🍥｜anime

[unkempt cape]

day 2 of trying and failing to make something decent (this is not too much rite)

[unkempt cape]

slightly better i think?????

[nova remnant]

Smol looking at stars

[nova remnant]

Grabby

I'll start on the tail

[gloomy scroll]

[tiny cypress]

I WILL SET THE SEA ABALZE

[coarse brook]

Soo... Anyone using ComfyUI? xD

Way to get kicked off a server quickly. spam all channels.

[terse ore]

I use comfy

Is controlnet tile basically just an upscaler?

[coarse brook]

So... I'm trying to learn Comfy, but I'm having issues finding all of the stuff i had in Auto1111, such as Embeds and styles.

Idk where, and what they are called in Comfy

[terse ore]

Embeddings? They are called embeddings You can use them, see the comfyUI GitHub page and it'll show you. Styles are workflows, just drag and drop the image in and it'll load it up

[green temple]

yes

[nova remnant]

from what I know it upscales by tiles, so you can do higher than what you would normally because its limited to sections

[terse ore]

I know with SDXL and controlnet it changes the facial expressions, does controlnet tile maintain the same facial expression?

[nova remnant]

depends on denoise

the higher the more likely it is to change stuff

[green temple]

indeed

[nova remnant]

lick

[terse ore]

The workaround I use is to zoom in and img2img a face and then overlay the output on the original small face. That way the facial expression is the same. But idk if it's the same for tile. I'll just have to try it

[green temple]

ah shit , i just noticed ... 💀

[nova remnant]

Notice what

[green temple]

nothing 😶

[nova remnant]

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

how dare you

[coarse brook]

Not in my UI it apparently isn't

[terse ore]

You have to put the embeddings in the correct folder. And then to use them, you have to put (embedding:embeddingname) to use it in the prompt box. You don't need to use .pt or anything, just the name itself.

Where embeddingname is the name of the embedding

[coarse brook]

I'm re-using my Auto1111 paths, which i have stuff in

Can't i have a seperate window, like the Lora, for adding Embeds?

[terse ore]

Yeah try that and see if it works. I couldn't get my embeddings to work like that so I put them in the comfyUI folder as well. B managed to get the other models and stuff working with the file paths could use both comfy and a111

I don't know of a dedicated node to use embeds. It's annoying because I have to go to the folder and copy paste the names

[coarse brook]

😐

[terse ore]

There is the comfyUI discord. I'll ask because I'm curious too

[terse ore]

"If you are using comfy-scripts it will auto suggest them for you as you type in the prompt" that's what a response was

[coarse brook]

Ye, i did DL that

[neon linden]

does anyone know how I could replicate this type of rendering?

the way light appears and so

[unkempt cape]

novelai

maybe kohaku xl model can come close too

[neon linden]

@unkempt cape thanks!

[winged oasis]

at the moment of the nuclear blast

[oak spoke]

totally normal high school girl

[gloomy scroll]

[unkempt cape]

wait holy crap this kinda nice tho (its the exact prompt used in the dataset but pics look different but sick)

[magic wave]

Damn she looks good

[green temple]

random smol 🥹

[nova remnant]

smol

(I can't even load images, downloading a game atm but smol still)

lmao

[green temple]

indeed

[coarse brook]

First Chibi made!

[dim spire]

I cut the background since it was meh, it's better without it

[green temple]

cool

[nova remnant]

smol

[coarse brook]

quite smol

[nova remnant]

booba

[nova remnant]

probably nai

its paid

a model

and its paid

search for novel ai on google

should be easy to find

its famous

[rough oxide]

https://novelai.net/

but looking at that

it doesn't rly look like nai

nai would have no problem with her pupil pattern

naming format isn't same as nai either.

example with nai

altho i used that as a base lol

without base, pupils come out better in nai

[nova remnant]

eh

looks like nai

if you just use a lora for a nai look it should be similar

[rough oxide]

nai has a specific naming format when u dl from it, that still has like a base name format

and it's not nai's

[rough oxide]

also it's way too many details to describe makina Lol.

in NAI all you need to do is say 1girl, makina nakajima, chainsaw man

it's probly a pony-based model

based on how it looks

wait wrong makina

xD

"1girl, makima (chainsaw man), chainsaw man" is all u need in nai

then details on scene

anyways, that's all to say doesn't look like nai.

[nova remnant]

shrug people do different prompting

[rough oxide]

just further points that it's not nai, since in nai you'd almost always adhere to the prompt guideline of writing the character name & series near the beginning to maximize the benefit. it's almost certainly a model where they had to describe every aspect of the character and not use the char name in prompt based on what's shown on the filename xD

[nova remnant]

people adhere to guidelines on the internet now ?

thats news for me!

[rough oxide]

anyways it's literally not nai style

[nova remnant]

nai has infinite styles

could be

but looks like generic nai so a lora for nai style or models like tponynai

[eternal temple]

A boy with short white hair in white pajamas showed a very happy and happy expression on his face.

[wooden sail]

Animated

[supple raptor]

From the filename I can tell you that the image was generated with Automatic1111's WebUI or a fork of it, such as forge or SD-Next. The artstyle could be NAI, KohakuXL, or Animagine without LoRAs, or possibly even some Pony based model with a LoRA. NAI is unlikely, though it is possible with an extension that uses the NAI api inside WebUI. I think the eyes would look better if it were NAI though. If I had to guess between Animagine and KohakuXL, I'd probably say Animagine because I've never seen KohakuXL make a mouth that looks like that.

