🍥｜anime

[nova remnant]

[steady grail]

i actually like that a lot. reminds me ender lilies a lot

[green temple]

thx ❤️

[steady grail]

i like how the model plays with the focus

with the focus only on the subject

[green temple]

most my gens are ( upperbody, depth of field)

maybe that`s why

[steady grail]

interesting that it generates legs with upper body. usually models gens half bust gens

and for what i tested, not every models rend well to depth of field

[green temple]

oh , i see

[steady grail]

even after so much time, models are still mistery

[green temple]

indeed

[steady grail]

even with my crappy skills in merging, some merges gens only monsters, some others are even better than expected

playing with this tech is like finding perls in the desert

[green temple]

i`ve never tried merging

[steady grail]

it's funny, but it's very complex

so many values to manage

[green temple]

then it`s not for me 🥲

[steady grail]

eheheh

[green temple]

i just know prompting, even my prompting is bad

[steady grail]

i'm more in love with the technical part of ai than prompting lol

[green temple]

😂

[steady grail]

if anything else, i spend more time on comfy workflows lol

too bad merging with comfy is even more complex

[green temple]

so you love complexity 😂

[steady grail]

i love to know

i love the act of learning the process that can make beautiful things

[nova remnant]

overcomplicanting the process*

mooning me

[stable crane]

lol

[green temple]

i love mooning people

[nova remnant]

@green temple moon colored smol

(those moons are lucky gens I just noticed)

[steady grail]

charming

[green temple]

ohhhh

noice !

[nova remnant]

Cookies

[green temple]

evil cookie

[nova remnant]

squish

👍

[green temple]

where

[nova remnant]

runs away

[nova remnant]

@green temple she is your new neighbor

Eat the cookie

[green temple]

i`m full thanks

[nova remnant]

She looks so friendly

[green temple]

no

[nova remnant]

Face of someone who ate too much ice cream

Guilty

[green temple]

she just woke up

she was asleep here

[nova remnant]

Extra smol

She's smol too

[slate narwhal]

This one was kind of random; I was trying to get white and blue colors, and it gave that translucent lightning/electricity effect.

[vital raptor]

I think she's summoning something..

[bitter latch]

weird way to hold up hair

blep

there back to cute

bleps

[bitter latch]

hmm are those wings or extra ears?

ok so pointed ears works well

huh did not prompt 2

donuts!

[bitter latch]

sleep.. thud

[brittle rampart]

Neko is just lovely!

[nova remnant]

https://media.discordapp.net/attachments/931925914470711346/1061076535525785680/321666B2-CC36-467F-B254-56055DEB1C8C.gif

rare eface post

[dawn jasper]

@lyric zealot How goes life?

[lyric zealot]

playing a lot of wukong

[dawn jasper]

@lyric zealot Wanna hang out?
Also you have Dual Monitors?

[lyric zealot]

I do have dual monitors, and im watching stream while playing wukong right now

[dawn jasper]

Groovy.

[dawn jasper]

Kanna Kamui?

[green temple]

yup

[dawn jasper]

Her clothes and design are noted to have an Ainu theme to them, by what I heard.

[green temple]

idk

[dawn jasper]

@green temple Who is this one?

[green temple]

my imagination

[nova remnant]

grabs your imagination

[green temple]

noooo get it back !

[dawn jasper]

How goes life?

[nova remnant]

https://tenor.com/view/cat-questioning-life-questioning-life-what-is-life-gif-4882578

Life.

[nova remnant]

[nova remnant]

I'm big lazy

Ate ice cream

[nova remnant]

She angy

Big (smol) warrior

[green temple]

[green temple]

big ( big )

[nova remnant]

It's just the armor

[nova remnant]

Nyoo

[dawn jasper]

This is beautiful.

[green temple]

thx

[dawn jasper]

@green temple How you keep having images like this?

[green temple]

i generate them with stable diffusion

[noble widget]

wdv is neat

[bitter latch]

goes rogue

[glass summit]

No anime specific model or lora

[unkempt cape]

public beta?

[rough oxide]

yeh param data says that's the v0.1 model.

