🍥｜anime

[gloomy scroll]

||Tbh I have never watched Ina stream but I own Ina nendoroid||

[bitter latch]

random mode engaged

"I have broken thru"

[nova remnant]

||please explain||

[nova remnant]

donuts

[bitter latch]

"hello can I come in?"

"boo"

huh...?

[gloomy scroll]

Yoink

[bitter latch]

ummm hmmm what?

demon donut?

crystal donut

ooops

[keen hollow]

crystally

[bitter latch]

random donut

[keen hollow]

have ya'll tried this "Krita" integration?

[bitter latch]

nope

[nova remnant]

I'm hungry what's she making

[dapper verge]

A divine celestial being, angel wings widespread and feathers glistening, a radiant blue eye in place of a face, glowing aura, standing in a kitchen, hilarious chaos as she attempts to bake cookies, flour covering the wings, dough stuck to the feathers, celestial light accidentally baking the cookies instantly, heavenly beams reflecting off mixing bowls, an ethereal spatula in one hand, cosmic light still shining through the kitchen window, a sense of divine confusion and amusing struggle.

cookies apparently

[nova remnant]

i want

[gloomy scroll]

https://tenor.com/view/aki-chainsaw-man-gif-27647022

[nova remnant]

bnuy

[gloomy scroll]

[south sinew]

What model u used?

[dapper verge]

It's pixart refined with aam xl

[south sinew]

What is pixart model or stable diffusion websites?

[dapper verge]

https://github.com/PixArt-alpha/PixArt-sigma

[south sinew]

Can i run this on phone?

[acoustic locust]

What model did you use?

[dapper verge]

You can probably find a pixart huggingface space, and you can probably run aam xl on civitai, but you'd have to run one refined with the other on a computer.

[south sinew]

Thx but can u give me ur settings

[dapper verge]

Sure I'll paste the json

[south sinew]

Is there on hugging face stable diffusion like automatic 1111

[dapper verge]

Do you use comfyui? That's what I've got the settings for. You can't refine a model with another on a1111

[south sinew]

I cant run a1111 on colab

I use TensorArt and seaart

TensorArt

Ur better

[dapper verge]

https://tensor.art/models/684315249756803520

So this is close enough. It should give you good results.

[south sinew]

I tried this the result is just noise

[dapper verge]

Hmm, not sure what's going on there

[south sinew]

Is there a lora or model have results like niji journey?

[dapper verge]

Hard to say. I like aam xl. There's also neta art model. There's also deep blue on civitai that's good.

Aam xl has that sharp line look.

[ivory pebble]

NovelAI

[glacial bane]

WOA- FINALLY A ZZZ FAN

[glacial bane]

MORE MORE ZZZ

[glacial bane]

Which character?

[ivory pebble]

Hasumi from blue archive

[glacial bane]

I dont like it.

I love it

[formal shoal]

a have a lot of zzz

but some are not nsfw

(most of them)

[glacial bane]

May you pretty please send me them

[glacial bane]

Everything

Pls

[keen hollow]

interesting colors in this one

[glacial bane]

Someone found an interesting outfit

[ivory pebble]

Just wanted to see what she would look like with that outfit, that's all

[glacial bane]

Suree sure next thing you'll know yor is wearing it

[hollow moss]

What model are you using?

I find it similar like this one

[glacial bane]

He says NovelAi the paid model

[glacial bane]

@formal shoal

[formal shoal]

I have a pixiv where i upload them

[ivory pebble]

Novelai

[glacial bane]

Oh umm send

[formal shoal]

Okey md

[glacial bane]

Ty

[formal shoal]

Give me a moment

[formal shoal]

I cant dm you

[lyric zealot]

[glacial bane]

Umm add me ig

[hollow moss]

Is this novelai?

[glacial bane]

She's sucha mommy

[keen hollow]

have no idea who that is but thought you'd appreciate those 😛

[keen hollow]

Don't know. But you can get to this with artist tags in SD. Check this post of mine from a while back for some artists to reference. #🍥｜anime message

[glacial bane]

Oh ty i really do ty tysm

[keen hollow]

@hollow moss , here's an attempt at my own versions:

[lyric zealot]

[ivory pebble]

Yup

[formal shoal]

what artist tags did you use to get that

[keen hollow]

teranekosu, ciloranko, ariimaw, konya_karasue, ask_askzy, ningen_mame

[formal shoal]

thanks bro

[nova remnant]

Too small

[keen hollow]

you bet

[ivory pebble]

I knew it

[formal shoal]

nah, id win

[rough oxide]

whaddabout winning

[rough oxide]

wut r we competing on

[formal shoal]

i made that image with the kashimo pose

of jujutsu kaisen

[formal shoal]

cool generation btw

[nova remnant]

Not cute

[rough oxide]

cat still no like pure anime screencap style ;_; lol

this one is moar stylized

ah yes i have no idea how to even start to do a pose like that xD

[formal shoal]

you have to use a lot of prompts to get anime screencap style

[rough oxide]

yea i can normally get anime screencap pretty easily with my setup rn

[formal shoal]

controlnet

[rough oxide]

ooo

i never played with cnets

[nova remnant]

Booba

[rough oxide]

what controlnet model is good for anime?

swarmui supports it but i never touched the feature xD

[formal shoal]

you use stable diffusion right?

[rough oxide]

ye.