Where did you get the image from?

[unkempt cape]

hmm doesnt animagine has like one style that it cant escape out of?

[supple raptor]

animagine is trained on at least hundreds of artists. it's quite flexible in style. one of the main goals with animagine 3.1 was to better replicate the most common artist combinations from NAI posts.

[zinc flax]

Sucks at anatomy though

[supple raptor]

yeah, I'm not a fan of it. the character choices in the dataset are highly opinionated.

[zinc flax]

Had to use multiple style loras just to fix the body

[unkempt cape]

real or fake

[supple raptor]

I like kxl-delta/epsilon a lot more for local gens.

[unkempt cape]

animagine having alot of styles is the 2nd most cap thing in ai image gen

i mean technically if you count the one pixel color change as new styles, yes

[supple raptor]

try it yourself. Here's some wildcard files with the artist tags ripped from their dataset. the "50up" version includes only artists that appeared at least 50 times.

[unkempt cape]

later

i saw a post on civitai about artists impact on this model

is animegine anatomy that horrible?

[zinc flax]

base animagine suuucks

[unkempt cape]

never used it by itself i always coupled it with loras

[zinc flax]

It's terrible without style loras lol

[supple raptor]

here's a couple from the same prompt with different artists on animagine3.1. just some random ones I found in my output folder.

[unkempt cape]

Wanted to train loras with neta but shit just doesnt allow ):

[unkempt cape]

ok i c

anatomy doesnt look that horrible whats everyone talking about

[supple raptor]

I was cancelling anything that had shit anatomy while I genned those.

it's a lot worse at anatomy than even KXL which is already not great compared to NAI.

[unkempt cape]

wow look just like anime episode

[unkempt cape]

don't compare anything to NAI

Maybe I can train my loras on kohaku

Hmmmmmmmm

[supple raptor]

they'll probably be less compatible with other random models than if you used animagine, but should be a lot better in that regard than if you used pony.

[unkempt cape]

Screw pony nahhhhhh

The compatibility is already like 0

I just want something that's less of a shit hole than animagine

[zinc flax]

Pony is still the better one

Unfortunately

Either choose pony or animagine 3.1 lol

train on pony, and gen on autism confetti

[unkempt cape]

Nah screw pony im going kohaku

[supple raptor]

just too much fuckery in Pony's text encoder and Animagine overfitted badly on some characters.

[zinc flax]

and the hashes

rest in rip.

[unkempt cape]

Im going kohaku if diffusers allow it. Pls pls pls 🙏

[zinc flax]

speak of the devil with animagine crappy anatomy lol

[iron star]

this is a 1.5 model,i can tell because of the thick lines,its probably meinamix

[supple raptor]

backgrounds are a little worse in general on kxl than animagine. but i haven't found many problems that shouldn't be fixable with loras.

[unkempt cape]

Me after clip cuts off my prompt leaving only a wall of score_x: 🗿

[zinc flax]

Idk if I'd recommend training on kohaku since it's essentially a merge

[supple raptor]

there are line thickness sliders for all 3 major models (animagine, kohaku, and pony)

[unkempt cape]

No thats animagine 100%

look at that mouth

[supple raptor]

diffusers will do that every time. it can only do 75 tokens unless you're skilled enough to write some custom code.

[unkempt cape]

still better than whatever this shit hole is

[unkempt cape]

And I suck at code

[zinc flax]

welp, good luck i suppose

[unkempt cape]

Whatever I just don't like pony. Kohaku if the diffusers god allows it

[iron star]

thats cool,there are line thickness sliders for every model out there even 1.5

[supple raptor]

it's not. the format it was trained in was just LyCORIS which was then permanently merged to the same base model for ease of distribution since loading LyCORIS can be handled differently in different UIs.

LoKr high factor, to be precise.

[unkempt cape]

there is a diffusers pipe that get rid of token limits

[supple raptor]

is it for training or for inference though?

[unkempt cape]

pipes are for inference bruh

im too lazy to figure that out tho

im fine with the limit (as if the model is going to listen past that amount LOL)

[supple raptor]

why are you using diffusers anyways? are you on a remote system that already uses it or genning locally?

[unkempt cape]

Free gpus

[supple raptor]

if you have the ability to set up docker wherever you're getting gpus from you could set up https://github.com/Stability-AI/StableSwarmUI

[unkempt cape]

ill see if its allowed

[supple raptor]

it installs comfy as a backend so you get a lot better flexibility

[unkempt cape]

heard docker a lot in my life, never knew what the hell it is

[unkempt cape]

comfy? Hell nah

[supple raptor]

don't worry, you don't have to interact with the comfy nodes.

you get an actual UI that builds the nodegraph automatically for you. it's not a whole lot different from using A1111 WebUI in practice.

[unkempt cape]

webui 🥶

ill try it on kaggle first, if not then i dont want to risk my colab lol

[supple raptor]

the comfy nodes are there on a separate tab if for any reason you ever did want to try your hand at it, but you also just have normal buttons, dropdowns and sliders for everything you need and can basically just select a model and resolution, then type your prompt and negative and hit generate.

[unkempt cape]

ok 👍

[supple raptor]

also has a nice built in image browser to look at previous gens and import the parameters from them.

