Decensoring Hentai with Deep Neural Networks

https://github.com/deeppomf/DeepCreamPy

by cosentiyes • 8 years ago

314 66 8 years ago

RankingMember• 8 years ago

Gotta give them some credit, that is a brilliant name.

EForEndeavour• 8 years ago

Next up in deep learning for porn augmentation: TenseHerFlow.

eberkund• 8 years ago

Are there any more examples? Removing a few lines doesn't seem that impressive and can be done with the PhotoShop pattern tool without any fancy neural net.

claudiawerner3• 8 years ago

There are examples posted on 4chan's /h/ board, that is to say, people using it for its intended purpose rather than only with examples, but I'm at work so I can't really browse it at the moment (and the threads are ephemeral anyway, unless you use an archive site).

morenoh149• 8 years ago

some threads are archived, your sibling for example

nathanasmith• 8 years ago

More examples and commentary found here: https://boards.4chan.org/g/thread/68220714#p68220714

Edit: should have mentioned NSFW

zawerf2• 8 years ago

I believe "Image Inpainting for Irregular Holes Using Partial Convolutions" has been superseded by "Free-Form Image Inpainting with Gated Convolution", especially for interactive use cases?

https://arxiv.org/pdf/1806.03589.pdf

http://jiahuiyu.com/deepfill2/

I think the most interesting part of this project would be generating the dataset. That is a lot of manual redrawing of vaginas/penises. Or the reverse, where you find already decensored stuff and add realistic censorship yourself. Either way, a lot of time will be spent tagging genitals to build a large enough dataset to make NN techniques work...

yorwba4• 8 years ago

It's not clear to me that you'd actually need to manually label individual images for training. The work of detecting and marking the censored regions is left to the user, so the model just needs to be good at inpainting. There are probably enough specialized sites to get a decently-sized training set with little effort beyond writing a crawler.

zawerf2• 8 years ago

You could be right, I don't know. Maybe the authors can tell us about their experience.

My intuition is that the dataset will be too imbalanced to learn anything useful. Even if you crawl only decensored images, the area you truly care about inpainting is still pretty small. If you don't focus on it somehow, it might learn how to inpaint anime-style geometry correctly (from the rest of the image) but produce "barbie doll" style anatomy.

gwern3• 8 years ago

I suggested that to deeppomf a while ago (I was thinking of simply using unlabeled anime images from https://gwern.net/Danbooru2017 with random area deletions to simplify the model & training process as much as possible) and his belief is that because genitals are such a small fraction of any images, and the rest of images vary so much while genitals are a fairly small narrow domain, a generic inpainting/denoising CNN will learn to inpaint pretty much anything else possible and neglect genitals specifically.

Presumably if you trained a really big inpainting CNN a lot, it would learn genitals (along with everything else), but it's understandable that he would try a much more targeted approach.

hnaccy• 8 years ago

There's tens of thousands of manually decensored manga, you could download and compare them to original to identify censored regions and use as training I suppose.

RIMR2• 8 years ago

Is it bad form to share one of these "datasets" here?

GuB-42• 8 years ago

Hentai is censored only for the Japanese marked, because it is a legal obligation. When it is exported, the uncensored version is normally used.

So you just need to find both versions in order to make your dataset. Add the large amount of decensored artwork and it is not that hard to build a huge dataset. It may even be automatable by parsing imageboards.

claudiawerner3• 8 years ago

Are you sure about this? How much hentai is actually exported? I get the impression that it's very little, and very little of the Japanese releases, due to the fact that most hentai sites have English translations provided by fan groups, and they're scanned translations of what's sold in Japan.

oh_sigh• 8 years ago

If only because censored only hentai is popular that someone came up with this idea. I'm not a hentai expert by any means but I'd imagine a lot of hentai is never exported and so no uncensored version is ever published.

corysama• 8 years ago

If there’s too much manual work, you aren’t using enough layers! ;) Seems the smart approach for a... “focused” AI would be to manually label a bunch of nsfw bits and use that to build a censor bot that auto-detects and censors the of-interest regions. From there you can generate a large corpus of censored-uncensored pairs with which to train the decensor bot.

huntie2• 8 years ago

Saw this a few days ago. It could use some work for mosaic censors but bar censors looked pretty good. I heard it was just some guy on /g/ who said he would make it and then actually did it, so I'm pretty impressed.

skrowl• 8 years ago

Finally, a use case for Deep Neural Networks that we actually need as a society!

a1369209993• 8 years ago

I wouldn't say need per se, since this is something that can and has been done in other ways already, but it's at least a use case with positive, rather than negative, value.

asianthrowaway4• 8 years ago

I wonder what will happen when neural networks can generate illegal content, like child pornography.

bcOpus• 8 years ago

It will vary by country. In a lot of Europe and the US the content will still be illegal. In Japan it won’t be. At some point the issue is going to come down to what you want to spend manpower and money fighting, people who actually endanger and harm children, or people who use computers to emulate it. It might not be a very clear cut thing though, because I frankly wouldn’t be shocked to find out that the overlap between people who want pornography of real and simulated children is significant.

Where it might be a little more clear cut is feeding the market that clearly exists for “teen” porn, which is currently filled by 18+ actresses. I suspect a non-trivial percentage of the non-child-abusing population would consume simulated 16 year old porn without ever considering the real thing. In that case you might see some changes.

