🍥｜anime

[brittle rampart]

multiple view and...?

[bitter latch]

actually no

it just did it

[brittle rampart]

well damn

[bitter latch]

haha

thats what I was going for. not very detailed tho

mirror mirror on the wall

turns random meter up

😄

[supple raptor]

oh nyo! what happened to hatate's wings?

[brittle rampart]

lost....

[supple raptor]

marisad~1

[gloomy scroll]

[brittle rampart]

https://tenor.com/view/chuunibyou-dekomori-sorry-please-regret-gif-8229773

[oak spoke]

howdy how's everyone?

[bitter latch]

oh getting my groceries is gonna be fun... no.. it's raining like a mofo

[fluid torrent]

Just uploaded my new merged sd15 model to civit. It's not cute lunar anymore, it's Sexy Lunar!

https://civitai.com/models/637083

[bitter latch]

I got pocky

https://tenor.com/1F8v.gif

[bitter latch]

hmm good pokkie

[bitter latch]

hmm almost two diff angles of hte same location

then again it is suppose to be ueno park

[green temple]

boice

noice *

[green temple]

can i use artist tags without turning on hires upscale ? because i tried to use it without it and it doesn`t seem to work

[bitter latch]

yeah, I do all the time. I dont' u se hi-res much

then again my faces come out wierd so.. shrugs

[green temple]

oh , ok

[bitter latch]

but doesn't have to be perfect to be good art...

[bitter latch]

listens to the outfield your love and can't help think of ed ed and eddy singing it

same guy I think

[bitter latch]

turned on the new addetailer that does more...

turned lily pads on right side into umbrella's

in water

[green temple]

ooohh

[green temple]

🌚

[lyric zealot]

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981

might be time to update soon

[vital raptor]

i havent looked at forge since they announced it was going off the rails, but as far as I understand it's missing a lot of 'normal' features you'd be used to

at this point it's just an experimental platform for features before they get put into a1111

[bitter latch]

yeah it's borked

[lyric zealot]

yup, i guess it's because they are changing all the basic stuff, that's why most things broke

[wooden coral]

There is this other forge fork I forgot the name off, you can update to that

[lyric zealot]

reforge

[vital raptor]

i'm kind of curious how swarmui compares with forge/reforge for performance on smaller gpus

I'm pretty sure a lot of the forge improvements were very similar to how comfy handles things, and swarm uses comfy for a backend by default, so it might be pretty similar

[nova remnant]

foooooooocus seems to be really good for smol gpu

[vital raptor]

yeah, i heard that too

I think swarm is a better option in the long run if you don't want to mess with comfy though because it gets all the same updates as comfy for newer model types and such

the ui needs some work, but it works pretty good

[lyric zealot]

there's this new thing in forge where you can get even faster speed for smol vram gpu

Using this option, you can even try SDXL in nf4 and see what will happen - in my case SDXL now really works like SD1.5 fast and images are spilling out!

[vital raptor]

yeah, 4bit quantization stuff

I havent really read into it much, but the problem is that you're basically reducing the noise 'quality' by doing that, like the model is made in 16bit, then it gets halved to 8bit and then halved again to 4bit

I haven't taken the time to actually compare the differences but my understanding is that it reduces overall detail

could be a reasonable trade off for speed if you like the results though

[lyric zealot]

probably not for you, since you already have decent/fast speed

only really applies to people with smol vram i guess

[vital raptor]

i mean yeah, i don't have a need for it

just wondering how much it reduces quality in exchange for speed

[lyric zealot]

interestingly

he said that in theory, NF4 will outperforms FP8

due to some stuff

i don't really understand, or could bother understanding

[vital raptor]

it's possible that the details it loses are different, so it appears aesthetically better maybe

[lyric zealot]

well if you wanna read, it's linked #🍥｜anime message

he explained it under Flux checkpoints

[vital raptor]

yeah, it's basically doing the math in a different way to leverage different parts of the hardware, so it ends up being (mostly) more efficient, the different maths used would explain the difference in results, although I wonder what 'outperform' means

they do say that the speed-ups are completely random and need more testing

it'll be interesting to see how flux develops over the next few months or so though, between performance improvements and some loras and finetunes, maybe we'll get a decent model

[junior quest]

@bitter latch got the old forge github link?

[vital raptor]

I like how the 'canons' are camera lenses 😄

[lyric zealot]

yup yup, i think there's already a few lora for them

[unkempt cape]

omfg

gpt 4o image is out 🥶🥶🥶🥶🥶🥶🥶

[lyric zealot]

[unkempt cape]

can any one go to the website and test if img 2 img possible?

[lyric zealot]

is that a paid thing or?

[unkempt cape]

free

like try to see if it can copy styles or someth

[lyric zealot]

well, i need to wait for limit to be lifted, used too much when debugging stuff

[bitter latch]

tried to make cortana from halo. didn't work but .... shrugs

[supple raptor]

for large scale finetuning it's usually done at fp32 afaik. we then half precision to fp16 for standard inference since inference precision is less important than training precision.

[brittle rampart]

new level!

