Machine Learning Makes Your Waifu

- makegirlsmoe.github.io/assets/pdf/technical_report.pdf
- make.girls.moe/#/
- waifu2x.udp.jp/

Other urls found in this thread:

github.com/hanzhanggit/StackGAN
mtyka.github.io/machine/learning/2017/06/06/highres-gan-faces.html
twitter.com/SFWRedditGifs

Fuck off already, this belongs on /c/ or something, not Sup Forums.

Free beta testing is over

>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.
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"?

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.

Something went very wrong.

NANI!?

CUTE!

Why do all these girls look like they came straight out of KAWAII UGUU land?

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.

7/10
rate my loli

The robots still can't beat us at producing true smug, I guess

I think it made it so smug that it's face melted a little

...

higurashi/10

9/10 my aesthetic

This time the hat came out non terrible!

>wearing a used condom for a hat

I would fuck her.

oh that's a good one

What has science done
and how will we atone for such a monstrosity

what did you do
WHAT DID YOU DO

She looks tsundere.

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 digits
don't roll a 94 you sick fuck

they can come close

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

>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.

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?

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

Dude weeb lmao

Technological limitations. The higher resolution, the higher probability of "chimeras" (fucked up images).

>She sees your dick

Wow. I want to be M for her.

>he wouldn't fuck his RNG waifu for ants

I just want to generate my waifu Cocoa, but she comes out weird.

Would you watch it?
Who is best girl?
How angry would the winner of MCbowl make you?

Roll

Needs more hair options. It should have a front and back option. Oh, and a fang option.

94 is best girl.

Trap and mom are either the same person or related.

with more gan wrangling
it is possible to get higher-resolution samples that are still coherent (albeit strange looking)

see
github.com/hanzhanggit/StackGAN
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

>Childhood friend wins MCbowl
Tsunderefag salt would sustain me for eternity

certified reverse trap

ive perfected the cyclops

grats

...

Yes, they forgot to add a skin color option

>network error
>tfw no qt ai gf

reality is bleeding/10

Try switching browser.

>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

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 Sup Forums

Cute

rate my gremlin

open mouth smile makes some really genki girls
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

>CMU
Sasuga #1 CS program

>input glasses
>don't get glasses
She's still super cute. I like what I got.

Is this someones thesis or something?

What if VN waifu desins are already generated using this method?

god tier taste brother

roll

this is the future we chose

>ME WAIFU
>ME SMASH
>proceeds to suplex the MC affectionately

Looks like from Initial D

each waifu is unique

ara ara

>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.

Which one, Sup Forums?

...

...

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