Anime technology

>>56830243
>>>/a/

That's really impressive. The best i have is a simple script to count the files in each of my cute anime pic subfolders, and return the value. And one to open a random pic of maki. Looking forward to hearing about anime related technology and programming—so have a bump on me

>>56830243
which k-on does it say is the cutest

>>56830243
>not training a neural net with that data to predict the cuteness of new images without needing to specify it yourself

File: Ranked Elo 1657 points.jpg (441KB, 1650x1220px) Image search: [Google]

441KB, 1650x1220px

>>56830284
beep boop calculating calculating
the result is: Akiyama Mio and in last place Mugi

mooo

>>56830319
haha ritsu owned

File: Ranked Elo 1557 points.jpg (465KB, 1409x2000px) Image search: [Google]

465KB, 1409x2000px

>>56830312
that's the next step
I wish the images had tags because then i wo-
actually i can get tags on the images so im going to use the scrores to make a booru searcher that finds the best image based on tags
neural network sounds hard

Mugi is NOT the least cutest

>>56830348
sorry its science

>>56830345
that's completely useless unless they make a "good_art" tag because a ton of stuff on the booru is garbo even if the tags are good

>>56830414
what if it finds some hidden artist in the ghost in the tags or something like that
its already a pain to search boorus so it would just help think of ideas for searches

Can you post the source code?

File: debian.png (1MB, 2133x1200px) Image search: [Google]

1MB, 2133x1200px

is this the new anime thread

>>56830569
no its really embarrassingly bad

>>56830243
Khorosho

>>56830578
It can't be that bad. Please?

File: Ranked Elo 1562 points.jpg (566KB, 1612x2277px) Image search: [Google]

566KB, 1612x2277px

>>56830658
you need to run it from vim or change the part that recoses the window
you need to remove the first empty line from the table generated from table generator
you need to change all the directories to yours

https://u.teknik.io/cr3n8.zip

changing the 'league' string to different folder names (only the parts before the first space lol) will restirct it to only those images
other options are 'random' 'new' 'cuttest' 'high' 'low' and I think that's all of them

goodluck getting it working

>>56830888
pls give nichijou images

>>56830243
i am working on getting caffe to work for my own programs so taht i can make it auto-tag my wallpapers

I am probably gonna hook it to some *bouro or somethign to get lots of pre-done sources to train on

Unfortunantly i dont relly get how caffe is put together. It seems to require putting all inputs and outputs into lmd databases and then uses that.

>>56830243
post the cutest image already faggot

>>56830952
she already did it's on top of the op image

File: 1454564942313.jpg (647KB, 900x1284px) Image search: [Google]

647KB, 900x1284px

>>56830888
>you need to run it from vim
dropped

File: Ranked Elo 1967 points.jpg (2MB, 3000x2210px) Image search: [Google]

2MB, 3000x2210px

>>56830952
here

>>56830976
i said it was terrible
you can change that to whatever title or id you give your terminal running it

>>56830319
hanging breasts are the best

your algorithm has good taste

>>56831038
pls don't be gross

>>56830243
you live an empty meaningless existence weeb

File: 1458961982846.gif (453KB, 426x755px) Image search: [Google]

453KB, 426x755px

What kind of technology do we need to make anime real?

File: 1468799113918.jpg (208KB, 1920x1080px) Image search: [Google]

208KB, 1920x1080px

>>56830243
this is some incredibly autistic shit, OP

but it's awesome. You're the pioneer of weeb software, creating new opportunities for scientists to achieve perfection in cute anime girls. We need people like this to move mankind towards the future where catgirls will finally become reality

File: a2b.gif (243KB, 1200x878px) Image search: [Google]

243KB, 1200x878px

>>56832539

>>56830243
I made a tool to keep track of my anime.
https://github.com/sora-chan/bklg
The code is old and shit though.

