[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vr / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / asp / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / qst / sci / soc / sp / tg / toy / trv / tv / vp / wsg / wsr / x] [Settings] [Search] [Home]
Board
Settings Home
/e/ - Ecchi


Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • There are 21 posters in this thread.

05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
06/20/16New 4chan Banner Contest with a chance to win a 4chan Pass! See the contest page for details.
[Hide] [Show All]


All work safe boards are now on the 4channel.org domain. Make sure to update your script blockers and whitelist the new domain.

There's now a setting option under Navigation to display the full list of boards on 4channel.org

The 4chan Vtuber Competition is over. Click here to see the winning entry!




File: Test 1_bars_orig.jpg (193 KB, 1063x1500)
193 KB
193 KB JPG
Read the article here:
https://motherboard.vice.com/en_us/article/43ekxg/machine-learning-algorithm-that-uncensors-hentai-porn-deepcreampy

DeepCreamPy is a Machine Learning algorithm for decensoring images.

The next images in the series will show the few I did on
+ bar sensors
+ mosaic sensors
+ hear sensors

I’ll post the original censored image, the marked up image, and the result.

Any of you tech savvy folks out there should get into this and help make this better. This is sheer win for this community.

GitHub page is here:
https://github.com/deeppomf/DeepCreamPy

Shown here is the original test I did with bar sensors.
>>
File: Test 1_bars_marked.png (1.92 MB, 1283x1810)
1.92 MB
1.92 MB PNG
All sensors must have clean edges, no antialiasing. They must be marked in RGB (0,255,0) for the program to pick them up.
>>
File: Test 1_bars_out.png (1.98 MB, 1283x1810)
1.98 MB
1.98 MB PNG
Here’s the first result. It’s pretty good from the picture perspective. I’m surprised it even seemed to add the urethra. There are artifacts when zoomed it, but from a whole picture perspective it looks quite good.
>>
File: Test 2_bars_orig.jpg (428 KB, 1126x1600)
428 KB
428 KB JPG
A second bar sensor test.
>>
File: Test 2_bars_marked.png (1.79 MB, 1126x1600)
1.79 MB
1.79 MB PNG
This one I grabbed a standoff from just the white bars themselves to see if it handled better. Here most of the vagina is covered, unlike the last.
>>
File: Test 2_bars_out.png (1.83 MB, 1126x1600)
1.83 MB
1.83 MB PNG
Result is still good from a whole picture perspective. It lacks some detail in the center, because there is nothing for the algorithm to interpret from, like the first image. However, grabbing a bit larger of a standoff from the bars yields little to know edge artifacts. But the tradeoff is the removal of detail to interpret from.
>>
Doing gods work anon.
>>
File: Test 3_mos_orig.jpg (1.21 MB, 2280x3476)
1.21 MB
1.21 MB JPG
This is a mosaic. Mosaic is still a work in progress according to the documentation.
>>
File: Test 3_mosaic_marked_ss.png (2.85 MB, 1030x1570)
2.85 MB
2.85 MB PNG
With the whole thing covered, it will be a challenge to train this algorithm well. There is basically zero to interpret from. The algorithm will have to draw an entire anatomy here. (Actually a screenshot b/c output was over 4MB.)
>>
File: Test 3_mosaic_out_ss.png (2.84 MB, 1024x1560)
2.84 MB
2.84 MB PNG
Result isn’t good, as expected. No anatomy to pull from. This needs the most work, but is a super hard challenge. (Again a screenshot.)
>>
File: Test 4_heart_orig.png (1.75 MB, 1736x2456)
1.75 MB
1.75 MB PNG
Now for a heart.
>>
File: Test 4_heart_marked.png (1.44 MB, 1736x2456)
1.44 MB
1.44 MB PNG
Again with the last, it’s covering the entire anatomy. The algorithm is expected to have a hard time generating the entire anatomy here.
>>
File: Test 4_heart_marked_close.png (196 KB, 1156x1098)
196 KB
196 KB PNG
I show the closeup here to illustrate the need for the connected line tool to grab standoff. With a threshold of 80, the select tool can’t grab the antialiased portions without reaching out into other aspects of the drawing (dark lines for legs, pubic hair, etc.) that may be present. A hard standoff has to be created.
>>
File: Test 4_heart_out.png (1.5 MB, 1736x2456)
1.5 MB
1.5 MB PNG
>>
Oops ...
The result has a lot of artifacts. The algorithm has to pull from the anatomy that’s there, which is dark lines and smooth skin. Not to good effect.
>>
File: Test 5_heart_orig.jpg (993 KB, 2280x1607)
993 KB
993 KB JPG
Heart with anatomy to pull from.
>>
File: Test 5_heart_marked.png (2.24 MB, 2280x1607)
2.24 MB
2.24 MB PNG
Marked up.
>>
File: Test 5_heart_out.png (2.24 MB, 2280x1607)
2.24 MB
2.24 MB PNG
Result is actually really good. Only a small bit of artifacts remain on the closest zoom level.
>>
File: Test 6_bars_orig.jpg (1.06 MB, 1600x2238)
1.06 MB
1.06 MB JPG
The bars fair the best when there’s anatomy to pull from.
>>
File: Test 6_bars_marked.png (1.26 MB, 1600x2238)
1.26 MB
1.26 MB PNG
Marked.
>>
File: Test 6_bars_out.png (1.28 MB, 1600x2238)
1.28 MB
1.28 MB PNG
Good results.

