🍥｜anime

[brittle rampart]

what mean thi is

[lyric zealot]

someone on reddit

[vital raptor]

try adding --disable-safe-unpickle to the command line (just for testing)

if it works after that, then at least we know if it's a bad/broken model causing it, or if it's something else

[brittle rampart]

nope... 😭

same errors

[vital raptor]

ok, next this --always-gpu

😛

and if that doesn't work, try --always-cpu

if none of that works, then i don't know, might have to report it as a bug

[brittle rampart]

WORKING!

[vital raptor]

it's only been out for half a day or so

[bold ether]

it's gonna be super slow that way

[brittle rampart]

with --always-gpu

[native halo]

how did u fix it?

[vital raptor]

yes, but if it was the only one that worked then we know where to start to try and fix the problem

but, since it works with the gpu flag, then it seems to have something to do with low vram cards not autodetecting properly

[bold ether]

you just need to indicate to set which device to use with --device-id x (assuming you're using auto1111), x being your gpu id (can be found in task manager)

[lyric zealot]

now add --vae-dir %A1111_HOME%/models/VAE^ --realesrgan-models-path %A1111_HOME%/models/RealESRGAN^ --esrgan-models-path %A1111_HOME%/models/ESRGAN

[vital raptor]

this isn't a1111

[lyric zealot]

if not you wont have your upscaler or vae

[bold ether]

stuff like that usually happen when extension are incorrectly written and refuse to use the same cuda device as the rest of SD

[lyric zealot]

we are using something alike to A1111, can be found here #🍥｜anime message

[bold ether]

you can always use set CUDA_VISIBLE_DEVICES=x then

[lyric zealot]

is your preview slightly different when compared to A1111

[vital raptor]

you mean during generation?

[lyric zealot]

yup

[vital raptor]

maybe a little

[lyric zealot]

same for me

[vital raptor]

seems to be more accurate

doesn't seem to affect anything though, images are seed accurate between them

[lyric zealot]

yup, it's still the same seed, so doesnt matter

[vital raptor]

oh neat, there's a new button under the preview image for the SVD tab, and also for generating an upscaled version

instead of turning on hires fix and generating again, it just does it for you

i was literally debating making an extension for exactly that feature 😄

[brittle rampart]

just add this is argument with --always-gpu

[brittle rampart]

ok try add this is

[lyric zealot]

[pallid moss]

What is the best anime checkpoint

[vital raptor]

whichever you like the most

[pallid moss]

Okay

[brittle rampart]

hmm this is a joke?!?!

where my 50% speedup!?

[lyric zealot]

something is definitely wrong

[brittle rampart]

HRes.fix step slowly 3x....

in general, my usual generation time has always been 3-5 minutes. and this is with After detaler

now here without Adetaler estimated time 11 minutes?!

[lyric zealot]

every parameters are the same?

[brittle rampart]

yep

same Upscaler same Prompt

[vital raptor]

some things seem to be broken currently, like caching models, probably from the memory use reworking

not sure if it's causing your problem, but it's causing some major slowdown for me overall

the actual generation of a base image is faster, but when it switches to hires, then switches to adetailer, there's significant pauses inbetween

[brittle rampart]

maybe i try now this --always-offload-from-vram

[vital raptor]

i'm still generating a sdxl image with hires x2 and adetailer in ~29s so i shouldn't complain too much

but i'm used to that happening in ~18s 😄

12-14 with TRT enabled, which doesn't work in this

[vital raptor]

my favorite SD1.5 screenshot 😄
batch of 100 at 30 steps each

[brittle rampart]

hey hey it`s really slowly

what happen....

this is look like i am use ControlNet.....

[native halo]

https://tenor.com/view/parks-and-rec-ron-swanson-computer-trash-garbage-gif-21878872

[lyric zealot]

is your model located on SSD or HDD

[brittle rampart]

hdd

[vital raptor]

lol, built-in LCM sampler

[lyric zealot]

i guess that might be the reason why

[brittle rampart]

bit in a1111 webUI this is work faster...

and model folder on HDD

[lyric zealot]

in a1111, you don't have to keep loading and unloading the model, but i guess this version, they are doing that

[brittle rampart]

but the loading of the model occurs before the start of the stage HRes.

but for me the process Hres itself is very slow

[vital raptor]

it's a bit buggy, might just not be good for your card yet

[brittle rampart]

crap!

[vital raptor]

I keep getting this crap randomly
Moving model(s) has taken 9.00 seconds
and my entire system is NVME drives and lots of fast ram

[lyric zealot]

[vital raptor]

i can move 200GB from one drive to another in less than 9 seconds

but, it's only been out for ~9 hours and it's a pretty good start for what it is

i could see this easily being the 'default' webui replacing a1111 soon

hmm

the --always-gpu thing might actually be causing your speed problem

it's suggested as a fix for keeping models in vram for 4090's, which means it might not be letting your card release the vram

using set CUDA_VISIBLE_DEVICES=x might work better

[brittle rampart]

It turns out I can't work with this version WebUI?

[brittle rampart]

x - this is ?

or set CUDA_VISIBLE_DEVICES=x

[vital raptor]

should be 0, but it might be 1.. depends on your system

[brittle rampart]

oki doki. try it!

[vital raptor]

should show in task manager

[lyric zealot]

go to your task manager > performance > GPU

[vital raptor]

under performance

[brittle rampart]

i have just 1 GPU

[vital raptor]

should be 0 then

[brittle rampart]

and remove --alway-gpu

[vital raptor]

yeah

[brittle rampart]

oki

but agane errors

[vital raptor]

try adding --xformers

i wonder if it's a pytorch thing

looks like --always-gpu is for high vram cards, not low ones 😄

fixed my issue

[native halo]

least buggy a1111 fork

[pallid moss]

Finally a face that looks good

[dusk seal]

Yep

10/10 would

[pallid moss]

Question, if I wanna create someone with sword do i need 2 loras like one for character and one for sword?