[formal shoal]

i use open pose

if better use real person images for the pose

[rough oxide]

do they work for sdxl based models?

this is wut i see on my ui for controlnet

but not sure if u need diff stuff for sdxl. i guess i should look into it

[formal shoal]

i thonk yes

[rough oxide]

(also oops now that i posted that i noticed i put 1girl twice in my curr prompt lmao)

[formal shoal]

maybe you have to download it

[rough oxide]

yea, doesn't come with any controlnet models prebaked

so i'd have to find one

[green temple]

[rough oxide]

https://huggingface.co/xinsir/controlnet-union-sdxl-1.0 know if this one is any good? saw it recommended somewhere and the desc looks good at least

[keen hollow]

surprised cat kid. Is that those artists?

[green temple]

yup

[keen hollow]

yeah it's an interesting combo

messy paint strokes and stuff

i think animagine was a good model for it

[nova remnant]

smol

[rough oxide]

kitties r kawaiii

that's one plus-sized elf

[nova remnant]

big

[nova remnant]

rejects

[nova remnant]

The image representation of the entire plot of Steins;Gate 0

[steady grail]

1 or 2?

[steady grail]

i love zero 😦

[nova remnant]

Me too but all the way until like episode 20 it's all burning around him and he's like "This is fine" then he kicks into the usual Okorin bad ass mode

[steady grail]

eheheh

after the hell he lived, i suppose that became his world lol

[nova remnant]

He's probably my favorite male character in anime

[steady grail]

same prompt, different res

totally agree

[green temple]

1

[steady grail]

yeah, i like the details too

i wonder if i can find a way to just import some of the creativity of the 2° mix

[nova remnant]

put (creative:1.5) in the prompt

👍

[rough oxide]

if only it was as simple as (creative:1.5) kekw

think this goes to a badware site now? might wanna update link/remove pin

[rough oxide]

what model is that one?

[mental marten]

#🍥｜anime message```yaml
best quality, masterpiece, score_9, score_8_up, rating_safe, bold line, medium breasts, mature female,

1girl, solo, nakano yotsuba, orange hair, green hair ribbon, blue eyes,

(bee (deadflow):0.8), ilya kuvshinov (style), wlop, realistic, (ligne claire:1.1), aogisa, (shion (mirudakemann):1.1), (cogecha:0.9), sho, (solipsist:1.1), (impasto:0.7),
Negative prompt: bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, (worst quality, low quality:1.1), jpeg artifacts, signature, watermark, username, white background, simple background, (gpo, aca, aer, api, fla, gcx, hll, hnj, gpc, fii, fey, fbv, evg, iew, ifl, igh, iwj, iwp, ixb, ixe, ixz, jaf, jbm, jfb, jsf, jyk, kmz, ksh, kxg, kzg, lbv, xsl:0.5)
Steps: 24, Sampler: Euler a, CFG scale: 7, Seed: 3442527685, Size: 768x1024, Model hash: 89e7c7518c, Model: tPonynai3_v55, VAE hash: 66f78ad136, VAE: sdxl_vae-fp16fix-blessed.safetensors, Denoising strength: 0.4, Clip skip: 2, ControlNet 0: "Module: CLIP-ViT-H (IPAdapter), Model: ip-adapter_sdxl_vit-h [68f9f08d], Weight: 0.15, Resize Mode: Crop and Resize, Processor Res: 0.5, Threshold A: 0.5, Threshold B: 0.5, Guidance Start: 0, Guidance End: 1, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Both", ControlNet 1: "Module: CLIP-ViT-H (IPAdapter), Model: ip-adapter_sdxl_vit-h [68f9f08d], Weight: 0.2, Resize Mode: Crop and Resize, Processor Res: 0.5, Threshold A: 0.5, Threshold B: 0.5, Guidance Start: 0, Guidance End: 1, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Both", Hires upscale: 1.5, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Version: f0.0.17v1.8.0rc-latest-269-gef35383b```

[lyric zealot]

^ this, but then im using ip-adapater as well

[rough oxide]

how do you do that view param thing

and ooo

i looked at tponynai3 a few times

but never tried it

[lyric zealot]

ahh that's my own bot but people can "install" that bot to your own user profile and be able to use those features as well

[rough oxide]

oooo

super useful, didn't know that was a thing now

[lyric zealot]

click on it's profile, and add app, then to your own profile

[rough oxide]

ye; just did that

[lyric zealot]

haven't seen much pnginfo bot around, so yeah

or should i say, one that can be installed to user profile

[rough oxide]

i think i found the ip-adapter, that's just revision right? never used that before

[lyric zealot]

anyway, if you ever intend to remove it, just go into your settings then Authorized Apps and remove it from there

[lyric zealot]

it's kinda like vibe transfer if you ever use Nai

[rough oxide]

got it; probly will keep it though for quick viewing of params so i don't have to like dl and drop in a metadata viewer lol

Ohhh

i absolutel ykno wut that's like then

so you can use an image and like percentage influence it?

....i have no idea what model folder that goes in though

[lyric zealot]

well, to use it, you can right click any pic on any server or dm, then use view parameters, the one with (public) means it's not invisible to anyone. The default is only visible to you

[lyric zealot]

yup, just one weight only tho compared to nai info extracted and reference strength

it should be a controlnet thingy, so under controlnet?