[unkempt cape]

ok

ill import the code from colab first see if that works. If the link pops up then crash then i know im screwed

I do hope it works tho, notebook interface makes me wanna die really bad

[supple raptor]

yeah that change on colab sucks

[unkempt cape]

cant blame em tho

[supple raptor]

yeah, unfortunately there were a bunch of businesses abusing it with thousands of free accounts running simultaneously

i know a lot of people moved over to paperspace, but that isn't free for the usable one. $8/m is fairly reasonable but that does rely on decent GPUs being available when you want them.

[unkempt cape]

just done w the lora training 13 epochs im too damn lazy to boot up kaggle to test

[supple raptor]

ah, it's a lykon made lora which means that it's most likely AAM https://civitai.com/models/84586/aam-anylora-anime-mix-anime-screencap-style-model

[supple raptor]

@cyan brook do you know which model this image is from? it's one of the previews for your 1.5 makima lora https://civitai.com/models/5373/makima-chainsaw-man-lora

Edit: not the replied image, this one: #🍥｜anime message

I doubt he's online right now, but he will probably reply whenever he does get online. It's 1yr old so he might not know any more, but if anyone does it would probably be him.

damn it. I did

Discord jumped to the wrong one when I tried to follow the replies back XD

[unkempt cape]

assssssh shibal

Lmao one of the squares looks like a trapdoor

we got trapdoor checker give us trapdoor chess

[unkempt cape]

ok maybe lighting a bit weird but DAMN

another artist

No one asked but basically i wrote code to automatically try every artists out. There were many other good pics... except animagine's anatomy broke so like 2 made it

ok trying out that ui thing now

[unkempt cape]

Crap its banned on both sites

[vital raptor]

you didn't need to write code for that, you can just run the X/Y script and use a list of names

[unkempt cape]

tf is that

[green temple]

like this ( i dunno how to use it )

[vital raptor]

you don't even have to use all x/y/z, you can just do X for S/R (search replace) and put in a list of artists

you can also do it from a file if you have a text list of the artists

[unkempt cape]

Ok but one question, What the fuck is it?

Another question: Is it anyway related to webuis?

[vital raptor]

It's a script, in Auto1111 it's in the scripts dropdown at the bottom of the controls on the main screen

and forge also

in comfyui you might need a custom node, not sure if that node comes with it or not

[vital raptor]

epic star wars and rescue rangers crossover 😄

[green temple]

i didn`t want this but its ok

[bitter latch]

added the ext that auto complets tags 😄

[bitter latch]

oops

someone's about to lose hands

haha

[vital raptor]

force wielding

[bitter latch]

hmm I wonder

had to do it

not sure whats with the hair

[bitter latch]

hehe

[bitter latch]

Sawe Waars

[vital raptor]

hmm, oddly A1111 seems to be pretty unstable on linux

[bitter latch]

DOH

haha

no eyes

[vital raptor]

maybe it's time to try out swarmui..

[dreamy star]

Moon light silent night

[vital raptor]

random words

[bitter latch]

ummm

[green temple]

bad hands but fine

[bitter latch]

almost...

[green temple]

🤣

[unkempt cape]

No shit bro if I have auto 1111 i wouldnt be using this stuff

[vital raptor]

🤷‍♂️ you're always asking for help in a channel full of webui users so I figured you were doing the same thing

[unkempt cape]

🤷

Don't remember asking for help

about the xyz thing at least

[vital raptor]

no worries, next time i won't bother

[unkempt cape]

ok

[umbral depot]

[unkempt cape]

Man wtf is going on sometimes i create this

And sometimes i create the worst abomination in mankind

[green temple]

or not ?

[vital raptor]

too much ankle

[green temple]

noooooo 😭😭😭😭

ok i will delete

[nova remnant]

[unkempt cape]

Can anyone give me a quick prompt to test stuff real quick

currently have no inspiration rn

[nova remnant]

I'll take two

[zinc flax]

Two popsicles

[nova remnant]

1girl, cowboy shot, solo, elf, pointy ears, dress, goth, eyeshadow, eyeliner, smug, white hair, short hair, green eyes, sitting, sofa
The elf goth test of all time!

[nova remnant]

The other one will come from me

👍

[zinc flax]

oh no

[nova remnant]

Yes I know the best place to buy them

Now I'm hungry for ice cream

there's one that's so good I want it

[zinc flax]

gogogo

Buy ice cream

[nova remnant]

the smug :3 is too smug

[unkempt cape]

Goddammit i hate my damn model. It trying to do smug was the worst thing ive witnessed

Had to take that off for something remotely fine

[vital raptor]

Most models have trouble with smug. Even pony models it doesn’t work on every style

[zinc flax]

Here's smug

[vital raptor]

meh, the upscaler in stableswarm-ui is garbage

[green temple]

indeed (enjoying 1080s) (sadge )

[green temple]

anyone knows the lora used in this image to make the outlines like this ?

i mean the style in genral

[nova remnant]

Why you on stable swarm ?