And how do i arrange them in prompt section

[dusk seal]

good question

idk

[dusk seal]

i'm just here to lurk pics

[brittle rampart]

😭

[brittle rampart]

LoRa "more_details" add darkness on pic. try use LoRa "add detail"

[brittle rampart]

and you mayby can try add LoRa`s "Hyper refiner" + "Hyper denoise" for more best result

[brittle rampart]

you mean just old a1111 WebUI?

[vital raptor]

Yeah mine says that too. Haven’t looked into it yet and now I’m at work

[brittle rampart]

hmmm... now i try activate this is

and generation finish with time 5 min

5 min wihtour ADetailer...

@lyric zealot how you add Extentions here?

[lyric zealot]

like how you did it in auto1111, but you cant link them together

[steady grail]

ohy

what are you using?

[brittle rampart]

i cant use it together?

[lyric zealot]

https://www.reddit.com/r/StableDiffusion/comments/1ajxus6/stablediffusionwebuiforge_for_low_vram_machines/ this

[brittle rampart]

m?

[lyric zealot]

just go to extension tab and add what you want

[steady grail]

oh, that's interesting

[brittle rampart]

ah oki

[lyric zealot]

it's gonna download a new copy for this webui

[brittle rampart]

understand.

[steady grail]

already played with it?

how much it's the improvement?

[brittle rampart]

but your Vcard good

why?

[steady grail]

4 gb is not goot at all

[lyric zealot]

yup, it's not bad. From 1 minute sdxl generation to about 45 seconds

depends tho

[brittle rampart]

really!& i think you have a 8Gb

[vital raptor]

I get about 1-2it/s improvement on a 4090. Supposed to be better for smaller cards

[steady grail]

not at all, i hope to upgrade something like this, but for now i'm stuck with 4 gb vram

[vital raptor]

I was more interested in all the other improvements they’ve made. The performance was a nice bonus

[brittle rampart]

hey hey now best time 4 min 27 sec... but

but i still want speedup 50%!

grrrrrr

[pallid moss]

I did in the second one

How do I improve quality tho

[brittle rampart]

very strange.. but the generation time dances back and forth!

[brittle rampart]

use HiRes.fix step

time generation without HiRes really faster! Previously it took me 40-50 sec. now 20-30 sec...

[pallid moss]

Wdym😭

[pallid moss]

Some youtubers told not to tho

Sorry for so many questions idk anything about these coding stuffs..even the YouTube tutorials were kinda complicated for me

[brittle rampart]

without this you will lose a lot of detail. which you won’t be able to “pull out” later if you do Upscaling through img-to-img

[lyric zealot]

if you can, always do hires fix

[pallid moss]

How about the settings be?

[lyric zealot]

depends, but usually denoising can be from 0.4 to 0.6, use around 0.5 to start with. Hires steps can be like 10. And upscale by, go for 2x if you can, if not 1.5

[lyric zealot]

obviously it depends on your own gpu

[brittle rampart]

this is my

I put the steps at least half of the steps of regular generation

Denoise intensity 0.42 max

[steady grail]

i get this 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 index in method wrapper_CUDA__index_select)

it's always a pain to change gui

[brittle rampart]

onee-chan!!! You, like your sister, won a prize!

😂

[steady grail]

what's that? lol

[brittle rampart]

thats our low Vram Card! add --always-gpu

[steady grail]

at the end of run but i suppose

[brittle rampart]

set COMMANDLINE_ARGS= --always-gpu

[steady grail]

oh, like the classic auto

ty

[brittle rampart]

yep

and like in comfy you can add paths to model folder

unrem it in bat file

[little flume]

。。。

[pallid moss]

This?

Will it make my faster too

[steady grail]

still the same error

maybe is the model

anywau, the bat should be like this? @echo off

call environment.bat

cd %~dp0webui
call webui-user.bat

set COMMANDLINE_ARGS= --always-gpu

[brittle rampart]

i add this is in webui-user.bat

now my webui-user.bat look like this ```@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --always-gpu --always-batch-cond-uncond --theme dark

set CUDA_VISIBLE_DEVICES=0

@REM Uncomment following code to reference an existing A1111 checkout.
set A1111_HOME='e:\3D_Projects\stable-diffusion-webui_v_1\webui'
@REM
set VENV_DIR=%A1111_HOME%/venv
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS%^
--ckpt-dir %A1111_HOME%/models/Stable-diffusion^
--hypernetwork-dir %A1111_HOME%/models/hypernetworks^
--embeddings-dir %A1111_HOME%/embeddings^
--lora-dir %A1111_HOME%/models/Lora^
--vae-dir %A1111_HOME%/models/VAE^
--realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
--esrgan-models-path %A1111_HOME%/models/ESRGAN
call webui.bat

[steady grail]

ohhh

that's it

ty+

[brittle rampart]

still 5 min... with HiRes + ADetailer... hmmmmm

[steady grail]

i try to set all the directory of comfy

so that i can use the 2 gui, with the same directory

[brittle rampart]

yep. like me

[lyric zealot]

i think there's a mininum of like 6gb vram at least

and if you have that cuda error, using always-gpu will basically make it the same as a1111

but im not sure tho

[brittle rampart]

should try to start xformers

maybe he can save the situation

@lyric zealot you work in Linux?

[lyric zealot]

nope, windows

[brittle rampart]

oki...

[brittle rampart]

look on this is
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/7115

[brittle rampart]

you mean i cant make faster generaition on 6Gb VRam?

[lyric zealot]

maybe, not sure

imma be honest, i have no idea what im talking about

so probably not

[brittle rampart]

wow 😱 ... this is my problem too
https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/56

wow!!! it really work with SDXL model on my 6Gb Vram!

but i have a some question about SDXL generation

[lyric zealot]

ask away

[brittle rampart]

@lyric zealot look this is my last generation on SDXL model.