Weirdos

>>56830319
Pretty good.

the best technology for us anime fans is a helium tank for suicide! As an anime fan i really recommend suicide to all people who like anime, just like me! Suicide is the fastest way to meet your wi-fu. My wai-fu is Sakura from Naruto, shes really kawayy hahahha lets commit suicide all toghether my fellow anime lovers

Only tech I apply to my animu is an old usb-2 dock station in which I plug various sata drives from time to time to archive stuff.
I stopped bothering to save and catalog anime images many years ago. It's life consuming. I even had an anime galleries website with ~50k monthly pageviews back in the early 2000s.

>>56830243
If you give me the dataset I can make a neural network to generalize on your concept of cute and post it on github for your own use...

File: painting.png (9KB, 410x292px) Image search: [Google]

9KB, 410x292px

i wrote a python script that does it's best to reproduce a cute anime pic using only circles of various colours and sizes
here's the one it's working on now, it's about half way done can you tell who it is???

>>56838719
kindof interesting. Please tell me that it is higher res than that

>>56838719
Ram

>>56838719
why are the circles moving?

>>56838743
it can scale up the circles arbitrarily
while it's running it uses a low resolution because comparing the difference between two images of high resolution takes a MASSIVE amount of time

>>56838761
what?

>>56838719
how does it determine what size circles to use?
does it like check the area around one point to see if the colors near it are similar enough to keep using it?

>>56838719
Ram?

>>56838774
the circles are moving

how long until I can put some helmet on my head that will measure how cute I perceive an anime image while building a profile of my preferences then feeding me the cutest images from the internet, producing a feedback loop that will bring me to the cuteness fixed point in a minimal amount of time

File: 3fc4408b9ea829c63f92101ed7e9e025.png (435KB, 1024x1024px) Image search: [Google]

435KB, 1024x1024px

>>56838778
it starts with random size and position, then changes it slightly, if the new position/radius/color results in an image closer to the original it keeps that change and starts over
>>56838758
>>56838780
you're right congratulations u win the original pic
>>56838800
i don't think they are

>>56830243
I wrote a script that used RSS feeds to download torrents off of nyaa and add them to transmission. But then I discovered flexget.

>>56830926
If you ever get that finished and working, could you please post it?

>>56838824
>randomly choosing
i would have thought that you would do edge detection, pick two random points on 2 lines of the same blob and use that as your circle

>>56838837
i am a ways off right now, but yes
i have their example code compiling on my system using my own build system
i dont quite get how they have things meshing
it seems to require a database to read from which will be terrible for processing MASSIVE amounts of images i think as it probably tries to load the whole thing into memory at once.

>>56830319
your algorithm is garbage

>>56838882
i tried something similar while ago but i wasn't getting very far with it and dropped the idea

i also thought about using edge detection to compare the images, it might result in nicer lines

>>56838949
His algorithm is a sorting algorithm in O(Nlog(N)). It's fairly effective since he seemed to divided them according to cuteness.

>>56839053
No I meant it's garbage because it put Mio in first place.

yui is objectively the cutest keion

>>56839067
then redo the algorithm and see if you can get a better result with the same efficiency.

I've been working on testing waifu2x on 1080p bluray video by dumping the video to individual frames, upscaling them individually, and reencoding. Trying to get good quality anime-specific 4k upscaling...

File: touhou kogasa hug nazrin plush.jpg (615KB, 1000x1000px) Image search: [Google]

615KB, 1000x1000px

it's too bad the algorithm can't take personality into account
yet

File: painting.png (54KB, 520x390px) Image search: [Google]

54KB, 520x390px

>>56839111
precomputation:
1: ritsu
2: azusa
3: mio
4: mugi
5: yui

OWNED

>>56839125
I tried that on something before too. Ran for a very long time. Anyways, i gave up when i couldn't get the audio back in sync. Looked alright though.

>>56839137
friend, i think you meant to type
1) yui
2) azusa
3) ritsu
4) mio
5) mugi
glad i could help correct your typo (V●ᴥ●V)

>>56839141
I'm running it on a university computer cluster, so 20min of footage only takes a few hours to run it on all frames

ffmpeg has weird options for framerate, i suspect i'll end up running into the same issues...