[nova remnant]

@lyric zealot stooooooooooooooopid

[nova remnant]

@rough oxide send me a cool (non anime screencap) image

[rough oxide]

i haven't played around a lot with the pony models yet

cuz i prefer anime screencap style xD

but

[lyric zealot]

don't do it, @nova remnant is gonna vibe transfer your stuff

[rough oxide]

whaddabout this?

lol

iz fiennnn

[nova remnant]

👍

[rough oxide]

been using animagine & animagine merges (currently anima pencil)

keep hearing pony is super gud tho but haven't found gud pony merge yetttt

i wish t-ponynai3 had like a recipe sheet

but i might try it

[lyric zealot]

i think it's a merge of generated output of NAI and pony

[rough oxide]

ahhhh

i never really got those like generated output ones

[south sinew]

Which one better?

[lyric zealot]

Training Instructions：Merge Lora used 94 pictures for v1, 119 pics for v2, 348 pics for v3, 474 pics for v3.5,which generated by NAI3 to train into the basemodel for fine-tuning,Pony supports all artist tags which ponyv6 already have, but there is no any addition artist tag from nai3. Using more than two artist tags may cause background crashes,At present, it has been found that characters that can generate Genshin Impact.I don't know the others.I haven't tested much for this model.,Marvel at its reproduction of the painting style of NAI3.The base model is a fusion model of T-anime-xl and ponyv6 and animage3, which has not been released

[rough oxide]

usually not as good quality as just like using danbooru or something

[rough oxide]

oooo

but it uses animagine as part of it?

only 3 or 3.1?

[lyric zealot]

are they same seed?

[south sinew]

Yea

[lyric zealot]

no idea

[rough oxide]

o yea

i see where u got that frm

but if it uses animagine then worth a shot if it can make images of that qual

[south sinew]

But my lora want more work cuz training on phone

[rough oxide]

btw is it ip-adapter_sdxl_vit-h or the plus one?

[nova remnant]

vibe stuff

[lyric zealot]

comparing the two, i think with lora is slightly better

[rough oxide]

cute bat wings

[south sinew]

Can u give the prompt i want try with mine lora

[lyric zealot]

im using CLIP-ViT-H

[rough oxide]

wait it's not one of these?

[lyric zealot]

and model ip-adapter_sdxl_vit-h

[rough oxide]

o wait u have to like

change the entire clip thing too?

xD

[lyric zealot]

something like that

[rough oxide]

hmmm

not sure where to get the clip-vit-h one then

[lyric zealot]

no no

i think you just get the model

preprocess is bulti-in

for controlnet

i think that's just some script and stuff

[rough oxide]

oh ok

[nova remnant]

I didn't prompt for much, its only like "1girl, cowboy shot, dress, maid, broom, looking at viewer, ponytail"
since its using image references

[rough oxide]

so then the only diff between these 2

is plus gets closer to reference image i guess

is how it describes it

[south sinew]

With mine lora:

[lyric zealot]

im not sure why i chose what im currently using tbh

[south sinew]

Guys what is best training tool in google colab

[rough oxide]

hmmm wonder if it's not controlnet section lol

o

clip_vision folder i think

[rough oxide]

for stableswarmui

[south sinew]

What is this

[rough oxide]

looks like separate folder just for that

[lyric zealot]

[rough oxide]

o i might've messed up do u actually use it

[lyric zealot]

you might have to find it then

[rough oxide]

like a controlnet

lol

[lyric zealot]

i certainly use it like a controlnet

[rough oxide]

hmmmm i'm not exactly sure how it works in this ui

[south sinew]

Is this a1111?

[lyric zealot]

see there's two more pic, that's the two ip-adapter pic

[lyric zealot]

that's stableswarm

[south sinew]

Oh

[nova remnant]

@rough oxide she the devil

[south sinew]

Try with AAM anylora screencap

[rough oxide]

ye; sorry just looking over it, it's swarmUI

[nova remnant]

I'm not using A1111 or local its NAI

[south sinew]

Oh

[lyric zealot]

stooopid

[south sinew]

Niji journey:

[nova remnant]

[south sinew]

Niji journey also

[lyric zealot]

she's gonna beat your ass

[south sinew]

WOW it's Friday

[lyric zealot]

https://tenor.com/view/its-friday-then-gif-20650284

[lyric zealot]

[nova remnant]

I recommend my vibe transfer for those 👍

[lyric zealot]

[south sinew]

I will try with another chkp

WTF HAHAHAHA

[lyric zealot]

[south sinew]

What midel u use

[lyric zealot]

@nova remnant is using nai for his last few posts

[south sinew]

I use TensorArt

[nova remnant]

https://tenor.com/view/darth-maul-gif-21544920

that broom

lmao

[south sinew]

😂

Ahm what is the model used in nsfw ai arts 🙄

In pixiv 🙄

Cuz it looks like anime drawing

[nova remnant]

nsfw

im innocent

[south sinew]

😂😂

[lyric zealot]

s-surely @nova remnant is the most innocent person here

https://tenor.com/view/when-the-banana-sells-a-million-dollars-gif-11873212726979639512

[nova remnant]

👍

[rough oxide]

i think nai's greatest value is in its vibe transfer stuff. i was able to do a lot of neat stuff w/ that. but now there's soooo many good checkpointsss. animapencilxl is absolutely my fave rn

[south sinew]

Who know the model

[rough oxide]

wut model?