[pallid moss]

My eyes hurting

[lyric zealot]

do you have the correct vae?

[brittle rampart]

what the rifiner i must use?

[lyric zealot]

refiner isn't needed for all models

[brittle rampart]

ahh VAE! crap!

[lyric zealot]

you can try turning it off for now

then use the correct sdxl vae

[brittle rampart]

i have only one this is

[lyric zealot]

yes

#🍥｜anime message

[brittle rampart]

oki try againe.

[lyric zealot]

your resolution also might wanna increase a little. Since most of the sdxl model are trained on 1024 or 768

so if possible minimum go for 768x768

[brittle rampart]

hmm oki

ratio must be 1x1?

[lyric zealot]

nope

im using 768x1024

but try to be more than 768

[brittle rampart]

I'll try to start small. gradually increasing

[lyric zealot]

[brittle rampart]

@lyric zealot i have in termonal some errors on Lora

[lyric zealot]

even if 512x768, should not be a problem, maybe just a little zoomed in?

[lyric zealot]

is the lora compatible with sdxl based model?

[brittle rampart]

of course no

[lyric zealot]

press the lora tab, and if you have sdxl model loaded, you will see all the lora that are compatible

[brittle rampart]

I was happy early... 😭

with the correct VAE the estimated time is more than 40 minutes

why.....

[brittle rampart]

hmmm 4 min 36sec. SDXL

but this is a low setting

9 min

i am really not understant how it work

but this is SDXL on my 1660 6Gb

[gloomy scroll]

I get 7-9min on my 1070 8GB if I don't have window with video playing focused or app with hardware acceleration on focused

[brittle rampart]

and i do not have any SDXL LoRa

[brittle rampart]

but you have 8Gb i am 6Gb

[brittle rampart]

and in old a1111 i have a more 50 min on SDXL generation

[gloomy scroll]

old

[lyric zealot]

it's definitely an improvement then

[brittle rampart]

yes

[lyric zealot]

#🍥｜anime message

[brittle rampart]

this is my last setting

[gloomy scroll]

Oh yeah I was supposed to retest that now that I figured out how to stabilize my generation time

[brittle rampart]

how?

[gloomy scroll]

Just figured out that browser I used with a1111 had hardware acceleration enabled and it slowed genning like 10min

All the hands

[brittle rampart]

hmmm and where in FireFox browser hardware acceleration setting?

[steady grail]

wow... it's fast

[gloomy scroll]

https://i.imgur.com/F9FOlHl.gif

[steady grail]

i mean, really fast

i must try the upscaling

[pallid moss]

gimme best argument for 4gb vram

[brittle rampart]

fast what?

[steady grail]

this forge

[brittle rampart]

and SDXL work on my 6Gb Vram!

[steady grail]

it's flexible as comfy, but it's very fast

maybe i should use comfy only for img2img

[brittle rampart]

👍

[gloomy scroll]

I guess forge decided to update

[brittle rampart]

what you mean?

[gloomy scroll]

It worked in the morning

[brittle rampart]

now my server forge working with xformers

[steady grail]

what did you do?

just updated?

[brittle rampart]

but i am not restart it more than 40 minutes

wooow! they fix it! 40 min ago

https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/56

try update now

[brittle rampart]

true

😭

[gloomy scroll]

Ok I fixed my issues with changing \ to / in a1111 path

[gloomy scroll]

Or maybe I just started a1111 bat file

[brittle rampart]

welll try it agane

[gloomy scroll]

Why does it even want original a1111 folder path if it doesn't take every model and extension from there

[brittle rampart]

I do not believe in that!!!! 2 min!

[lyric zealot]

you can get it to use your models/lora/embeddings and such

extension probably not tho

[gloomy scroll]

It didn't take upscaling models and vae

[lyric zealot]

  --vae-dir %A1111_HOME%/models/VAE^
  --realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
  --esrgan-models-path %A1111_HOME%/models/ESRGAN```

go add this yourself

and it should work

[steady grail]

oh, that's where the codes for vae and upscalers

[lyric zealot]

include a ^at the end of your Lora to become Lora^

[steady grail]

do you know the difference between realesrgan and esrgan?

[lyric zealot]

so it should look something like that

[lyric zealot]

no idea tbh

[lyric zealot]

most of my upscaler are in ESRGAN

[steady grail]

damn, one letter on the left

i'm very bad with multi tasking

[lyric zealot]

can your show your command line args?

[steady grail]

i will never became plain pilot

[gloomy scroll]

I jsut used defaults but I still don't understand why have seperate arg for each model folder when there is root model folder

[vital raptor]

it's lazy because checkpoints, loras, vae, can all be safetensors

so they use different directories instead of looking at each file to decide what it is

[lyric zealot]

well, because sometimes there might be people like me where i stored my most commonly used models on SSD

[vital raptor]

probably for easy performance

[lyric zealot]

and that's why my ckpt-dir is totally not in auto1111 folder

[gloomy scroll]

I have just symlink

[lyric zealot]

symlink works for a1111

[vital raptor]

none of my files are in the a1111 folder, i have it all in it's own folder so any UI can be pointed at them

[lyric zealot]

but you cant use the forge and direct it to a1111. Since it will not look at the symlink there

[gloomy scroll]

Forge at least checks models folder even if it's symlink

[lyric zealot]

[native halo]

where --controlnet-model-dir at

[vital raptor]

in settings

the flag was removed because controlnet is built-in

so there's a spot for it in settings

[native halo]

i prefer a command since im an arch linux user, im removing a star from the review

[vital raptor]

lol

why are you using webui then?

just commandline SD

[native halo]

tryin to remember 2k lora names on cmd

[brittle rampart]

guy`s who know how work this is and why?

[vital raptor]

HyperTile optimizes the self-attention layer within the Stable-Diffusion U-Net and VAE models, resulting in a reduction in computation time ranging from 1 to 4 times, depending on the initial resolution and tile size. The implementation is exceptionally straightforward.