[vital raptor]

trying it out

I recently installed linux on my main PC and was having trouble with A1111, so I've been playing with comfy and swarm

[zinc flax]

Don't think any lora was used in this

Looks like any 1.5 model lol

[green temple]

oh ok

[vital raptor]

comfy has all the features, and it's fast, but it's such a pain to actually do anything without spending 3 weeks making workflows for everything

swarm has a UI I really like, but it lacks some pretty basic functionality, like a proper upscaling/hires method.. but it has the comfy backend so it's fast like comfy, and has the potential for custom nodes for features, but that just means i have to make workflows again

I'll be trying forge next, see if its "totally not comfy" backend will behave better than a1111 but still give me the features I'm used to having

or maybe i'll just take the time to get a proper working workflow in swarm and stick with that since it has the best UI i've tried so far

[nova remnant]

Kek

[nova remnant]

loraforsmol:1.0

[vital raptor]

😄

[supple raptor]

Huh? Swarm supports every upscaler that A1111 supports and more since it's just comfy. Just use the one you like.

[vital raptor]

I'm just talking about the built-in one

it's not configurable and just destroys the quality, i still need to play with the custom workflows in it

[supple raptor]

Hires fix is under refiner. I can snag a screenshot later but you can probably find it faster if you search "highres" or "hires" in the #🐝｜swarm-ui channel.

There's also built in adetailer using the <segment:> command

[vital raptor]

i saw the mention of the 'refiner' being basically hires fix, but it doesn't seem to have any of the same upscalers or a setting for a folder for them

haven't checked yet to see if there's a way to add them or not

looks like there's a directory for them, just no way to custom define it like all the others

well, that was simple enough, just had to put the models in the right folder manually

[supple raptor]

👍

[green temple]

( wolf ears of furr, fluff, fox tail : 1.5) 👍

[nova remnant]

so much weight

[green temple]

here it`s like 1.7

[nova remnant]

even more

[nova remnant]

grabby tail

[cyan brook]

it has been asked in the comments. I didn't remember then and I sure don't remember now 😄

[terse ore]

What resolution to use in controlnet tile if the image is 1344x768?

[oak spoke]

there's imposter among us!

🤣

[supple raptor]

Ah, sorry to bug you then. I didn't dig into the comments. I just saw it was yours

[nova remnant]

the camera is too high

[unkempt cape]

hmm finally coded kaggle to train sdxl but man

idk the hell is happening but steps are halved

[unkempt cape]

nawwww loss=nan wth ??????

[zinc flax]

Maybe lr too high?

[steady grail]

lr?

resolution? rl?

ress?

[steady grail]

choid? you have exp in training?

[unkempt cape]

I have exp failing at training

[unkempt cape]

Bro the recommended value was 3e-4 i bumped it down to 3e-5. Also colab worked fine

[zinc flax]

3e-5 is waaay too low for xl I think lol

[unkempt cape]

Nah it works 100%

[zinc flax]

I use 1e-4

[unkempt cape]

Works 100% Best? idk but it works 100%

[zinc flax]

Undercooked probably

[unkempt cape]

How to tell if undercooked?

cause maybe askshually

[zinc flax]

If there's no significant change to the gen at all

[unkempt cape]

Eh there was changes ig, sometime it resembles well

just god damn unstable

[zinc flax]

Odd

[nova remnant]

math talk

[unkempt cape]

sometimes it makes real high quality stuff

and sometimes it makes eyes burn

im changing the model since i heard by people here that the current stuff im using was horrible

[zinc flax]

What did you use for the base model anywau

[unkempt cape]

anima v3

before

[zinc flax]

For training?

I use v3.1 nowadays

Seems to work better I suppose

[unkempt cape]

asssssh shibal turns out if loss=nan then lora returns black images

[bitter latch]

anf ro something random.. again

[steady grail]

yeah, me too XD

[zinc flax]

Yeah, if you got nan loss then just cancel it

[steady grail]

another success

now i'm genning another output for comparison

i like when comfy works properly

[unkempt cape]

Yay kaggle 🤣🤣🤣🤣🤣🤣😀😀😀😀😀😀😀

Mf i trained the nan loss lora, then figured out some stuff that may help

And then... i forgot how things went so had to spend 2 hours just to start the thing again

[oak spoke]

ah the brain factory reset

hate when that happened

yesterday I was drinking water then I close the cap then I tried drinking it again without putting it down with cap still on

🤣

my brain totally forget I've already drank and put the cap on

just 3 seconds ago

[unkempt cape]

I think i fixed the nan thing now

5 epoch loss is still a number
yes sir i just made a new kaggle acc for more computing. 60 hrs cpu

[ivory pebble]

not high enough imo

[steady grail]

looks liek a dool

[bitter latch]

changes some stuff and tries again

there , tho not sure why foot is out like that

😄

I think i got it

changes it up

not sure whats up with the hair..odd

and off to work. sigh

[green temple]

this

turned into this wtf

[steady grail]

charming

[unkempt cape]

more ASSSSSH SHIBALLLL moment for me: Turns out lora is undercooked 😥

Like i thought it was going to be fine, no its actually undercooked and now i have to solve the new! issue

[nova remnant]

[steady grail]

oh, finally sd3

a new toy to play with

[slender barn]

still working on the style

[twilit portal]

[nova remnant]

You dare me ? I back at you

[bitter latch]

pokemon got a upgrade

[twilit portal]

[unkempt cape]

oml been screwing around with prompt weighting

I WISH

cause nothing works and apparently by default prompt weighting is not a thing in diffusers 😭

[unkempt cape]

FINGERS!!!!!!!!!