[stable crane]

@zinc flax thanks for the lora ❤️

[zinc flax]

O is that lk149

dang that's clean

[stable crane]

it is

yeah it turned out very clean lol

[zinc flax]

I haven't mixed it with other loras but that semi thick lines seems to transfer well lol

Looks great

[stable crane]

oh I had some good mixes too

[stable crane]

this is part lk149

[zinc flax]

Neat

[stable crane]

oh and those are mixed with another artist

[zinc flax]

Super clean

Might upload some other styles soontm

[nova remnant]

very big

very clean

[zinc flax]

Glad you liked it lol

[stable crane]

I'll use it a lot

[formal shoal]

Whats the lora?

[formal shoal]

🍋

[stable crane]

lk149

[nova remnant]

@lyric zealot is there a way to do the vibe transfer thing on a1111

i was trying to make a lora to mimic an artstyle but its going horribly so im thinking i could use vibe transfer

[lyric zealot]

Ip adapter is the closest thing to vibe transfer

So try that

[nova remnant]

ip adapter??

whats that, how do i use it

[lyric zealot]

Yup, it’s likely under controlnet if you have that

You will need to get the model, I recommend the vit-h one

I’m not at pc right now, so can’t give more info

[nova remnant]

oof

controlnet

isnt that hard to install

i always have trouble

[keen hollow]

accidentally deleted a bunch of models from my drive today... whoops

[lyric zealot]

I think for auto1111, it’s just an extension, but you might need to download additional stuff for it to work, like the model

[nova remnant]

i tried to install it like i usually do but it gave me an error

[lyric zealot]

It appears that your stuff didn’t fully download

[nova remnant]

fuuccckkkk

i keep forgetting my internet is ass

i dont think ill be able to install controlnet

[lyric zealot]

Maybe a better way, is to download the controlnet yourself, then once that downloaded

Put it in your extension file

[nova remnant]

i shall try this

[lyric zealot]

or even try using method that download by chunks, so if your internet dies, you still can continue when you have internet

like example torrent

[nova remnant]

torrents always take like 4 days with my internet

[nova remnant]

what do i do if it doesnt show up

i downloaded it

extracted it

its not there

[keen hollow]

flux in Forge now... woo

[bitter latch]

but is the new forge actually working?

[keen hollow]

works for me

just did a clean install to test this out

(and keep my older forge install untouched)

[bitter latch]

hmm good plan

but mine works. so i'm leaving it alone

welcome back to dog poker show

hhaha

woof woof

to many arms....

[nova remnant]

@lyric zealot how do i install controlnet models/ install this https://github.com/TencentARC/T2I-Adapter

[bitter latch]

ah ha!

[nova remnant]

i cant even get controlnet to actually show up tbh

it shows up here

its not here though

[lyric zealot]

here is where?

[nova remnant]

extensions tab

[bitter latch]

I don't feel so good mr stark

😄

https://tenor.com/uwIlHbDI8dZ.gif

.. bored

puts in miku hatsune in Wildcard background, random theme, some styles

ooooh "Flat papercut style"

ooh I gotta say I like this style

gets serious.. hmmm ponders theme

[nova remnant]

so i downloaded two versions, on the first version whenever i manually install it a1111 pretends it doesnt exist, and the second one just makes a1111 break and it wont start up

[nova remnant]

i cant figure this out 😭

[lyric zealot]

[nova remnant]

[lyric zealot]

https://github.com/Mikubill/sd-webui-controlnet this might be the one for a1111

[nova remnant]

yeah thats the second one

it just makes it to where the python stuff freezes

it wont start up

it just stops at a certain point and no a1111 no matter how long i wait

[lyric zealot]

what's the error?

[nova remnant]

it doesnt even show an error

it just

starts to start up and then it stops

[bitter latch]

hmmm......

there we go

not sure why her dress is melting

wow this adetailer makes gens take way longer

[bitter latch]

hmm no

blah, arms all screwed up

[lyric zealot]

[bitter latch]

goes to bed

[nova remnant]

okay okay i think i got it, how do i do this ip adapter thing

can you walk me through it

[bitter latch]

https://letmegooglethat.com/?q=how+do+I+use+Ip+adapter+for+stable+diffusion+

😄

[nova remnant]

i deadass looked it up and understand nothing

[supple raptor]

1. download the ip-adapter models you want some are available here
2. place them in .\extensions\sd-webui-controlnet\models
3. click the blue refresh button to the left of the model dropdown menu in the controlnet section of your webui
4. set Control Type to IP-Adapter and make sure the correct model is selected in the Model dropdown.

[junior quest]

@bitter latch got the old forge link?

[nova remnant]

bet

[supple raptor]

old forge? like a previous commit?

[junior quest]

commit doesnt work if you install current one right?

[supple raptor]

forge repo is here, and you can browse the commits to find the point in it's history that you want https://github.com/lllyasviel/stable-diffusion-webui-forge

[nova remnant]

like this??

[junior quest]

this one works? I updated and messed up

errors

[supple raptor]

yep, then you drop your image in the field a the top of the screenshot. you'll have to experiment to figure out the best settings for those 3 sliders you see near the bottom since i have very little experience actually using ip-adapter.

[nova remnant]

fuck yeeaaahhh

thanks nubby

i can do like a sort of style transfer like this right?

i was trying to make a style lora but its uhh

not going great

i heard you can do a sort of vibe transfer in a1111 using controlnet and ip adapter

[lyric zealot]

#🍥｜anime message

^ this is my settings

[supple raptor]

github repos utilize version control, which means once you clone the repo you can simply switch to any previous version or "commit" simply by running the command git checkout [commit hash] (replace "[commit hash]" with the hash of the commit you want to switch to. the hash can be found by clicking the button circled in the 1st screenshot and the hashes are shown in the circled column in the 2nd screenshot.