[unkempt cape]

is it the same stuff posted here? like wdv-tests

[opal cradle]

I most admit the ai is getting more better

This is what I was able to generate

And this

And this what you guys think about them

Used Tensor ai

[wooden coral]

looks a little low quality compared to whats possible now

[steady grail]

maybe low denoise

[bitter latch]

looks like LCM

[steady grail]

yeah, that's a possibility too

[gloomy scroll]

Is this good enough

[zinc flax]

Gotta love 1.5

[wooden coral]

Who goes outside without pants

[gloomy scroll]

I saw clip that would be perfect for this but not in this pure server

[noble widget]

yep

[wooden coral]

[gloomy scroll]

[bitter latch]

well it's not a hammer or a bat but here ya go

[noble widget]

https://cdn.discordapp.com/emojis/956658029821059155.webp?size=48

[bitter latch]

for when you really wanna get rid of someone...

[indigo oasis]

I think that's excellent I like the colors and art style

it almost looks like pixel art but it's not

if that makes sense

[glacial coral]

any good Flux loras for anime? or do i stick with Pony/SDXL?

[nova remnant]

Loras there probably are

But don't think anything that would be really good

Like a finetune

[supple raptor]

Anlatan just opensourced NAI||v1|| ||https://blog.novelai.net/novelai-diffusion-v1-weights-release-en-e40d11e16bd5||

[zinc flax]

We're so back

To 1.5

[steady grail]

nai was based on 1.4 lol

[zinc flax]

To 1.4

[nova remnant]

Download

Rename it to v3

Boom naiv3 on your PC

[zinc flax]

Let's pretend to be surprised and didn't know anything about the nai leak

[steady grail]

there was a leak shared on torrent? didn't know

[rough oxide]

essentially it's been out there forever. that's how stuff like anything v3 exists

[nova remnant]

[nova remnant]

happy (she wants food)

@green temple boo!

I put "scarying you" but I think shes the one scared lmao

[green temple]

no jumpscares please

[nova remnant]

many smols

big 👍

[keen hollow]

hmm.. can't seem to use loras with nf4... just slows to a crawl

[supple raptor]

forge or comfy?

[stable crane]

lol

loving these hats

[bitter latch]

hehe

well there's a skull on the hat but....

[indigo oasis]

that actually looks drawn man, with upscaling and denoise it would be great

[indigo oasis]

what model was that from?

[vestal crypt]

Mostly novelAI, but I tweaked it with a model I merged myself, plus GIMP.

[indigo oasis]

oh very nice

I'm looking into this one called Animagine XL 3.0

[keen hollow]

Forge. I found some GitHub threads on it, just still a little buggy. You can get it moving by shifting the memory use but it's a little janky.
Swapping over to straight dev instead of nf4 is more consistent.

[unkempt cape]

yk on second thought

animagine 3.0 wasnt as horrible as i made it out to be

its decent askshually

i think my half baked lora just fucked it over but its decent enough on its own

[unkempt cape]

welcome to the world of paid models. If its good, its paid

[supple raptor]

janky is on brand for Forge to be fair, but i've had a better experience using the gguf models than nf4. on 12GB vram i'm running:
model: flux1-dev-Q5_K_S.gguf from here: https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
text encoders: clip_l.safetensors and t5xxl_fp8_e4m3fn.safetensors from here: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
vae: ae.safetensors from here: https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main

[keen hollow]

yeah I've seen others tout success with gguf, too. Have not tried that. Have heard they're not as fast as nf4 but still faster than dev.

Why Q5?

[supple raptor]

Q5_K_S and Q6 were recommended for me to try based on the amount of vram I have and in the few test images I ran on both, Q5_K_S came out better.

[keen hollow]

interesting... shouldn't Q6 technically yield better results?

I'm also team 12GB VRAM. 3060. 🙂

[supple raptor]

i think that should be true. might just be down to the parameters used to quantize this version at Q5_K_S were better than the ones used for Q6

[keen hollow]

you find lora results ok on this?