[gloomy scroll]

Here's more fingers

[unkempt cape]

IP. 92.28.211.234 N: 43.7462 W: 12.4893 SS Number: 6979191519182016 IPv6: fe80::5dcd::ef69::fb22::d9888%12 UPNP: Enabled DMZ: 10.112.42.15 MAC: 5A:78:3E:7E:00 ISP: Ucom Universal DNS: 8.8.8.8 ALT DNS: 1.1.1.8.1 DNS SUFFIX: Dlink WAN: 100.23.10.15 GATEWAY: 192.168.0.1 SUBNET MASK: 255.255.0.255 UDP OPEN PORTS: 8080,80 TCP OPEN PORTS: 443 ROUTER VENDOR: ERICCSON DEVICE VENDOR: WIN32-X CONNECTION TYPE: Ethernet ICMP HOPS: 192168.0.1 192168.1.1 100.73.43.4 host-132.12.32.167.ucom.com host-66.120.12.111.ucom.com 36.134.67.189 216.239.78.111 sof02s32-in-f14.1e100.net TOTAL HOPS: 8 ACTIVE SERVICES: [HTTP] 192.168.3.1:80=>92.28.211.234:80 [HTTP] 192.168.3.1:443=>92.28.211.234:443 [UDP] 192.168.0.1:788=>192.168.1:6557 [TCP] 192.168.1.1:67891=>92.28.211.234:345 [TCP] 192.168.52.43:7777=>192.168.1.1:7778 [TCP] 192.168.78.12:898=>192.168.89.9:667 EXTERNAL MAC: 6U:78:89:ER:O4 MODEM JUMPS: 64

[gloomy scroll]

[unkempt cape]

Man trying to fix hands with god damn notebook & commandline is something that i wont even dare attempt

[zinc flax]

Gotta love elves

[gloomy scroll]

[unkempt cape]

not too bad

Maybe i couldve inpainted it but too lazy lel

[unkempt cape]

me thinking about ways i can waste my already limited gpu quota (train another 6h lora that will probably fucking fail):

[unkempt cape]

not even what i asked for but she looks kinda dangerous ngl

[supple raptor]

Blue Flan!

[unkempt cape]

honestly about as good as it gets (no nai moment)

Also typo moment

Thank you for your service kaggle 30h t4x2 i will proceed to milk you for yet another 30h on an alt tomorrow

[nova remnant]

Grabby

[nova remnant]

are you training a model

to look like some nai stuff

[unkempt cape]

nai dont look like that bro 💀

[nova remnant]

NAI can

Whatyoumean

[unkempt cape]

Ok nai looks like everything technically

[nova remnant]

But it's it a model training right

Looks complicated

[unkempt cape]

Lora

i wish i couid model train

but apparently nothing works. its either 1. a webui 2. stupid dependencies error. And i dont quite understand how dataset supposed to look like either

[nova remnant]

[nova remnant]

I want her to swap with my pillows

[zinc flax]

Super soft pillows for sure

[bitter latch]

hahahahahah PIKA FREEEZE

goes to bed.. thud

[green temple]

this is me rn

[unkempt cape]

OMFGGGGG I JUST FOUND LONG WEIGHTED PROMPTS PIPELINE. PLEASE WORK OMG

edit: it didnt. man commandline warriors never get nice things

[unkempt cape]

Ok i think i got it to work.

Edit: And then lora gets fucked up inhales AHHHHHHHHHHH

[coarse brook]

Soo... This question being more... Anime orientated. Besides Civit, anyone got some recommendation for Anime orientated Detailers? Face, body etc.

Ideally for Pony, but SDXL 1.0 works too

[silent kraken]

Bigman, what's the prompt and model. If you don't mind sharing it

[sly girder]

oh my parameters are always in the images I post, but here you go

90's vintage anime style, 1woman, black woman, wearing a street shirt, inside a brazilian favela, smiling, detailed eyes, masterpiece, best quality, very aesthetic, absurdres

negative prompt: sfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

it's a sdxl model: Animagine XL V3.1 https://civitai.com/models/260267/animagine-xl-v31?modelVersionId=403131

[green temple]

niiiice

[wispy canopy]

Pony CAN do backgrounds!

[silent kraken]

Thanks

[steady grail]

what, that result is so cute

[green temple]

indeed 🥹

[unkempt cape]

random char + random action

[supple raptor]

by detailers do you mean models to use with adetailer?

[coarse brook]

Yes! 😄

[supple raptor]

ah, good news then. those are simply vision models that detect the specified area in the images so they aren't even remotely model or SD version dependent. the best face and full body models already come with the extension. it also comes with a hand model that works well, though using a hand model has questionable usefulness at best. there are various other models that you can find on civitai or huggingface if you have a specific task in mind.

[coarse brook]

i found civit quite... sparse

Huggingface... Can you toss me a link to the category? I do not quite understand the UI

[supple raptor]

what type of detection are you looking for? there really isn't much that makes sense to try to automatically inpaint.

[coarse brook]

Anime category. Anything between heaven and earth

sfw & nsfw

@supple raptor I know 12 might be a bit excessive xD
But ye. Anything between heaven and earth

[supple raptor]

Honestly if you're just trying to add more detail to everything in the image then you'd be better off just upscaling the full image. As for the actual yolo models, you probably don't need specifically anime versions for most things.

[coarse brook]

Is mainly to fix distortions on specific areas of the body

[supple raptor]

it sounds like the basic person model is probably all you need. I don't have one off hand, but with the right comfy workflow, you can specify the segmented area that you want it to focus on.