Reply edit It’s already outlawed. In the US for example pornographic drawings of children are outlawed. In the same way it’s illegal to try and sell cocaine stimulants, it’s illegal to try and sell child pornography simulations. I doubt the A.I. itself would be illegal, just its output.

asianthrowaway4• 8 years ago

But how would you outlaw it? A neural network is simply a set of coefficients, would certain coefficient combinations be illegal? It's an interesting question.

PS: I am obviously against child pornography, I'm just curious about the implications of being able to generate illegal content and how governments would deal with it.

int_19h2• 8 years ago

Reply edit It’s already outlawed. In the US for example pornographic drawings of children are outlawed. In the same way it’s illegal to try and sell cocaine stimulants, it’s illegal to try and sell child pornography simulations.

https://en.wikipedia.org/wiki/Ashcroft_v._Free_Speech_Coalit...

int_19h2• 8 years ago

Same thing that happened with 3D-printed and home-CNC-machined guns - there'll be a moral panic, and a demand to ban the tools.

crimsonalucard• 8 years ago

This is genius. Give the network a bunch of photos with a persons' face censored (donald trump, for example). What would it replace the face with?

userbinator• 8 years ago

What would it replace the face with?

Judging by the examples, correctly-coloured yet featureless skin.

HONEST_ANNIE• 8 years ago

I predict that the porn industry will be the first industry to suffer mass unemployment due to AI.

Phase I: Deep learning combined with realistic physics simulation will eventually replace human actors. At some point amateurs can record their movements, then use variational autoencoders to improve their features (bigger boobs, thinner waist, smooth skin) and replace professionals. This is 'Uber phase' for porn and drives down the wages.

Phase II: Eventually AI can mimic complete scenes and adjust the 'plot' to the current state of the observer. Tight sexual biofeedback loop will form.

Phase III: The rapture of the nerds. Fully interactive VR + toolset. In the context of the Fermi paradox, Phase III is the Great Filter.

----

https://en.wikipedia.org/wiki/Great_Filter

dna_polymerase• 8 years ago

What? Think about it, call center employees are going to suffer first on a large scale. NLP is getting better and Google Duplex already showed us that human voice is no problem anymore.

It's way easier to remove call centers than coming up with physics simulations that replace humans. Also I think porn is incredibly cheap to produce so everything must be automated to beat the current production cost.

RIMR2• 8 years ago

I am going to call this the Great Stimulus.

dmitrygr• 8 years ago

"Diamond Age" has a subplot to the tune of this

failrate2• 8 years ago

Porn has always been the earliest adopters of technology. While porn actors may be individually replaced by deep agents, actual porn sales will do fine, especially since extremely targeted pornography will be in everyone's budget.

A market for verifiable human pornography will exist alongside the new porn.

Also, have we already dubbed virtual ML pornography "deep throat"?

setr• 8 years ago

I thought it was deep fakes?

TeMPOraL• 8 years ago

Also, have we already dubbed virtual ML pornography "deep throat"?

No, that term is already established in the industry and means something else entirely.

hhjinks• 8 years ago

especially since extremely targeted pornography will be in everyone's budget.

This is why the ML pornography excites me. All my twisted fetishes that nobody caters to can finally be fulfilled.

anonytrary4• 8 years ago

There should be NSFW examples in the Readme. I saw some practical (NSFW) examples on 4chan and it doesn't seem to work very well. Most hentai doesn't contain simple lines and blobs covering homogenous skin area. There is detail under those areas that the algorithm does not reproduce at all.

benatkin2• 8 years ago

Not when Microsoft is about to acquire GitHub!

anonytrary4• 8 years ago

Microsoft serves abundant amounts of NSFW content on Bing, so I can't see this being a valid concern. I highly doubt Github is going to be censored.

flyingfences3• 8 years ago

On the contrary, the examples I saw on /h/ looked pretty impressive for the most part.

anonytrary4• 8 years ago

Well, I guess we have different definitions of impressive. The de-censored areas did not reproduce genitalia properly at all.

matachuan• 8 years ago

The project name is GOLD

anon_cow1111• 8 years ago

I wonder how this deals with the stippling seen in black and white printed pages. As someone familiar with digital drawing/editing, any slight imperfection in the pattern becomes super obvious and it's even worse with stippled gradients.

I'm sure it'll be a useful tool with refinement, but this is still a solution to an artificial problem. There's no reason for Japan's current censor laws to exist in 2018.

flyingfences3• 8 years ago

FAQ says they currently make no attempt to deal with that and it straight up won't work if you try.

Brian_K_White• 8 years ago

At least for the most common forms of obfuscation, pixellation and fuzz, why do you need AI for this? It seems conceptually simple to interpolate from several similar but moving frames, how a small obfuscated feature travelled across a huge pixel boundary to cause the small change in the entire oversized pixel.

You'd need I, be it AI or the normal kind, to invent a guess from a single frame. But watch frames that each differ a little over time, and you should be able to deobfuscate pixellation and fuzz with plain math.

Conceptually simple. Not saying I could knock it out over the weekend, just that there doesn't seem to be a mystery about how to proceed that puts it into "throw some ai at it and see if anything happens by ai magic" territory.

sheharyarn• 8 years ago

This is exactly why AI was needed (and subsequently created) in the first place!