- https://makegirlsmoe.github.io/assets/pdf/technical_report.pdf
- http://make.girls.moe/#/
- http://waifu2x.udp.jp/
>>161162406
Fuck off already, this belongs on /c/ or something, not /a/.
Free beta testing is over
>>161162406
>serval methods
SUGOI
did they actually expect anybody to read this and get baited
because it worked on me
Why doesn't anyone ever include the best hairstyle?
This will never take off without an option for smug
have they considered vectorizing the images first and then doing machine learning on that? it'd probably give cleaner lines.
>>161163760
Have you read the technical documentation? It has trouble learning complicated stuff like "drill hair" and "glasses" do you think it could ever manage something as subtle as "smug"?
>>161163938
Smug is basically just small eyes with a large smile. It seems easier to teach than drill hair, but I suppose more challenging than glasses, probably.
my friend made this beauty
Something went wrong.
>>161166458
Something went very wrong.
NANI!?
CUTE!
Why do all these girls look like they came straight out of KAWAII UGUU land?
>>161162406
Don't! You're only making it stronger!
Oh god what the fuck happened to your brain?
A CUTE
Data science, putting character designer out of work since 2017
Rate my smug.
>>161167296
7/10
rate my loli
>>161167296
The robots still can't beat us at producing true smug, I guess
>>161167348
I think it made it so smug that it's face melted a little
>>161167345
higurashi/10
>>161167345
9/10 my aesthetic
This time the hat came out non terrible!
>>161167558
>wearing a used condom for a hat
I would fuck her.
>>161167173
oh that's a good one
What has science done
and how will we atone for such a monstrosity
>>161167790
what did you do
WHAT DID YOU DO
She looks tsundere.
>>161167759
i need a brown girls option
Alright I edited one of the templates so that the winner of the MCbowl is now the one whose number corresponds with your last two digitsdon't roll a 94 you sick fuck
>>161167348
they can come close
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
>>161168158
>u havn a giggle sempai
cute
Now how do we get to fuck them though
This technology only seems useful for lazy character desginers.
Cute track team captain.
Rate my fabricated wife.
>>161168390
If we don't support the technology in it's infancy we will never get to fuck them.inb4 pedophile joke.
Why is the output image so small?
What's this expression trying to convey?
>>161168466
It's because the AI at the moment isn't good enough to make the images larger.
If the images were larger the errors would be more glaring at the current level they have
>>161168537
Dude weeb lmao
>>161168466
Technological limitations. The higher resolution, the higher probability of "chimeras" (fucked up images).
>>161168537
>She sees your dick
Wow. I want to be M for her.
>>161168466
>he wouldn't fuck his RNG waifu for ants
I just want to generate my waifu Cocoa, but she comes out weird.
>>161168070
Would you watch it?
Who is best girl?
How angry would the winner of MCbowl make you?
>>161168814
Roll
Needs more hair options. It should have a front and back option. Oh, and a fang option.
>>161168814
94 is best girl.
>>161168814
Trap and mom are either the same person or related.
>>161168573
with more gan wrangling
it is possible to get higher-resolution samples that are still coherent (albeit strange looking)
see
https://github.com/hanzhanggit/StackGAN
http://mtyka.github.io/machine/learning/2017/06/06/highres-gan-faces.html
but its difficult and not really standardized yet
so might be a while before someone incorporates it to this project
>>161168814
>Childhood friend wins MCbowl
Tsunderefag salt would sustain me for eternity
certified reverse trap
ive perfected the cyclops
>>161169402
grats
>>161162406
>>161168044
Yes, they forgot to add a skin color option
>network error
>tfw no qt ai gf
>>161168429
reality is bleeding/10
>>161169719
Try switching browser.
>>161168466
>As Appendix 8.2 shows, the resolution of available training images are not high, making generation of high resolution facial images from the GAN model directly is difficult. Here we try to build another Super-Resolution network specically for generation extra high resolution animation style images to overcome the limitation.
>Image (b) in Figure 12 shows the 2x upscaled image with waifu2x 10, which is blurred. [12] gives a comparison of GAN-based super-resolution model and traditional MSE-based super-resolution model. Their result shows that GAN based models can bring more high-frequency details to the upscaled images than MSE only models. This is preferred in our situations, so we choose to implement SRGAN[12] as our super-resolution model. Image (c) (d) show our attempts of training a SRGAN. (c) is trained with a low adversarial loss weight and (d) is trained with high adversarial loss weight. We observe that as the weight of the adversarial loss increasing, the upscaled image looks sharper. However, this would bring more undesired artifacts to output images. We observe that the visual quality of anime-style images are more sensitive to extra artifacts than real photos. We hypothesize that it is due to the fact that color/texture patterns in anime-style images are much simpler and clearer than real photos, any artifacts would largely damage the color/texture patterns and make the result looks dirty and messy. This obstacle limit the usage of GAN in our super-resolution networks. Discouragingly, we failed to nd a model balanced well between the sharpness level and the artifact strengths, so we choose not add the super-resolution model to our website for now, and leave the exploration of anime image focused super resolution for future work.