Lora results on nf4 were not as expected... much better results on dev

Downloading and will give it a try.
I can get dev rolling around 4.3-4.5s/it. (512x512)

[supple raptor]

I haven't tried running straight dev, but the few loras that i've tried have seemed to work as expected.

[keen hollow]

ill give it a whirl... let's see

[supple raptor]

heres one using the InkSketch lora from Civitai

Flux clearly doesn't know who 2b is

[keen hollow]

I grabbed that lora yesterday, haven't tried it yet. 😄

gosh darnit

what's the dang clip folder location startup parameter? 😐

--clip-dir ?

[supple raptor]

I'm not sure, I just put those in the intended directory in my forge install and put the others in my remote folders that were already configured.

[keen hollow]

i could do that, but that's no fun 😛

gonna see if this works

and the answer is "no" 🤣

there's a "--clip-models-path " -- will try that

[supple raptor]

that sounds like it should be the one

[keen hollow]

well forge loaded, but generation still failed with "You do not have CLIP state dict!", bummer

[supple raptor]

you have to select the two text encoders as well as the vae in the vae dropdown. it's a very janky solution, but it works.

[keen hollow]

yeah they don't show up unfortunately

[supple raptor]

here's how mine is currently set

[keen hollow]

why do you drop your gpu weight that low?

[supple raptor]

i was doing other stuff earlier and needed some vram free

I think it can handle around 9.5-10k without other stuff running.

[lone hazel]

anyone know how i can create something like this? im new to stablediffusion and im just looking to make similair images but ive been struggling for hours

[keen hollow]

ugh ill move the clip files to the vae folder... so messy

man it's amazing how much faster nf4 is to dev...

[keen hollow]

just about any anime model could do that if you're ok to add text in post

[supple raptor]

mine are in ./stable-diffusion-webui-forge/models/text_encoder

[keen hollow]

yeah I didn't see a way to specify a customer text_encoder folder

[lone hazel]

can i DM you?

[keen hollow]

I only use custom folders because I run multiple instances all sandboxed.. it's a whole thing 🙂

[supple raptor]

I'm not sure if they'll load properly from the VAE folder. if they do, you can always make a subfolder in there to keep them organized.

[keen hollow]

yeah I did that, but it breaks my comfy install unless I move 'em back 🙂

ok here we go .. moment of truth

https://tenor.com/view/load-loading-april-fools-gif-5435835

yep, it's alive!

2.5s/it versus like 1 with nf4, but it is what it is

[supple raptor]

ah, i haven't tried getting flux going on comfy yet at all. i have access to flux pro so didn't see a reason to run dev or schnell locally until we had lora support and forge got that first.

[keen hollow]

I started on comfy just because that was the first place to try it... comfy is a last resort for me. 😐

[supple raptor]

the problem i had with nf4 was that it would fail to apply the lora 3 times each time i changed it's weight and then have to offload more of the model to continue which took like 2 minutes before it would even start generating so even though my it/s are lower on the gguf models, my time to generate is faster.

[keen hollow]

yes, exactly, there's some jank with it right now

and frankly, the loras for me were clearly working, but not working nearly as well as they do on dev.

ill try with gguf and see how they look

[keen hollow]

no need to DM - ask here! lots of experts. 🙂

a poor first attempt 🙂

that's on flux: "A black and white drawing, anime style, of a man on the left dressed as a knight with a cape, and a woman on the right wearing a modest white dress. The man is saying "Loser"."

one of the actual anime models will be better (and easier to get set up, and faster)

[supple raptor]

i used joy captioner

[keen hollow]

had to flip to "Automatic (fp16 Lora)" but seems to be running

we'll see how it turns out

[lone hazel]

ive just been struggling, i have no clue what model to use

and what prompt, im new to stablediffusion

[keen hollow]

well, it worked.. how about that

gonna try higher res

[keen hollow]

You might want to start with a basic youtube tutorial to get it all set up.

Then grab an anime model like Blue Pencil.

[zinc flax]

First thing first though, do you have a beefy pc lol

[keen hollow]

or just ask nubby.. apparently nubs has flex pro. 😄

[zinc flax]

Cute

[lone hazel]

okay thanks ill try this

[lone hazel]

idek, not really

i did watch some tutorials

[supple raptor]

what model is your GPU?

