I would like to give my contribute to this interesting discussion.
I must say, first of all, I work in the R&D department of an IT company, and I am a JAV lover.
I have knowledge of the GANs and I played with those for a while. I can confirm demosaicing using a GAN is technically possible, and the code is even very simple. Also, I think the result would be good, much better than the deblurrying effect we can see in the videos posted abive.
However the whole process which leads to have a demosaiced JAV is highly time consuming and there are some difficulties.
The first problem is the maximum dimension of a single picture the GAN would be trained can't be too big, because the training process would be impossible. Realistically speaking, if one does the job on a PC, let's say 256x256 px is possible, 512x512 is not, I guess. This means, the GAN will be able to demosaic, after being trained, any picture which is 256x256 px, so that means when the genitalia is in very foreground, it can't be demosaiced. If one wants 512x512, that would have to be run in a cluster, something Universities only have access. And you understand Unis can't run that kind of stuff.
Second problem is the groundtruth images. Even though there are actually a number of leaks one can use for groundtruth images (really 30+?), it's not like one runs VLC to produce thousands of frameshots and that's finished. Each image has to be cut, adjusted, and moreover, theoretically both censored and uncensored images should be the same exact image, except for the mosaic part. So, that means the censored and the leak should come from the same video, otherwise the uncensored pic wouldn't be exactly a groundtruth. You understand, the whole process of set up and preparing the dataset (1000 pair of images minimum) is something only a team can do. It's not a one man job (even though a nerd).
Third problem. Assuming one did all that above, assuming the GAN worked good and demosaiced video looked great, the whole software is not something that can be distributed. The reason is, it would be a system of different software, where the manual intervention is still important. It can't be a single program one installs, open a JAV, click a button, and et-voila the JAV is uncensored.
Any way, the idea of using a GAN for that job is still so interesting, exciting and fascinating. The ultimate heaven for all JAV lovers 