[nova remnant]

its good with artists and stuff

but not much more than anime/danbooru tags and whatever its special tags are

also kinda dumb sometimes need to keep weighting stuff

[rough oxide]

yeaaaaa

one of the prompts i was workign with i had to use so many weights on nai

[south sinew]

The model artkoikoi uses

[rough oxide]

artkoikoi?

[south sinew]

Yea in pixiv he post nsfw content

[rough oxide]

they probably mix some anime screenshot model (animagine xl3.1 base for example) w/ loras

but they dun have params on their posts

so unless they say what model they use

[south sinew]

He's not the only one who uses it, it's similar to the anime

[rough oxide]

then an upscaler on top of that

[south sinew]

The model also

[rough oxide]

if u like this screenshot style i can tell u what this model is

[south sinew]

Not look like his arts

[rough oxide]

ye, like i said they probly apply lora too

i don't use loras

[south sinew]

His style better than Niji journey

[rough oxide]

what i'm saying is you have to combine a model + additional things like loras/embeddings/controlnet to achieve like perfect quality pics

but it's some anime screenshot base model

then also they use an upscaler of some kind

[south sinew]

It's like a scene from chainsaw man

But its ai with his model

[rough oxide]

it could also be NAI cuz i notice power's eyes are messed up

i know it can handle makima's eyes but not power's eyes

and nai allows nsfw

it could be any number of things really

you'd have to ask them directly

[brittle rampart]

Why not just ask the author directly? but I’m sure that in addition to the generation itself on the model + Lora`s + Upscaler, in each of his work there is a significant share of post-processing. only on one model such quality “cannot be achieved”

[wheat falcon]

Thanks, I removed it

[brittle rampart]

If you look at models on CivitAi, you see these marks. This is not entirely clear to me.

[nova remnant]

feet

[rough oxide]

checkpoint are full models, all things baked in, sometimes pre-baked with loras/embeddings/etc. for example animagine xl3.1/pony is checkpoint

[nova remnant]

Pony IS a SDXL fine tune

[rough oxide]

pony means that

[nova remnant]

But the way it was trained, made it not be compatible with normal SDXL loras and stuff

thats why there is a tag/mark on the civitai site

[rough oxide]

whoever made that model trained that specific checkpoint ontop of pony for instance

[nova remnant]

to indicate that is compatible with that stuff

[rough oxide]

ye; pony's tagging system is far different than any other model's as well; so it's just notifying u that model uses pony in its merge

and in most cases it'll probably be a major part of the merge/continued training

[brittle rampart]

It turns out that when using such models, should I adhere to the methodology of constructing a prompt using the "Score_N" tags?

[rough oxide]

yes when it's the pony checkpoints

no when it's something like an animagine ckpt

[brittle rampart]

It turns out that all I need to know is that if this model is based on a pony, then I need to build a hint using the "Score_N" tags

[brittle rampart]

😉

[lyric zealot]

the best thing to do is to visit that model page, and see how the author is prompting, since some of the merge or even finetuning requires much lesser score_tags

[brittle rampart]

Certainly. Moreover, the authors sometimes explain some of the nuances of using

[nova remnant]

its tomato juice..

[brittle rampart]

of course of course! it's just ketchup 🤣

[south sinew]

Navia from genshin impact with my lora

[tidal sundial]

https://www.youtube.com/@PureFxAi

[brittle rampart]

One-Punch Man

[unkempt cape]

ads ?

[steady grail]

well, the content is ai after all

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1264935821736411237/757b3fb5-dcaa-433a-8ad8-c8b1027c6f2a.png?ex=669faec9&is=669e5d49&hm=7889777cc8c618ebbd4eb0193e531cb2a82e17520982d193b48e8286f16164ea&=&format=webp&quality=lossless&width=266&height=393

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1264940515971956767/705ba765-e33e-4b70-928c-b0e90c076df7.png?ex=669fb328&is=669e61a8&hm=f76a08b0a008de1e82a7947614580c9d8a057327cb2b501f1893ea785596945b&=&format=webp&quality=lossless&width=266&height=393

[steady grail]

a big chicken cooking chickenn legs...

[bitter latch]

better

[bitter latch]

well that came out interesting

[summer night]

[green temple]

prompt

[brittle rampart]

(khyle:1.6), 1girl, solo, (preteen inugami korone:1.3, hololive, hold positive pregnancy test:1.15), sunset, lens flare, closed eyes, >_<, happy smile, head tilt, (spoken heart mark from left), from above, dynamic pose, dynamic angle, intricate shadows, best quality, amazing quality, very aesthetic, absurdres

[green temple]

didn`t work

[brittle rampart]

Yes. It doesn't work the first time. 😁

[nova remnant]

[brittle rampart]

https://tenor.com/view/ouch-gif-12136515515962044163

for what?

[nova remnant]

For being silly

[brittle rampart]

in just a few words?! 😁

[nova remnant]

her arm

[brittle rampart]

it `s Magic!

🤣

[green temple]

[brittle rampart]

😝 BAKA!

[green temple]

baka wa omae da !

[brittle rampart]

not me ! he!

[nova remnant]

big bonk from big elf

[brittle rampart]

I admit - I was wrong! where to put the cross?

🥴

[nova remnant]

smug cat mouth

[nova remnant]

What did I do

[green temple]

she is scolding you cuz of booba

[nova remnant]

steals as my vibe transfer

[nova remnant]

im innocent

[green temple]

big still

not what i wanted but still ok

[brittle rampart]

"no Booba" team

[nova remnant]

police is here to take you

[brittle rampart]

nobody is at home!