>>56838824
>it starts with random size and position, then changes it slightly, if the new position/radius/color results in an image closer to the original it keeps that change and starts over
that is an genetic algorithm, much like this one: http://alteredqualia.com/visualization/evolve/

I once wrote a version of this that ran on OpenGL

triangles are much faster to render than circles

>>56839155
ui pls go

>>56830319
Mugi has bigger tits than Mio, plus she's plumper and has a better ass.
>Not wanting to have Mugi as your sugar mama

>>56839228
true but i need them to be circles for reasons

>>56830243
If you'd label your charts then maybe I would understand what you want to show me.

File: Yui_Azusa 68.jpg (407KB, 1000x1000px) Image search: [Google]

407KB, 1000x1000px

>>56839235
>only caring about her body
shallow, pathetic, and despicable

>>56839232
:)

>>56839281
Moeblobs don't have a personality

>>56839228
also rendering takes almost no time at all, it's comparing the image that kills it

test

>>56839305
That's the thing, my implementation did everything on the GPU.

It first renders the image in a texture.
Then runs a shader that computes the per-pixel differences between the new render and the best candidate using a customizable function resulting in a floating point texture.
Now I just need to apply a reduce algorithm in order to average all the differences to get a fitness measure.
Fortunately glGenerateMipmap already does this and is implemented in hardware so I call it and read the smallest mipmap level wich is just one pixel (one floating point number).

Also I did some extra stuff to accelerate the generation.
For example based on the best candidate it generates N mutations and render each one following the steps above but intercalated. This takes advantage of queueing behavior of the GPU drivers.
Mutations that produce a worse result are straight out discarded while the top mutations of one generation are mixed if they are compatible (if they don't overwrite their changes) and retested before moving on to the next generation.

>>56839528
>tfw no gpu

>>56830243
https://www.youtube.com/watch?v=z23Gzp0RTOg
Now get your dakimakura.

>>56839235
>Mugi has bigger tits than Mio
bullshit

Any tips on writing a script to download a selected manga from madokami?

>>56838719
Literally generating moeblobs, nice.

>>56830243
Does anybody here scrape the *boorus? I've been thinking about making my own local copy of gelbooru because the website seems to be getting shittier and shittier

I found https://github.com/Oros42/BooruMirror but it's some PHP garbage. I wonder if there's some cleaner solution out there. I also want to be running this as a cron job so it should be easily resumsable and continue scraping where it left off.

>>56838824
share script? I wanna try it with some of my pics

>>56838837
tfw too dumb to use flexget. I've been using rss for deluge for far too long and I have to manually put in both the subscription and the torrent itself. flexget seems too complex for me. wanna dumb it on step by step for me senpai?

>>56842184
Have you tried Hydrus? I've only used it lightly, but it's designed for boorus/imageboards and has autistic level tagging abilities.

>>56839137
wrong
>>56839155
wrong
>>56839235
wrong
>>56839281
yuiazu best pairing

File: 1474421889878.webm (88KB, 226x222px) Image search: [Google]

88KB, 226x222px

>>56830243

1. Mitsu (Mio and Ritsu)
2. Mio
3. Mugi & Yui & Ritsu & Ui
4. Jun and sawako and azusa

tfw no ui imouto

File: 1412891881050.jpg (357KB, 850x1300px) Image search: [Google]

357KB, 850x1300px

beep beep
best slice of life show coming through

>>56830243
2 bay usb3 dock, shared through windows to my network so i can watch my shit on my tv or my laptop.

File: s7wg.jpg (118KB, 850x1142px) Image search: [Google]

118KB, 850x1142px

>>56830243
>liking k-on this much

>>56843040
it's literally mankind's greatest achievement

>>56843050
this desu all other anime is shallow and worthless

>>56842656
It doesn't seem like hydrus can scrape sites like gelbooru at all?

Also, I got the BooruMirror thing to “work” but it's sort of weird. Basically it just scrapes “new” posts via the RSS, and only when people actually perform requests.