So those of you who know how, jump to the GitHub page and start helping out! Hearts and mosaics need a lot more training data.

Cheers,
Caged
>>
File: Salute.jpg (3.31 MB, 2490x3510)
3.31 MB
3.31 MB JPG
You're doing god's work, anon. Glad to see it getting better and better.
>>
Making sure anons see this. :D
>>
>>2256146
How can I help with training this thing, simply by feeding it images to uncensor?
>>
File: Test.png (746 KB, 724x1024)
746 KB
746 KB PNG
>>2256146
>>
File: Test.jpg (464 KB, 724x1024)
464 KB
464 KB JPG
>>
File: Test.png (754 KB, 724x1024)
754 KB
754 KB PNG
>>
>>2256169
>The bars fair the best when there’s anatomy to pull from.
And you know what that means.
Japan is going to ban bars when this explodes
>>
>>2256521
you could help by feeding it decensored images and their censored version

I'm going to post a few pics I decensored using it.
>>
File: yoshikadu 71456253_p4.png (1.42 MB, 2238x1600)
1.42 MB
1.42 MB PNG
>>
File: yoshikadu 71456253_p5.png (1.43 MB, 2238x1600)
1.43 MB
1.43 MB PNG
>>
>>
two more
>>
>>
>>2256164
Is there no way to feed it other pussies so it uses those as samples for the missing one?
>>
>>2256680
It has been trained with thousands of uncensored pussies, but sometimes there just isn't enough information to go on. It doesn't have an anatomical model behind it.

It's better with horizontal bars than vertical ones, for example, because there's more information between the bars to tell what's going on.
>>
>>2256146
Ok, I know python and will check out the code but I've never done anything with tensorflow. I only see instructions to use it but not to train it. First time I see something of value out of machine learning instead of marketing shit.
>>
>>2256688
To train it you should probably talk to the developer first so you don't have to start from scratch. He might still have a thread on one of these boards, last time I saw it it was on /d/ I think.
>>
>>
>>
File: dikk0 70422948_p0 - ENpq.jpg (3.66 MB, 4972x6118)
3.66 MB
3.66 MB JPG
>>
>>
>>
>>
>>
>>
This is revolutionary! Fuck, I can't wait to see how far this goes.
>>
File: improvement.png (133 KB, 512x256)
133 KB
133 KB PNG
https://github.com/deeppomf/DeepCreamPy/releases/tag/v1.3.0-beta
DeepCreamPy v.1.3.0-beta is out!
Updated neural network! Fixed the neural network skipping mosaic censors that were covered by bodily fluids. The tradeoff is while before you could be really messy with coloring the mosaic censors, now you need to be more precise with your mosaic colorings.
>>
>>2256782
very good, I wish there was a gimp tool that helped with selecting the mosaic tiles better.