[junior quest]

yoshh thanks, will try

[nova remnant]

hmmm i dont have those models, let me look for them real quick

[lyric zealot]

well, if what you have works for you, then why not

[junior quest]

what kind of style?

[nova remnant]

the style from this manga/anime

ive cleaned so many of these images of text

just to make a lora

but its meh

[supple raptor]

you want the .safetensors versions here https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models ignore the .bin versions.

[nova remnant]

downloading them rn, thanks 😭

[zinc flax]

Maybe you could try using datasets from the anime instead?

They have similar style

[nova remnant]

i actually have some images from the anime in my dataset

[nova remnant]

these

have a mixed bag of images from the anime, like official promo art and bluray covers

and images from the manga, specifically colored manga art, taken from manga reading sites, zerochan, or ones that i found just on google and cleaned up

this is how the latest version of the lora came out its meh

[junior quest]

why not try to create a lora

[zinc flax]

Pain

[junior quest]

ohh

[nova remnant]

frrrrr

this was like my third try

ive tried using only like 20 of the best quality images

[zinc flax]

Time to try another setting lol

[nova remnant]

ive tried using everything i had

ive tried just trying to prompt the style, and just lettting auto tag do its thing

[zinc flax]

Gonna try training it for practice

[nova remnant]

cant find this one

[lyric zealot]

that's the preprocessor, comes with the controlnet

im using some outdated forge with built-in controlnet

[nova remnant]

i found something that looks similar enough

[nova remnant]

this image came out better

[nova remnant]

Misewy

[keen holly]

hello there, i am pretty new to stable diffusion (i use sd1.5 with A1111) and i wanted to create a Lora for my OCs, so I am doing it simply by generating my character with different poses and i plan to then train it on civit ai. But I have two main questions : 1. Do I have to put some psychological traits of my OC into the prompts for the dataset or is the physique description enough 2. what can i do if i want to create a secondary outfit/a transformation ? Thanks in advance

[nova remnant]

[nova remnant]

steals her cookie

[nova remnant]

(where is my pizza outfit)

stupid model

[bitter latch]

and now for something different

sci-fi and fantasy

[nova remnant]

caatt

[steady pumice]

shes a cute

[bitter latch]

there thats good

wow this new merge works good

haha I had t-shite instead of t-shirt

no styles prompt: Detailed Face, Chibi slim petite thin Girl , blue braided curly hair, camisole T-shirt and simple skirt, black background

haha instead of making her Miku, it put it on the tshirt , kinda

ok going back to bed. thud

[nova remnant]

@lyric zealot look

That is some Smuggered slime

[lyric zealot]

that slime is totally you

[nova remnant]

I'm innocent

[lyric zealot]

[lyric zealot]

super insane prompt of "somt"

[brittle rampart]

super versatile!

[lyric zealot]