Not bad
>>161169843
there's a lot of times when it comes out 3/4 of the way to the perfect anime waifu but then there's something just kind of shooting off into space or it doesn't understand how chests are supposed to work at all
I've done it /a/
Cute
rate my gremlin
>>161170040
open mouth smile makes some really genki girls
>>161170140
I feel the warp overtaking me
Make another chart for the abominations.
top tier semen demon trap
STOP GENERATING WAIFUS YOU'LL CAUSE WAIFU MARKET INFLATION MAKING EVERYONE WORTHLESS
>>161162406
>CMU
Sasuga #1 CS program
>input glasses
>don't get glasses
She's still super cute. I like what I got.
>>161162406
Is this someones thesis or something?
What if VN waifu desins are already generated using this method?
>>161163703
god tier taste brother
>>161168814
roll
this is the future we chose
>ME WAIFU
>ME SMASH
>proceeds to suplex the MC affectionately
>>161171021
Looks like from Initial D
>>161170319
each waifu is unique
ara ara
>>161171021
>I won't let you hurt mc-kun!!
Spot the mysterious antagonist
Brain problems
I'm not sure if she's got problems or if she's insanely smug.
>>161168070
Which one, /a/?
>>161172085
Canon winner in your VN is the Reverse Trap apparently. Vampire is best girl though. Let's see what I get anyway
Help
>Ask for drill hair
>keeps giving me girls without drill hair
>>161162406
Soon all anime will be created by a computer algorithm and even Kyoani will go out of business as uncompetitive.
well she's cute in a 'i got hit with a shovel one time' sort of way
Best waifu coming through
this is your antagonist
>>161168070
I edited it to make the numbers less horrendously arbitrary and also added a need option
actually do they have a version for men?
this sort of thing would be great for /tg/ campaigns
>>161173490
You can set hair to short and both colours to black and it will generate men usually. There are hardly any samples for other types for men due to anime itself
>>161173569
That's a cute tomboy girl you scum
>>161173617
fuck you here's Sephiroth.
>>161173569
I said men. I already drew a mustache on one
I feel like even my players might catch on if everyone has facial hair in the universe
>>161162406
When I see the really good generated characters, I have trouble believing it. I keep thinking it's just pasted up something the machine scraped off the web, or maybe re-colored the hair or something like that.
Intellectually I know what is going on (well, mostly- I didn't totally get the whole process described in the PDF), it's just harder to process on the other side.
reminder that installing safari tech preview for this is genuinely worth it
chrome takes ~5s for me per image on my mbp, but stp is <1s
>>161173490
set blush to off, hair to short, keep clicking until you get lucky
Just gonna drop this here
>>161173851
sometimes a little too manly tbqh
>>161174018
More like retarded
>>161173679
that looks super retarded
Not bad
>>161174080
cute and pretty
A CUTE
A friend of mine made this for me.
Pretty good
Oh god what did I do?! WHAT DID I FUCKING DO?!
>>161174239
>>161174245
The Tard sisters.
>>161174239
A little oversmuged
>>161174231
Ojou-sama, please step on me
>>161174018
>just be confident bro
>Anon.. people in my class have been really mean to me and have been making fun of my ears. They call me hurtful things like "dumbo" and tell me to jump off the roof and flap away
>Does it really look that bad?
>>161174239
>>161170356
use the same input (fix the noise) and turn glasses off and post results
Call me maso-chan
>>161174477
>retarded elf
>>161173893
Some are good
Some are not
The chins are all fucked up.
>>161174567
What do you mean?
>>161172569
it doesn't understand what drill hair is, they address this in the report
>>161174540
turn glasses off, idiot, it doesn't understand what glasses are
what the fuck is that thing on her head supposed to be?
>>161174748
horn. she's a demon loli
What do you think of my new hairclip?
Was she fucking decapitated?
>>161174239
Here's what I get with a bit of imagination.
>>161174820
Nice. I wonder if one could get in legal trouble adapting these outputs and using them for something
Everyone in this thread can fuck off for I have successfully managed to generate a catgirl and thus have won
>>161174880
Less successful this time but still.
>>161174876
I think you should be alright- make some changes here and there at the very least (something you'll have to do anyways).
Not sure what im looking at
>>161175132
You are like a baby, watch this.