[brittle rampart]

then should it be turned on during generation?

[gloomy scroll]

I need to remember all the settings I changed

[vital raptor]

ah, i see what it does

[vital raptor]

use it if hires fix goes OOM

with 24GB vram i still can't do 3x hires on sdxl, but with HyperTile turned on, it worked

[vital raptor]

still maxed out my vram plus some, but it didnt crash

and finished pretty fast, 1m 22s

3072x3072

[brittle rampart]

and what parameters?

[vital raptor]

i just left the defaults

[gloomy scroll]

My vram doesn't hit max

[vital raptor]

try with sdxl hires x3 and adetailer

[brittle rampart]

hmm i am still have a 9 min SDXL generation... but ...

[vital raptor]

9min sdxl is pretty good for a small card

[gloomy scroll]

[brittle rampart]

ah Yes! but this is 1.6x HiRes.fix only

and 8 step

[lyric zealot]

you are right

i didn't oom

taking ages tho

[vital raptor]

yeah

[gloomy scroll]

I got just under 3min gen with these settings was just missing ADetailer

[vital raptor]

it slows down because it's overflowing into system ram, but it gets the job done eventually

i maxed a couple times during a 1min 3x hires gen 😄

[brittle rampart]

what about --xformers - in bat file in forge?

[vital raptor]

i guess i might as well see how far i can push this

[vital raptor]

try it

doesnt help mine any, but it might for you

[brittle rampart]

oki

[lyric zealot]

so fucking long for this

7 minutes

dafug

[brittle rampart]

SDXL?

[lyric zealot]

SDXL 3x upscale

yup

[vital raptor]

😄

[brittle rampart]

3X upscaling! dreams.....

[vital raptor]

GPU go brrrrr

[lyric zealot]

https://tenor.com/view/anime-angry-slow-loop-koharu-gif-25050167

[vital raptor]

apparently 4x is too much

[brittle rampart]

my SDXL generation in forge

[steady grail]

exactly what i was looking for, ty

[vital raptor]

she can't take much more captain!

i rarely see this card maxed out, that's impressive

[lyric zealot]

time to get a new card

[brittle rampart]

That's it... I've arrived... I'm running out of disk space...🤣

[vital raptor]

yeah, time for a H100 or something

lol my pc is so laggy, I typed that 2 minutes ago

Having to use discord on my phone 😆

[brittle rampart]

SDXL - good thing but still not for me...

[vital raptor]

On my 3080ti it was taking 40s for sdxl and I gave up on it until I got the 4090

I don’t seem to have patience for pretty pictures

[pallid moss]

Gimme something to create

[vital raptor]

A girl with half blond and half blue hair with red eyes wearing a hoodie and bunny ears

Sitting in a box

[gloomy scroll]

Instructions not understood got neko

[native halo]

steven seagal

[vital raptor]

With cat ears

[lyric zealot]

slight horror

[lyric zealot]

slight horror

[gloomy scroll]

[pallid moss]

Gave me error

[vital raptor]

Ffs my gpu is still maxed out on this gen. Almost 10 minutes so far

[pallid moss]

Why my gfpgan doesn't work

[gloomy scroll]

[lyric zealot]

[lyric zealot]

more horror than before

[pallid moss]

Smash

Whatever new thing i do stable diffusion is always downloading something

[vital raptor]

This is why my ai folder is 650gb

[pallid moss]

I only have Ae, Photoshop, Premiere pro and stable diffusion

[pallid moss]

Still my cdrive is about to fill

[vital raptor]

Small drives are small

[brittle rampart]

and alot of games! 😂

[gloomy scroll]

I have Skyrim

[pallid moss]

Deleted everything after finishing tbh

[brittle rampart]

with 2000 mods

[pallid moss]

Not to mention editing materials

[gloomy scroll]

Casual numbers

[pallid moss]

What checkpoints u guys using

[native halo]

custom mix

[gloomy scroll]

Soushiki for 1.5 & autismmixSDXL_autismmixConfetti

[pallid moss]

Lmao

[vital raptor]

Well this stupid gen is still going. Guess I’ll go give my dog a bath and see if it’s done after that

[pallid moss]

I can only create like 10 image in stable diffusion in my lifetime with this speed

[pallid moss]

[brittle rampart]

hmm forge really faster... well try add some my fav extention here...

[nova remnant]

How much faster is it for you

[wooden coral]

SVD

[nova remnant]

[brittle rampart]

50-60%

[wooden coral]

Literally just makes it bounce

[brittle rampart]

and i can make SDXL gen

8-9 min

[brittle rampart]

SD 1.5 in my case 2-2.5 min. (not forge version it was 4-5 min)

[nova remnant]

[pallid moss]

Which model will be best for something like this?

[nova remnant]

I don't gen much with my PC anymore but might try forge just to see

[pallid moss]

I also got 4gb vram

[gloomy scroll]

Forge increased my speed on 1070 8GB about 40%

With XL model

[pallid moss]

What is that

Tell me

[gloomy scroll]

https://github.com/lllyasviel/stable-diffusion-webui-forge

[pallid moss]

Also which model will be best for something like this(a lil more realistic ig)

[native halo]

idk search for game cg illustration models in civitai

[brittle rampart]

1 min 34 sec.

[pallid moss]

Why my arg doesn't work

It only works when I only put ---no-half

[brittle rampart]

not

[pallid moss]

Even --xformers--medvram was showing error

[brittle rampart]

you must start without any arg

[pallid moss]

How did you increase speed

My ram is getting eaten very fast

[native halo]

you buy more ram

[vital raptor]

SDXL hires x3.75 (3840x3840), 58m 27s

[gloomy scroll]

I can't get forge to save to my old image folder

[lyric zealot]

go to settings > paths for saving > output directory for images

then just paste the whole directory in there

make sure the format for saving is the same as your a1111

[gloomy scroll]

Yeah but it gets error when I try to save to that folder

[wooden coral]

An hour for a single image

[lyric zealot]

what error

[brittle rampart]

how!? just try use new WebUI

[gloomy scroll]

why is it using that temp directory

[pallid moss]

How do I use new webui

[vital raptor]

yeah, pinned my 4090 at 100% the whole time.. not worth it

[lyric zealot]

did you change the format for saving to be the same as your a1111?