i see you playing with your metadata

⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣴⣂⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣠⣾⢇⠀⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⠔⢉⠾⢜⢆⠀⠀⠀⠀⠀⠀⠀⠀⠀⡠⠊⠰⢅⣌⡄⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡰⠁⠀⡎⠀⠀⢹⡄⠒⠒⠒⠒⠒⠒⠒⠮⠄⠀⡇⠀⠘⣧⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⠎⠀⠀⠀⢁⡀⣀⠔⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠑⠢⣸⠀⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⢠⢔⠏⠀⠀⠀⠀⠁⠁⢀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠑⢄⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⢤⠞⡣⠋⠆⠀⠀⠀⠀⠀⡄⢰⠀⢗⠄⠀⠀⠀⠀⠀⠀⠀⠄⠀⠀⠈⢂⠀⠀⠀⠀⠱⡀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⢀⡴⠋⠀⣐⣃⠀⡄⠀⠀⠀⠀⢰⠘⠈⡆⠘⡈⢢⠀⠙⡄⠀⠀⠀⠈⡎⡄⠀⠈⡆⢄⠀⠀⢠⢹⡄⠀
⠀⠀⠀⠀⠀⠀⢠⡞⠁⠀⡘⢆⡀⢰⠅⠤⠤⢄⠀⡌⡀⠀⣷⠀⡇⠀⢣⠀⢳⠀⢀⡀⣀⠙⣡⢀⠔⠻⡸⡆⢨⠉⠐⡾⡀
⠀⠀⠀⠀⠀⢠⡟⠀⠀⢰⠁⠀⠀⡆⠐⠁⠐⠊⢠⣧⣦⣳⢸⢀⡇⠀⠀⡖⣼⡆⠘⠘⠐⡄⣭⠊⢀⡆⢁⢹⠀⠀⠀⢸⢃
⠀⠀⠀⠀⠀⣎⠇⠀⠀⡎⠀⠀⢸⠁⠀⠀⠀⠀⢰⠁⣠⠏⣸⣾⣤⣀⢀⢀⡼⣧⠞⠟⢿⣷⣿⣿⡿⣦⣼⡎⠀⠀⠀⣸⡾
⠀⠀⠀⠀⠀⢻⠀⠀⢀⠃⠀⠀⡌⠀⠀⠀⠀⠀⢸⠔⣡⠖⢛⣉⡙⠻⠊⠈⠀⠀⠀⠀⠘⣹⣈⣿⣷⣄⠙⣇⠀⠆⢀⠇⠇
⠀⠀⠀⠀⠸⢸⠀⠀⢸⠀⠀⠀⠇⠀⠀⠀⠀⠀⢸⡞⠀⣄⣹⣿⣿⣷⡄⠀⠀⠀⠀⠀⠀⣿⣿⣿⣿⣿⠄⢹⡀⠘⢳⡀⠀
⠀⠀⠀⠀⠀⠯⡆⠀⢸⠀⠀⢰⠀⠀⠀⠀⠀⠀⢸⡄⠀⣿⣿⣿⣿⣿⡇⠀⠀⠀⠀⠀⠀⠹⢿⣿⣿⠟⠀⠁⠳⡀⠀⢣⠀
⠀⠀⠀⠀⠀⠘⣷⠀⠘⡀⠀⢸⠀⠀⠀⠀⠀⠀⢸⠑⠄⠘⠿⡿⠿⠛⠀⠀⠀⠀⠀⠀⠉⠀⠀⠀⠀⠔⠎⢄⢢⢳⠀⠈⡆
⠀⠀⠀⠀⠀⠀⠈⠣⡄⣇⠀⠸⠀⠀⠀⠀⠀⠀⠘⡇⢂⠢⢋⠣⣢⠀⢄⠀⣀⣀⣤⣤⢤⣤⡤⡔⠈⠎⠬⠦⠦⠂⡇⠀⢰
⠀⠀⠀⠀⠀⠀⠀⠀⠈⠸⡄⠀⡆⢀⠀⠀⠀⠀⠀⣧⠑⠁⠑⠑⠁⢀⣿⠿⠿⠙⠚⠭⠿⣾⣽⣻⠀⠀⠀⠀⠀⡰⠀⠀⢸
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡧⠀⢃⠸⠀⠀⠀⠀⠀⢡⠆⠀⠀⠀⠀⢸⠁⠀⠀⠀⠀⠀⠀⠀⠈⠿⠀⠀⠀⡠⠎⠀⠀⠀⢸
⠀⡀⠀⢀⠠⡤⠤⡤⣄⠀⡇⠀⠘⡀⠆⠀⠀⠀⠀⠘⡜⣄⠀⠀⠀⠀⠱⢄⣀⠀⣀⣀⠠⠤⠂⣁⢠⢔⠈⠀⠀⡄⢠⠉⣾
⠀⡿⠊⠐⠁⠀⡜⠁⠀⠀⡧⢀⣀⢱⠘⡀⠀⣠⢀⡎⢫⣮⣻⣶⣶⡄⠀⠀⠀⢀⢠⣴⣶⣮⡉⠀⢸⢸⠀⠀⠁⠉⠁⣰⡌
⡜⠀⠅⠂⠂⢾⠀⠀⠀⢀⠟⠸⠘⠀⢧⠣⠀⠊⠈⠀⠙⣿⣿⣿⣿⣿⣤⢀⡀⣀⣿⣿⣿⣿⡟⢆⣌⠌⡀⠀⠀⠀⡰⣹⠁
⡇⠀⠀⠀⠀⢸⡄⠀⠀⡸⠀⠀⢰⠀⠀⠣⡕⠀⠀⠀⠀⠸⣿⣿⣿⣿⣿⣯⣉⣼⣿⣿⣿⣿⣇⢌⠣⡰⠀⠀⡠⠊⠀⠁⠀

[brittle rampart]

Ouch! Why did it stretch her out so much!? but it's funny.... 😁

[lyric zealot]

looks perfectly fine on my screen

[brittle rampart]

yes.... cute cat!

[nova remnant]

Don't use me to react

@lyric zealot

[gentle pelican]

hi

[lyric zealot]

[brittle rampart]

🖖

[zinc flax]

cuute

[nova remnant]

eep

[bitter latch]

umm.. ok. then.. turns down clipskip..

incoming !

haha

[green temple]

cool

[nova remnant]

Grabby tail

[bitter latch]

odd hand placement(?)

haha

BIG head

[bitter latch]

ummm. wtf

nom nom

hmm... o.. k

[bitter latch]

shaggy and scooby as anime

[bitter latch]

haha

[fluid torrent]

werid mahsup

[bitter latch]

thats better. Shaggy as a bunny

or at least a fursuiter

adds "cthulhu mythos by H.R. Giger"

hmm five nights at Freddies?

[steady grail]

cute, reminds me of taiga

[bitter latch]

"i'm awake.. dammit"

[bitter latch]

CHARGE!

character sheet

there we go

[nova remnant]

No cookies for you, go away

[lyric zealot]

[bitter latch]

grumpy

haha

[lyric zealot]

hugeeeee

[zinc flax]

Massive, even

[bitter latch]

donuts

love this effect. not sure how it happens tho

goes back to bed.. thud

sleeps

Zzzzzzz

[green temple]

niicee

[bitter latch]

installed new forge on new directory, looks good so far.. so far.. but new stuff to learn

will have to play with after work. sigh

[junior quest]

how? kohya ss got flux support? or cmfy lora training?

[bitter latch]

meh 15 still can't do text well. oh well

break over back to work

[bitter latch]

oh this thing comes with alot of styles. nice

tho I may just say with curretnt version i'm used to that works the way I want it to

[lyric zealot]

What do you mean 15?