[nova remnant]

Bang bang

Misery is here

takes out the goth prompt

[brittle rampart]

all in anticipation

[nova remnant]

the bot doesn't want to work

gonna have to wait for the goths

@nova remnant

big

[nova remnant]

Goth women are my weakness

[brittle rampart]

Smol!

[brittle rampart]

colorful

[nova remnant]

[nova remnant]

Smol on water

[south sinew]

Sdxl look like anime 💀

[rough oxide]

like... base SDXL?

[bitter latch]

in the army now!

there

[nova remnant]

no, the cookies I ate it all

don't make that face to me

[nova remnant]

would

[lyric zealot]

would what

[vital raptor]

eat her

extra donut

[lyric zealot]

[brittle rampart]

and do it quickly and quietly

[lyric zealot]

[vital raptor]

realistically, aqua is annoying af, i wouldn't be anywhere near her 😄

[rough oxide]

just summon a familiar to grab the donut for you.

[lyric zealot]

she's just there to get bullied

[brittle rampart]

should I contact Megumin?

[rough oxide]

i just did an img of komekko but it wasn't gud enuff

xD

not enuff komekko data.

[vital raptor]

just send Darkness after the donut, at least she'll enjoy it if she gets caught

[brittle rampart]

in any case, Kezuma will end up cleaning up the mess. 🤣

[vital raptor]

that reminds me though, i haven't watched the newest season yet

[rough oxide]

D'OH nuts successfully acquired.

[south sinew]

anyone know the name of this model?

[rough oxide]

if it's not in the metadata it can literally be like any combination of model+lora lol

the quality looks a bit lower than the curr models tho

[south sinew]

So do you know the model?

[brittle rampart]

you can look in the metadata of the image file. Of course it should not be a jpg format. and hope that the author did not delete the metadata.

[rough oxide]

there's literally so many models it's impossible to tell. if you put the time into it you could probly get similar results with a screencap model + lora for ganyu. but here's some lora-less ones from the model i use.

the only way to tell for certain is if there's metadata

the image you posted has no metadata at the very least

it's been mentioned before but your best bet if the metadata is wiped is to ask the creator

but it's probably an older model using a ganyu lora cuz the eyes are incorrect

[brittle rampart]

at least an image search yields this source. but I'm not good at Korean(?)

https://www.fmkorea.com/index.php?mid=genshin&category=4367003870&document_srl=6985646077

[south sinew]

How this

[brittle rampart]

pretty similar

[nova remnant]

anime

[vital raptor]

yes, that is the name of the channel we're in...

[nova remnant]

[lyric zealot]

[nova remnant]

Big

[south sinew]

🤨

[brittle rampart]

in some places and parts!

[rough oxide]

smol is bestu

[nova remnant]

what's that on the floor

[brittle rampart]

This is a Yokai Summon Card!

😁

[nova remnant]

[rough oxide]

couple of messy gunblade attempts and a random kiara because why not.

[nova remnant]

is there like any quality or style modifiers for pony diffusion? It's all I can find loras for anymore and it's always the same exact thing every single time I generate an image with no variation. No way to add more details or anything

[brittle rampart]

there are lora`s that add details

[nova remnant]

You can use artist tags

Not much from prompt only other than that and the score, source anime, negatives..

Merges with different styles and loras pretty much all you need

[brittle rampart]

won't add-detail-xl really work?

[nova remnant]

pony says use clip skip 2 this pony detailer I found is saying use clip skip 1 lol

I don't think clip matters for sdxl models

Just always 2 >.>

also what do people use for inpainting? Pony inpainting doesn't seem to work well, quality destroyed

[nova remnant]

Pony

You're doing something wrong

it always does this stuff

what am I looking at

smudges, discoloration etc

hmm

maybe it's an issue with a lora

[brittle rampart]

Perhaps the settings of the mask and the inpaint itself are not configured correctly

[nova remnant]

using stable diffusion has increased my desire to learn to draw

lol

Being able to just say screw inpaint and do it yourself

[zinc flax]

either it's too overcooked or just pony being bad lol

[nova remnant]

only does it with inpainting though

[brittle rampart]

what a straightforward and tough approach! But what about creativity!?

@nova remnant are cats afraid of injections?

[vital raptor]

it's not "because pony", it's because you need to tweak mask and blur settings for inpainting and just keep generation with different seeds until it does what you want

sometimes generate with a smaller mask, get part of it right, then mask a different part

inpainting can take a long time depending on what you're fixing

[nova remnant]

Im innocent, no need for injections

I like how Pony seems to get hands right more often than 1.5 was... 1.5 would draw this mass of 90 finger blob then you'd inpaint it for 5 hours before giving up

lol

(arms behind back) was my favorite tag

[vital raptor]

the fun part about pony is that it's also really bad at hands (not as bad as 1.5 though) but it depends more on your artist tags and such

like, one artist has perfect hands, another artist has terrible hands

[nova remnant]

well yeah finger length problems, sometimes 6 fingers instead of 5 but that's a lot easier to work with than the fleshmass handblobs 1.5 dishes out

[gloomy scroll]

https://tenor.com/view/shikanoko-shika-shikanoko-nokonoko-koshitantan-snknk-groom-gif-2511096066649814404