>>161175214
what the actual fuck
>>161175132
Looks like a puchina
I dare you to make a cuter girl
>>161175298
>bad inbetween from 2004 . jaypeg
>>161175409
Challenge accepted
These will be the faces that haunt my existence
>>161175409
>>161175409
You made me post this.
>>161175409
>>161174986
>>161174880
ah yes true, congratulations
I don't have any nekololi applaus pics so have a question mark tail one instead
>>161175409
Easy
>>161175409
I FEEL THE WARP OVERTAKING ME
IT IS A GOOD PAIN
>>161175607
>>161175506
>>161175409
And here's a loli, with a hat.
UGUU am I cute senpai?
>>161162406
So this is what the future feels like
Jesus fucking christ, I have opened the portal to hell.
I think we broke it.
SMUG
>>161175816
>>161175816
>>161175409
>>161174748
She's a oni, a cute oni.
this looks like some yugioh face warp shit
>>161170670
>>161163703
the best hairstyle
Best girl in the thread with the potential exception of the catgirls
>>161175409
GET FUCKED EVERYONE
Rate my wife
>>161176294
0/10 I can see their melted eye through their hair despite their effort to hide it
>>161176338
Rude.
>>161176434
Kizuna AI?
>>161176434
Rinzuki
>>161176440
Why don't you try making a few that don't look like trash?
>>161176486
that's impossible, all anime is trash
>>161176486
Blame the AI, not me
>>161163703
>>161170670
>>161176111
Actually this is the best hairstyle
>>161176476
This one is more Rin like
Make your own Isekai template
>>161163938
It's complicated.
Neural networks form features based on pixel information. Those pixels build up to form lines, then features, which are also composed of pixels.
You couldn't realistically do machine learning on vector images because vector images don't have fixed sizes. You could do it on the rasterized output of a vector image, but how would you map the vectors to the colorized regions of space?
tl;dr shit's complicated already, involving vectors makes things even more complicated.
>>161176554
With a few adjustments, this girl looks pretty good.
these are my favorites so far
Do you love me anon-kun?
pretty coherent result
>>161176824
it looks like there's a flat faced cat coming out of her shoulder
Extra spooky
>>161176714
>You couldn't realistically do machine learning on vector images
not using this architecture, no, but it's not categorically impossible
for example sketch-rnn produces svg output based on the quickdraw dataset
https://github.com/hardmaru/sketch-rnn
if someone made a large, good-quality anime vector dataset, figuring out a way to make more is feasible
>>161176796
cute sketches anon
i appreciate you
>>161176796
Thanks draw anon, she had fucked up eyebrows but I fixed her
>>161176934
>https://github.com/hardmaru/sketch-rnn
forgot to link the actual demo
https://aiexperiments.withgoogle.com/sketch-rnn-demo
>>161175409
>>161176714
the images are composed of discreet lines, that's what forms the shapes we perceive
you should do machine learning on lines so that the robutt is trying to understand shapes or something instead of getting it to try to figure out what the fuck is going on with anime images
but I'm pretty much on top of mount stupid here
>delicious brown still isn't an option
This must be a sign
>>161177189
>even robots understand the appeal of reverse traps
Truly we live in the best reality
>>161177219
It doesn't understand what to do with clothing at all, does it
I think that's because it's too complicated, there's no tags associated with it, and the selector pressure wasn't very good at judging the chest area.
>>161177312
I wouldn't say that
>>161176486
>>161176294
>>161176265
>>161175824
>>161175409
>>161175169
>>161174986
>>161174748
>>161173569
>>161172806
>>161172073
Have decent clothes, to name a few
>>161177243
>D-don't leave me onii-chan . . .
>>161168907
Because the trap's actually your dad?
well... she looks like a fun person to be around.
>network error
Strange, didn't get that yesterday. We might be overusing it to death. Like all the other ones.
I would tongue her asshole
What emotion is this expression trying to convey?
>>161177396
Most of those are just really lucky noise
>>161177572
It's the fourth doctor
This is the closest I've gotten to an ahegao.
Now that I think about it, it would be fun to train this AI with nothing but porn.
>>161177926
Find them 40000 images of uniformly cropped and positioned porn faces and your dream can come true
Neat color scheme
>when you finally get glasses that aren't shit
>>161168466
Because computation time
Cute! I would mating press this cutie.
Now I'm wondering whether this is actually a published paper
This just proves that all anime is generic sameface trash that can be easily replicated by a fucking bot.
Give it a few months, some fresh otaku-pandering LN author will use this for character design "inspiration" for isekai harem LN #39204187, perpetuating the cycle of eternally generic weeb trash that infests the industry to this day.
What happened to her?
what the fuck I got a boy
Now this is what I call a chimera!
>>161176684
>>161178293
>look at me. I do not even remotely understand what the program does
>>161178346
the AI knows you are a fag
>>161167790
Sleep tight, pupper.