[native halo]

your folder name path is too dam long

[brittle rampart]

#🍥｜anime message

[wooden coral]

An hour with the 4090? how

[gloomy scroll]

You sure?

[vital raptor]

was playing with the new Forge UI and it has HyperTile built-in, makes it so you can hires fix more than your ram normally allows

[vital raptor]

probably better settings to use, but i was just testing it

[gloomy scroll]

The damn gradio folder isn't even in that temp folder

[native halo]

whats the full path name where you wanna save the imgs?

[wooden coral]

Another UI to install

[gloomy scroll]

D:\AI - Images

But problem is with temp folder

[wooden coral]

Maybe it doesnt like space or -

[native halo]

😔

[vital raptor]

it's a fork of a1111 from the controlnet/fooocus person, it has some nice updates but most of the improvements are supposed to be memory stuff for lower end cards

[native halo]

try with D:\AI\Outputs

[gloomy scroll]

I would expect that they didn't change how a1111 saving works

[wooden coral]

Yeah ive read through it, only thing that interested me was the increased res you can generate at

[gloomy scroll]

That isn't the temp folder even so it isn't the problem

[wooden coral]

You never know

[vital raptor]

yeah, that's what i was testing, I can do a 3x hires sdxl in ~1min

4x won't go at all though

[gloomy scroll]

And the error is csv error so maybe I should delete that 50MB log file

[vital raptor]

so, you should be able to copy your ui-config.json and config,json from your a1111 folder to your forge folder and it'll have all your settings except the ones that changed

[gloomy scroll]

Pls

[vital raptor]

but, make a backup first because i dunno if it will work 100%

[gloomy scroll]

log file was too large for it to update it

[vital raptor]

lol

[gloomy scroll]

But it's only 180k lines

[lyric zealot]

that's what she said

[pure oyster]

9/10

[gloomy scroll]

https://tenor.com/view/yoink-cat-dragged-gif-1263419341748157821

[lyric zealot]

i downloaded stealthpnginfo on this webui as well so go ahead

[gloomy scroll]

Oh yeah I need to install it again

Maybe I just check if I can make forge to use original extension folder

But now back to work

[lyric zealot]

https://tenor.com/view/anime-anime-girl-horimiya-hori-wave-gif-18306417171641386768

[vital raptor]

that might work, but you'd have to uninstall controlnet from a1111

since it's built-in on forge, you can't have it installed

[pallid moss]

The git clone installation was so fast man

Took 2 sec

[lyric zealot]

that's just downloading it

so should be fairly fast

[pallid moss]

Wouldn't that automatically put it in stable diffusion?

[lyric zealot]

what do you mean by that?

[vital raptor]

not just the hand..

[pallid moss]

I ran the link on git bash

I'm dumb asf idk anything

[lyric zealot]

im confused af. You are trying to install the new webui right?

[pallid moss]

Yes

[lyric zealot]

then click this #🍥｜anime message

[pallid moss]

It told to run the code on git vash

Bash

[lyric zealot]

i mean, you can do that, or run the one click version

but if you already git clone it, then open the folder and you should see update.bat and run.bat

run update.bat first

then run.bat

[pallid moss]

Oka

[vital raptor]

the update.bat isn't in the git clone for some reason

also isn't needed with a fresh clone

[lyric zealot]

ahhh

[vital raptor]

but, git pull if you need to update

[lyric zealot]

if you clone that shit, then you just run webui-user.bat

[pallid moss]

I just did

Git clone (the link)

And it said done

[brittle rampart]

I understand why Cnet is built in here...
​this was done by the same author

[vital raptor]

yeah

[brittle rampart]

then I will hope that it will also work quickly!

[pure oyster]

added shading and post process effects

[lyric zealot]

nice

[brittle rampart]

sad... Civitai Browser+ cant run in forge WebUI

[wooden coral]

This Lora is kinda hilarious

[vital raptor]

is that supposed to be a body pillow, or a body in a bag?

[wooden coral]

Shrinkwrap

[vital raptor]

lol

[wooden coral]

https://civitai.com/models/289027/shrink-wrap-style-sdxl?modelVersionId=324992

Its not made for anime models I feel

[vital raptor]

that dude makes so many loras

[wooden coral]

Yeah and there are a few epic ones in there

Fried egg is my fav

Kinda works

[brittle rampart]

this is look like anime doll figure in package for saleing

what is this? first for making video but second? Zero123?

[vital raptor]

it's the new 3D model generator

https://discord.com/channels/1002292111942635562/1184555609232777337

[gloomy scroll]

I can make 3D model out of my waifu?

[vital raptor]

if you have the minimum 24GB vram, sure

[brittle rampart]

if you have a lot VRAM!

[gloomy scroll]

[brittle rampart]

i was right

[vital raptor]

i've been meaning to try it so i can send the model through my resin printer 😄

[pure oyster]

my amd card can't run sdxl despite my 24gb vram 😦

[brittle rampart]

interesting Tile VAE in forge WebUI is it necessary or does it no longer make sense to use it?

[vital raptor]

It uses it when it needs to

Automatically

[pure oyster]

Anyone know what version these things should be?

[vital raptor]

generally, whatever it installs

a1111 is on version 1.7 though

[magic wave]

More Kobeni

[brittle rampart]

tile vae not work for me...

[gloomy scroll]

What error?

[pure oyster]

Best girl

[normal stirrup]

Пьеро

[magic wave]

Yes

[pallid moss]

Quality so high

[pallid moss]

How?

[pure oyster]

I do my own lighting.