[nova remnant]

too many horns

[gloomy scroll]

https://tenor.com/view/しかのこのこのここしたんたん-shika-noko-noko-koshitantan-とらこ-虎子-you-deer-gif-14524014536672925431

[nova remnant]

I really gotta train my own lora for this girl though. The available ones consistently screw her antlers up. In fact I spend less time with the hands than I do these antlers lol

[gloomy scroll]

https://tenor.com/view/shikaanime-deer-sika-shika-shikanoko-gif-16894366684372908798

[nova remnant]

why does that remind me of rick astley

I don't think it is lmao

[gloomy scroll]

https://www.youtube.com/watch?v=doZzemrY7dQ

[lyric zealot]

https://tenor.com/view/shikanoko-nokonoko-koshitantan-my-deer-friend-nokotan-anime-gif-3888990587758024677

[nova remnant]

[nova remnant]

they are the same

[brittle rampart]

it is impossible!😁

[brittle rampart]

thinks about whether it’s dangerous for her to have her hair down

[brittle rampart]

and yet they are similar

[nova remnant]

[vital raptor]

i have a lot of images of this particular character..

[brittle rampart]

we know it!

What tokens can be used to describe a print on clothes with an inscription?

[nova remnant]

b

[nova remnant]

mouse!

Uninvited guest

[brittle rampart]

yeep

[nova remnant]

[jagged axle]

Beautiful picture anime.

[dapper verge]

Got it off an ad for Emily in Paris. Prompt: A stylish woman wears a polka dot top with big pink roses on her dress. She has red lips and a fancy black hat with feathers. Behind her is a light pink Eiffel Tower. Pink rose petals float in the air. The background is bright white. The woman looks to the side with a curious expression.

[brittle rampart]

@nova remnant they came for you. It's time for treatments!

[unkempt cape]

neta art 2 out maybe
hopefully its good

[unkempt cape]

not bad 👍

like not sure if this will make me spend 23 hours a day for sdxl or not but its pretty cool 👍

[unkempt cape]

hold on wait a minute 😳😳

neta 2 does not screw around bruv

guys drop your firest nai prompts here and lemme test

HEJEJNEISBSHFUEBDONDBFIEJDNDODJNDIJDNEKWKD

well it did mess alotta things up but OMG

[unkempt cape]

improve and have it make more natural poses and nai out the door

prompt stolen from nubby

[junior quest]

nai?

[rough oxide]

i tried neta a bit but wasn't super impressed. one day the next big thing will be here ;_;

isn't it just a finetune of artiwaifu or is it more?

[fluid torrent]

Copy pasted from the civitai page of neta:
" In V1.0, the default poses generated by our characters are relatively fixed, while the new version enhances the sense of dynamics, diverse character camera expressions, and more precise prompt control.

On the basis of increasing more stylistic diversity, more stable anatomy has been maintained, especially the probability of having five fingers (instead of four or six) on the hands is now higher!

Further training methods for issues such as Rectified Flow and Noise Offset have been better deployed. The current picture has a strong sense of epic, and can now show very bright and dark scenery (as shown in the figure below)."

and also: "In Neta V2, we re-trained based on ArtiWaifu Diffusion V1.0 (AWA1)."

so, Yes & Yes

[rough oxide]

maybe it'll be good in some model merges or something, base model didn't seem that impressive to me when i tried it within like the first hour of it coming out

[ivory pebble]

Looks kinda fried ngl 😭, but maybe that's just the art style

[rough oxide]

was happening to me when i tried it too lol

[ivory pebble]

Yeah that was the main issue with arti as well, also it couldn't really do any complex positions like NAI or Pony

[ivory pebble]

Now make her smile

[rough oxide]

that base prompt is super stylistic

but how did you get crossed arms to come out xD

might need portrait gen i guess

[ivory pebble]

Should be pretty easy to do with "crossed arms", no?

[rough oxide]

that's in the prompt

o

i just had to emphasize it more

[ivory pebble]

Yeah that usually helps

[green temple]

why no crossed arms

[rough oxide]

just wanted to see it with the model merge i was using xD

now there isss

crossed arms normally just comes out, no idea why it didn't on that one xD

[fluid torrent]

I think I found some good prompt bases for cute 2d anime in SD3.

just put (2d anime style, cute anime:1.3) in the POS prompt
and put (3d render, cgi, real, photo:1.1) in the NEG prompt

[rough oxide]

different prompt now but here's a smile with crossed arms xD

[rough oxide]

just base sd3 or something like artfusion sd3 anime?

[ivory pebble]

Cute

[fluid torrent]

base sd3

[rough oxide]

o but alizee lora

[rough oxide]

wait whai are u using a lora for like 3d chars for an anime image thooo xD

[fluid torrent]

I use stableDiffusion3SD3_sd3MediumInclClips

in a Comfy workflow

[rough oxide]

ahhh

that explains why the parameters are so hard to read lol

[fluid torrent]

lol, could be!

are you using forge or A1111 or something?

[rough oxide]

o no just reading it off here lol

it names a lot of models

i use swarmui for making images

[fluid torrent]

hmm, that might be information overkill due to the Lora Info node I've added to the workflow. It holds a list of lora names, just to fetch info from civitai or something

[rough oxide]

yea

this is what i see, 1 sec

[fluid torrent]

swarmui ... is that comfy, but covered so you dont see the spaghetti?

