Generate and post. Let's go!
I never thought a machine-based algorithm would create my new waifu.
it means you're not allowed to join the fun
I wanted a delicious tomboy, not this piece of shit.
this kind of tomboy? with hair so short like a dude?
This one is probably the most clean girl I've generated so far, not much distortion
what if you get pic related as your SSS grade gacha pull?
>your loli makes this face when you show her your penis
Most style-transfer work thus far has preserved feature locations in the original image, making it unsuitable for real <-> anime translation
For example, the nicest looking style transfer work I've seen is https://arxiv.org/pdf/1705.01088.pdf, which gives you results like attached
Another SotA related work is CycleGAN
but it isn't quite suitable for the same reason–it doesn't naturally recognize spatial distortions.
The closest is probably this recent paper on generating photorealistic images from segmentation maps
If you could get a segmented dataset of anime pics as well (which seem doable, given that the only parts you care about segmenting have lines already drawn around them already), that would probably be enough to start building a network that took in photos, segmented them using normal object segmentation, then filled the segmented regions with plausible anime stuff.
Will likely be a while before anything like this actually happens, though; there is not a whole lot of engineering effort going towards making anime real at present.
this generator is a goldmine for retarded anime reaction images
A shame about the bottom portion of this one, everything else turned out fine though.
Site name is a lie. Clearly there is a cheat code to make boys.
>Paper authors are all Chinese
Will China be the first to make anime real?
This is pretty cool
There's the new /a/ at it again. Is no one here austistic anymore? They're clearly clueless as to the true meaning of moe if they think that facial attributes are what matters. Is this the /a/ where people don't know at least know that there is a difference between cute and moe? Also half of these look like dogshit because the AI learned from danbooru and there are a ton of horrible looking fanarts on there. Also they're all Chinese.
>learned from danbooru
pic related, it's me reading that
Did you read the fucking report? They specifically said that danbooru was shit and too varied to give good results
They used some game design place
Jesus Christ, where are all these retarded summerfags coming from
not sure why you're replying to me instead of the baka who didn't read the report, but, yeah
all the girls made from that trash look weird and ugly
They're not all bad anon
Most of them are tho
I never expected to get master mangaka level drawings, it's still better than anything I could ever draw myself.
Repost, but I also narrowed it down to one between one-eyed mexicano mustache and slotharella with a very delicious eyehole.
She doesn't look half bad when enlarged.
The source images are small to start off with. It says in the paper that they tried enlarging them with waifu2x, but that with all the noise in the final images resulting from the combination, even decent-looking ones didn't turn out well when upscaled. The solution would probably be to get a better source image set.
The URL of the generated image contains a long base-64 string, and running that string through a decoder (like https://codebeautify.org/base64-to-image-converter) reveals the top half of my waifu's face, i.e. from the mouth and above.
She actually turned out pretty well. Next time I guess I'll make it more random, but I need more cute blonde girls in my life.
I don't think you understand how code works
That long string is probably a raster encoding of the image, I'd expect modifying it to be about as effective as changing the pixels manually (i.e. it would defeat the purpose of generating them algorithmically).
Make some tweaks in Photoshop. Interestingly, the original picture had 2 different uniforms on either side of her body - I took one half and duplicated to the other in order to make this pic.
One noticeable flaw of the algorithm is that each of the two eye shapes appears to be independently generated, and unfortunately I don't think it's something that can be fixed with minor tweaks (simulation models aren't my forte).
Thanks, have a cute girl while I'll get around to generating even more.
I am the bone of my templates
Code is my body and algorithms is my blood
I have created over a thousand waifus
Have withstood pain to create many templates
Reading their paper in more detail now (though my knowledge of machine learning is cursory at best so my conclusions should be taken with a grain of salt).
Apparently the data labels come from image tags, which means the algorithm is optimizing for ease of tagging rather than art quality. If the creators could add a feedback button for users to rate the quality of the output, and incorporate that feedback into the loss function, the algorithm might be able to generate better images over time.
Also regarding the face issue:
>we only need to sample some points in local regions around real images [...] this can be done by adding a gradient penalty term to the generator loss
Not sure what they mean by "local regions", but if they have some way of drawing boundaries around the eyes/hairs/mouth etc. they could adjust the metric to be something other than (linear) Euclidean distance, which would improve the consistency of the facial features.
Yet those waifus will never be in anything
so as I pray, Unlimited Template Works.
If not for the mismatched eyes, some of these could probably pass off as drawn images.
Pretty good, but her glasses seem to be fusing into her face.
On average, the smug faces seem to be the best-looking ones.
Maybe it's because slight facial distortions become excusable when paired with a smug expression?
Actually this is the best looking one.
You can also use import and export to save the noise and even share it, here is the noise I used to generate sort of Catgirls.
Oh wow, I just noticed they added an import, export button for the noise, nice.
And about running locally, I meant running it with a local web server, on my pc.
I messed around with that noise in MSPaint a bit and made it worse.
it just looks at the top row of pixels of the noise pic. You can use all colors but it will ignore the red value and set it to 255. So you only have to change blue and green.
faggot here turned out pretty well >>161188824
this reveals all sorts of interesting things about this
like how it doesn't understand the difference between ties and ribbons and other such spaces in the center of the chest and also how it doesn't understand how shoulders work at all
I used 2 images with the same settings but different noise as a base and then I swapped 4 bits in the noise code with those of the other pic
here is what happens with the noise
It's essentially a neural network; it'd be a stretch to claim that it even understands faces at all.
>you lived long enough to see computer generated waifus
We're really living in the future. Imagine what we'll be able to do in a few years.
In the future, there will be computer generated shows and movies.
I converted a generated waifu into noise, then fed her back into the AI. Further testing needed.
Do you know that you can subtract/add different noise vectors to get a certain property?
This expression is clearly trying to convey her superior intellect.
Oh god. This is like an apocalypse.
I can imagine that a year or two later, when the AI can generate images with higher resolution, some amateurs will actually use them as sprites in their VNs.
finally got the interpolation script to work in safari
at a cursory glance it looks like an on/off thing
i guess i could do first half of frames with, second half of frames without?
i'll set up option interpolation after i get auto-gif generation working, definitely worth playing with
finally got in-browser, ping-pong'd gif generation working, took forever because i am a dummy
posting two more now that they're easy to make
now to get breakfast and work on option interpolation
her smug is distorting reality itself
Do one that makes a collage from an interesting subset of the different options with the same seed please
I want to see how the waifus vary with stuff like hair length from the same seed
I guess all eight combinations of blush/open mouth/smile in particular are a must
How could even come to that conclusion?! She's clearly looking forward to going to school with her best friend!
I was hoping for something more Kyouko ish but this will have to do.
This generator can create some really believable VN heroines.
Apart from the wonky eye, she looks pretty cute.