor, roop, or faceID) smooths the face to much

[fair oxide]

interesting, i have to try that on my end

im actually gonna try it now

@ornate elk just to be sure, you used these models?

[deep canyon]

cool

[ornate elk]

yes, and no lora, as it doesnt need one

[fair oxide]

right... ok im gonna try that image with that settings

[ornate elk]

what was the resolution?

[deep canyon]

640x640

[ornate elk]

strange that shouldnt happen with that res

[ornate elk]

everytime you get that error, you need to restart the webui to clear the vram

[fair oxide]

i'm not using adetailer when using ip-adapter btw

[deep canyon]

i did :(

im gonna relaunch

[ornate elk]

i use hires fix to smooth it a bit

[fair oxide]

yes me too, hires fix is active

[deep canyon]

this shit dont work lol

[ornate elk]

what are your settings?

[ornate elk]

yea as i thought

hires fix on 0

[deep canyon]

oh

it was default figured it was fine

[ornate elk]

disabale hires fix and it will work

[onyx solar]

You're pushing the resolution up when you have that on.

[deep canyon]

i mean its a 24GB card

[ornate elk]

on windows its very limited, because directml is poorly optimized in vram usage

rocm would be much better, but we need to wait a few months for that on Windows

[deep canyon]

can i force it to use more?

like how i did on fooocus

[ornate elk]

nope

it uses all of your vram

[ornate elk]

you can use hires fix, but just use thse settings:
Resolution 512x512, 512x768, or 540x960 (for FullHD)
Hires steps: 10,
Denois 0.5
upscale by: 2
Upscaler: Esrgan4x as example, or latent bicubic

@fair oxide you see what i mean, left face full, right face-id

that was the input image

[fair oxide]

yep i see it, you are right about face id smoothing out too much

[ornate elk]

face-id-plusv2

better than face-id plus

[fair oxide]

nice, i'll grab that one too

@ornate elk can you offer some explanations on these 3 parameters ?

i assume controlweight would refer to more closely folllowing the source the more weight you have?

[ornate elk]

weight is how strong the controlnet itself should be applied, (like lora weights)
starting step is when should the controlnet comes to action in %
Same goes for the Ending Step. (if set to start 0 and end 1, it will force the controlnet until the last step)
so reduce the end step so the face doesnt get too forced on an image if you want to upscale for example

[fair oxide]

great thanks!

[ornate elk]

also with face full you can switch to an unrealistic model and it still gets the face right:

same input image from above

[deep canyon]

how many loras do you need to make something not look like deformed dogshit

[ornate elk]

no lora

you need a good negative prompt first

some quality tags in the prompt too

did you tried dreamshaper v8 ?

[ornate elk]

for example, add these to the negative prompt:
disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, , mangled, old, surreal, text, blurry, b&w, monochrome, conjoined twins, multiple heads, extra legs, extra arms, fashion photos (collage:1.25), meme, deformed, elongated, twisted fingers, strabismus, heterochromia, closed eyes, blurred, watermark

[deep canyon]

oh ok

i was just using something called "easy negative" because i thought it would do it for me lol

[ornate elk]

thats an embedding (textual inversion), that will work too if you downloaded it and put it into the embeddings folder, and then added it to the negative prompt

[deep canyon]

i did but everything looks like shit

[ornate elk]

feel free to post an example

[deep canyon]

like wtf is this
a couch? a leg?

i dont get it

[ornate elk]

do you use an SDXL model?

[deep canyon]

whats that

[ornate elk]

a model thats 6gb large

[deep canyon]

definitely not

[ornate elk]

okay, then what are your txt2img settings?

[ornate elk]

disable refiner, you dont have a refiner model so that wont work

also not needed

[deep canyon]

this shit is just nightmare fuel

[ornate elk]

did you downloaded a community checkpoint?

[deep canyon]

nope

[ornate elk]

ok, like i said download dreamshaper v8 as a good start

the 1.5 base model is bad

[nova ermine]

I'm looking for "restore faces" option, but I don't find it, how can I find the option?

[deep canyon]

is that under collections? models?

found it

dreamshaper XL

or am i stupid

[ornate elk]

get v8, not sdxl
https://civitai.com/models/4384/dreamshaper

[deep canyon]

must not be on civitai

oh

its just called dreamshaper

oops

sorry im a total end user with this shit

[ornate elk]

its hidden, to get it back go into Settings - User Interface- Quicksettings, and there add Face_restoration,
then hit apply and reload ui

[ornate elk]

no problem at all 🙂

@deep canyon here is an easy example:

[deep canyon]

"C:\Matrix\stable-diffusion-webui-directml\repositories\stable-diffusion-stability-ai\checkpoints" right path?

note says put unCLIP checkpoints here

[ornate elk]

nope, stable-diffusion-webui-directml/models/stable-diffusion

[deep canyon]

ty

[nova ermine]

This stuff?

[ornate elk]

then you can click the reload button near the models(checkpoints) dropdown

and then you can select it

[deep canyon]

got that part down already

im ascending from end user

[ornate elk]

uprank 😄

[deep canyon]

they all seem to be coming out deepfried

[ornate elk]

can yo show your settings again?

[deep canyon]

actual image is good

but then deep fries on last step

[ornate elk]

thats strange, did you changed anything?

[deep canyon]

no

was coming out deepfried before too

but whole image was fucked then

[ornate elk]

can you paste an image in here?

[deep canyon]

hmmm seems to be coming out fine now

strange

maybe its just RNG to get deep fried ones?

[ornate elk]

make sure to not run Wallpaper Engine or Epic Games Launcher in the background, also whitelist the wbeui in any browser adblocker

also sometimes a simple restart fixes it

[deep canyon]

this thing will not stop making women lol

i even just put "boy" as the prompt

still women

lol

[deep canyon]

yea had to take wpp engine off startup and quit using it because it kept crashing my shit

[ornate elk]

yea a common error ^^ had crashes tooo with it when using SD

[deep canyon]

it will crash me out when im just gaming

or itll just straight up brick my explorer

[ornate elk]

yea, i disable it everytime before starting a game

[deep canyon]

i dont get why it oversaturates stuff so often

[ornate elk]

extremely oversaturated? like before?

[deep canyon]

mhm

it gave me 2 images that werent

now its back to oversaturated

[ornate elk]

do you run any other programm or multiple browser tabs in the background?

[deep canyon]

yea

discord desktop with github pages and civitai

nothing else really

except HWInfo and Aquasuite and a music player

[ornate elk]

minimize all other programms, and close the civitai tab xD that site is a browser lagfest

these oversaturations come from lagging, when some other programm uses to much resources while generating, then the image bugs out and get corrupt.
happened to me a few times, but its very rare normaly, but i dont have anything open, also using firefox with maybe one other tab

[deep canyon]

yeah it only happens on the last step too

like it looks fine before

aaaaand I just closed auto by accident

[ornate elk]

xD

[deep canyon]

looks perfect then fries at the end

what the fuck

[ornate elk]

can you paste an image in here, so i can check the mata data?

make sure to upload the original output

its in the outputs/txt2img folder

[deep canyon]

i think the lora i was trying to use is broken

i removed it and its not oversaturated now

[ornate elk]

yea

that can cause this too,

you didnt said you used a lora xD

i would have figured it out in seconds with that info

[deep canyon]

my bad

[ornate elk]

no problem, you have to check if a lora is made for 1.5 or sdxl

you find that info on the civitai page when downloading the lora or model

they have a little info box

Base: 1.5 or SDXL

[deep canyon]

says 1.5

[ornate elk]

okay, than it can work with dreamshaper v8,
but loras still can be "overtrained"
then you have to adjust the Number after the lora, lower it to 0.5 and try again.