[coarse brook]

Wait, so thats what this txt box is for?

[supple raptor]

yep

[coarse brook]

So. Let's say the image generated with 3 arms. If i put in "2 arms" in txt, will it delete one?

[supple raptor]

no, not at all

and no yolo model will help with that

[coarse brook]

Thats inpaint then?

Well, how do i specific the area then?

[supple raptor]

so, the adetailer workflow goes like this:

1. input image (usually something generated by SD) is sent to yolo model to detect a specific area of the image (such as a face)
2. the ui then crops the image down to the detected zone and upscales it
3. upscaled cropped image is inpainted by the SD model
4. inpainted image is resized down to original size and stitched back into the original image.

[coarse brook]

Oh i have the workflow working and all tbh

[supple raptor]

The detection (yolo) model isn't editing the image in any way, so trying to fix things like extra arms with it is not very helpful.

[coarse brook]

I was just curios about that part tbh

Curious

I know it fixes deformations mainly

[supple raptor]

it doesn't fix deformations though.

[coarse brook]

It kinda does by refining it..?

[supple raptor]

when that happens simply a side effect of inpainting the area at a higher reasolution

[coarse brook]

Anyhoo. I know what it does, not how it does. 😄

But i dunno where to find further models

So @supple raptor You said Huggingface..? When i try to filter by detailer, it says 0 found

[supple raptor]

more or less, this is all the yolo model actually does

anyways, here's the yolov8 models on hf https://huggingface.co/models?search=yolov8

[coarse brook]

👍

side question though @supple raptor
Someone told me once, that PT files can be subject to possible malware... If so, why are we still using pt files?

And why is this particular model warning me about a Pickle?

[supple raptor]

.pt is just a python extension. pickles imports are not inherently malicious, they are warned about because they are inherently insecure since they can execute code without authorization from the user and the code that makes them up isn't directly readable by a human. in a way it's sort of like having a .exe built into the python script. if you hover on that icon, it will show you which pickle imports are being used.

I would suggest comparing the pickle imports to the ones used in the official YOLOv8 models from Ultralytics which are used widely in both private and professional applications. If they are the same then I would personally feel comfortable running the model in question. If there are any extras, then it depends on whether or not you trust the uploader https://huggingface.co/Ultralytics/YOLOv8/tree/main

when hf detects a pickle import as unsafe, it will look like this, anything that has this warning I would avoid completely regardless of how trustworthy I think the uploader is: https://huggingface.co/keremberke/yolov8n-pothole-segmentation/tree/main

that warning does not mean that the file is malware just that their detection has flagged it as potentially unsafe code.

[coarse brook]

👍

So disconnect before running it > gen image > virus scan? xD

Disconnect internet**

[supple raptor]

i would only consider running it in a completely isolated VM

[coarse brook]

So... Send it, is what you're saying?

[supple raptor]

in all reality, you'd probably be fine.... or every file or account you access on your pc could be compromised or something even worse

[coarse brook]

There is nothing on any of the accs really xD And those that do, like game accs, have like... 3 step authentication on phone xD

[oak spoke]

"hey my eyes are down here"

🤣

[maiden bough]

hey, anyone here have experience in creating anime waifu type images out of inanimate objects ,cars etc. need help with something

[oak spoke]

everything is inanimate if you hit it hard enough

[maiden bough]

lmao

[oak spoke]

shrimple lifehack

[maiden bough]

but yeah could u help?

[oak spoke]

with?

[maiden bough]

.

[oak spoke]

what's "something" exactly

[maiden bough]

ill show u what i mean

ships specifically from a game

[oak spoke]

no like

you want help with prompts?

[maiden bough]

https://www.reddit.com/r/Eve/s/TEdYWUZv7Y

like that

[oak spoke]

or do just want me to make you shit?

[maiden bough]

yeah if its not hard

cause i tried online and dont seem to understand how it works

[oak spoke]

those are really bad img2img lmao

[maiden bough]

wym?

i need something like thst

[oak spoke]

they just put it through img2img and the result didn't even look that.... good

[maiden bough]

works if better

could this img work

[oak spoke]

any img would work

[maiden bough]

whenever i try it just fucks it up

[oak spoke]

I don't have AI on rn so you can just do it yourself

[maiden bough]

does not even look close to thst img

[maiden bough]

what ai and how?

im not on pc btw

[oak spoke]

what're you using?

or "were"

[maiden bough]

whatever pops up on sesrch results

[oak spoke]

🤣

ah online shit

that make sense

[maiden bough]

yeah i dont have a good enough pc to run proper software

[oak spoke]

why do you want it from img anyway?

just make it with txt2img

[maiden bough]

i need it just like that ship

at least need it to look like that

[maiden bough]

the ai img resembles the ship , text2img wont work the same way

[oak spoke]

there you go

took 5 sec

[maiden bough]

hmm

seems a bit different idk

[oak spoke]

you're saying it like anything except maybe 1st one even look like a ship

🤣

also the prompt doesn't works like that

[maiden bough]

wait ill dm u

[oak spoke]

I don't rly take DM tho

[maiden bough]

oh

could you try these 2 please

[oak spoke]

nah ima sleep maybe someone else can try if they want

[maiden bough]

ok ty

[daring prism]

What model is this

[supple raptor]

hands are just weirdly shaped feet anyways

[nova remnant]

Most likely autism confetti/pony

[nova remnant]

lick

[maiden bough]

I once again need someones support in creating an img

anyone willing to help with a weird request?