[pallid moss]

Oh okay

[pure oyster]

For more complicated changes you can just very coarsely composite stuff in and inpainting will blend it in almost perfectly.

[gloomy scroll]

Is this some 2d game character idle animation?

[signal panther]

Looks dashin'

[brittle rampart]

    Traceback (most recent call last):
      File "E:\3D_Projects\stable-diffusion-webui_v_forge\webui\modules\scripts.py", line 798, in process
        script.process(p, *script_args)
      File "E:\3D_Projects\stable-diffusion-webui_v_forge\webui\extensions\multidiffusion-upscaler-for-automatic1111\scripts\tilevae.py", line 716, in process
        if devices.get_optimal_device_name().startswith('cuda') and vae.device == devices.cpu and not vae_to_gpu:
      File "e:\3D_Projects\stable-diffusion-webui_v_forge\system\python\lib\site-packages\torch\nn\modules\module.py", line 1695, in __getattr__
        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
    AttributeError: 'AutoencoderKL' object has no attribute 'device'```

[gloomy scroll]

I would guess forge has its own

[brittle rampart]

you mean hypertile?

[gloomy scroll]

At least I don't think that Tiled VAE would increase Forge's performance

[brittle rampart]

well I was hoping to try to SDXL speed up a little

[gloomy scroll]

I would say that Forge is already using all known reasonable speed up methods

[brittle rampart]

oki-doki...

then i can uninstal Tile VAE extention. inForge

guy`s and what is this?

Crap! I really can’t get used to it, the generation speed is just over 1 minute 😁

[wooden coral]

I just fed the image to SVD tbh

[gloomy scroll]

Not many years and you can get image right after clicking generate button

[supple raptor]

Which gpu do you have?

[supple raptor]

Its a different method of fixing higher than native resolution generations. I've yet to get it to give as high of quality as traditional highres fix so I just don't use it.

[brittle rampart]

thanks

[patent shard]

she is angwy

[supple raptor]

https://cdn.discordapp.com/emojis/1093687388905161019.gif?size=64&name=pat&quality=lossless

[brittle rampart]

why did you offend her?

[gloomy scroll]

[gleaming flame]

hi! doing lots of anime, or half-anime stuff lately... this time not ghibli style. How do they look?

[patent shard]

i give her milk

[magic wave]

Those look good :o

[patent shard]

she is full now

[magic wave]

Makima

[nova remnant]

Bretty

[nova remnant]

Google bard is pretty good in generating images, follows descriptive prompt very well

Draw a photorealistic dreamy cottage next to a stream. There's trails of smoke coming out of it's chimney. Colorful wildflowers around under the sunny blue sky.

[brittle rampart]

sometime. you still need to adapt the prompts it generates

[nova remnant]

Depends on users expectations but in general it adheres to prompts nicely

[brittle rampart]

I would rather say that he gives good IDEAS

ideas connected into fairly meaningful and well-thought-out sentences and phrases

[nova remnant]

What do you mean by that?

Umm.. bard can help with ideas too, but that isnt what im talking about its image generating capabilities

[brittle rampart]

you give him a couple of basic words or phrases and get good sentences that reflect their essence

[nova remnant]

Nothing new for those who have been using bard all along, but this image drawings compared to bing is impressive

Btw they are calling bard now gemini ...

Obviously since its now powered by gemini

[brittle rampart]

oki

[nova remnant]

Well the water it shows is not a stream, also missing the flowers here that i wanted to capture. Is that bing btw?

[brittle rampart]

yep this is your Draw a photorealistic dreamy cottage next to a stream. There's trails of smoke coming out of it's chimney. Colorful wildflowers around under the sunny blue sky.

[nova remnant]

If that's bing its seemingly missing the prompt context

[brittle rampart]

I just took what you wrote.

[nova remnant]

That's what im saying ... Its not following the prompt how i wanted the image to be

No smoke out of chimney, no stream, no wildflowers, and the cottage looks like a barn

[brittle rampart]

and made a couple of changes (increasing the weights on those tokens that should reflect the essence) regarding the model used.

[nova remnant]

Are you using that on SD or bing?... Lemme scratch that, i see you're using sd model with weights and by the look of image dimension.

[brittle rampart]

yep. on SD 1.5

[brittle rampart]

of course . I also added tokens that, to my taste, reflected the theme of the entire illustration. As I said earlier, Google bard gives just good IDEAS.

and the results I received were more than satisfactory 😉

[nova remnant]

I was referring to how closely bard adheres to the prompt its given, not individual tates or ideas they like.

[brittle rampart]

but does bard generate images?

[nova remnant]

You can try with some of your ideas to see how well it follows prompts

I posted the image on top that bard generated

Bard is using Imagen 2 for generating images

Yep, powered by gemini

[brittle rampart]

I thought Bard was a text chat

[nova remnant]

they recently added the image gen feature

[nova remnant]

It was until a couple of days ago

somethin sus is amongus

[brittle rampart]

Well, then I didn’t understand you from the beginning. Sorry.

[brittle rampart]

but for me it’s still just a chat

[brittle rampart]

where img generation?

[nova remnant]

ask the chat bot to draw you something

"Draw a" ....

[brittle rampart]

I can't help you because I don't know how to create images yet.

🥴

[brittle rampart]

this is function work only on english language

[nova remnant]

This is the level of ridiculous censorship btw

[brittle rampart]

well not sure what it understand me

[nova remnant]

lol I don't use bard I use local SD models

Me too, i was just trying out bard

[brittle rampart]

me too

[nova remnant]

the freedom of local is just too good.

[brittle rampart]

That’s why i am said that Bard is good at generating IDEAS

[nova remnant]

Absolutely

[nova remnant]

makes sense

[nova remnant]

using that prompt modified slightly on my local workflow

[brittle rampart]

I like the houses based on the idea from the Bard

[nova remnant]

SVD or AnimateDiff?