[deep canyon]

still overcooking

[ornate elk]

what lora is that?

[deep canyon]

im not at liberty to say 💀

[ornate elk]

okay, then try an other lora to check if that works

[deep canyon]

even 0.3 its cooked af

[ornate elk]

its rare but loras can be just broken

[deep canyon]

i tried on v8 and the one it said it was trained on

even cooked @ 0.1

[ornate elk]

then its incompatible

or just broken

[deep canyon]

its the one it was trained on

so im leaning towards broken

other people have it working tho in comments so idk

[ornate elk]

you can share me the lora in private if you want that i take a look

[deep canyon]

idk its pretty gay

literally

[ornate elk]

xD ah okay

[deep canyon]

(not derogatory just actually gay)

[ornate elk]

i dont mind

have seen mostly everything on civitai xD

[deep canyon]

i think i figured it out

the checkpoint got updated

[upper narwhal]

why is it now saying "RuntimeError: Not enough memory, use lower resolution (max approx. 640x640). Need: 0.2GB free, Have:0.1GB free" i was just working ealier

[deep canyon]

reverted to a previous version

not cooked

[ornate elk]

ahh okay

[ornate elk]

whats your GPU and whats inside your webui-user.bat?

[upper narwhal]

2080ti super max q

[upper narwhal]

wdym waht inside my webui-user.bat

[ornate elk]

when you right click and edit the webui-user.bat

[upper narwhal]

@echo off

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

call webui.bat

[deep canyon]

yep got it working

baller

now i can make my homie regret showing me stable diffusion

[ornate elk]

hahaha 😄

[ornate elk]

okay, at the line set COMMANDLINE_ARGS=
you need to add:
--xformers --medvram-sdxl --no-half-vae
then save and relaunc the webui-user.bat

that will increase the performance alot and lower vram usage

[upper narwhal]

thank you

[ornate elk]

Np

[thick igloo]

Hi again, thanks for your rapid and helpful response. It seemed to install this time without error but I'm still getting an error when I try to generate an image. It says RuntimeError: Unspecified Error and this is what the code looks like. Thanks!

[ornate elk]

Hey, would need to see the whole cmd output

[thick igloo]

I tried to generate a couple times and added a new model on the last one.

[ornate elk]

Can you close and relaunch the webui-user.bat?
Then try the 1.5 EMA pruned model

With default settings

[thick igloo]

I only have 2 gigs of VRAM so I put the --lowvram condition in but was I supposed to add --no-half in addition to --no-half-vae or --no-half instead of --no-half-vae? My GPU is AMD RX 560

[ornate elk]

Ah okay, thats hard on the edge

Then you should download a 2gb model to test. Like Dreamshaper v8

[thick igloo]

Thanks

[ornate elk]

I can't guarantee that it will work with 2gb

[harsh token]

Could someone help me figure a problem out. So I'm currently using a SDXL 1.0 as my main model. But sometimes there will be character LoRAs I want to use that doesn't exist for SDXL, only for 1.5. Obviously I can't use it on SDXL. Is there a good workaround for this problem? I've tried making the image first in SDXL then using that output in img2img and switching to a 1.5 model with the LoRA, but the result are just not good.

[hollow trench]

@ornate elk what kind of prompts are you using that you found have helped custom trained models with dreambooth look more like the person youre trying to get

[lament fog]

the workaround is doing the same training to create a lora for SDXL, if you consider that a workaround

[boreal bison]

@ornate elk and @tardy flax thank you guys for all the help installing SD locally

[thick igloo]

I had to update graphics card driver again and thought it was almost going to work. I could see the image loading up with dreamshaper8 and it got to 90% but then it hit a runtime error. Could nto allocate tensor with 268435456 bytes. Not enough GPU video memory available. Might try once more with all other programs closed.

Brought down height and width slightly and success! Quite slow but I think it's as good as I'll get. Thank you.

[harsh token]

Well not really

[frozen sandal]

hello! I have a question and would love some help ^^

basically I made 2 loras, each of a different character. I'm trying to do images with these two characters together in the same image, but they keep mutating into each other (swapping genders, clothes, hair colors, eye colors) and a lot of extra heads/arms/legs/hands are generated as well.

I've tried using control net and I've tried regional prompting, but nothing seems to be working for me and I can't find much info online

Could anyone please provide me with a prompt or other solution that might work? I'm using SDXL and this checkpoint https://civitai.com/models/261336/animapencil-xl

any help would be very appreciated as I've been unable to figure this out and it's been driving me crazy for a few days!

Thank you in advance

[lyric surge]

hello! I have a problem... my load image nodes arent working properly, i cant choose a new image from a path anymore after some updates today
[12:12 AM]
my browser console says:

Error calling extension 'Comfy.smZ.WorkflowImage' method 'setup'
Object { error: TypeError }
​
error: TypeError: callerLine.match(...) is null
​

anyone maybe knows how to fix this?

id rather not reinstall comfy 😦

[upper narwhal]

i dont know if this is the right channel but the anantomy and everythign fuck up at the end and things are getting worse and worse the more i use this

[onyx solar]

That doesn't really help unless you share things like examples and screenshots of your configuration.

[upper narwhal]

yeah prolly not best since im making some nfsw content currently

[upper narwhal]

i mean i do request the help though

[unique glacier]

@upper narwhal DM me the images with the mess up anatomy and i'll see what i can do to advice.

[lyric surge]

is this normal that with some SDXL checkpoints i only get noise output in my animatediff workflow? while others work just fine?

i mean i get that some work better than others, but here with animatediff some checkpoints produce absolutely nothing, just random noise

omg i see something! clip skip helped

still really bad tho

[ornate elk]

Where?

[ornate elk]

No problem, the image quality is epic!

[boreal bison]

Thanks! 😎😊

[ornate elk]

I didn't custom trained a model with dreambooth yet, would recommend training a lora or using a faceswap tool or ControlNet IP-Adapter

[ornate elk]

Nice good to know, what you can try too is to install the Multidiffusion extension with Tiled VAE support.
With tiled vae enabled it can stop the crashes most of the time. So you can use higher resolution.

[thick igloo]

Is there a tutorial somewhere for the Multidiffusion extension?

[ornate elk]

You dont need one. Its literally installing the Extension, via the Extensions tab.
Click on Available, Load from, to get a list.
Then search for Tiled Diffusion. Then install.
Then restart the webui-user.bat.
And then you only need to enable the Tiled VAE Option and test it with default settings

[thick igloo]

Thanks!

I'm a noob hehe.

[ornate elk]

No problem at all

[minor pumice]

what's the problem ?

why could'nt drean,it says i have no permission

dream

[ornate elk]

The bots are currently under maintenance #1047610792226340935

[minor pumice]

Means all of the bots are under maintennance ?

[ornate elk]

Yes

[minor pumice]

Thanks

[dim egret]

👌

[viral dock]

all of my generations are coming out as odd smudges, i just started using stable diffusion today but i couldn't find anybody else talking about this. if it helps i'm using sdxl base with comfyui and default settings, but no matter the prompt it always turns out like this. it generates the same smudge look when i use my nvidia or my cpu so i'm unsure where to go from here. if anyone has an idea what i'm doing wrong i'd really appreciate it!

[ripe burrow]

🤝

[ornate elk]

Hey, did you tried the standard sdxl workflow from the comfyui github ?