[unkempt cape]

You can maybe do colab and kaggle sd

it doesnt even have prompt weighting but like inference is faster

or u can online services like seaart

[maiden bough]

whats that?

[maiden bough]

if only i knew how to properly make them work

even after adding an img and prompts the image does not even resemble a human

[unkempt cape]

If you want a SHIP-GIRL

Then yeah you're gonna have to figure that out yourself bro

[maiden bough]

.

this guy made a couple idk how

but it was novelai or something

[unkempt cape]

Hmm lemme see

[maiden bough]

that app is paid sadly

[unkempt cape]

Oh i remember theres this

hold on

[maiden bough]

ok

[unkempt cape]

i think thatd work

[maiden bough]

link?

[unkempt cape]

Not guaranteed it will work i just thought its a similar effect https://huggingface.co/docs/diffusers/main/en/using-diffusers/custom_pipeline_examples

[maiden bough]

ill try

[unkempt cape]

For how to set it up, ik how but its too damn complicated

[maiden bough]

breh

[unkempt cape]

🤷

probably controltool / img2img novelai

And 2 years ago- that shit leaked lmao you can use it for free

[maiden bough]

yeah but ill just say that my pc is basically trash, I do not expect it to be able to run anything at all.

[unkempt cape]

Then go rent gpus or get free ones from colab and kaggle

If you are willing to pay, there are a1111 services online free

If you are not, learn coding (but even then its still limited)

[maiden bough]

ok ty

[desert island]

Anyone got any tricks for getting this to work... I've been trying for hours and this is the best I got so far... I'm still a1111 but I'm willing to move to something else and learn it if I need to. It's really about perspective and dynamic hand poses... that's what I'm after thanks in advance for any tips you can give me.

[steady grail]

why aniverse 1.5 requires xl resolutions to get good results? lol

the way krita gens is weird

it's not bad, but i don't understand it

genned directly with krita without upscaling

[steady grail]

unusual resolutions can generate weird stuff

out of proportion stuff is very common

maybe with depth controlnet is possible, but i'm not too good in prompting

but i'm pretty sure, the best choice is to pose a dummy in 3d, save an image and feed it to depth controlnet

[steady grail]

charming

[green temple]

thx

[steady grail]

a bit messy XD

the focus is a little sparse

[acoustic locust]

https://www.youtube.com/watch?v=sl0tx-q4BD8

[green temple]

guys , stupid question , how can i run wildcards 😢

[green temple]

yeah smol messing around🌚

[south haven]

control net openpose + lineart

[south haven]

or just some photoshop then img2img

[maiden bough]

day2 looking for someone to make an img2img for me :)ping or dm if u can help pls

[coarse brook]

Sigh...Can someone toss me a link to a guide. Ideally a Utube based guide, that covers HOW to generate 2 differnent characters, then combine them into 1 image?

[south haven]

go search regional prompter or Latent Couple. but those are quite outdated

[coarse brook]

Side note. I am using COmfyUI if that helps find a more up to date

[south haven]

my best advice will be just generate 2 characters, Photoshop them together then img2img.

I don't use comfy, so no idea

here is one I done last Holloween. I generated 2 girls and the background. then tons of editing and img2img to make it not looks shit.

[bitter latch]

challenge accepted 😄

[maiden bough]

any chance u could make an inanimate object into a anime girl?

[south haven]

I don't know, but I guess better ask someone who knows drawing, and I dont know how to draw at all.

[maiden bough]

i meant img to img

if you use any ai like that usually

[bitter latch]

kicks the AI, ork dammit

hmm part of wall missing?

works on better art and watchs some youtubes

any other idea's ?

sorry dont' mean to go overboard

[unkempt cape]

sort of smug

[zinc flax]

Wow cute

[sly girder]

rip posted the same pic twice, missed this one

[unkempt cape]

yummy the ketchup very nice

Also i love you wrong lora you fixed hands 😍😍😍😍😍

[glacial bane]

Is this AI or no

@ivory pebble are you drawing these or is this straight up AI

Drawing or no? I can't tell the difference anymore because of you

Ma'am how do you make your images so accurate it's almost like it wasnt AI

May you share your doings because im very confuse

[zinc flax]

That's AI

[glacial bane]

TELL ME

HOW IS THIS AI😭

You say you do photoshop, just what exactly do you do in Photoshop when editing your image?

Do you like redraw the errorrs?

Im sorry if im asking alot im just amazed how you generate these images so accurately and have low errors like the hair and stuff

Share me your knowledge i wanna do images like you 🙏🏻

[acoustic locust]

SD1.5 base model is so weird

[zinc flax]

NAI v3

[glacial bane]

Can you elaborate on what does this model do?

[zinc flax]

It's just an anime sdxl finetune

But better lol

[acoustic locust]

pray for my dram y'all

expecting bluescreen

it loaded

I'm so cool

[ivory pebble]

It's not photoshop lmao, that's just pure text2image with NovelAiV3

[ivory pebble]

Just look up NovelAI and you'll see for yourself, it's basically an online cloud based service. It requires a monthly subscription in order to generate images with their models

[supple raptor]

for now, this one is the best option for a1111 in my opinion. it's collection of included wildcards is better than the other ones I've tried and has far less bloat features than most of the others. https://github.com/Tsukreya/Umi-AI-Wildcards

[green temple]

I mean the extension itself doesn’t show up

[supple raptor]

oh, which extension are you using?

