That user who cockteased us with the hentai decensoring NN finally released it

That user who cockteased us with the hentai decensoring NN finally released it.

Attached: file.png (1830x1470, 1.5M)

Other urls found in this thread:

github.com/deeppomf/DeepCreamPy
imgur.com/a/v41LmGm
imgur.com/a/tyjKay9
github.com/MathiasGruber/PConv-Keras
github.com/keras-team/keras/issues/6176
opengameart.org/content/mermaid
paintschainer.preferred.tech/index_en.html
stackoverflow.com/questions/47068709/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u
pixiv.net/member_illust.php?mode=medium&illust_id=71159739
a.pomf.cat/wwizda.png
ghostbin.com/paste/z4nq8
twitter.com/NSFWRedditImage

github: DeepCreamPy

Truly based.

If that user is here, there's nothing listed about training.
I assume we can keep training using deepmindbreak?

God bless that user

github.com/deeppomf/DeepCreamPy

Does this technique recolor the green areas based on some sort of average smear and blur method? The resultant image in the OP looks like the smear tool in GIMP.

someone post real world samples

Mosaic: imgur.com/a/v41LmGm
Bar: imgur.com/a/tyjKay9

Not all heroes wear capes

based

Wow these are actually decent holy shit.

Has anyone gotten this to work on a non-included-sample image? Every image I try always results in "Found 0 censored regions in this image! No green regions detected!"

read the docs, you need to paint the censored areas with pure #00FF00 green so it knows what to modify

I did you baka. Perhaps it is a color space issue. Pure green in what colorspace. sRGB? RGB?

It doesn't say what to name the images.
How the fuck is it supposed to tell which is the image to decensor and which is the image with the green?

Attached: 1519615943977.png (925x351, 48K)

Place them in different folders. Original goes in decensored_input_original.

Holy shit it actually works well.
Truly doing God's work. God speed.

big if true

Attached: pork.jpg (192x192, 13K)

>Deep Leaning buzzword actually good for something

>He believed that because "AI" and "machine learning" became buzzwords, actual AI and actual machine learning had no merit.
NPC tier.
You don't deserve your lewds uncensored.

>It does NOT work with Black and white/Monochrome image

Attached: INTO THE TRASH.jpg (700x479, 42K)

Where's my apology?

Attached: APOLOGIZE.png (981x264, 35K)

I for one never doubted you for a single second.

>find color ten times more erotic than bw
>99% of it is western
Life is suffering

ImportError: `load_weights` requires h5py.
>Still hasn't fixed this issue.
And yes I have h5py installed.

Well work user. How long until the computer can make whole picture?

Don't know what the fuck you're on about.
95% of ecchi and hentai art is colored.

Training code is from github.com/MathiasGruber/PConv-Keras
Neural networks
I thought srgb is the default for image editors?
see github.com/keras-team/keras/issues/6176
Years

>colorspace
mate, you're entirely overthinking it
it's just reading raw rgb values
strip the color space info if you think that it'll help

>save marked-up image as png
>put in ./decensor_input/
>if it's a mosaiced image, put the original (as png) in ./decensor_input_original/
make sure the names for the files in each folder the are the same

the shortcut for mosaiced images is broken, so just make a batch script with "decensor --is_mosaic=True", no quotes in it and use that
otherwise, just run decensor.exe

managed to get a mediocre result out of it with one test image -- three censored areas, one dick and two pussies totally censored
one puss was entirely ignored despite the program saying it did recognize the censored area
the other puss that did get uncensored is mildly fuzzy looking, but it's much better than before
and it really didn't like the guy's dick, it did a hilarious job of trying to redraw it, the edges of it still have a bit of mosaic and it's a bit lumpy
it did do a decent job of trying to draw veins on it

this
it's mostly doujins that are an issue
most individual pics are actually colored

l-lewd

Attached: 1538569547199s.jpg (235x249, 8K)

so, as neat as this is, sometimes it doesn't seem to even make an attempt on mosaic images
marked an image correctly, it detected it, it said it decensored it
result showed no change, comparing it with the original image reveals zero differences at all

the purely content aware fill mode always produces some kind of results, if not always great ones

>srgb is the default for image editors
Neat it works when using sRGB. Arigato user.

I think I'll need some examples

give me sauce on that semen kraken

check'd

opengameart.org/content/mermaid

I admit the mosaic decensor is nowhere near as good as the bar decensor. The mosaic decensor was added second to see if a single neural network can do both types of decensors.
The neural network uses the green region as information about what's censored for bar decensors but it can't for the mosaic decensor.
In the future, I think I'll train separate models for mosaic and bar decensors.
See

31yo boomer here, i downloaded all the shit that is in the requirements.txt now what the fuck do i do with all that shit? i just want to test this TECHNOLOGY.
can some1 do a braindead proof tutorial to this shit?

it's happening again.

