I know there’s some good descramblers over the years but man, let me tell you it’s still my dream to watch something like kuroinu uncensored.

  • RA2lover@burggit.moe
    link
    fedilink
    arrow-up
    2
    ·
    1 year ago

    It’s already achievable with current workflows. The only missing tool is a censorship detection/marking neural network.

    As far as i know, no one has worked on this, but the problem doesn’t require the network to be particularly large or costly to train. Rather, the problem is in the lack of a readily-available dataset to train it.
    You could build this sort of dataset by gathering a few hundred censored and uncensored releases of a given JAV / h-anime, then running a difference filter on the censored and uncensored versions. The availability of censored/uncensored releases seems to be much greater for JAV, but i can see how a model trained exclusively on JAV would potentially have issues if used on h-anime.

    After this network is trained, the rest would be using it to predict which regions on a given frame have been censored, and then use that as an inpainting mask for an existing image generation model.

    The inpainting approach would allow the inpainting network to be changed as developments on temporally-stable image generation improve.