Also what's the resolution?

[viral dock]

yep, first image here is perfectly standard and other two are just "half orc standing"

as for resolution its 512x512 as is standard

[ornate elk]

Then something is wrong

For sdxl a standard resolution of 1024x1024 is needed

[viral dock]

my apologies i didn't know

annnd instant improvement

[ornate elk]

Np, 1.5 models are for 512x512

[viral dock]

thank you so much i apologize for my ignorance

[ornate elk]

No problem, the default sdxl workflow should use 1024 by default

So if it was set to 512x512 that would be strange.
But good that it works now 🙂

[viral dock]

actually it doesn't hold up

i thought it did at first

it's still smudges just slightly higher resolution

damn

is it a hardware issue? a download issue? should i try reinstalling?

[ornate elk]

Do you have sdxl and the refiner model?

[viral dock]

yep

[ornate elk]

Then delete all nodes and load that image into comfyui:

That contains the default sdxl+refiner workflow

[viral dock]

i guess i wasn't using the right defaults? i just used it as installed, i didn't realize there was another set entirely

let me try this

[ornate elk]

Here is a helpful site for basic workflows:
https://comfyanonymous.github.io/ComfyUI_examples/

[viral dock]

ran it exactly as it was

[ornate elk]

Strange

Which sdxl model did you downloaded?

[viral dock]

sd_xl_base_1.0.safetensors

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

[ornate elk]

Okay

[viral dock]

honestly i'd think it was my gpu's fault or something if the cpu didnt do the same thing

[valid gorge]

I'm looking for a way to monitor progress, memory usage etc directly from the A1111 UI, anyone know any extensions that would do all or some fo these?

Using the UI trough --listen on another computer and I can't access the comman line easily

[ember tartan]

can you help me please

[viral dock]

has there ever been anyone with the same issue?

[ornate elk]

You should install auto1111 via my install guide in the pinned messages of this channel.
Maybe that webui will work for you

[viral dock]

alright

thanks for all your help!

[ornate elk]

There is an extension for that:
https://github.com/vladmandic/sd-extension-system-info

[ornate elk]

To less infos to help

[valid gorge]

Percect, thanks! Will check it out.

[viral dock]

it's not the webui, it's something about stable diffusion itself apparently

just installed auto1111 to the letter

[ornate elk]

Screenshot all settings pls

In txt2img

Also what model (checkpoint) did you used?

[viral dock]

This one again

[ornate elk]

Try an other model please

[viral dock]

any recommendations?

[ornate elk]

This one:
https://civitai.com/models/4384/dreamshaper

Its 2gb

[viral dock]

there's absolutely no blurring or artifacting on this one, so i guess it's just hardware limitation on my part or something?

sorry, i'm still new to this

but thank you so much for your patience and help

this one works great

[ornate elk]

What's your GPU ?

Good to hear that the model works

[viral dock]

its definitely an older GPU, nvidia geforce rtx 2070 super, but what tripped me up was that the cpu produced the same results just slower

[ornate elk]

Maybe the model is broken. Try this sdxl model:
https://civitai.com/models/133005/juggernaut-xl

Also for sdxl use the 1024x1024 resolution

[viral dock]

this one works just fine, i guess i just downloaded a broken model or something? i'm honestly confused as to how that's possible, but point is it's working now and thank you

[ornate elk]

Could be a corrupt file then

You can also redownload the 1.0 one

[worthy portal]

Hello to all friends
I wanted to ask a question, in the control net and openpose extension, I also downloaded the control_v11p_sd15_openpose [cab727d4] model and put it in the extension, and even when I click on the run preprocessor option, it gives me the pose of every photo I uploaded, but when I write a prompt, the image It does not create a pose. I don't know where the problem is, I tried almost all the options, but it still only works with the same prompt, what should I do to solve this problem?

[storm vector]

keep getting this error with "next view" when I try to turn batch Images Into a video, says there Is a video file missing but Im pretty sure thats the one Its supposed to create out of the Images

[empty walrus]

#1047610792226340935

[storm vector]

??

[valid gorge]

I see you are using DreamShapeXL_Turbo model, it's an SDXL model, right? You need SDXL ControlNet models for SDXL models. Try if that helps.

[ornate elk]

You installed the webui inside a onedrive folder. Thats not recommended and cause path and permission issues

[scarlet prawn]

hello, could someone tell me how to have stable distribution on Kaggle

[valid gorge]

You could also try a different workflow, first create a base image with an SD model + ControlNet to get the right pose, then use that image as a base and do some refining with inpaint using the SDXL model.

[valid gorge]

I remember reading somethig about on their website

see if this helps https://youtu.be/dpM02YMj8FY?si=uXrdlXIIW1vcMcAI

[worthy portal]

These are the models I have, I think they are all sdxl right?

[worthy portal]

Oh yes, the problem was with the models
I tried another model and it made a for me in the controlnet pose

[worthy portal]

Thank you very much

[valid gorge]

I only recognise the DreamShaper, sorry. Check the source where you downloaded them from. Civitai, Hugginface etc

[worthy portal]

My problem is solved, I will go the rest of the way by myself 😘

[valid gorge]

Great! 🙂

[sonic ferry]

ComfyUI question, is there like a steps to CFG ratio to follow? I'm getting decent images this way, first KSampler 10 steps, 3 cfg, 0.86 denoise, upscale to 1.5 before going to second KSampler 12 steps, 3cfg, 0.50 denoise, before going to last upscaler of 1.5x on Ultimate SD Upscale. With that much Denoise on second KSampler, is it worth doing say 25 steps on the first and 5 cfg? then second at 35 steps 7 cfg?

[graceful crescent]

i have everything installed, but i keep getting this error

how do i fix it

[ornate ruin]

Does someone know if there is a way to avoid this? Would making a low res image and then progressively upscaling it step by step work?

[fair oxide]

@ornate elk just to share this observation with you, most of the images I generate using SDXL models, those are fine with 512x768 resolutions along with hires fix as usual workflow.

[ornate elk]

Your python version is to old.
Check my install guide in the Pinned Messages in this channel

[ornate elk]

What's your GPU?

And what's inside your webui-user.bat?

[ornate ruin]

RTX 4060 Laptop/ 8GB VRAM

[ornate ruin]

Pardon?

Not sure what you mean by that, excuse me..

[fair oxide]

a file named under root webui folder

to see the content of that file you'd have to right click and edit to open in notepad

[ornate ruin]

Should be this one then, yea?

[fair oxide]

that should basically show something like this ...

[ornate ruin]

Ahh this one 1 sec

I had --xformers in as well until just today when I decided to try it out without it

[fair oxide]

--xformers is crucial to optimize performance greatly if you are using nvidia cards

[ornate ruin]

So shall add it back in there?

[fair oxide]

wait for @ornate elk he was going to suggest something im guessing

[ornate ruin]

I've heard that results can vary when using it and trying to recreate images from copied generation prompts. But yes no worries I'll wait thanks 🙂

[fair oxide]

well to recreate an image it requires using same seed and settings, but xformers is a built in technical feature to optimize SD performance for nvidia GPUs regardless you want to create or recreate an image

[ornate ruin]

But if it's important, I'll add some background, I'm trying to make an image as high resolution as possible (if possible the same resolution as my phones screen) for a wall paper or so 😃 those don't have any quality yet. -

[fair oxide]

these are the conventional resolutions you should set your image size to then hires fix by 2x to upscale

[ornate ruin]

I have the exact generation datas, same checkpoint version, Lora version applied, clip skip, seed etc. But kept ending up with a different, disturbed image. So I figured its either --xformers or hirez fix isn't transmitted when coying an images parameters so I'll just have to figure that out my self

[fair oxide]

that could be anything about the settings such as LoRA, source image (if used), embedding and so on, but not xformers

[ornate ruin]

Ah ok.. I've set it to the exact parameters as my screen ratio decided by 4 and then wanted to upscale x4.. maybe I have to use another ratio then tho ..