[lone hazel]

its a 2060 super

[supple raptor]

that's plenty capable. very comparable to what I had when i started doing this and we have a lot better optimizations now.

[supple raptor]

https://tenor.com/view/touhou-marisa-kirisame-master-spark-toss-gif-25660086

[lone hazel]

great good to know

[supple raptor]

i started with an rtx 2070 and now have a 3080 12gb.

[nova remnant]

What you mean she looks exactly like her

[lone hazel]

yeah im a youtuber that makes manga videos and i wanted to experiment with AI images for thumbnails, just thought it was worth learning

[supple raptor]

yeah thumbnails seem like a good use for ai images

[lone hazel]

could I DM you?

[supple raptor]

better to just ask questions in here that way other people can chime in if they know something I don't

[nova remnant]

emote name

why does it exist

lmao

[gloomy scroll]

Because Nubby created it

[supple raptor]

because i made it

[nova remnant]

[supple raptor]

i don't even remember what she was supposed to be doing, probably pouting, but it definitely just looks like she gotta pee

[lone hazel]

okay, i'll just post a screenshot. You know you saw what i wanted to create but im getting something totally different. You see anything obvious that im doing wrong?

[supple raptor]

i'd recommend to start by installing this extension https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

that will autofill tags from danbooru's tagging system as you type in the prompt. danbooru's tagging system is what most anime models are trained on and are what will work best for prompting on them.

if you haven't installed extensions before, it's easy. just go to the Extensions Tab then the Install from URL sub-tab, paste that link and add .git at the end then click install

[lone hazel]

okay got it installing it rn

[supple raptor]

once it's installed you'll need to go to the Installed sub-tab and click Apply and restart UI so that it can take effect

[lone hazel]

okay doing it rn👍

[supple raptor]

this is another useful extension that will allow you to input an image and have it give you a list of the danbooru tags that it would use to describe the image (no need to add .git to this one since I already did) https://github.com/picobyte/stable-diffusion-webui-wd14-tagger.git

[keen hollow]

ok... so the results are in.
GGUF runs at about 5.65s/it
Full dev closer to 7.3s/it
(896x1152)

Not a tremendous speed increase, but some. Jury is out on the gguf quality, need more tests.

[lone hazel]

thanks, installing that too

[supple raptor]

while you're at it. grab this one too https://github.com/Bing-su/adetailer.git

that one will detect faces in your image and automagically inpaint them at a higher resolution to improve the details

[lone hazel]

alright thank you also getting that one

[keen hollow]

adetailer rules

[lone hazel]

this one seems really useful, so how do i exactly use it?

i want to try it on the image i showed earlier

oh actually im reading it rn i see theres a guide

[supple raptor]

it will make a new tab on your UI called Tagger. drop your image there and click interrogate image and it will work. there are some options that you can play with if you want, but it will just work ootb without any fiddling

[lone hazel]

yep just did it thank you

[supple raptor]

since you're a little on the low side for vram i would recommend using either that checkbox that says Unload model after running or manually clicking the Unload all interrogate models button when you're finished running it

those models aren't very big, but every little bit of extra vram can help

[lone hazel]

i see thank you

so i used the tags that it found but the image is still off, could it be my settings or model that im using?

[keen hollow]

maybe try animagine if you want to use tags

[lone hazel]

okay i'll download it

[supple raptor]

another thing you can do is to try using artist names if you know fairly popular artist that draws in a similar style to what you want. not all models know the same artist names, so you may need to experiment to find what works on the model you're using.

[lone hazel]

you mean like manga artists?

[supple raptor]

manga artists or artists who make fan art

are you familiar with danbooru?

[lone hazel]

im not

[supple raptor]

it's a very large anime oriented image repository. here's the relatively sfw version of the site. you can go to the tags tab then change the category to artist and order to count and it will give you a list of artist names along with the number of images by them on the site. in general, the more images an artist has, the more likely their name is to work as you'd expect on any given anime model. most anime models are trained on data directly from this site https://safebooru.donmai.us

[lone hazel]

oo i see

and i guess i need to put their name in the prompt right?

