/cgl/ - Cosplay & EGL

>>10780849
there are ways to work around its limitations

trying to generate a cutsew

>>10781117

i tried to make you a 4chan ita bag, cgl

>>10781152
the AI can handle other shapes, it just cannot count to 4

chocolate, vanilla, and strawberry cupcake handbags

>>10781875

>>10781877

>>10781880
last one

i wonder if AI like this could be used to teach or illustrate the differences between such niche fashions. it would first need more training since its current knowledge is narrow. still, it is surprisingly close with these 4.

I tried to get it to generate Ouji coords

>>10782014
Last one

>>10782045
interestingly, the above post and this one shows heavy preference to a single design. I wonder if it's basing it off some single reference set of images somewhere on the internets.

wall-e thinks these are itai as fuck.

>>10782047
The search won't know what "itai" means. Gotta use more common terms.

>>10782014
>>10782015
nice

>>10782508
variants, clockwise: honey lavender, raspberry peach, vanilla mint, blackberry + blueberry

>>10782602
it struggles with shapes sometimes

>>10782604

predictive text craiyon made doglita

gobelin for the forest goblins

>>10783178
more gobelin fun

>>10783191
putting prompts together

>>10783325

it knows larme

>>10782612
I love this, it totally worked

>>10783696
it is amazing what the ai can understand

>>10784001
second attempt.

>>10783338
my take on a similar theme

>>10784007
nice

this is the least unsettling result of the batch

vivienne westwood ita bag with kamikaze girls theme

what do you think, gulls? did the ai understand the assignment?

lmao

>>10784554
>deamholic

fuck

>>10784554
attempt #2 with minimal changes

trying to keep it simple and see at what point the complexity becomes too much

>>10784580
attempt #3

it refused to process the wig request until i changed the wording, and some colors it would not generate.

>>10784647
i think this is because of the extra details (material + pattern) i added to the clothing items. the order of items matters less when the prompt is simple.

>>10784647
at the limits of complexity with this formula

Has this ever happened to you?

the data set includes images from Tokyo Fashion, Harajuku Fashion Walk, and wear.jp for sure. the ai's ability to reproduce the average wear.jp fashion snap is top notch, and we know other image sets have used wear.jp for machine learning.

so then why is the ai's knowledge of japanese brands so restricted? why does it recognize (and give extraordinary weight to) "angelic pretty", but it has nothing for "baby, the stars shine bright"? it does not even return an "everything but the girl" album, just literal baby dresses with stars on them. if it is because AP is more popular, then why is "metamorphose temps de fille" recognized but not btssb? is it due to AP and Meta being easier for the ai to recognize visually? is it because, despite being a generic name like "innnocent world" or "victorian maiden," the brand "angelic pretty" is the most iconic usage of its words, while "baby, the stars shine bright" is too ambiguous (along with VM, IW, MM etc.)?

is there an internet photo site that has hashtag categories for AP and Meta, but not for any other lolita brand? or did the data set only include certain tagged brands, to save space?

as for possible issues with uniqueness or vocabulary: "hime gyaru" is recognized as distinct from "gyaru," with small variations, but "ganguro" returns something completely off, despite its iconic distinctive look and unique name. (what the ai returns for "ganguro" is a thing you should see for yourself lol.)

the gaps in its knowledge are strange. it does not recognize "moi meme moitie", but it has a rough idea of what "egl" means, which is just wild. it can produce good yosuke shoes with the right prompts, but no other gothic jfash brand i have tested will show.

also it cannot understand "h.naoto" (or any variation thereof) but it knows hypercore, which is just disrespectful.

all of the brands and styles i mentioned are on the websites the ai definitely knows. so why does it only recognize a sliver of them?

>>10785007
(ctd.)

what the ai can do already is amazing. it knows much more about the particulars of jfashion than what i expected from a free-to-use image generating bot. i love it. but its logic confuses me. i suspect human interference, probably eliminating whole categories to save space, but idk how the categories to save vs categories to exclude were decided on. does anyone have any insight?

in closing, here is Mana from Moi dix Mois in egl that is sadly too unlike Moitie

neat

>>10785091
small improvement

>>10781117
Middle top row totally referenced this image for the poses

>>10785169
awesome

trying to prompt border prints

>>10784714
>Otakon
Yes literally

>>10785108
Oh the ai used a betsey johnson necklace for this lol. Look up You Give Me Butterflies necklace.

nothing abnormal here

finger fringe is the future

>>10785169
That's wild that you recognized that, nonna

>>10784696
I love these, would absolutely wear

Anyone here ever try disco diffusion? I've been trying gothic lolita related prompts and it's been interesting

biblically accurate angelic pretty

i tried to generate a lolita dress like a faberge egg. i cannot do it to my satisfaction.

challenge: make a prompt for the AI to reliably make faberge egg lolita.