[fair oxide]

think of xformers as performance related not image manipulation.

[fair oxide]

for most tasks you'd be using 512x768 then upscale by 2x using hires fix .. that's the usual convention, but you can also use other larger resolutions, in fact with SD 1.5 models you can even use 512x2048 and the output would be symmetrical. you can experiment with those set dimension i showed above and mix them up

[ornate ruin]

I see... Interesting

Well, to start things off I added --xformers back now and deleted the venv folder the get a clean first boot of automatic1111 (I think)

[fair oxide]

you didnt have to delete venv folder to add back --xformers

[ornate ruin]

Oh.. anyhow... 😅

[fair oxide]

only reason you may need to delete venv folder when you have lots of updates patched over a period of time and bugs are getting in the way

[ornate ruin]

During this boot bow I can see it suggests to upgrade. Do I just add '-- upgrade pip' behind where I also wrote --xformers and the restart again?

[fair oxide]

typically you should not update webui everytime you launch it, that could break things over nightly builds, but to actually update you would go into webui root folder and open up cmd from there and type in git pull but other than webui updates you are encouraged to update your extensions within the webui which you can do by going into extension tab and check for updates then apply and reload.

[ornate ruin]

Sitting in this menu for a good couple of minutes now.

[ornate ruin]

Okay 👍🏻 I'll do this update another day then.

[graceful crescent]

thats normal

[ornate ruin]

It's done now.lol

[ornate elk]

Your card needs: --xformers --medvram-sdxl --no-half-vae
For the best performance

[ornate ruin]

But they are all sure to not decrease the outputs? What's the downside to using them?

[ornate elk]

there is no downside in using them, xformers had an issue which was that you cant recreate an exact 100% same image, but that seems fixed with the 0.0.20 version,
the command args dont reduce the quality in any way

[raw pier]

New to the SD world, and I am starting with what I believe to be a simple task. The task is using personal self images and generate realistic portrait images. I am on a local installation of the kohya trainer and Automatic1111/dreambooth to generate the text-to-image. The training data set is all in the (512X512) resolution with the the naming convention "subject(1), .... ,subject(n). Using this trained model, I then merge with a citivai mode (50/50 blend). The resulting generated image looks terrible (non-human, poor resolution, insert any undesired trait of a realistic iamge . This issue to be independent of the prompt itself. Any suggestions? (investigate a new data set with higher resolution, trainer issues, citivai model selection)

[ornate ruin]

Okay that's crazy. I just wonder then, why isn't it fundamentally recommended to use all of those commands even if you don't 'need' them if they don't have a downside? Basically just to make 'extra' sure your experience is the best it can be

[ornate elk]

yea, thats a good question, the devs dont want to force any of these arguments, to keep it clean at install

for example, if you would install Auto1111 to use it with your cpu, you would get errors because xformers is for nvidia only

or --medvram-sdxl is only usefull for 8-11gb vram gpus

so it wouldnt make sense to add them for all

[tacit light]

Does AMD automatic1111 still not work with Olive optimization? I keep getting all kinds of different errors

[ornate elk]

it works with olive partly but isnt recommended to use

[tacit light]

I could get the Ui to work and it would be 15s per iteration so I tried the amd community post and followed it but it doesn't work lol

So for amd cards its just not usable until more support is added officially?

[ornate elk]

if you got the --skip-torch.cuda-test "error" thats fixable

im using SD Auto1111 directml fork with my Radeon RX7900XTX, without Olive, on Windows

[tacit light]

Yeah i got around that by adding the arguement

[ornate elk]

yea thats the wrong approach, adding that will use cpu only

[tacit light]

with Olive optimization I could not get it to work but generally I got both Auto and tiger to work

[tacit light]

so It is not using the gpu at all? I saw 100% gpu usage in task manager

I installed the directml btw

[ornate elk]

if you want a usable webui with the best support but with mediocre speed, you should follow my install guide, its in the pinned message of this channel

[tacit light]

what's the average speed you get on your card per iteration? I only have a 6800XT

[ornate elk]

if you want something really fast, but with less support i would recommend the Shark webui. Much better than Olive, but has the same problems (converting models, for every resolution, no multi lora or extension support, etc)

[ornate elk]

i can check, one moment

[tacit light]

I understand but I just want to know your speed per iteration so i can see if its worth it or not

I gave up on Olive version and was just going to find some good online generator like getimg

Also does Controlnet work with the set up you have? I usually use that a lot

[ornate elk]

Linux is still the best option for AMD righ now. But we will get rocm support on windows in the comming months.
i would say its pretty worth it with a 6800XT to use it localy

[tacit light]

I can try your guide, what about controlnet?

[ornate elk]

getting 3-4 it/s

[ornate elk]

controlnet works completely

[tacit light]

awesome

and I can also add other models from civitai?

[ornate elk]

yes, every model will work

and any lora

[tacit light]

alrighty, Thanks

[ornate elk]

only sdxl can be slow as its a 6gb file

[tacit light]

will go with that then.

sdxl is too much for my pc

[ornate elk]

nah, it should work

but not to fast

[tacit light]

I ran FOOCUS yesterday, it worked but it was terrible lol

[ornate elk]

yea foocus added amd (directml) support a few months ago, but didnt added other optimisations

auto1111 directml fork from lshqtiger is the best we have right now in terms of use and support

the next best is comfyui with directml

then Shark webui for 20it/s speed (but lacks features)

[tacit light]

sounds good, I used comfy last year on my laptop with a 1060 6gb lol

I do love the interface of comfy

okay, I will follow your guide and post here if I get any errors

[ornate elk]

Yea sure 🙂

[mellow plover]

Is the bot down?

[tacit light]

#1047610792226340935

[tacit light]

do i need --medvram?

[ornate elk]

Yes

[tacit light]

I got 16gb doe

[ornate elk]

I got 24 and need it to 🥲

[tacit light]

loool

[ornate elk]

Directml is not optimised in vram usage thats why

[tacit light]

RIP

[ornate elk]

We will get Rocm support in the next months so everything will change to the good

[tacit light]

ok mine is installing now

yeah hopefully we get the support soon

and it actually works

unlike olive

[ornate elk]

yea i think it will

[tacit light]

ALRIGHTY it works and is 4seconds per it

[ornate elk]

Make sure to test 512x512, euler a, 30 steps

Also important: make sure wallpaper engine and epic games launcher dont run in the background

[tacit light]

so for models from civitai I just paste into models/stable-diffusion right?

[tacit light]

yeah they're closed

[ornate elk]

Yes

[tacit light]

okay, Can I make higher resolutions as well?

I like 1024x576

16:9 ratio

[ornate elk]

Could work. I don't know your cards limit. You have to try.