15 steps?

https://media.discordapp.net/attachments/1270939143626686506/1270939144335659030/eliv_eating_plate.gif

[bitter latch]

SD1.5

[lyric zealot]

Ahhh, I think sd1.5 will almost never be able to truly do text

[bitter latch]

sad

[lyric zealot]

Correct me if I’m wrong tho

I’m just saying stuff without knowledge

[lyric zealot]

Have you tried flux on nf4?

[bitter latch]

nope

just got the new Forge installed in it's own Directory.

working from so I keep getting interupted. will have to check things afterwork. but not sure what the upside will be besides text tho.. even for xl. I get results I like with 1.5 so..shrugs

[lyric zealot]

I also got the new forge working yesterday. For flux, I think the major point of it is text and also having such high quality model on a based model

[bitter latch]

tho I do like the extra styles

but then again I like making my own so..shrugs

[lyric zealot]

I haven’t really explore those new styles on new forge, but looks sick

So copy and paste your old style.csv over

Not sure if it’s style or styles

That will not replace all those new styles that was integrated on new forge since it’s a different file

[green temple]

is there a new forge

[lyric zealot]

A new version of forge which most likely breaks a lot of stuff

I got it on a totally new place, to test it out

[green temple]

oh , so my 6 gb vram🥲 will explode

[fluid torrent]

preview for my landscapes and scenery series for civitai

[green temple]

wooow !

[bitter latch]

nice

[lyric zealot]

depends, cos there's new stuff like using nf4 kinds of model

[bitter latch]

i'm prob screwed using the new stuff with my 4gb

Need to get me another system

[green temple]

ok

[bitter latch]

but I dread reinstalling and moving stuff over..sigh

[lyric zealot]

try it then let us know

if you are interested, https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981

there he explains about nf4, but also on how to get your top speed (in theory) without using stuff like TensorRT

but there's a lot of work to be done and test to do so

[bitter latch]

back to using my old setup 😄

I copied over the styles from the other to my styles sheet 😄

[slate narwhal]

Yes, comfy lora training. Ostris trainer in this case.

[green temple]

[bitter latch]

using my own styles 😄

ooh these models don't like low step or cfg. eesh

or could be the clipskip 3

[bitter latch]

very nice, I should look at those

[fluid torrent]

its not posted yet, im building

any requests for landscapes or scenery to include ?!

[bitter latch]

i'm tired of fooling around, I need direction

[fluid torrent]

its all done with sexy lunar

[bitter latch]

I'm doing mine with a Lunar merge with two other models for 2.5 and a 3D one 😄

so yeah it gets confused

[fluid torrent]

like it, very dreamy abstract but semi-real-anime-ish ... or something 😄

[bitter latch]

thx

huh...

weird

[bitter latch]

trying out some of these new styles between stuff

wow

[bitter latch]

haha love it, added a ink style I had made with a lora .. interesting

[fluid torrent]

ill share this one here, too much focus on subject so it wont go into the landscape and scenery series

[bitter latch]

removes some of the photo prompt styles

oooh

[bitter latch]

there

[bitter latch]

now lets try that last one with a non anime model.... 😄 .. .kinda

[fluid torrent]

https://civitai.com/posts/5431330

👆 Landscapes and scenery series

[bitter latch]

thanks, will check it out

[fluid torrent]

(3d, real, photo:1.5) in the neg ---> back to anime

[bitter latch]

think the model i'm using isn't working right for what I want. giving weird stuff inside pic

interesting bridges

-adds in the futuristic -vaporwave style-

-falls over-

[bitter latch]

borrows "3d, real, photo:1.5) " and puts in neg

[bitter latch]

huh.. ok

[bitter latch]

gets rid of new forge and keeps old

I am loving htese new styles

[bitter latch]

almost time for work day to end woot woot

ponders a theme for a series of images. hmmmmmms

maybe?

been pondering making a comic strip or something.. ponder ponder

or make a visual novel using this. I have the software. several different one's actually

work done

falls over

[green temple]

oh... that was nagatoro 🌚

[bitter latch]

uses same everything but turns it sidways

[bitter latch]

woot

[bitter latch]

I can do that , kinda

[slate narwhal]

nice one. 💖

[bitter latch]

ah ha

goes rogue

haha oops

[slate narwhal]

Reminds me of my old SD 1.5 gens:

[bitter latch]

hahahah

[dapper verge]

i assume these are flux. are you using a particular lora?

[slate narwhal]

Yes, they are, and yes I'm using a Lora, softserve_anime

Link: https://huggingface.co/alvdansen/softserve_anime

[bitter latch]

Peter.. I dont' feel so good....

[dapper verge]

very cool, thanks. was looking around on civit

[slate narwhal]

I think they didn't post it there on civitai yet

[dapper verge]

yeah i'm not seeing it there under flux models

[slate narwhal]

Yeah, there's lots of good stuff on HF that are not on civit

[bitter latch]

ok more anime

ups the effect from 1.1 to 2

waits

yeah .. no

[dapper verge]

looks great.