Attached: omrskw.png (449x174, 52K)

Hey, let's use these NN's to cure diseases!

Attached: 1536523356909.gif (480x270, 236K)

>31yo boomer

Attached: nipples.jpg (111x211, 20K)

Why is /a/ so productive?

use an AI to colour it first
paintschainer.preferred.tech/index_en.html
>the absolute state of humanity

>README clearly explains how to use it

RTFM

im retarded. blaming the 2nd bottle of beer i guess

no man should be this based

>github
>lewd
Gitlab you fucking morons
GITLAB

>Literally just checks each pixel for (0,255,0) value and goes from there
Novel, but hardly an effective use of NN. The fact it has to search for an exactly perfect green pixel in the first place tells me this guy doesnt really know a lot about ML based algorithms, and this image is an ideal test case rather than a real world use.

Any good suggestions to improve the model?

github.com/deeppomf/DeepCreamPy
I don't know m8, looks like github to me.

He's implying that it should be hosted on gitlab instead because github has a tendency to take down such projects

Fucking awesome. Gonna test this later.

Proprietary botnet

I don't think you understand what those words mean

porn is a powerful incentive

>no linux binary
dropped

retard

Anime is not a very time consuming hobby after you clear your backlog and start on the seasonal treadmill.
Have to do something besides tickle my pickle and cumbutton until my cock is raw and my balls are sore.

Attached: mfw_got_my_keyboard.gif (480x270, 502K)

>cumbutton
?

Attached: Butts.jpg (1084x771, 327K)

the problem is clearing the backlog. Idk if i'll ever catch up.

Using TensorFlow backend.
2018-10-25 23:25:10.263151: I T:\src\github\tensorflow\tensorflow\core\platform\cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2. Should I just compile it myself?

so basically automated content aware fill

It's not an issue. See stackoverflow.com/questions/47068709/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u

I've tried it out and it's... alright I guess. The bar decensor works much better than the mosaic decensor but still leaves noticable marks and green patches in the image. I guess I shouldn't expect it to be perfect

>still leaves noticable marks and green patches in the image
read the docs, you're explicitly fucking up and it'll tell you what you're doing wrong

it's mostly just a content aware fill system

How do I install this without downloading python etc

Can't it just be a program

>the fucking state of Jow Forums
read the rest of the page until you see the binary download link you fucking brainlet cretin

>The fact it has to search for an exactly perfect green pixel in the first place tells me this guy doesnt really know a lot about ML based algorithms
rule number one. don't train things you already know or can define

Attached: 1517413734663.png (231x224, 11K)

Is this not against github’s own CoC?

I'm curious why you made the green image the model input, instead of just using it to build a mask over the original. Or am I misunderstanding things?

Nope. I checked their community guidelines.
For bar decensors, the green region is converted to a mask.

Is the guy who made it here? Are you the same guy who wanted to sell this product at comiket? I absolutely would purchase it.

Im approaching the end of it because im fairly picky and i have no idea what am i gonna do in my nonproductive free time other than browsing this site

Attached: IwM78gl.jpg (600x540, 34K)

What does your dataset look like / do you have any training code up?

pixiv.net/member_illust.php?mode=medium&illust_id=71159739 into a.pomf.cat/wwizda.png
>can't be censored if it doesn't exist
It's shit.

Not him but almost certainly a boatload of censored-uncensored pairs either made by the artists themselves or on the tons of tons of requests for that drawfags fill every day.

Yes, but a trip to Comiket is very unlikely.
I will set something up for donations.

This runs into an error on my computer, don't know what to do to fix it.
Traceback: ghostbin.com/paste/z4nq8
Using the pre-built binary on windows 7 (64 bit).

Honestly just run it from source instead of the binary.

It shouldn't be too hard to write an other tool that recognizes and marks censor bars, right?

Test post

Depends, but it's probably easier to write than this one.

humanity BTFO, make way for the AI revolution

I know what I'll be working on this weekend: integrating both systems in a complete hentai greatness pipeline

God bless him.

Neat.

Attached: MIR4BP0IHQQO8KVGAJS9SFU4BO7NTV4Y_0.jpg (512x712, 125K)

it's doable. just takes a lot of time to prepare dataset.

yep. i hope i reach the end soon so i can finally be productive.

how is it different than photoshop's content aware fill

>after you clear your backlog
That's a myth.

I'll try it on some pony porn later.

>mfw Decensoring-NN for JAV is only a matter of time

Attached: latest.png (184x208, 35K)

It's another episode of pedophile bullshit on Jow Forums

the way it should be.

Great job!

Attached: 1476311607265.png (420x322, 135K)

but WILL IT BLEND?! XD

retard