>>161178436
brutal
>>161178530
goddamn
>>161178530
What is that abomination?
>>161178530
>Thank you for not aborting me.
>>161162406
Smug
>when she isn't impressed
>>161178530
Horrifying. Good reaction image material though.
uhh.
am I kawaii uguuu
>tfw too intelligent for anime
>>161177674
>oh, it was just a movie?
DUDE, WEED LMAO
Onii-chan! Dinner's ready
she looks familiar
Made through the express effort of fucking it up.
my wife
Onii-chan
>blush on
>glasses on
this is all i need
>>161179454
Hi Yuki!
that's it?
Let's see if I can get Alice-chan.
penis?
>>161179793
This is how I feel after 50 attempts.
>>161179772
huh if you import "improper" noise you get fucked up results
cute
Getting close.
Experiments successful?
Once you've found noise that has very limited results, all you need to do is find hair colors and styles that it doesn't have much to pool from, then details adjust from there.
finally, a girl I can be content with.
cute
>>161179872
very cute. Me likes!
kill...me..uguuuu
>>161180150
>3 year old's crayon drawing
>Everyone in this thread gets isekai'd as the girl they posted.
I think she turned out very cute.
>>161180225
Please no, I was only trying to create mutants.
>>161180225
yes please, If I become pic related, I can finally lose my virginity easily
l-lewd
cool to see they got the ZUN touhou style
this girl's hair is covering most of the space
>>161180225
196 files, 84 IPs. What do with all these extra bodies?
>>161180421
they become part of your harem
some abstract shit
>>161180448
>anon sees this monster everyday coming up to him and being all over him
>>161178530
I can picture this girl being really short and have gigantic hair like Kokonoe Rin
I tried drawing mine
smug
>>161180421
You can shapeshift between them and they act as extra lives if you die as one you auto revive as a new one
>shapeshift into one of my many eldritch horror forms
I got a madoka look alike
Weird taste in clothing, this bitch.
>>161180748
MOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOT
>>161180748
HIROSHIMAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
>>161180748
MOOTWOOOOOOOOOOO
>>161180748
NAGASAKIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII
Orgasming as she's generated.
shit, this is so addicting. I cant stop generating and saving ones I like.
This is pure genius, I would have never assumed that using machine learning, to create 2D girls, would be so addicting. I just can't stop generating more in hopes of creating the ideal girl.
>>161180406
she's just broad shouldered
>>161180920
Why are anime girls so magical?
dork
>>161181050
It sure is, Its pretty helpful to get a quick character design to build off of.
Rate my harem.
Onii-chan, what are you doing with my panties?
Is there a clothing generator like this? It would be nice for drawing.
>>161181517
I see you
Is it possible to get a good one with glasses though?
Computers today generate anime girls. In the future, they could generate whole anime shows and movies.
>>161181877
I want to deposit my cum inside her cunny
how to create something can make your san decrease?
Found perfect waifu immediately, thanks OP
The closest I could get to Alice-chan.
This is so fun and addicting. You never know what you'll get.
>trying to get Osaka
>got a hybrid between Dio and Hisao
I got Mikuru
>>161182184
I got Tsutsukakushi without hair clip.
She's indeed literal perfection.
>>161182388
Opps, forgot to attach the picture.
mine keep coming out with hair clips
what am I doing wrong
>>161182435
Post her noise, lads.
Betcha can't get bigger eyes
I got Natsumi from Non Non Biyori
Dawww.
duude.
DAWWWWW
>>161182687
I wanna hug her
I'm rather impressed with some of the results. I could see these sorts of tools being quite useful for inspiration as least.
Short black hair and eyes break the mixer.
>>161182754
>>161182754
Nonsense. In the future, whole shows and movies will be automatically generated.
>>161182754
Holy shit, she looks like a Monobeno character.
Why can't I get enough of anime girls?
>>161182767
I don't believe you.
Original character, do not steal.
>>161182890
Yeah, getting some real good results.
>>161182767
I believe it.
>>161182950
Can't believe I came out.
>>161182950
Scratch >>161182961, this is literally me, holy shit.
>>161172119
this is a good one.
>>161182991
Good with glasses, anon.
>>161183032
Is that you, Yupiel-sama?
>>161163703
Just curious, but what is this hairstyle called?
Tucked ponytail?
Help me /a/, I've been generating cute girls for the last 4 hours.I can't even display my creations, since we've just gone pass the image limit.
>>161183248
time for a new thread?
>>161183248
There is no help for you.
>>161182754
Do you have the noise image for that?
New thread.
>>161183370
>>161183905
Regrettably no, unless the data was cached somewhere.
>>161167790
KILL MEEEEE
>>161178290
They published at Comiket 92
>>161162406
So this is the future.