[bitter latch]

orders dinner and breakfast (for tommorow)

[slate narwhal]

I really like it. And Flux adapts the Lora to the stuff in the base model after prompting, even if they seem weak in the base model, this didn't happen with Pony or even the base SDXL, the more I use Flux the more I like it

[bitter latch]

oh who am I kidding I'll probb eat it all sometime tonight

[dapper verge]

I'm liking that even if you throw danbooru tags at it, which flux would work with but it would look like really low score pony, this makes look good again.

what would probably look killer, is this with 1.5x upscale with a pony anime model.

with light denoise

[bitter latch]

hmmm bat-spider woman?

Bidar woman?

Spater woman?

haha

The Spitter

see's himself out

[bitter latch]

-goes on patrol outside gotham-

[slate narwhal]

I haven't found a good solution for upscaling Flux yet. I'l wait for a bit. I hope the whole model fine tuning people are talking happens, and then a Ponyflux or some other variant maybe becomes a possibility

[slate narwhal]

Very nice, yeah, it's a good idea for certain cases where the higher quality is not crucial

[dapper verge]

well that one was flux -> pony anime sdxl a couple times until 2.3k res from original 1344x768. upscaling flux with itself works well, but it's ~5 minutes just to double the res. it's just not worth it.

[bitter latch]

HALT!

[slate narwhal]

I will try this later. It really does make things better overall for these, the eyes would still need some correction though. I've seen upscales from Flux to SDXL that I didn't like, the quality loss is too much

[bitter latch]

hmm not doing well on face...

thats better

[bitter latch]

bat cat

[green temple]

oh.. nice

[terse ore]

I wonder how well flux will do with img2img+ controlnet

Usually SD has issues with straight lines, such as brick walls and stone tiles floor

[fluid torrent]

this sd15 model does look artistically nice, but indeed the tiling doesnt make sense

source image for img2img

reflections also could need a fix

g'night chat!

[terse ore]

Good research, Ive been trying to find a decent img2img workflow and although there are some decent outputs, inpainting each component of the image at high denoise is always better than denoising the image entirely. But, it takes so much longer

[bitter latch]

hmmm shrugs

enough of that

blah

[stable crane]

very sharp

[bitter latch]

it tries.. but..

[bitter latch]

some of these xl models are huge, 14 gigs.. really?

xl

hmmm handles buildings better

sigh...

haha

[stable crane]

lol

[bitter latch]

tries a diff model, vae and adjusts prompt

ooops

huh htf do you get the words?

scratches head

[stable crane]

flux

[bitter latch]

meh

[bitter latch]

ok founds some lorea's

lora's

hmm looks .. meh

huh.. ok... I got a werfault and well that wasn't good, things started crashing

so no more xl on this toaster

[bitter latch]

much better, predictable results

love this style

ah ha

much better

tweaked the style

hmm ok

blinks

[bitter latch]

donut?

donut fashion

[nova remnant]

been mia for a while here's a recent

[long berry]

genshin impact lenny running for president

[nova remnant]

steals her halo

[nova remnant]

[nova remnant]

[dapper verge]

digging the new fcs_style_v3.1 lora for flux on civit. 0.75 strength.

[stable crane]

cute!

post more

[stable crane]

oo

that last one is really nice

[stable crane]

the style is pretty legit for anime

[dapper verge]

yeah i've been refining pixart with pony-anime models, but this is by far the cleanest and most prompt following of them all.

[stable crane]

and it's pure flux+lora?

[dapper verge]

yeah

[stable crane]

where'd you find it?

[dapper verge]

https://civitai.com/models/180891/flat-color-anime-v31?modelVersionId=724241

[stable crane]

oh I couldn't find that

ty

[rough oxide]

hmmm so are the loras coming out now good for flux? i find flux can already do anime-style stuff super well.

the stuff in the flat color lora seems like stuff that should just be possible on its own unless i'm mistaken

[glacial coral]

well unless it makes a good impact i dont think so i generated the image above with my macbook and it does really well

i only see rn the realism lora being a game changer for some

[rough oxide]

this is without the flat color lora

similar prompt they used for the main image they show

still totally nails anime vibe without any lora so idk what the flat color lora really adds in this case

[glacial coral]

Yh

[rough oxide]

eventually as always finetunes/loras will be good but i'm just not seeing any good ones like 'right now' i guess

[glacial coral]

Exactly my case rn using draw things application on mac instead of ComfyUI or Forge only thing I use is realism lora if I'm going photorealistic images and other than that it's just prompting and stuff

[rough oxide]

ye. prompting can get you nearly anything you need in flux except like hyper-specific character traits which a potential flux.weeb would solve (in terms of anime)

[glacial coral]

im very happy with text

thats what i can say also

[bitter latch]

no lora's

well it almost said flux

[glacial coral]

U running it off 4gb of vram?