[supple raptor]

yep, their name is counted as a tag just like all of the ones that describe the image

[lone hazel]

ah i see

you know any artists that do like manga style?

[supple raptor]

some other tags that could be useful for getting something similar to the image you posted are comic, greyscale, and monochrome

[lone hazel]

yep thanks👍

[supple raptor]

oh, potentially sketch and graphite \(medium\) too.

i don't have any specific artists off the top of my head, but you could search comic on the site and click some of the images you like to see the artist name. artist tags are shown in red on the left side. purple is copyright (like the series name), green is character names, blue are all the general tags that describe the image, and yellow are meta tags which can usually just be ignored for our purposes.

[lone hazel]

okay i see ill try this

[green temple]

i tried to download "tagger " and i can`t find it on my UI tab

should i download something else to run tagger ?

[supple raptor]

It should show up after clicking Apply and restart UI then refreshing the page once that finishes, but if not then restarting WebUI completely should solve it. Sometimes both Forge and A1111 are funky about adding new extensions without completely restarting.

[green temple]

oh , ok i will try

[gloomy scroll]

I just use PI-chan hf space

[supple raptor]

that also works. it's the same code running both. PI-chan version just has 10000% more Chen.

[gloomy scroll]

https://media.discordapp.net/attachments/829123747603677245/1156244416940490792/chenstarejif_1.gif?ex=65144406&is=6512f286&hm=244a6a5e06cc798ce62ef989dff3afc5d8c12d8292caa424db83d8b4895f8bfe&

[supple raptor]

https://tenor.com/view/chen-excited-touhou-bkub-anger-gif-8545301

[nova remnant]

wtf

looks like those terraformers cockroaches

[gloomy scroll]

https://media.discordapp.net/attachments/163382741491122176/1203958996277395466/chenspin.gif?=&width=300&height=300

[supple raptor]

https://tenor.com/view/chen-touhou-marisa-chase-pest-gif-26682906

[green temple]

nooooo

[supple raptor]

somewhere an alternate universe exists where the portable versions of all of these UIs don't exist and people just git clone the repo and use a regular python venv for their environment. i want to live in that universe.

[bitter latch]

invisible person? what?!

[keen hollow]

oof... you would hate my setup 🤣

i don't mind git-cloning, but I use sandboxing because of all this untrusted code and it's a whole thing...

[supple raptor]

the thing is that the portable versions aren't any safer than git cloning and pip installing the deps to a python venv. the only thing they protect the user from is accidentally installing the wrong version of python before they start, and as soon as a dependency issue comes up it's a lot more difficult to fix than it would be if they were just running a regular python venv. sandboxing isn't a bad idea, but also not something that probably 99.99999% of users have even considered, yet alone know how to do properly.

[bitter latch]

and finally they arrive at the castle

oooh

[keen hollow]

yeah the portables just auto-download a bunch of stuff regardless. I get it.

Sandboxing helps me run various versions of everything (python included) pretty easily, while also providing some extra security layer.

[bitter latch]

love how all these could be the same planet

hmm same location different times?

random girl

[bitter latch]

and random student from that village?

hmm adding random style from a ext on each on makes it interesting for sure

new student lost

changing one artist and suddenly we are in town

oh look they are still lost

and i'm not prompting for them

change to clipskip 3 and

stepping thru...