[mental marten]

#🍥｜anime message

[rough oxide]

pretty much

[fluid torrent]

lol, yeah there is a lot of noise info in there. Perhaps I'll throw out those extra nodes, to make my images easier to share

im all for prompt sharing

[rough oxide]

it's fiiiiine

if it's made in comfy then

[fluid torrent]

or generation data in general

[rough oxide]

when someone drops in comfy they get the data

like full node workflow i think

[fluid torrent]

yes, but also some confusing unconnected nodes

[rough oxide]

but ye for example mine are likeee

[fluid torrent]

😄

[mental marten]

#🍥｜anime messageyaml { "sui_image_params": { "prompt": "1girl, klaudia valentz, atelier \\(series\\), yoracrab kou hiyoyo, (smile:0.8), crossed arms, light particles, night, futon, making a mess of covers", "model": "animaPencilXL_v500", "images": 8, "seed": 534584020, "steps": 36, "cfgscale": 5.72, "aspectratio": "1:1", "width": 1024, "height": 1024, "sampler": "euler_ancestral", "batchsize": 4, "clipstopatlayer": -2, "automaticvae": true, "wildcardseed": 534584035, "negativeprompt": "", "swarm_version": "0.9.1.1", "date": "2024-07-23", "used_wildcards": [ "character", "artist" ], "original_prompt": "<wildcard:character>, <wildcard[2-4]:artist>, (smile:0.8), crossed arms, light particles, night, futon, making a mess of covers", "generation_time": "82.78 (prep) and 40.84 (gen) seconds" } }

[rough oxide]

lol

[fluid torrent]

that is very good human readable

[rough oxide]

i think swarm itself is what makes it more readable

cuz under the hood it uses comfy

well, like u said, w/o all the spaghetti

i used to make comfy workflows and liked the ui a looooot

but swarm is just more manageable for me nowadays

[fluid torrent]

do you also make (merged) checkpoint and loras?

[rough oxide]

back in wd era

right now i use this merged ckpt https://civitai.com/models/261336

[fluid torrent]

(stole your prompt from the example...) sd3 with those added anime base prompt

[rough oxide]

for not being an anime model it's surprisingly a little better than i thought xD

still waiting for sd3 large though

or before that even

wdv model

but for the various styles i like, anima pencil is a near perfect merge already

i feel like it's even better than base naiv3, cuz i always had to use vibe transfer for naiv3

[fluid torrent]

sounds nice indeed, im going to create some room on my harddisk

[rough oxide]

ye, also you can get good results without paragraphs of text most of the timeee (and no neg prompt needed)

so super overfitting maybe? XD

but i think that's fine

kohaku xl epsilon rev2/rev3 are also pretty good for style

lol. i guess kinda funny how i don't really see the appeal of neta art but the merge i use has neta art as part of its recipe in the merge xD

[lyric zealot]

i think it's also partly because of how the reading of the comfyui meta data was coded. Could probably improve on the readability of it

[brittle rampart]

White mouse supports! We condemn!

[rough oxide]

yeaaaa; i was wondering why it'd come out like that sometimes when i used the bot

now i know it's comfyui xD

kinda glad swarmui makes it super readable

[lyric zealot]

ahh, i think it's because comfyui metadata is usually super long, and will have to send it as a yaml file instead of just an embed

[rough oxide]

ooo. that's a problem with discord itself then

i think

unless it broke it up between messages or something

might get spammy if it did that tho

[lyric zealot]

yup, that's why it's probably made it to send yaml file on purpose

[green temple]

wtf 😂

[brittle rampart]

fantasy ai

[green temple]

fantasy smol

[green temple]

angry rat

[brittle rampart]

Judgmental.

mouse

[green temple]

indeed

[brittle rampart]

She can't even get angry...

[green temple]

kawaii anger

[brittle rampart]

bzzzzz

[nova remnant]

Angy

[brittle rampart]

https://tenor.com/view/no-bugs-bunny-nope-gif-4095597578532971838

cute

[nova remnant]

No u

[fluid torrent]

Here's a cute younger version of Caitlin !
Curious as ever, though not so innocent
Many years she trained pet animals

[nova remnant]

Mouse girl of the witch woods make delicious food for me

[nova remnant]

rejects her

[brittle rampart]

yeah... fly agarics... grass...

[green temple]

remind me of tomo-chan 😂

[nova remnant]

shes saying swears out loud

[fluid torrent]

Next up: BAD anime teeth vs GOOD anime teeth

BAD anime teeth.

GOOD anime teeth.

[green temple]

all bad

[nova remnant]

both looks anime

[lyric zealot]

[nova remnant]

both will eat @lyric zealot

[lyric zealot]

i mean

[nova remnant]

She's...

Heh

[lyric zealot]

[nova remnant]

Wet

[lyric zealot]

[nova remnant]

[nova remnant]

I want this one 👍

[lyric zealot]

not for sale

[nova remnant]

[nova remnant]

[lyric zealot]

[lyric zealot]

her fingers

[green temple]

noooo , big

[brittle rampart]

bit a big too

[green temple]

here is smol

[brittle rampart]

yep!

[supple raptor]

It's a yaml because there isn't a standard way to determine what each thing is since the order of values can change as well as the name of the values. Comfy metadata is extremely cursed to try to parse. You could make a workflow that would make it easier, but then everyone would have to agree to use that same basic workflow for the parsing to work as expected.