>easy mode
a dress with jeweled egg elements
>normal mode
a dress with a lolita silhouette (i.e., not egg-shaped) with faberge egg-like elements
>hard mode
a lolita wearing the normal mode dress
>impossibiru mode
hard mode but it has to look like an actual person and not a funky homunculus

wig game too strong

>>10787267

consider:

lolita fashion lesbian dating sim

how well does craiyon know its two recognized lolita brands?

>>10788098

>>10788099
using the prompt:

>ultra hdr 1080p close-up shot of high-quality fabric with highly detailed border print : "print name" designed by angelic pretty

cat tea party, sugar dream dome, crystal dream carnival, and misty sky. for misty sky i took out "border".

>>10788120
ditto for meta: wonder carousel, twinkle journey, fruit punch soda, and nostalgic matryoshka doll

>>10788098
it turns out that the ai can sometimes guess which dress does not have a print. Meta's dresses whose names are also physical descriptions make good prompts.

fake AP prints are fun for the whole comm

you could generate prints using common AP words or phrases, then have people guess which words were used. was it "melty dream diner", or maybe "sugar cream parlour"? a "candy theater" or a "bonbon circus"?

>>10788415
dropped the damn pic

>>10788415
i actually need "cherry chocolate whip" to be real

>>10788415
i need "cherry chocolate whip" to exist irl

wait what

>>10789486
Algorithm thought you meant Impact Wrestling

Best timeline

get into the future

>>10789486
the pictures alone are funny but the misinterpreted prompt makes it hilarious

30% of my comm

no, craiyon

let's try something else

>>10791071
still Bowie, ok.

thanks for trying, craiyon

>>10789986
embrace tradition

>>10788797
more AP word salad mixing

jesus christ

he was ahead of his time in so many ways

do you think this is funny, craiyon?

>>10791868
i cannot keep these results to myself any more

try "moi meme moitie", pls, and tell me if you get this consistently unsettling shit

>be me
>be AI
>recognize bodyline on sight
>cannot tell moitie apart from war crimes

que the fuck

>>10791869
I love the atepie images, he just wants to be a designer

>>10791876
craiyon keeps showing me naked cavorting soldiers and dead bodies when i prompt it with variations of Moitie*,

so to cleanse my eyes, your comment inspired me to prompt pic related. bottom left looks cute, bottom right looks like it popped straight out of an Eiko Ishioka exhibit.

*when i tried moimememoitie, it showed me The Hague lol

Advanced AI art (midjourney) for you noobs.

"sweetlolita angelic pretty fashion brand with sheep print"

Anytime I do angelic pretty and sweetlolita together, I get fucked up looking mascots.

The future of ita threads is here

I just got accepted into dall-e 2 beta

>>10791972

>>10791959
amazing

>>10791973
awesome, congrats!

>>10791959
>Anytime I do angelic pretty and sweetlolita together, I get fucked up looking mascots.
how much control do you get with midjourney? have you tested for other brands?

i have played with Starry AI. it lacks a lot of creative freedom, so even if the images are high quality, i did not enjoy it.

>>10792022
>pic unrelated

i literally cannot stop generating prints based on nothing but brand name + brand-style word strings. craiyon's unlimited iterations and opportunities to fine-tune make its quality constraints hardly an issue.

shipping delays got out of hand

>>10792328

>>10792330

>>10791959
>>10791973
can they understand multi-subject prompts? craiyon tries, but fails

>>10782047
>>10782091

>>10792533
perfect

>>10792029
As someone who's been waiting 6+ months for their seamail parcel containing a Meta DD, I can relate

Is it just me, or is the algorithm getting smarter and recognizing lolita terms better due to all the people here making lolita prompts?

>>10793033

>>10793034
theoretically it does not work that way. however, there was a recent update that improved overall quality for all results, which may be what you have noticed.

what if decora, but zumiez

sXe sk8er grrrl

>>10794027

>>10794029
damn what's her skincare routine

>>10792458
yes. very well, but there's some tricks to getting it to work right. you'll end up writing a big ass fucking paragraph and re-iterating until you get something you like, which it gives you a lot of flexibility.

I'm not fully sure it registers lolita well though. I've gotten some lolita esque art, but nothing fully "this is lolita"

It does make high quality art, however.

>>10794118

>>10794122
based

>>10794121
cool

y'all got any of them accidental nightmare results worth sharing?

>>10794118
>I'm not fully sure it registers lolita well though. I've gotten some lolita esque art, but nothing fully "this is lolita"
i wonder if that's due to censoring, to avoid users abusing the AI for harmful shit.

craiyon will give a sweet lolita dress if it's by AP, even if you don't specify "lolita" at all, and you can fiddle with design elements (pic related). if different AIs do censor the word "lolita" i hope they still include the japanese fashion datasets that use brand and print names.

>>10794152
actual pic related

>>10794147
Looks like hardboiled eggs glued onto the dress

>>10794292
face like an angel

i would buy it

a familiar face appeared

>>10796425
looks like crystal from old

Craiyon & DALL·E mini are over. StableDiffusion is the new big player in town

>>10797034

>>10797037

>>10797034
>Craiyon & DALL·E mini
same thing

nice pics, how much control do you have over the result?