But I know the upscale limits mostly

[tacit light]

okay, How do I get controlnet for img2img

need that for fine tuning hehe

[ornate elk]

Also important. Every time you get an Out of memory (vram) error, you need to restart the webui to clear the stuck vram

[tacit light]

ok

[ornate elk]

go on the extensions tab, click on Available, click on Load from, then you get a list

type sd-webui-controlnet

then install

then restart the webui

[tacit light]

alrighty

[ornate elk]

also get the Tiled Diffusion extension

it can help you later on

[ornate elk]

for the controlnet models its important to use the smallest that are available, like these:
https://civitai.com/models/38784/controlnet-11-models

there are multiple on that page so here is a quick overview how to get them with their config file

put everything in models/controlnet folder

[tacit light]

alrighty

all done

thank you frfr ong

[ornate elk]

No problem 🙂

[tacit light]

one more question if a model on Civitai has XL in their name, It is not going to work with mine right?

also where do I get SD 2.0 is that usable?

[ornate elk]

it can work but is slower

[ornate elk]

oh nope, stay away from that xD

it will work but its so bad supported

you also need config files for some 2.1 models

not worth it

[tacit light]

i see

screw that

what about XL?

[ornate elk]

its good but still 1.5 is the most used as it has more community made loras and model variants

you can download an sdxl model like Bluepencil xl and try it out

[tacit light]

so stuff like this would be slower comparitively ?
https://civitai.com/models/260267?modelVersionId=293564

[ornate elk]

yes it would be slower and could crash when going to high in resolution

[tacit light]

i see

[ornate elk]

but if you do like 1024x768 for example that will work good

[tacit light]

yeah I usually do that or 512x512

[ornate elk]

sdxl models are trained on 1024x1024 resolution

1.5 is trained on 512x512

best resolution for portrait is 512x768

or if you want a FullHD image later then use 540x960

[tacit light]

i see

[ornate elk]

upscaling by 2 and you have HD

[tacit light]

How does tiled diff work?

is it like old blender rendering

[ornate elk]

its tiled rendering, not needed for now

but later on Tiled Vae can help in situations where vram is limited

[tacit light]

i see

okay so it is working however images are mostly ugly rn, when I used get img I had to make longer prompts so this is also similar to that right?

like detailed prompts otherwise you get ugly images?

[ornate elk]

yea you need some good quality tags, like high quality, highres, detailed, etc,
and more important some good negative prompts

for example add these to negative and create again, it will be much better image
disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, , mangled, old, surreal, text, blurry, b&w, monochrome, conjoined twins, multiple heads, extra legs, extra arms, fashion photos (collage:1.25), meme, deformed, elongated, twisted fingers, strabismus, heterochromia, closed eyes, blurred, watermark

[tacit light]

okay, what about hires. fix and redifiner

[ornate elk]

you can completely ignore refiner

hires fix is very nice to get much better quality outputs, but it slows down the process and use a lot of vram.
so if you want to use it then i can give you the best settings

[tacit light]

maybe later

what's the best place to get prompts like Lexica was nice until they added their own model to it

[ornate elk]

Civitai has a site with only images

[tacit light]

im getting really weird color issues with the model I sent above

is it because its XL model?

[vocal burrow]

what's your model and VAE ?

[ornate elk]

maybe because you used 512x512 with sdxl

[tacit light]

higher than that it ran out of memory lol

[vocal burrow]

sdxl is trained for 1024x1024

what are your settings and command line args ?

[ornate elk]

try 768x768

[ornate elk]

AMD

[tacit light]

also what sampling method should I use

there's too many lol

[ornate elk]

euler a is good, DPM++ SDE,, DPM++ 2M SDE Karras also

[vocal burrow]

for testing, Euler A is probably the best speed / safe compromise

[tacit light]

it works!

yess

m not going back to 512 lool

[ornate elk]

Euler a is better than Euler 😄

but nice image

[ornate elk]

you may cant use hires fix with sdxl models cuase of the high vram usage of both together

but you can also upscale in img2img without vram limit

[tacit light]

haha Im really happy now I can use it finally on a decent speed

yeah

[ornate elk]

yea like i said, its worth on your gpu

even with some limits

[tacit light]

that should work but SDXL model doesn't like 512 lol

[tacit light]

yess ong

[ornate ruin]

I understand that. By the way, is Automatic1111 using VRAM only? I have only 8GB VRAM but 64 GB RAM if that's any helpful for anything

[ornate elk]

the image generation part is only done with vram, you can only store some models in ram for quick switch if they arent stored on an ssd

[ornate ruin]

Aye okay

[tacit light]

@ornate elk Will the Ui save my last prompt?

or settings

How do I not start from scratch each time lol

also what is Lora, Am i able to use that with AMD set up?

[ornate elk]

Loras are smaller files trained on a specific character or style. It helps the model to create that char / style.
Yes you can use loras.
Check for their base version.
1.5 loras only work with 1.5 models and sdxl loras with sdxl

[ornate elk]

Yes you can click the blue/white arrow to get the last settings used

[tacit light]

ok

[tacit light]

cool

ty

[waxen fulcrum]

Stability API is down.

[nova ermine]

I've Installed Posex

According to the tutorial, this interface should be shown

However, I just see this

I don't see the option to modify any pose

How to fix?

[ornate elk]

Make sure the webui is whitelisted in any browser adblocker

[nova ermine]

I've finally downloaded another one which the tutorial man offers in case that doen't work

[ornate elk]

Posex is pretty outdated, guess its broken

[nova ermine]

But what exactly would be the problem if webui is not whitelisted?

[ornate elk]

Visual glitches and issues in img2img tab or some extensions dont work

[nova ermine]

Ok

SD is pretty tricky to learn

[ornate elk]

Yea it takes a while, depending on what your goal is

Also with new extensions and features the learning never ends

[vocal burrow]

SD itself is quite easy, learning its dozens of extensions on the other hand.

Start simple, play with it and add extensions to your workflow one by one.

[nova ermine]

I have some aims with SD, I guess I will have problems, so I will come here to ask experts for help. But now I'm watching a course on YT

[vocal burrow]

As you just learned, be mindful of outdated tutorials on YT.

Text tutorial are usually easier to update and therefore often most up-to-date.

You can't go wrong by simply reading the official manual / github page of whatever you're using.

[nova ermine]

I got some experience last summer, but I left SD and I have returned

[nova ermine]

Maybe I should look at text tutorial, but I get easily lost with that, I'm not sure where exactly to look for text tutorial

Also I am not so good at following text tutorial

But I'll consider

[vocal burrow]

reddit, stable-diffusion-art, github page of whatever extension you're using, etc

[nova ermine]

I guess I will come here again if I have doubts with text tutorial

[ornate elk]

This channel for everything technical, for prompting there is #📝｜prompting-help

[cinder moth]

which webui i can use instead of AUTOMATIC1111, im getting new errors with AUTOMATIC1111 every second

[viral dock]

you could try comfyui like i did

[cinder moth]

is it working properly?

[viral dock]

yes it is, here is the install guide which is also in pinned messages

[ornate elk]

What errors?

[cinder moth]

I've been getting cuda out of memory errors, now its not allowing me to use sdxl based checkpoints

i've had enough of it

[ornate elk]

What's your GPU and what's inside your webui-user.bat?
You mostly missing the right performance settings

[cinder moth]

RTX 3070 ti

[proper gust]

#1047610792226340935

[cinder moth]

i've deleted the diffusion so. Don't have the bat

[ornate elk]

Oh xD

Yea you just missed the needed performance settings in the bat

Then sdxl would have worked

[cinder moth]

Can you give me tutorial link? I can always download it back

[ornate elk]

You can checkout my install guide in the Pinned Messages of this Channel. There is everything you need to know and tweak at the install.