[green temple]

[unkempt cape]

dude this is first model to somewhat does nai's combo styles. Like anywhere from 60% to 90%. No other models do that ever

poses are still a problem but this is a big leap

[unkempt cape]

so does the original pic actually, i'd say like style is 75% on point (last time was like 9%)

[rough oxide]

to be fair the model merge i use has neta in it, so it might be better than i give it credit for, but as a pure base model w/o finetuning i don't like it much personally

[unkempt cape]

wtf already?

[rough oxide]

neta v1.0

not the new one yet

[unkempt cape]

no the old one is alright at most lul

[rough oxide]

i mean i tried the new one as a base model too

wasn't really impressed kindaaaa

but i do minimal prompting style

[unkempt cape]

I just go on ze booru and picked off random nai prompts

pose is no where as good but style is getting close

few more months bro we will get a model that kicks nai out the window, trust

[rough oxide]

nai diffusion v3 was amazing when it came out but i feel like some merges surpass it in quality (not quite base models yet though... but maybe with wdv?)

[unkempt cape]

ok rlly??

[rough oxide]

well, for exact poses and getting every single detail you put in the prompt though, naiv3 is decently good

but for aesthetics i prefer other models

[unkempt cape]

can i see some names?

[nova remnant]

any way to fix these hands? Regenerating inpaint for a half hour isn't coming close to a desirable result

[rough oxide]

https://civitai.com/models/261336/animapencil-xl https://civitai.com/models/399873?modelVersionId=546178 these 2 are my fave rn for style

guessing a future anima might implement neta art v2 in the merge too

[native halo]

yes draw good hands on top of it

[nova remnant]

but I use AI art cuz I have the artistic skills of a neanderthal

lol

[rough oxide]

adetailer models for specific things like face/hands help a lot

[native halo]

adetailer for hands suck it will only give u more noodles

[rough oxide]

but in the first place is that even an anime model?

[nova remnant]

It's Pony

[rough oxide]

o

i'm not very well-versed with pony. kinda looked like a stable cascade/sd3 output

[nova remnant]

but I kind of want to ditch pony cuz it has no control over the art details it seems

[native halo]

it has if you use loras

[nova remnant]

1.5 had an add detail lora that did an amazing job

the detail loras for pony aren't even working for me

[unkempt cape]

used kohaku before, it has some cool styles but it doesnt know too many artists. Base style, i hate smol, so no to me lol

[rough oxide]

yeaaa i primarily use pencil right now

merge of a bunch of animagine xl 3.1 and artiwaifu models

so can handle pretty much everything animagine can in terms of artists/chars

[nova remnant]

does it adhere to prompts well? It's one thing I do like about Pony is it doesn't require excessive complexity and weighting to get it to do what you want

[rough oxide]

there's a version that mixes with pony as well

same author

https://civitai.com/models/432249/ponypencil-xl

handles prompting just fine tho

in all 3 of anima/pony/blue pencil

i don't rly like pony pencil style tho

i just... don't like pony style i guess xD

or it's more like, in the ponypencil version, you can't get the anime screencap look anymore

it's always stylized

or flat colors in general just wouldn't come out well in pony pencil from what i saw

(this one's animapencil)

it adheres to prompts as well as animagine does, which is... enough but i think kohaku did poses & such better

and neta art v2 also maybe does poses & such better

but what i dislike about both neta & kohaku is they have prompts like

[nova remnant]

ugly

[rough oxide]

waaaay 2 much textttt

[nova remnant]

but do you actually need that much

[native halo]

easier to just get a good base pony mix and then use a style lora

[rough oxide]

when i tried both neta and kohaku, i found that yeaaa those long prompts of things ppl posted weren't just for show Lol

[nova remnant]

I blame @native halo

[rough oxide]

you need the rating tags like on novelai, you need a bunch of tags for everythinggg

tho nai does good with some minimal prompting stuff

so not nearly as necessary at all

[nova remnant]

nai without a billion quality prompts is scary

[rough oxide]

^meanwhile this is like one sentence of prompt

lol

[native halo]

yes but if u use generic prompts on nai u wont get certain styles so generic promps = generic pics

[rough oxide]

and no neg prompt

[rough oxide]

ye. i prefer models that vary a lot more creatively even within the same promptttt

used to use waifu diffusion but long lull waiting for next release xD

sounds like it's almost here tho

[nova remnant]

steven seagal from nai

[native halo]

this is wrong,his irl muscles are bigger

[rough oxide]

before waifu diffusion i totally used generative GANs like this waifu does not exist/this anime does not exist XD

they were a lot of fun

total unpredictability

looking back at them now tho, the quality is sooooooo bad

#🏞｜general-with-images message just looking back at my old posts here too, lulw. crazy how much sd has evolved in 2 yrs

[nova remnant]

damn

some old stuff in there

[rough oxide]

yehhh

back when 512x512 was max and there were no finetunes

textual embeddings still in their infancy xD

or i think u could probly have pushed it to 640x640

but anything past that = total quality loss

[nova remnant]

I bet some people were saying that that was the PEAK of AI

[rough oxide]

i probly was

[nova remnant]

like some people today about sd3 lmao

[rough oxide]

lmao

at that time

cuz it wassss

until like dall-e 3

[nova remnant]

I remember first dalle

was fun

[rough oxide]

ogod

Lol

[nova remnant]

first or second

[rough oxide]

dall-e 1 was so low qual it's so funny looking back at that too