[bitter latch]

not flux no

[glacial coral]

Was gonna say

[rough oxide]

it's been quantized down a lot

but i prefer full-scale always

[glacial coral]

Still 4gb vram on flux is a bit yk

[bitter latch]

haven't even tried it

[glacial coral]

I mean I'm saying this like I'm not running it off a m2 pro macbook pro 😭

[bitter latch]

was trying xl yesterday but got a error of werfdafault and everyhing started crashing

so.. nope

[glacial coral]

I think the draw things application I'm using does all the heavy work for mac optimisation but I'm surprised I've got flux running on a 16gb ram model of my m2 pro

[bitter latch]

hmm I like it

[glacial coral]

It's great

Flux just takes like forever tho to genrate a image but I'm fine with it

[bitter latch]

I had added " art by Flux Stable Diffusion" at end of that.. takes out and tries again

tho dont' think it'll make a diff.. shrugs

[glacial coral]

For me apart from it mainly just not adding any text I can't really complain about text

[bitter latch]

(medium view), white braided hair, chibi petite 1girl, little horns, wings, hoodie, smirking, head tilt, simple background, 2.5 flat, elegantly detailed, pastel chalk painted fine art, soft colors, soft shading, realistic, photo, movie, soft edges, multi-shaded colors, anime overtones, line art style

[glacial coral]

It's the best model I've seen when it comes to generating text

[rough oxide]

it'll mess up if u want like a whole ass paragraph

and some sentences might miss a word

[glacial coral]

What UI u using

[bitter latch]

Forge, old version

[rough oxide]

for some reaosn i could never get it to output the word "better" in one prompt i used

[glacial coral]

I mean yh I expect Flux Pro to do that correctly

[rough oxide]

i used api of flux pro for a loooong time

similar issue ish

flux pro isn't that big of an upgrade over dev really

[glacial coral]

Ah i see well it's new so I expect it to massively improve

[rough oxide]

a lot of the times i prefer dev outputs even

but there's some things it does a lot better on

[glacial coral]

How much do I yet agree with "if midjourney was free" statement I'm not sure as I use MJ quite often

[bitter latch]

now .. adds lora's for epi, dpo, more detail and mdjourny4

[glacial coral]

Ngl I do use shi ton of MJ cuz I enjoy making quick generations

Cuz getting this quickly saves me time iwl

[rough oxide]

midjourney is worse than flux imo. flux really took everything by storm

[bitter latch]

someone made a lora that suppose to mimic midjourney but shrugs

it does at more detail tho

[rough oxide]

now if NAI was foss....

[glacial coral]

Depends I do prefer MJ more if we excluding the price and it being censored but that maybe me bias

Tho both have advantages and disadvantages

[rough oxide]

i'd have a lot of fun messing around with that even though i think as a base model flux is better with general anime images that don't need hyper-specifics

[glacial coral]

So pick ur poison

[bitter latch]

personally I like to experiment and just go random

[rough oxide]

flux wins in elo too though. aside from trying to get like hyper-specific painting styles or something

it does like everything-ish better afaik

[glacial coral]

True but can create caused asf images sometimes for zero reason

[bitter latch]

love these new styles from the new forge I .. borrowed. 😄

[glacial coral]

MJ also does that tbh but it's not as bad icl from what I've seen but I will also say I do like the new photo editor they just recently added to MJ

[glacial coral]

That's actually really nice

Part of me forgets Leonardo.ai exists

Even tho I enjoyed that tool also

[bitter latch]

adds in the style psychedelic.. ooh trippy

[dapper verge]

did you see mine from above? that new flat color anime one is rather good. yes flux can do anime, but I've found it's not reliable. about 40-50% of the time you ask for anime and it still gives you a photo. with this lora in place it's far more reliable. here's some with it: #🍥｜anime message

[glacial coral]

Iirc Canva bought leonardo.ai or sum and integrated it

[glacial coral]

Yh that's nice

[rough oxide]

i've literally never had that problem...

but i usually use more than just "anime" once randomly in the prompt

something like "anime episode screenshot" is really powerful

my outputs are always anime like 100% of the time

[dapper verge]

I've had good luck with anime still. anime screencap is also good, although more geared towards retro anime.

[rough oxide]

was getting ready to post these on 4chan to reply to sum1

[bitter latch]

huh ok was not expecting that

[rough oxide]

but like i have never had it not do anime properly

also instead of booru terms like "anime screencap"

screenshot works better i believe for these kinds of models

at least from my exp

[bitter latch]

I find ---- Anime Retro studio , ---- works too

[dapper verge]

So look at this. on the left is anime still in front of the prompt. basically her face is anime, the rest isn't. on the right is with this lora with my settings, it's all anime. face, body, background, everything.

[rough oxide]

what's ur prompt?

i think it has to do with your prompt

[dapper verge]

Futuristic robot woman, illuminated by soft blue-green light, with intricate metal and glowing wires. Long, flowing black hair with tiny lights woven in, creating a magical effect. Holds a round, transparent orb filled with vibrant, colorful underwater plants and small glowing fish, nested in her chest. Surrounded by dark, underwater-like environment with gentle light particles floating around. Background has mysterious, shadowy shapes resembling ancient coral reefs. Soft glow highlights her smooth, metallic robotic parts. The scene feels serene, enchanting, and otherworldly.

[rough oxide]

it doesn't even mention anime at allll

lol

[dapper verge]

i was putting anime still in front of it for the one on the left.

i have a concat box in front of it.

which either has anime still, or the trigger words for the lora.

so, see if you can work your magic for the one on the left to be all anime without the lora.

[bitter latch]

goes back to basics

[rough oxide]

i mean part of it is your overall prompting style

also 1792x1024 is such an odd resolution

takes even longer than 1920x1088 to gen

xD

[dapper verge]

shrug, it all works great for me.

as you can see from the gens i pasted up above.

[rough oxide]

doesn't look anime at all

i mean

doesn't look

not anime

at all

xD

it's ur prompting style

[dapper verge]

..... so eh.... you gonna paste the resulting prompt or no?