[cinder moth]

God bless the person who invented the recycle bin

lets try

[ornate elk]

Your card needs --xformers --medvram-sdxl --no-half-vae

[cinder moth]

thanks hope it works properly

lets try

[gaunt rune]

Hello there!
I am trying to use (guided by yt video) MeshGraphormer Hand Refiner inside a ComfyUI but i got an error.
I am using Comfy-ui installed by Stability Matrix on 6900XT and ryzen 5900x and windows11. I already found that this problem exist because of amd gpu - my question is - is that possible to fix that, or i just have to swallow sad true and avoid of using MeshGraphormer on amd gpu?

Error message:
Error occurred when executing MeshGraphormer-DepthMapPreprocessor:
new(): expected key in DispatchKeySet(CPU, CUDA, HIP, XLA, MPS, IPU, XPU, HPU, Lazy, Meta) but got: PrivateUse1

[ornate elk]

Yea that means it expected cuda or CPU but got directml, so that won't work

[gaunt rune]

Is there a way to redirect it to using cpu?

[ornate elk]

If its a controlnet model and preprocessor you can run controlnet on CPU only mode

[cinder moth]

        size mismatch for model.diffusion_model.input_blocks.0.0.weight: copying a param with shape torch.Size([320, 9, 3, 3]) from checkpoint, the shape in current model is torch.Size([320, 4, 3, 3]).

im getting this error

[ornate elk]

When did you get that?

[cinder moth]

just now

[ornate elk]

When switching the model or when generating?

[cinder moth]

switching

[ornate elk]

Mostly your webui is not updated then. Check the version numbers at the bottom in browser

[cinder moth]

version: v1.7.0  •  python: 3.10.11  •  torch: 2.0.1+cu118  •  xformers: 0.0.20  •  gradio: 3.41.2  •  checkpoint: 6ce0161689

[ornate elk]

Hmm looks good

Do you have the full cmd output ?

[cinder moth]

i've tried it twice so its the same error 2 times;

it looks like a mess

i mean it can be cause of the model maybe

let me download an another one

[ornate elk]

Yea try an other one

Like Juggernaut XL

[cinder moth]

Good model, downloading it

[past solstice]

When we can expect the server to start the services again?

[ornate elk]

Idk, but we well get that information first in #1047610792226340935

[past solstice]

Hehe, yea keenly waiting for it. I need to make a creative Fun Friday LinkedIn Post.

[fair oxide]

@ornate elk guess what, google bard can now create images like dalle3

[ornate elk]

If you dont mind waiting times, you can use the stable horde bot in browser, for free

[ornate elk]

Yea heard about it too

Is an account needed?

[fair oxide]

just your usual google account that you use for gmail

Draw a photorealistic image of a majestic golden lion standing proudly on a rocky cliff overlooking a vast savanna landscape bathed in the warm glow of the setting sun. Include tall grasses swaying in the gentle breeze and a few zebras grazing peacefully in the distance.

draw a visual representation of the following keywords in landscape format: a girl, beach, golden hour, surfing, ocean waves, orange sky.

kinda coarse but it got the context

[ornate elk]

Yea a good start

[fair oxide]

yeah, its powered by gemini

[wraith bronze]

Are there any benefits from using ComfyUI over Automatic1111? or just different workflows? I tried ComfyUI and it was a bit confusing so I couldn't give it a fair comparison and don't know if I should invest the time in it since I am pretty comfortable in 1111.

[ornate elk]

Very interesting, I can't download these images on my Smartphone. Not even if I open them in browser

[fair oxide]

oh im not sure im on pc

[somber token]

Draw a candy’s enterprise

[ornate elk]

What file type are these images?

[wraith bronze]

Thank you! It was quite easy to start generating, maybe it is worth looking into.

[fair oxide]

png

[fair oxide]

i dont buy into the idea of walking the extra length when ai should be doing the heavy lifting

[wraith bronze]

If it add flexibility I am not against it

[fair oxide]

i have tried adapting to comfyui, the wiring is impossible

[ornate elk]

Its highly customisable with individual workflows making it good if you want the full control over everything.
But if your familiar with auto1111 and just want to change some settings then there is no need to switch.
You can also have both installed and comfyui can use the models folder of auto1111, so no need to copy models over

[wraith bronze]

I don't like the rats nests that can happen xD I get UE5 nightmares

[ornate elk]

For me comfyui is to much work to just get some images. You can waste hours by making the perfect workflow. And when downloading other workflows you never know on which preference they are created on. Or if they are better ones.

[wraith bronze]

Full control sounds nice especially when if I get comfortable using it.
I feel there are a lot of shots in the dark with A1111 which might be lesser if I get comfortable in ComfyUI.

I don't mind spending hours, I usually play with A1111 while at work during downtime

[fair oxide]

there is an obvious pattern where txt2img are going if you look at dalle3 and now google imagen2, those will never require you to manage clutters of wire like comfyui and where google and dalle goes from here is to maintain ease of use

auto1111 is quite intuitive to say the least

also consider midjourney for ease of use... if users are meant to deal with underlying technicalities that defeats the purpose of AI

[wraith bronze]

I like to have a lot of control ^^ but I am no picasso 😛

Or at least to know enough to have the level of control I feel I need

[ornate elk]

Midjourney and dalle are for the normal people who just want to get some instant output with nothing to worry, like apple users for example

I like having more Control over it, especially in the terms of digital Privacy

[fair oxide]

what you are currently missing should be provided through the webui interface, unless of course you want meaningless credibility by doing what should be automated

[fair oxide]

well comfyui interface design kills creativity for me and encourages more of a mundane tasks that Ai engine should be taking care of

[ornate elk]

But yea webuis like Fooocus are nice, because they do the heavy lifting mostly, if you just want some good outputs

[fair oxide]

a1111 has a nice balance of control and ease

although not complete

[ornate elk]

Yea auto1111 is the perfect mix

[fair oxide]

the LLM has been out long before text 2 image, and there are still lots of developments and improvements going on with LLM

so im guessing in time we will see more intuitive designs with text 2image

[ornate elk]

Yea for sure! In 2024 it will change a lot

[fair oxide]

hoping so, for sure

[wraith bronze]

I think I will spend some time with ComfyUI, even if I am not super excited about the rat nets and debugging that mess 😛

[ornate elk]

Okay, I can't download these 3 images because of the file name lenght

[fair oxide]

yeah the filenames are EXTRA long lol, i pruend them for you here

[ornate elk]

Works now instantly thx

[fair oxide]

here is my take on what comfyui offers at present, the control ppl seek with it will be obsolete as Ai tech advances

[wraith bronze]

That is probably true, and when it comes to that I won't mind using those tools as well.

[wooden iron]

taking too much time 108s for per /it
is this normal

[ornate elk]

Depends on the settings and GPU, but normaly not

[wooden iron]

i dont know about the setting but i am having amd gpu with 12gb vram

but still looking for knowladge how to minimize time

[ornate elk]

Hopefully you did the steps required for amd to work in comfyui

By installing the Directml backend

[wooden iron]

yes yes

[ornate elk]

Okay good

Have you tried 512x512 ?

[wooden iron]

that is fast

but my images get pixelated

so i want some bigger quality image thats why i choose 1024

x1024

[ornate elk]

Did you used an sdxl model for that too?