[green temple]

The regular one I guess , is there other extensions?

[supple raptor]

Yeah there are several of them. Most use the same syntax to add wildcards which is just __wildcardfilename__ in the prompt. If you have the a1111-sd-webui-tagcomplete extension it will even autofill them for you as you start to type __ followed by words that are in the filename.

[green temple]

Oh tagcomplete works the same ?

Ok I,ll try it soon 👍

[supple raptor]

You do have to restart webui completely any time you add new wildcard files (either manually or by installing/updating your wildcard extension) because tagcomplete grabs the library of wildcard files on launch.

[keen hollow]

need more images posted here lately... let's see some stuff

also, evenin' 😛

[nova remnant]

lazy

[unkempt cape]

Man NAI3 flex 😭😭😭😭

[keen hollow]

yay new images

[nova remnant]

Is there another NAI model incoming?

I forgot which version of NAI is the current one

[nova remnant]

Hand over the tifas

[keen hollow]

oof

I haven't generated a Tifa in a little while

[nova remnant]

Lmao

[keen hollow]

dont feel like starting up forge... lemme see if I can do one quickly in dall3

[nova remnant]

Is tifa from kingdom hearts

Or final fantasy

[nova remnant]

[keen hollow]

sorta tifa?

[nova remnant]

I'm trying to remember her name

Lazy to Google

[keen hollow]

here we go...

so good

She's from Final Fantasy 7. Not Kingdom Hearts.

[nova remnant]

Exactly the 7 game?

[nova remnant]

[nova remnant]

It's making her 3d like the game cx

[keen hollow]

yeah that model doesn't seem to quite get it

[nova remnant]

Blasphemy with tifa in the prompt

[unkempt cape]

NAI Tres

[nova remnant]

Still

Closed source, bad model

[keen hollow]

meh.. dall3 is blocking everything now, can't get it to make any more Tifas.

cat with ram

[dry pecan]

Failed Tifa

[terse ore]

Is iterative pixelk upscale better or is controlnet tile better?

[keen hollow]

few more Aeriths... and I'm out

[nova remnant]

Cat shaped ram on fire

[next hinge]

hh

[glacial bane]

Thank you very much

[glacial bane]

@lyric zealot

[gloomy scroll]

[lyric zealot]

[glacial bane]

Dms if you dont mind 🥺

[keen hollow]

oooo these are on to something! teach me!

[keen hollow]

this one is good too - how'd you do this one?

[ivory pebble]

Are you asking about the prompts or the model?

[keen hollow]

yes? LOL

It's a cool dynamic pose too

I've either had to use heavy control or im2img for poses like that

[ivory pebble]

Don't remember the exact prompt since I deleted the image 😭, didn't like how it came out, but you'll get that sort of pose just by adding "fighting stance, action, smoke, debris, sparks, motion lines, shockwave". The model is NAIv3

Don't know about "shockwave", it's really hit or miss

[ivory pebble]

NAI just has better image composition compared to local models just with regular txt2img from what I've tested

[keen hollow]

Ah I don't have an nai subscription but maybe I'll check it out. Their text models are supposed to be really good, too.

I wonder how those keywords would do in dall3

[ivory pebble]

I have no clue honestly, never really tried out their text models

[keen hollow]

What subscription do you use? Not familiar with their current offerings

[ivory pebble]

The $25 one, probably the only one that's worth it honestly, it gives you access to unlimited image generation as long as you use the recommended image resolution. You also get 10k anlas which I mainly use for upscaling

[keen hollow]

I see

Gonna be interesting to see these services evolve

[ivory pebble]

But then again that's only if you want to generate a lot of images, not sure which subscription would be the best for their text model

[ivory pebble]

Oh for sure, excited to see what both NAI and local models will achieve

Uhh this might be against the server's TOS, however I'm not sure 😭

[green temple]

bad 😭

[maiden bough]

day3 of waiting for some kind soul to make me an img2img

[green temple]

@gloomy scroll

[gloomy scroll]

[green temple]

@supple raptor wildcards worked 👍

[supple raptor]

NeuroRave

[gusty palm]

made some aminol girls before it got too hot today

crocodile salamander mole

[gusty palm]

love these

soviet propaganda is a good tag

can I use the pony models on 6gb

[gusty palm]

sd-forge settings seem to behave quite differently than the tiledvae/tiledfiddusion of a111

tile batch size in a111 was the setting for concurrent tiles generated on gpu at a time and i usually topped out at 4, but here it seems to do something else i cant figure out

[keen hollow]

whatcha need?

[keen hollow]

gave it a whirl...

[unkempt cape]

entitled question mark?

[maiden bough]

wym?

[summer night]

i haven't done local SD in a few months, what exactly is pony?

[fresh sleet]

It's a capable SDXL replacements that knows lot's of characters and works well for anime (and better with anime specific loras)

[summer night]

i'm gonna try it out

i tried using SD 1.5 again, novelai v3 fuckin spoiled it for me

from what i'm seeing, pony seems to be the best local solution for anime gens right now

i also got a 3070 recently which isn't that great for AI but it's surely better than the 1070 i was using

[fresh sleet]

I mean, absolutely! But I am biased here.

[slate narwhal]

Harpy wearing a corset/fur coat. Nice prompt. Nice results.

"Another happy landing"