[rough oxide]

literally just

"anime episode screenshot"

at the front

ok public param posting bork

but literally just

[rough oxide]

but u have to use more powerful words than just "anime"

anime is too general

[dapper verge]

if all you're doing is adding 3 words in front, there's nothing wrong with my prompting style. I just have to put the 3 words in front.

[rough oxide]

it's not just that, the way it's worded, although i used that kinda prompt for this

cuz i used joycaption for that to prove a point

xD

[dapper verge]

the backgrounds on these water ones aren't anime. only the foreground is.

[rough oxide]

which flux model r u using

[dapper verge]

dev fp16

[rough oxide]

the backgrounds seem plenty anime to me

[dapper verge]

not anime.

this is anime

[rough oxide]

i dun have a problem with top one at all. i've seen anime use backdrops like that plenty

[dapper verge]

i'll grab another one to try

with your words in fornt

[rough oxide]

episode can be replaced by movie too for more cinematic look usually

and this is digital anime illustration

[glacial coral]

I actually love with the most basic prompt u can get that try "anime screencap" or sum alongside

[dapper verge]

yeah anime still and anime screencap have worked well, i've just found them to be unreliable in their style. they kinda bounce around with what kind it gives you. with that lora it's far more consistent.

[glacial coral]

Ye rn I'm resting difficult percentages on the realism lora to see the "perfect" value and I'll test that lora after

[rough oxide]

2 more for just "digital anime illustration"

in a resolution i prefer more xD

[glacial coral]

I like the hair in the 2nd a lot more and the eyes

[dapper verge]

my favorite combination in pony anime/hentai model, with 2 different pony loras stacked on top for a really great sharp anime look. realistically that's not going to happen for flux probably for a good while unless I train it myself.

[glacial coral]

Iirc pony is the best model to come for anime generations imo

[rough oxide]

when i was using pony i just used https://civitai.com/models/429471/knk-ponymerge and stacked char loras ontop of it

knk already has a good anime look

they came out near perfect

[glacial coral]

Ik some1 that's gonna get triggered that the AI images came out good (they don't like AI much)

But eh I like AI a lot it's one of my fav things to use

[rough oxide]

that was knk ponymerge with a p&s lora

lol

[glacial coral]

That actually looks like it's a screenshot straight mid episode

Good job

[rough oxide]

i wanna be done with pony era tho.

nai foss/wdv/something

for now it's flux

i don't really like pony as a base model

like at all

because part of the training i heard was they stripped like

[glacial coral]

Ye only if it was as fast to genrate for me like Pony it would help me a lot

[rough oxide]

all of sai basic training

u can't even get a leek in pony

[dapper verge]

https://civitai.com/models/352581?modelVersionId=501154 https://civitai.com/models/432483/sinozick-style-or-style-lora-or-pony

[rough oxide]

like miku holding a leek lol

[dapper verge]

I was using these 2 in combination

[glacial coral]

I'll save these for later it looks cool

Ty

I need to explore AI more iwl

[dapper verge]

in this combo

[glacial coral]

Ah I don't use ComfyUI

My autism has issues with focusing on it

(That sounds more weird than anything)

[dapper verge]

that's probably ok with this since it's just simple model/lora/lora, which automatic and forge supports.

[dapper verge]

gives great output like this though

[rough oxide]

when i was using pony i had 300gb of loras... i really just want a good base model ;_;

[glacial coral]

Using base pony or another pony model?

[dapper verge]

That one is a combo of 1 pony model and 2 pony loras. prompt was: (sinozick style, flat color, dark theme, g4n1m3), 1girl, robot, cyborg, glowing, blue light, green light, metal, wires, long hair, black hair, hair ornament, lights in hair, magical, holding orb, transparent sphere, aquarium, underwater plants, glowing fish, underwater, dark background, particles, coral, serenity, enchanting, otherworldly, smooth metal, metallic skin, bioluminescence, science fiction, futuristic, ethereal, mysterious, shadowy, soft focus, gentle lighting, intricate details

[glacial coral]

I'm surprised my 1tb storage macbook still has 316gb left after everything I've downloaded

[glacial coral]

Ty

[dapper verge]

well, at least with comfy you can just do a model merge. then it's a base. 🙂

[rough oxide]

the real problem with pony is it has none of the fundamental SAI knowledge.

like it lost so many common concepts

flux is now the model with like all teh concepts™️

so you can do things that aren't even possible in more finetuned anime models like bring specific items or things in and have them go anime w/o anything else

but alas

it still needs work on sum stuff like

_< face n what not

[dapper verge]

sure, although i've only been messing with pony for a few months, i've been impressed with how much it knows.

[glacial coral]

I'm so used to using flux now I forgot how fast in comparison pony gens images

[rough oxide]

i use a lot of my images commercially for work n stuff

so i need moar flexible models

so u get to a point where flux knows miku and about leeks, but not >_< face xD

[dapper verge]

yeah, flux will definitely be that.

no question it's the future.

[rough oxide]

ye, once the finetunes come.

o lmao someone else posted it on 4ch

if it actually happens this is big

and yeh pony stands no chance vs that, naiv3 finetunes would be insane

[glacial coral]

Cool

[rough oxide]

nai can still provide a bit more style than flux on average for simplistic prompts

flux is more superb in long-prompting tho

[glacial coral]

Agreed