[wooden iron]

no i didnt use sdxl model i used this one

[fair oxide]

@wraith bronze since you have already looked into comfyui with certain likings and limitations, have a look at SwarmUI

[wooden iron]

i use these one

[wraith bronze]

Oh thank I will, the more to try the better ^^

[wooden iron]

and get this result ( its 512x512 )

[fair oxide]

np, SwarmUI takes comfyui node orientation and offers a more easy to use interface unlike all the wiring of comfyui

[ornate elk]

Thats okay, most people use lower resolution because the 1.5 models are trained on a resolution of 512x512.
Then you need to upscale the image to improve the quality

Generating at 1024x1024 won't help with 1.5 models. It will give you a bad output

[wooden iron]

https://tenor.com/view/pop-team-epic-i-dont-know-i-dont-understand-i-dont-get-it-nod-gif-25890418

thankyou thankyou

[flint swift]

Imagine

[wooden iron]

thankyou buddy

[wraith bronze]

I'll look at it when I get home, right now I am remoting into my machine at home, Discord is not popular at work! 😛

[fair oxide]

create a portrait photo of mona lisa in modern art style

[flint swift]

Prompt: Knight with bare buttocks running across the field to attack. Rear view. Funny. Anime.

[fair oxide]

@ornate elk unlike BING and dalle3 i dont see any limits or caps to using google bard to create images

[ornate elk]

Ah okay, maybe at first they will not limit it

[fair oxide]

oh no, i dont want them to set any tokenized limits lol

holding high hopes for google 🙂

draw an anime girl at a boardwalk on a sunny afternoon with candy floss in her hand. she is wearing a yellow t-shirt, black pants, and red sneakers.

i had to re prompt shorts from pants otherwise it wont generate image

[vocal burrow]

SD1.5 is trained for 512x512, SD2 for 768x768 and SDXL for 1024x1024.
You can generally ask for about 30% lower/higher res for each of them wih little to no visual artifacts.
Also, yes you can get much higher resolution for each of them but it involves upscaling.

[fair oxide]

a girl inside a plane flying above the mountains

ok i think i should stop with all the images

[wraith bronze]

Maybe we should port SD to minecraft redstones?

[wooden iron]

it took year to complete thi que with sdxl may be my machine not build for sdxl

[vocal burrow]

asking for 512 height with an sdxl model ? chances are high your output will be trash.

[wooden iron]

so what is the best size for that

[vocal burrow]

#🤝｜tech-support message

Here's a list of some possible resolutions for SDXL:

• aspect ratio: 4:3 - 1152x896
• aspect ratio: 2:3 - 832x1216
• aspect ratio: 9:16 - 768x1344
• aspect ratio: 21:9 - 1536x640
• aspect ratio: 1:1 - 1024x1024
• aspect ratio: 3:4 - 896x1152
• aspect ratio: 16:9 - 1344x768
• aspect ratio: 5:4 - 1152x896
• aspect ratio: 3:2 - 1216x832
• aspect ratio: 4:5 - 896x1152
• aspect ratio: 9:21 - 640x1536

[ornate elk]

That would be awesome haha xD

[fair oxide]

you could try 512x768 (width x height) with sdxl if you want to save up some resources, but best not to use 512 for height as @vocal burrow pointed.

im going to test that out right now on sdxl model

[wooden iron]

which one is best for performance
upscaling latent or upscaling image

[trail hazel]

Draw the nine circles of Hell are like big mountains covered with snow. little devils wander around each of them

[fair oxide]

512x768 and flipped on SDXL along with hires fix by 2x

[trail hazel]

How to use?
I forgotten(

.

[wooden iron]

what is you pc configration

it came out fast in your case

[fair oxide]

well for one i used lower res as i just mentioned that saves up memory and im using rtx 3060 gpu with 12gb vram

[wooden iron]

right now its stuck into ksample for 4 min

[fair oxide]

im using automatic 1111 btw

[wooden iron]

i am also having decent graphic card 6750xt 12gb ram

[fair oxide]

is that AMD gpu?

[wooden iron]

yes yes

[fair oxide]

ahh ok, AMD gpu compared to nvidia specs are bit on the lower end in SD optimization

[wooden iron]

okay

[ornate elk]

Wouldn't recommend using sdxl with AMD 12gb GPUs, its slow and you have a better time with 1.5 and upscaling

If you want to try out auto1111 for AMD, checkout my install guide in the Pinned Messages of this Channel

[wooden iron]

look at that my ram and disk got bottle neck

[ornate elk]

Is that right now while you generate?

Because your GPU usage is at 0℅

[wooden iron]

this happen i dont know why when i stuck at ksampler

it always my ram and disk who got bottle neck

https://tenor.com/view/anime-animegirl-angry-angrygirl-tissue-gif-25440728

[ornate elk]

You should install any webui on SSD, and then store only the models on HDD

[ornate elk]

Mostly your windows pagefile is stored on the HDD, you would need to check that and switch it to the SSD

[fair oxide]

@ornate elk as typical as it is, bard too is sensitive to prompts over ridiculous words

[wooden iron]

i install window on ssd you can see it c disk

[ornate elk]

I know, I'm not talking about windows

The pagefile is a file that is used for caching (virtual RAM), it gets used when your Ram usage is maxed out.

That file should be located on C by default. But for some systems its stored on their HDD, slowing down everything

[ornate elk]

Ahh okay xD

[wooden iron]

thankyou thankyou

[fair oxide]

an example

[ornate elk]

Ah well the typical over censoring

[fair oxide]

i forgot to replace her with him, but it worked

[wooden iron]

problem solved

[fair oxide]

lol

[wooden iron]

but there is another problem

[ornate elk]

What was it?

[wooden iron]

me do not know

[ornate elk]

Okay what did you changed ? xD

[wooden iron]

page file and you can see there is no ram or disck bottleneck

[ornate elk]

Thats a common error for the Directml Backend. Not enough vram for that resolution

[ornate elk]

Perfect

So it was on the HDD as I thought

[wooden iron]

yes it was

[fickle nimbus]

你好

[wooden iron]

can anyone help please
i am using amd gpu 12gb ram , confyui
while using sdxl 1.0

[ornate elk]

Lower the resolution to 768x768

[wooden iron]

ok

look at that

everything all right

half way to go

https://tenor.com/view/yay-anime-girl-kawaii-gif-18081573

again

https://tenor.com/view/colere-gif-12911656620295315199

[steady mountain]

So, memory leak (python) on automatic1111 on Ubuntu. Any advice on how to keep it from accumulating?

[ornate elk]

Have you restarted before you tried?

[wooden iron]

no

[ornate elk]

Because every time you get that error you need to restart comfyui

[night tiger]

Hey guys, if we ignore the benefit of Vram for training, is there any scenario where rtx 4070 is better than rtx 3090?

Is there any latest benchmark for stable diffusion I can check?

[wooden iron]

you dont need to you can check it on website based on cuda core

[proper briar]

J

[night tiger]

I remember there were some accelerations released later which boosted the performance for the latest rtx. since I have been away for a while, I thought I should ask if those accelerations were only for rtx 4000 or for old generation s too.
Can't remember the name of the acceleration or if its released by nvidia themselves.

[thorny crater]

how upload image

[tacit light]

Can I add google LUMIERE to my AMD ui or is that not yet available?

[short tulip]

Has anyone gotten instant-id in auto111 webui to work well? It works but the result isnt very good. colors are usually washed out, weird contrast, etc.

[true adder]

#🤝

[surreal flare]

How long the bots will be down?

[unique glacier]

Can you show an example of said mess up image?