I could set up a cron job to do the update automatically, but I sort of want to get an existing archive of all images on gelbooru as well (which the RSS feed does not include)

File: 1471269711940.jpg (54KB, 630x900px) Image search: [Google]

54KB, 630x900px

>>56843050
>>56843103
how about no

>>56843050
This. Only a retarded straightshit would hate k-on

>>56843274
how about gtfo out of my thread idiot

File: 1467311829059.jpg (136KB, 561x814px) Image search: [Google]

136KB, 561x814px

BEEP BEEP
BEST SLICE OF LIFE SHOW COMING THROUGH

File: a70d9209d9d495b260f4820ad46d7f657a9d086a.png (150KB, 500x350px) Image search: [Google]

150KB, 500x350px

>>56839137

>>56843160
What? It's specifically for that.
Here: https://hydrusnetwork.github.io/hydrus/help/getting_started_subscriptions.html
I even see gelbooru in the screenshot.

Reminder Mugi got BLACKED

>>56843482
No, she's a lesbian.

File: Yui_Azusa 12.jpg (175KB, 600x672px) Image search: [Google]

175KB, 600x672px

mugi is boring

not as boring as your dumb posts nerd but i saved that pic

File: Bp77wm4CYAAXtXs.jpg:large.jpg (79KB, 600x861px) Image search: [Google]

79KB, 600x861px

>>56843520
your boring idiot

File: Yui_Azusa 87.jpg (444KB, 1000x1000px) Image search: [Google]

444KB, 1000x1000px

>>56843540
:)
>>56843547
no u

>>56842470
Flexget has really good guides and documentation on how to use it. You can find the install guide here: http://flexget.com/Install

It goes step by step, all you have to do is follow along.

File: yande.re 336664 disc_cover koshigaya_komari koshigaya_natsumi non_non_biyori.jpg (3MB, 1680x2074px) Image search: [Google]

3MB, 1680x2074px

BEEP BEEP
ACTUAL BEST SOL COMING THROUGH

You should get banned for posting anime anywhere but /a/ just like /mlp/

>>56843455
Oh, I see. From the home page it just looked like a “make your own gelbooru”-type of thing, without any reference to the ability to scrape other sites.

Incidentally, I also found https://github.com/fake-name/DanbooruScraper which seems to be at least a step up from the php garbage

File: 1472795227029.png (12KB, 880x131px) Image search: [Google]

12KB, 880x131px

>>56843867
Anime Website

>>56843903
Doesn't apply now that Hiroshima is in charge.

>>56843867
anime website

>>56843867
anime website

File: 1473418015445.png (52KB, 652x480px) Image search: [Google]

52KB, 652x480px

>>56843919
You mean the jap dude who founded 2ch? If anything, that reaffirms the anime foundations of this site.

>>56843878
Okay, I managed to get danbooruscraper to run but it doesn't work at all, seems to be bitrotten; assertion failures, stalls, and the code is also a piece of shit. Finally, it seems to be geared around the assumption that I will only want a dump of the site as it is now and never again (it has no way to re-run or update, and the maximum ID is hard-coded)

tl;dr piece of shit. I'll try hydrus now, but it also seems to be broken to the core... fucking “script”langs man

>>56832539
Virtual reality?

>>56844303
Hydrus is still in active development. I do remember having to screw around with libraries and jazz to get it working, but after that it was great.

>>56844592
Yeah, getting hydrus running seems to be a royal pain. The entire thing is built like shit from head to toe as far as I can tell

It doesn't even print error messages to stderr, it prints them to crash.log in your db dir - which is fantastic if it fails before it even manages to find the db dir (e.g. because of an import failure)

the code is also unreadable because it has fucking whitespace for blank linse as well, which my editor marks as an error (as it rightfully should)...

