I believe "Image Inpainting for Irregular Holes Using Partial Convolutions" has been superseded by "Free-Form Image Inpainting with Gated Convolution", especially for interactive use cases?<p><a href="https://arxiv.org/pdf/1806.03589.pdf" rel="nofollow">https://arxiv.org/pdf/1806.03589.pdf</a><p><a href="http://jiahuiyu.com/deepfill2/" rel="nofollow">http://jiahuiyu.com/deepfill2/</a><p>I think the most interesting part of this project would be generating the dataset. That is <i>a lot</i> 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...
I predict that the porn industry will be the first industry to suffer mass unemployment due to AI.<p>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.<p>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.<p>Phase III: The rapture of the nerds. Fully interactive VR + toolset. In the context of the Fermi paradox, Phase III is the Great Filter.<p>----<p><a href="https://en.wikipedia.org/wiki/Great_Filter" rel="nofollow">https://en.wikipedia.org/wiki/Great_Filter</a>
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.
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.
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.
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.<p>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.<p>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.
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.<p>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.
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?