Also, do you think eventually there will be a way for the bar decensoring to work with semi-transparent bars? It should allow reconstructing much better what's under the bar, when it covers most of it like in this example.
>>
Because right now it just removes any sort of detail why I try.
>>
>>2256862
I think that's pretty easy to add. Will have to come after the UI is done.
>>
>>2257007
Really? That's awesome.

Seriously, what you're doing here is indistinguishable from magic to me.
>>
>>2256146
that's cool and all but can you use imagemagick before you upload them here, ie:
convert decen.png decen.jpg
>>
>>2257147
What? Why? And why imagemagick specifically?
>>
>>2257177
because they're big ass meguhbytes
>why IM
what else for bulk image stuff? its optimised and simple
>>
File: 8720.png (2.65 MB, 1463x1807)
2.65 MB
2.65 MB PNG
Fuck, this is awesome. If you really look closely side by side with the original, there's some imperfections, but overall it looks fantastic.
>>
I made a few more
>>
>>
>>
this turned out a bit blurry bit still better than the white bars
>>
>>
>>
>>
>>
not a lot of detail here
>>
File: karlwolf 9861890 - CMNP.jpg (2.05 MB, 2514x3467)
2.05 MB
2.05 MB JPG
this looks pretty nice
>>
>>
>>
>>
>>
again a bit blurry
>>
>>
>>
>>2257180
>because they're big ass meguhbytes
some people like the quality you know, just convert it yourself if you have a toaster from the 80's
>>
>>2259009
guess you're right, there could be some visible degradation
>>
I get a CPU error. something to do with ATX. I've tried both see and regular version. Anyone been able to fix this.
>>
>>2259252
I'm guessing (because you're a brainlet and didn't even post the message you complain about) but that's not an error, that's an info message saying that the program doesn't use all of your CPU's capabilities.

Most likely you did something else wrong.
>>
Made a few more now.
>>
>>
>>
>>
>>
>>
File: n.jpg (2.5 MB, 4093x2894)
2.5 MB
2.5 MB JPG
>>
>>
Bump. Anyone else playing with this?
>>
>>2263635
Tried to, doesn't seem to work with my PC, or something. Hopefully, it'll be a standalone program, one day.
>>
>>2263635
I'll give it a try later. I've never used it before though.
>>
>>2263679
Tried using it, figured out I don't have an AVX type CPU. Luckily there is a sse version of the program and it actually converted my images for me.

Unfortunately it just pixelated the area I colored pure green with. I'm probably doing something wrong, maybe putting too much green or the area is too large to decensor with quality. I'll give it another try later.
>>
Back with a few more. Probably last time I'll post if it continues that nobody cares.
>>
File: ijigendd 72016788_p0.png (667 KB, 1000x1000)
667 KB
667 KB PNG
>>
File: leaf_lsd 71908515_p0.png (1.07 MB, 750x1175)
1.07 MB
1.07 MB PNG
>>
File: linupers12490 72018130_p0.jpg (2.47 MB, 2480x3508)
2.47 MB
2.47 MB JPG
>>
File: user_jpmm5733 70190602_p1.png (3.29 MB, 1748x2480)
3.29 MB
3.29 MB PNG
>>
File: add_furnace 71616698_p2.png (1.04 MB, 985x1540)
1.04 MB
1.04 MB PNG
>>
>>2256680
>>2256681
Maybe have to feed it with those images where the heart is just translucent enough to make out details.
>>
>>2257147
>compression in my porn

go FUCK yourself
>>
File: 1517516439982s.jpg (8 KB, 250x187)
8 KB
8 KB JPG
>>2264868
I'll try
>wish me luck





Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.