[unique glacier]

from the looks of it. your cfg might be too high

what did you set it at?

wait no. it might be vae problem

[short tulip]

I've tried a range between 4 and 6

[unique glacier]

best if you post your image here with the metadata

[storm sentinel]

Does anyone know how to install comfyui to a docker server

[unique glacier]

to do so, drag the image from your folder into this chat.

[short tulip]

Does it require a vae? I think it depends on the model, in this case I've tried Juggernaut and Albedo, both sdxl

[unique glacier]

i have no clue till you share more info.

best way is to send one of the mess up image to this chat from your folder.

dont copy paste the image to here.

[short tulip]

followed this to a T. https://github.com/Mikubill/sd-webui-controlnet/discussions/2589 same settings. tried cfg between 5 and 6. Only difference might be the main model

[unique glacier]

please send this. Would be really helpful.\

[short tulip]

Grabbed a different image since I was using a personal one. I even grabbed the similar positive and negative prompt and steps (also tried lowering the steps) from an example form the model page.

The colors are a bit weird, since this model has VAE baked in, I dont think I need to provide another on

@unique glacier

[nova ermine]

This is an image I've made in Blender using two 3D models of a video game. I want to make an img2img of this image to generate another image should be similar, but has a better appearance.

[nova ermine]

I'm looking for a model which allow me to get a result wich has the most similar appearance in style to this image, which is an official render of the character, but I don't know what model to download. Anyone could help me with that?

[fair oxide]

see if this fits https://civitai.com/models/118086/3d-animation-diffusion

[nova ermine]

Cool!

I'll check it out

[restive sequoia]

Noice

[ornate elk]

Nope

[ornate elk]

Yea it works okay

[short tulip]

What models have you tried with it, every one I try has not great results, usually ends up blurry and odd colors and contrast.

[ornate elk]

Tried it with Dreamshaper xl or juggernaut xl

[short tulip]

I tried with Juggernaut and it doesnt come out very well. Pretty much followed the examples to a T.

[ornate elk]

Its not that good but if you lower the cfg to 4 its ok

[fair oxide]

an example of instant id but i used cfg scale bit high in this one. can you show your output for reference?

[ornate elk]

Im not at the PC to share examples

[short tulip]

I posted an example if you scroll up

[short tulip]

Is your output supposed to be that bright? I can get some similarities to the face, its kind of everything else that falls off.

[fair oxide]

looking at your post, are you using any lora, embeddings?

[short tulip]

no

[fair oxide]

yes its supposee to be like that given the settings i used

[fair oxide]

try the control weight and end step both to around 0.8

[short tulip]

0.8 helps a little. Ill mess around with some stuff. Thanks for the assist

[fair oxide]

sure, you could try lowering the control weight a bit more

[ornate ruin]

53 seconds at 15 hires steps
45 seconds at 10 hires steps

Every other setting isn't changed in those two generations.

X2 upscale, esrgan_x4 0.38 denois, euer a, 41 sampling steps, 1x1 batch, 512x768 size, CFG5, realistic Vision 6

Are those timers acceptable on the faster or slower side for single images generations?

Thanks in advance:)

I also am using --xformers --medvram-sdxl and --no-half-vae

[fair oxide]

would be helpful if you mention your gpu spec

[ornate ruin]

I didn't want to mention it because I thought there is like an average estimation time that images generally take to generate. Like 30 seconds for an image at 30 steps being good, 20 seconds very good, 50-60 seconds being acceptable but +2 minutes not good

• if that makes sense 😅

But sure -
RTX 4060 mobile 8GB VRAM
I7-13700K
64GB RAM

[fair oxide]

all that would largely depend on your gpu power, sd parameters come later

[fair oxide]

i would say fairly ok with that gpu

[ornate ruin]

Whilst browsing I saw a comment like this and thought - wow is his GPU that fast or do I have to sort out some issues with mine?

[fair oxide]

but to optimize your generations, you could go with 30 steps and 10 steps for hires

[ornate ruin]

But then there was another guy who said it took him +40 minutes for a single image because he's generating on a potato PC, so I wanted to check where I stand with mine 😂

[ornate ruin]

Most checkpoints work best at around 30 steps anyways iirc no?

Like that there wouldn't be too much benefit/different from 30 to 70 or so

[fair oxide]

your gpu is pirmary when it comes to speed then comes ram and cpu and after that it would also depend on the models you are using and other settings along with lora in some cases

30 steps is ideal cause in most cases you won't notice strong artistic differences going above 30 steps

[ornate ruin]

Okay 👍🏻

So checkpoints do alter generation time and even loras can be a factor for that, I thought of that before but the biggest different so far I saw was from different samplers

[fair oxide]

if you are extremely picky about fine detail which is usually a rare case scenario you could try steps in between 30 - 50, anything above is waste of time

[ornate ruin]

For details that most often the only Lora I actually use is one to increase/adjust details so I think I can safely stay at 30 yea

[fair oxide]

yes they do and to be specific SD checkpoints and XL checkpoints vary greatly, but not to be mistaken each sd or each xl would vary in resources

[ornate ruin]

Ah okay I was just gonna ask if SDs are faster than XL haha thanks

[fair oxide]

XLare more resource demanding due to its design and size

btw @ornate ruin speaking of steps, there are LCM lora and turbo models that can produce pretty good results with as little as 6-8 steps

[ornate ruin]

Really hahah I might check that out for fun if I stumble upon one of those on Civit

I have like 20 checkpoints installed rn bc I install all checkpoints and loras that I see if they were used on an image I like 😩

[severe kindle]

Its also common knowledge that on the free tier if Google 'catch' you using an SD GUI then they'll ban you.
Colab is there to run interactive script not GUI's
Diffusers script are safe, AUTO, Comfy, InvokeAI etc not so.

[fair oxide]

8 steps and cfg scale 2

[ornate ruin]

I don't even know what that means to be completely honest

I'm just running SD Automatic1111 locally on my PC is he not speaking of the same thing?

[autumn osprey]

¯_(ツ)_/¯

[ornate ruin]

Oh that's actually a good image, nice

[fair oxide]

yes its leveraging LCM technology, you can learn about it here https://huggingface.co/latent-consistency/lcm-lora-sdv1-5

cuts down generation time more than half and speeding up your workflow

[ornate ruin]

Is that something someone can learn? I thought it's just a Lora/checkpoint you'd use. LCM technology makes it sound un-beginnerfriendly

[ornate ruin]

I don't do much anime/cartoon stuff but how long did that image take you roughly?

[fair oxide]

you can learn about its application and the files you need to grab

[ornate ruin]

Ah alright

[severe kindle]

No he's talking about running Auto111 on the server Google provides for AI developement, there's a free tier which people where abusing to do SD remotely for free.
Google clamped down on it, if you pay them they allow it

[fair oxide]

2 seconds probably, it was one of my older gen, not now

[ornate ruin]

Okay no I don't do that.. thanks for clarifying

[ornate ruin]

Neat

[fair oxide]

np, but LCM is not limited to anime, you can apply that on realistic images too

@ornate ruin this for example is also with 8 steps and 2 cfg scale

[ornate ruin]

Sick

[fair oxide]

took me about 2- 3 seconds at best

[ornate ruin]

i have just completed this one and it took like 50-60 seconds rougly

working on improving the time, thanks ill most likely take a look into LCM next

[fair oxide]

that timing sounds fairly reasonable on your rig

and SD setup

[ornate ruin]

Okay that's good to know tho 🙏

[nova ermine]

Is there doccumentation on that model to read? (I am not so able for searching that yet)

[fair oxide]

all of the detail on using it on that page.