[vital raptor]

svd

[nova remnant]

did you see the SVD-XT 1.1 update on huggingface?

https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt-1-1

[vital raptor]

no, just the base SVD model, playing with the feature in ForgeUI, just threw an image in it and hit generate

[nova remnant]

nice

I'm waiting for the video models to get out of research/experimental phase before I play around with em

[supple raptor]

"butterfly" has the word "butt" in it. obviously you're trying to prompt lewd images

[nova remnant]

lol

[vital raptor]

also 'cute' is only used with certain kinds of images..

[nova remnant]

You'd be even more shocked to know that wasn't it, it was the word cute that got caught up in woke net, after removing "cute" it generated the image following the exact context

[supple raptor]

"anime girl" is enough to get caught in the Bing/Dall-E filter pretty often so it really doesn't surprise me

[nova remnant]

AnimagineXL v3 really loves the low cut dresses unless you specifically prompt against it lol

[nova remnant]

Yeah but i had high hopes with googles bard

After all bard seems to proclaim not censoring too much, at least in text chat, i was excited about it with image gen as well

[nova remnant]

D3?

[brittle rampart]

what?

SD 1.5

[nova remnant]

the particle effect always makes me think D3 lol

[brittle rampart]

you mean effect depth of field?

[nova remnant]

With Ai technology they figured out what is aesthetic and attractive about feminine expressions and they chose to censor that, this woke culture is cancer.

[brittle rampart]

but it seems to me that Forge interprets prompt much better

[nova remnant]

@supple raptor its basically killing its own functionality

lol

that's why I do everything local, fully uncensored and complete control.

Yeah same , but this is shocking what's going on with society and the woke culture

I didn't care what society thought in the 90s, 00s, 10s and certainly not in the 20s

@brittle rampart

You happy with the forge UI

[brittle rampart]

yep!

[vital raptor]

SVD XT 1.1, not really much difference besides a little increase in stability with more frames, still not really useful for much

[nova remnant]

It creeps in slowly and we are seeing some of its horrible reflection

[vital raptor]

The motion it does is random, what it decides to animate is random

[nova remnant]

what is svd xt 1.1

[vital raptor]

stable diffusion video

[nova remnant]

[vital raptor]

https://discord.com/channels/1002292111942635562/1176560067961696276

[brittle rampart]

this is who have a lot of VRAM..

[vital raptor]

yeah, it maxes out my 24gb

[native halo]

you can get better results with vid2vid + controlnet and it doesnt eat a lot of vram

[brittle rampart]

the main question is that I don’t know what this is for? generate 20 sec video...

[native halo]

its text2video

[brittle rampart]

Understand. but why?

[native halo]

for the memes

[brittle rampart]

just for fun 🤣

[brittle rampart]

redhead miku!

[nova remnant]

buttafly

[nova remnant]

[gloomy scroll]

Time to go enjoy my comfy pillow

[nova remnant]

https://tenor.com/view/pat-cat-pat-pat-gif-25250101

[vital raptor]

I think the idea is eventually it’ll be able to generate longer videos

But for now it’s just a neat gimmick

[acoustic locust]

bard tends to forget what it does

[acoustic locust]

Model?

[lyric zealot]

should have pnginfo

[lyric zealot]

1girl, hatsune miku, alternate costume, field, sunset, butterfly,

<lora:[diffLora] filmgrain:1>
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, negative_hand-neg, text,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 543032961, Size: 768x512, Model hash: ec37a01a01, Model: anmitsumimimi_v10, VAE hash: 63aeecb90f, VAE: kl-f8-anime2.ckpt, Denoising strength: 0.4, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.1.2, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x-AnimeSharp, Mask blur: 4, Inpaint area: Only masked, Masked area padding: 32, Version: f0.0.7-latest-44-g7359740f```

[nova remnant]

wish the SEGS hand detailer worked as well as the face detailer

[nova remnant]

[vital raptor]

FreeU is interesting, small u-net modifier (built-in with ForgeUI) that tweaks specific values, generally makes the image more saturated and adds a bit of detail in the lines. I went a little heavy with it in this example to really see what it was changing, but it seems it can help with models that feel a bit washed out (like mine sometimes) and has 0 performance impact from what i can tell

Left image is without FreeU, right image is with it

another example, first is normal gen, second is with FreeU at 'reasonable' values, third is cranked up saturation/detail

seems like it also promotes a little better prompt adherence (her hair colors) as part of the added detail

[brisk valley]

Could I know what this model is?

[vestal crypt]

NAI V3
you have to pay to use it

[brisk valley]

I hoped it might be AidXL

welp too bad

thanks

[brittle rampart]

colorful!👍

[nova remnant]

spiked hair girl

[native halo]

here u go,more spikes

[twilit portal]

🐱

[steady grail]

ohy guys

[nova remnant]

From behind, from below

@lyric zealot

[native halo]

(bodysuit:2.5)

[nova remnant]

You should stop with the dirty thoughts and read more books!

[native halo]

thats a big book

[nova remnant]

[nova remnant]

you can call me chair

👍

[vital raptor]

I also like books

[gloomy scroll]

🍑

[nova remnant]

@native halo she has great personality

[nova remnant]

Small personality

[nova remnant]

Meow cows!

[vital raptor]

apparently this farm has a lot of cats, go figure

[patent shard]

how four ear help them to survive

it rather become hard to survive

[nova remnant]

One for higher frequencies, other for low

[native halo]

shush

[gloomy scroll]

^Time when Nai didn't want to follow prompt

[gloomy scroll]

Pajama party?

[lyric zealot]

it's my super old gen

[gloomy scroll]

So you want old gen?

[lyric zealot]

time for random old gens

[brittle rampart]

oki -doki. old gen

[craggy turret]

Alright

[craggy turret]

How the fuck

Do I remove the light

[brittle rampart]

light?