[steady grail]

cool gens

[bitter latch]

thanks

ooh it finally did the underwater part of the prompt.... kinda

[green temple]

🌚

[nova remnant]

Those are some heels

[green temple]

Lucky heels 😶

[nova remnant]

Post smol

I'm lazy

[green temple]

Pc is off

[nova remnant]

[nova remnant]

Smol shoot him

[green temple]

Noooooo don’t shoot

[green temple]

dies from smolness

[nova remnant]

Smolsion

I was cutting my nails

I did 5 of them

Lmao

Just noticed

I stupid

I left the indicator one in my left hand and thumb on right, something caught my attention in the middle of it

[glacial coral]

Guys question what's the best way to get consistent anime character rn I don't think u can do that with flux so if you guys know any models and a quick guide for doing consistent characters that would be great

[nova remnant]

Loras are the thing for that

[glacial coral]

I mean for characters that are random generated than character you get from Loras or am I lost? 😭 (like a character u random gen and liked and u wanna have a consistent of that character on other generations)

[nova remnant]

Well you can do loras with those random characters

There are other ways to try and help you generate them again or at least similar but nothing i know is as consistent as a lora

I think controlnet ip adapter is one of the better ones

[glacial coral]

I may need a tutorial for that then

[indigo oasis]

Before and after I turned someone's VRchat screenshot into a real woman. ✨

[nova remnant]

Tifa looking different

[keen hollow]

ahhh messing with flux... guess I need to go back to Tifas... 😛

there's a Tifa lora for flux already, but it isn't that great

[nova remnant]

Is that using the lora

Or just random

[keen hollow]

what do you mean?

the images above are not Tifa lol

Ill try one with the Tifa lora, one moment

eh... not great

[nova remnant]

Looks like some Chinese knock off

[brittle rampart]

👍 😍

[keen hollow]

There's some flux Tifas

[bitter latch]

there we go

"oh did I not mention...? I'm part of a special group, and you are my target this evening my dear."

there go

just a simple turn and HS

sets vae to none

huh...

the gangs all here.. ooh boy

[steady grail]

just redownloaded comfy, instead of using krita. damn how much i love that gui

every time i open it, i get new ideas of workflows

[bitter latch]

the Mafia Assassins'' creed

[steady grail]

pretty close

[bitter latch]

Charlie and his angels

or demon's?

there more defined

ok... he keeps popping up..

Go ahead .. say it one more time.. I dare ya

[bitter latch]

thats it! you get a cut.. you get it cut and You mofo you get two!

[indigo oasis]

someone PLEASE generate Tifa at a gas station with a styrofoam cup

I need to see Tifa Wockhardt

[indigo oasis]

what's Miku doin

[bitter latch]

about to slice someone I would say

ok of these

up next 😄

[bitter latch]

you mean like this?

[indigo oasis]

it just looks like she has a coffee

https://tenor.com/view/drinking-lean-cup-looking-over-drank-hip-hop-gif-13897851

more like this

[bitter latch]

tries again

[indigo oasis]

I don't have whatever LoRA that is, idk if I have a model that knows characters by name

I will try it myself though

[bitter latch]

It's not a lora

hmmm

[bitter latch]

well.. ok

almost

@indigo oasis Better?

[indigo oasis]

definitely better

I could probably inpaint that to be how I want

[bitter latch]

damn she's tall

and why is that jacket janky?

[bitter latch]

not sure what the weapon is

Tifa the Jedi

final fantasy is strong in this one

moves on

ponders

[keen hollow]

I don't get it. That's oddly specific. 😄

[nova remnant]

fairly smol

[indigo oasis]

I'm trying to make a bad pun by making Tifa look like a soundcloud rapper

her name is Tifa Lockhart and Wockhardt is a manufacturer of codeine cough syrup

[keen hollow]

ah

[indigo oasis]

basically the equivalent of "drip goku"

https://tenor.com/view/drip-goku-drip-goku-dripping-dripgod-gif-25245981

[bitter latch]

umm ok...

PUNKY!

yep anime punky

haha no nose

[indigo oasis]

I need to make more anime art

[nova remnant]

@green temple smolinabox

[bitter latch]

hhah

o.. k

da fuk

well, it's not a box but ok...

[twilit portal]

HELP 🤣 WHYS THOMAS RUNNING AWAY FROM HIS TRAIN 😂

[indigo oasis]

idk your method to getting the same character every time but your "oc" is very cute

[unkempt cape]

(nvm)

[vestal crypt]

@indigo oasis @unkempt cape I trained my own lora to keep the character consistent, no magic here
And yes this is an OC.

[unkempt cape]

wait wtf nai allowed loras?