File: painting.png (14KB, 410x292px) Image search: [Google]

14KB, 410x292px

>>56842450
it's a bit embarassing but ok, i'm not sure how much explaining it needs so i tried to make it just werk
i put some default values in so if you just save something named 'src.jpg' in the same dir as the script and run it then it should work
keep in mind it's really slow especially if you use a large source image, and larger numbers = more accurate end pic but more time to get there
http://pastebin.com/GdsrV1h7

also, the ram pic finished and this was the result
for some reason it never painted her shoulders

>>56844631
holy shit, and when I thought this piece of shit couldn't get any worse

>the git commits
>the fucking git commits
https://github.com/hydrusnetwork/hydrus/commits/master

just end me now. Nobody can seriously be THIS bad at programming. I think it's time to make my own. Fuck's sake

>>56844732
oh i for got it also requires pillow
https://pillow.readthedocs.io/en/3.3.x/installation.html

>>56844732
>>56844782
and again, wait, i just saw an error
fixed
http://pastebin.com/YA99YWTS

>>56844834
http://pastebin.com/6ABGFXWy
ok there, i actually tested it this time

>>56839298
YOU don't have a personality

>>56844740
>I think it's time to make my own.
More choices is always better. It's worked okay for what I've done with it, but I think the python gui gets bogged down a bit.

>>56845953
I'm in the process of writing my own ATM, in case you want to contribute. (I'm writing it in Haskell, and using Xapian as the tagging database + scalpel as the scraping backend)

>>56846058
I've not touched Haskell, but I'd be interested to look at it anyway.

>>56843867
Try killing yourself. That way you will never have to see those evil animes again.

>>56846292
Well, I gave it a start: https://github.com/haasn/hsbooru/blob/master/HsBooru.hs

Currently only the gelbooru scraper is implemented, and it doesn't download the actual files or store the tags yet.

Spent most of my time trying to find a good solution for the tag database but I settled on xapian which seems like a perfect fit. I will be integrating that next.

>>56839528
dude, i love it
i am in a graphics class now and would like to see your code

>>56830319
Mio is best!

>>56844867
Hey dude, nice try but you were getting bit by some small things that are not apparent in python
http://pastebin.com/qBFhdvTH

I modified the Painting class to us image.load() instead of the SLOW image.getpixel()

Also, more applicable to more programming in python and other things- making a list of errors and then adding them at the end is terrible as a lot of allocation can happen. Instead, if you are only going to add them at the end, just add each value as you go

Also - please comment your code
you will forget in 2 days what something is without them

>>56848931
i just had an idea after looking at this run
why not look at a rectangle the size of the circle for the comparison instead of the whole image?
then the number of comparisons is actually a lot smaller per pass, unless i am not really seeing all that is happening when i just look at it every now and then

bedtime bump

>>56838824
>it starts with random size and position, then changes it slightly, if the new position/radius/color results in an image closer to the original it keeps that change and starts over
That sounds incredibly inefficient.

>>56849604
I challenge you to find a more efficient way than a genetic algorithm.

>>56849736
Convolutional Neural Autoencoder

>>56848931
wow thanks that is so much faster
>>56849061
i tried this earlier, the problem is that the circles tended to all group together into places that were already similar to the source image. and in turn left big swathes of empty space

File: scraper.png (36KB, 1431x1062px) Image search: [Google]

36KB, 1431x1062px

>>56846292
>>56848057
Well, we're up and running (basics only)

Still no real front-end though, so right now I'm hard-coding values and using queries straight from GHCi

Won't take much more work though

>>56836797
>https://github.com/sora-chan/bklg
Nice, did wrote something (even more shitty) like that but with anidb ed2k hash support.

Gonna give mpdhanyuu a try too, looks interesting.

File: size.png (29KB, 893x566px) Image search: [Google]

29KB, 893x566px

>>56850440
And now we wait

at this rate it will only take me about half a year

>>56830243
>>56846358
>animes

Fuck off; are you american? The plural of anime isn't "animes" you dumb fucks.

>>56852229
I hope you're throttling. Place like that will ban your IP if you burn to much bandwidth and they have a competent admin.