[craggy turret]

Light

These damn random rimlights keep plaguing the results I'm getting from this model

That and it keeps cell shading it

[brittle rampart]

correct in Inpaint

[craggy turret]

💀

So what about the cell-shading then

I've been trying to replicate the style of this result from an older model of mine, but the cell-shading keeps getting in the way

[brittle rampart]

I don't fully understand what exactly you're talking about

[craggy turret]

Pretty much just the straight dark shading

[brittle rampart]

obscure light and shadow,
soft shadow,
chromatic aberration

try

[craggy turret]

One sec,

[vital raptor]

Can also try putting cell shading in the negatives if you haven’t already

[craggy turret]

Obscure light and shadow in negative or positive?

That damn

l i g h t

[brittle rampart]

all in positive

[craggy turret]

Hm

[native halo]

light position is random,only way to make the room darker is with a dark lora or get lucky with a prompt like (dark room:1.5) in pos and (sunlight:1.5) in neg but its rng or with cdtuner extension just increase contrast and boom

[brittle rampart]

The question is not about lighting the room, but as I understand it, about the transition between light and shadow.

[craggy turret]

Half of it is the shading

The other half is the random ass light coming from nowhere

[brittle rampart]

"Ambient occlusion" in prompt will further soften sharp transitions between light and shadow

[native halo]

it depends on the model too,some models tend to create imgs with light coming from either left or right side

[brittle rampart]

the occlusion effect does not depend on the direction of the light.

[craggy turret]

l i g h t

[brittle rampart]

Light?

[craggy turret]

The random light

Still present here, although the shading certainly is better

[brittle rampart]

may i ask your model and prompt for try ?

[craggy turret]

Animage

[brittle rampart]

this is SDXL?

[craggy turret]

Yeah

[brittle rampart]

https://civitai.com/models/260267/animagine-xl-v3

[craggy turret]

Yeah that's the one

[brittle rampart]

oki-doki. try it. what about prompt?

[native halo]

all of the sample images from that model have a light present from either left or right side

so model issue

[brittle rampart]

but this model initially has pronounced spots of light and shadow

[native halo]

model issue

[brittle rampart]

up lighh source

[native halo]

yes,that one has light from above

[brittle rampart]

not one

[native halo]

yea that one has light from behind and from above

[brittle rampart]

I’m certainly not sure that I can work quickly and well with the SDXL model🥴

[native halo]

😔

[craggy turret]

So cooked

[brittle rampart]

but still, what exactly did you want from the position of the light source?

[craggy turret]

I just wanted it gone

[brittle rampart]

prompt?

[craggy turret]

best quality, masterpiece, aesthetic, highly detailed, professional, detailed eyes,

1girl, petite, long ashen-silver hair, hair between eyes, red eyes, black oversized t-shirt, black and red earrings, pale skin, gothic interior, livingroom, standing, night time, moonlight, dark:1.5, caustics:1.1, dimly lit, gloomy background,

HDR, character profile picture, upper body, highres, looking at viewer, depth of field,
obscure light and shadow, soft shadow, ambient occlusion
<lora:add_detail:0.8>

Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, oversaturated colors, bad composition, too bright, border, cell-shaded, random lights, poorly drawn hands, outline, comic:1.5, rim light, chromatic aberration, sunlight:1.5, (worst quality:1.3), minimalist,

[nova remnant]

It's actually hard to remove light

You'll need something more than txt2img

[craggy turret]

I want the light either removed, or just toned down

To not be a solid white line

[native halo]

then use a diff model cuz that one has the light present

[nova remnant]

Do you need quality prompt for that model?

[craggy turret]

I've just been harvesting random quality prompts from the samples to test them out

[vital raptor]

Make your own prompt from nothing. Only add things if they’re needed

[craggy turret]

I initially did,

[vital raptor]

Any one of those words could be heavily associated with lighting and cause the problem

[craggy turret]

Then wanted to try to get rid of the solid white lines

So I started picking up some of the prompts from the samples that had less of the white lines and cell-shading to see if they'd make a difference

(They didn't)

[vital raptor]

Also with XL almost none of the negatives are doing anything and nothing about quality really matters

[craggy turret]

Noted

I've been using my 1.5 ultra hybrid model for so long that I never got onto XL until now

[vital raptor]

Yeah, I just into it a couple months ago since I had a nice custom 1.5 model. Finally broke down and made an XL one

I wonder if the images my model makes are closer to what you want. I haven’t decided to release it yet but I can tell you what’s in it for maybe a model that works better lighting wise

[craggy turret]

Yeah I've just been trying to replicate the style of my old model

[vital raptor]

I’m not at home right now to post an example but if you search this channel for images posted by me they’re almost all from my model or a model I used in it

Oddly enough, I’ve been trying to figure out how to force my model to do more hard flat cell shading and it never works 😆

[brittle rampart]

and without any lamp

and light source up

[nova remnant]

Fastest blindfold runner

[brittle rampart]

peekaboo!

[nova remnant]

Put a blindfold on her

[native halo]

thats not a bodysuit

[brittle rampart]

say this is @craggy turret 😁

[nova remnant]

I think that's enough SD for today

[native halo]

just one more bodysuit 🗿

[craggy turret]

Damn, all that's left now is that solid white block of light on her head

[nova remnant]

Paint it black and img2img :p

[craggy turret]

Yeah that's probably the best way to fix it

[nova remnant]

Aren't there dark Loras?

[brittle rampart]

(illustrattion of a cute preteen girl standing at gothic interior livingroom:1.25), petite, long ashen-silver hair, hair between eyes, (glowing red eyes), black oversized t-shirt, black and red earrings, pale skin, ((night time)), dimly lit, gloomy background, occlusion moonlight,creeping fog on the floor, mist,

[nova remnant]

There is for 1.5 idk xl

[brittle rampart]

(dark horror gothic theme:1.3)

[native halo]

cuz xl doesnt need really need it but that model sucks so it probably needs one

[brittle rampart]

@craggy turret try to play with this in SDXL

i am really cant do it faster on my 6G Vram 😭