[vestal crypt]

no, I inpaint it with my local SD.

[vital raptor]

@indigo oasis keeping the same character is usually just as simple as having the exact same section of the prompt describing the character every time. I generally put the entire character description, and artist tags first, then only change things after that for poses, clothing, expressions, backgrounds, etc

[nova remnant]

Ratty Luna

[magic wave]

Yis :o

[nova remnant]

Many box girls

[nova remnant]

[nova remnant]

Grabby ears

[bitter latch]

hehe

[indigo oasis]

YO GUYS

YOOO

TIFA WOCKHARDT

we're getting very close

I must be using this model wrong, because it's supposed to have incredible graphics, but I feel like it's just passable

could also be my style selection

but the recognition of DBR tags is pretty good

[indigo oasis]

looks like she got a latte or something there

[keen hollow]

you wanted a styrofoam cup 😛

[indigo oasis]

yeah...

[mystic kayak]

Anybody know what the best base may be to match with this lora? without restricting nsfw?
https://civitai.com/models/379247/touhou-memories-of-phantasm-style
So far ive used ClearVAE(SD1.5) for vae at least
AnythingElse V4 and Original AnimeFinal for base but just wanted to check if theres anything better. Plus the formers hard to get working.

(if this is something i ask here and not tech support that is)

[zinc flax]

Try something like hassakuhentai I suppose

[stable crane]

doesn't look like you're highres fix'ing

[indigo oasis]

you mean like inpainting or upscaling or what

[stable crane]

upscaling

quality boosts immensely after at least 1 pass of upscale

[indigo oasis]

yeah can't really be bothered right now

none of these are perfect

but it might be worth trying on a couple

[stable crane]

no upscale = explains why you see it as just passable

[indigo oasis]

is it like an actual redraw, not like Waifu2x upscaling?

[stable crane]

are you using A1111?

an actual re-diffuse

not those older upscaler models

in A1111, it's a checkbox called "highres fix"

[indigo oasis]

I'm running SDXL in Fooocus

[stable crane]

ah I have no tips then

[indigo oasis]

it does have an enhancement section with upscaling, let's see what I can do

[stable crane]

I've never used that one

[indigo oasis]

idk if I have to add any additional prompts or just let it work its magic

it might pull from metadata

[stable crane]

there's a ton of ways to upscale an image with SD though

the super-resolution methods are worthless though

[green temple]

indeed

[stable crane]

lately I've been doing

• initial gen
• 1.5 upscale via high res fix at 0.56 denoise
• 1.5 tiled controlnet upscale at 0.5 str and 1.0 denoise

this is a regular sdxl gen -> 1.5x tiled controlnet upscale
the left kinda sucks, the right is OK

crisper but not much better

here's sdxl+highres fix (already 1.5x bigger and with more detail), then 2nd pic is 1.5x again with tiled CN, last is adetailer pass on face (not always an improvement)

can kinda only appreciate it at full resolution

[indigo oasis]

it looks better in some ways

[stable crane]

way better

but could be you just don't like the default style?

[indigo oasis]

deleted her straw though, the wheels still aren't engineered right on the cars and the girl in the back now has some kids face printed out on paper and glued to her face

[stable crane]

that's how it goes

[indigo oasis]

are prompts super important when upscaling

[stable crane]

usually I use same prompt the entire way through

[indigo oasis]

that's what I could have done is reran it 1:1

or inpaint after this to fix weird parts

[stable crane]

if you want to fix the wheel you probably have to inpaint

but I find it just as satisfying to roll the dice and hope for a better pic

the front of the car is also probably too messed up

unless it's 2 identical cars in a line

[indigo oasis]

I mean this is one of the best ones, after I got cyberbullied into adding pants it all went to hell for some reason

[indigo oasis]

Kia Stinger limo edition

[stable crane]

you got cyberbullied? 😄

[indigo oasis]

yeah

this model does not like giving detailed faces when you force it to zoom out enough to show legs

maybe I can just re run those too they looked ok

[stable crane]

you can try adetailer concept (it automatically inpaints just the face)

face quality tends to go down across all models with distance from camera

[indigo oasis]

this looks like you were on PCP when you saw this

I think that one needs a redo with the same settings

upscale

[stable crane]

yeah once you get used to it, the first diffusion at sdxl resolutions is kinda painful to look at

so some UIs have a feature to do the upscale in that initial generation

[indigo oasis]

ok I didn't know that!