>>56853087
My network bandwidth is shit, so I'm effectively already being throttled.

Also, I'm not really worried about IP bans. If they ban me, I'll try again with a different IP and a slower download rate.

File: progress.png (267KB, 1000x716px) Image search: [Google]

267KB, 1000x716px

>>56850134
i made the change this mornign to try a small search box around the last stroke instead of the whole image, and so far the results look good

this is roughly 250 strokes in

also i am noting that it doesn't like very detailed images

File: Anime 5.png (10KB, 1517x61px) Image search: [Google]

10KB, 1517x61px

>>56830243
Not much I just changed all my folder icons to anime icons.

>>56853400
shouldn't the circles be smaller still?

How do you determine "cuteness"?

>>56853606
in my opinion, yeah it should be
but in the code the default max size circle is 60/650 * image width
i will change it when i get back to my computer and run another test

>>56853640
Using circles to define curvature, the rounder a object is the more cute it likely is. Other factors play in such as color, size, and position.

>>56853811
Damn, that's some computer vision shit. I could never get into that, must be really complex.

I had a shit grade when I tried to make a crude facial recognition prototype in college, lol.

File: EPSON004.png (445KB, 250x606px) Image search: [Google]

445KB, 250x606px

>>56854096
I actually took a image processing class where we utilized opencv. My last project was kinda crude, but it was to show how a computer can detect shapes on some old drawings I did for /jp/.

>>56854213
yeah I used openCV as well, but I still can't get my head around all that shit. I never took any classes on it though. The most I do now is some simulation stuff with behaviour (rain, cells, movement, etc)

>>56830319
That's not cute. That's just lewd.

>>56850734
>Gonna give mpdhanyuu a try too, looks interesting.
Would appreciate any feedback and suggestions, it's my latest project.

>>56854516
Will do, it's gonna run almost 24/7.

>>56843867
anime website

>>56854902
>anime website
>90% is not about anime

>>56843867
Anime website.
Also, there's more interesting and cool stuff in this thread than there are in any other threads.
Makes you think, huh.

>>56854979
Who are you quoting?

>>56855026
you're mum

>>56855038
I am not mum.

>>56830243
>MAL
:^)

File: 1459438937140.png (168KB, 671x603px) Image search: [Google]

168KB, 671x603px

>I Want to Generate Moeblobs too but I'm Scared of the Command Line?!
What should our heroine do first?

>>56856586
>install gentoo

File: 1408421871600.jpg (219KB, 501x1200px) Image search: [Google]

219KB, 501x1200px

>>56843867

>>56854516
>>56850734
What is mpdhanyuu? Even google doesnt give me any results

>>56830243
Why the fuck are you archiving that much garbage?
You won't watch it ever again, it's useless...

File: googlefornewbs-mpdhanyuu.png (98KB, 779x660px) Image search: [Google]

98KB, 779x660px

>>56857200
Try harder at googling

http://lmgtfy.com/?q=mpdhanyuu

File: painting.png (36KB, 500x716px) Image search: [Google]

36KB, 500x716px

well this is one with 5000 circles
i think i have had my fun with this one

>>56854354
wrong idiot

so i discovered a pretty big bug in the moeblob script here's the fix
now it keeps randomizing the circle until it finds a good spot to start and then it starts mutating it
also uses floating point for the comparison so it doesn't get stuck once changes start becoming small
http://pastebin.com/n6UunMqZ

>>56860603
>>56860603
ugh wrong one
http://pastebin.com/wxqESNgT

>>56860629
i merge the change in with my version, and it seems to help some

currently i am thinking that i might change the mutation method

- have it check each direction for error
- have it check both larger & smaller radius
- pick the smallest error of all 7 (not changing included)

>>56838596
you'll need all 10,000 images too right
internet is spotty right now, can you email me or something and ill get back to you
it's atleast 5 gigabytes

[email protected]

>>56860603
what if you started with a very downsized version of the image for comparisons to make processing faster and then slowly stepped it up to the full resolution over time