I thought I was doing something wrong

[stable crane]

I can recommend A1111 or ComfyUI

A1111 is easier

[indigo oasis]

I'm rerunning it with 2x upscale but I didn't put an input image, idk if it actually started from the previous result

oh wait it's got it

I think the way I did it for the last one it actually upscaled twice and gave me one that was a real high res

[stable crane]

lol

[indigo oasis]

yeah that's clean

tifa wockhardt

thank you for the advice

[stable crane]

yw

[indigo oasis]

now I'm actually mad because I lost the logs to something I should definitely upscale

I upscaled it with W2x thinking that was the same result

[indigo oasis]

I only played the first two FF games where it was mostly medieval stuff but the new titles do have cars and stuff don't they

so my "fan art" isn't too far fetched

[nova remnant]

Grabby

[green temple]

grabby eyes

[nova remnant]

Yes

[green temple]

oh

[nova remnant]

Eyes

[nova remnant]

Eyes

[green temple]

evil eyes

[nova remnant]

I'll put happy in the prompt

Don't worry

@green temple

[nova remnant]

[green temple]

smol is invading the server

[nova remnant]

Big

[green temple]

blurry hands

[brittle rampart]

It would be cool if the dog’s ears were positioned the same way and his head was tilted!

[glossy eagle]

Feast your eyes on this soul shredding anime photo. you will be up all night Your welcome

[nova remnant]

Fufunya

[bitter latch]

how you do thta?

[wooden coral]

Flux

[bitter latch]

blah

[wooden coral]

blah

[bitter latch]

now I gotta figure out how to do it without that 😄

[wooden coral]

https://tenor.com/view/you-don't-invincible-omni-man-jk-simmons-that's-the-neat-thing-gif-4390372980992221432

[bitter latch]

switches mode from lewd to research mode

[wooden coral]

Here have more

[wooden coral]

Not as bad as I thought

[bitter latch]

oh i'm just getting started

[brittle rampart]

The doctor said - AMPUTATION! 😁

[steady grail]

damn lol

[brittle rampart]

only amputation!

[gloomy scroll]

But Yue wanted dinner

[brittle rampart]

donuts again?

[bitter latch]

zapped?

oooh

haha

oops

[brittle rampart]

oh Yui... ok, ok... the doctor gave permission... but only after the amputation

[bitter latch]

showoff

[wooden coral]

[bitter latch]

bah work keeps interrupting my art

[bitter latch]

getting interesting results

@wooden coral

ooooh

not the same but I liike it

I LIKE IT

@wooden coral

[pale current]

The picture is quite cool, are there any other character styles?

😅

[junior quest]

prompt?

[wooden sail]

No need for prompts when you have style-transfer

[junior quest]

ehhhhhhh!? how did you do that?

dm pls

[wooden sail]

Just controlNet reference_only

[junior quest]

dat works?

for me nope

[wooden sail]

Only works in Forge. Because different implementations

[junior quest]

new one?

Im using old forge

[wooden sail]

I have the recent version. But old one should work too

[junior quest]

hmm I tried old one didnt work

[wooden sail]

Model used: AutismMix_confetti

[junior quest]

btw what if I use multiple control nets with multiple images with same styles?

[wooden sail]

My prediction: It won't work

Very unstable but somewhat working

It's a type of style-transfer not visual prompt

And without using controlNet (same seed)

[wispy canopy]

HE STOLE MY PROMPT!

lovely output there boto ❤️

[nova remnant]

[wispy canopy]

here's some of the OGs from which boto "borrowed"

https://tenor.com/view/meme-our-now-gif-21036569

"our prompt"

[wooden coral]

[bitter latch]

trying to recreate those glamour photos you would see but anime-ish

[bitter latch]

eh close enough