4 ms·

Implementing Photoshop 2.0 using GAN

22 points by pranoy 9y ago

otakucode 9y ago
This has a pretty terrifying aspect to it... have we already passed the point at which no image seen on a screen can be trustworthy at all? How long until such a technology can easily process a command like "disrobe everyone in this photo" and produce a very believable facsimile of what the people in the photo would look like? Should such be seen as a violation against those people? The list of potential abuses of such technology is so very long and interesting and filled with pitfalls.
- Chris2048 9y ago
  You can already add a beard or a bald head to a person's face. Which authority will rush around preventing the creation of fake nudes?
- brokenmachine 9y ago
  That sounds amazing, I can't wait!
- joshvm 9y ago
  This would probably be a grey area. We're also long past the point where an image can be absolutely trustworthy (without digital forensics). Certainly digital 'fake nudes' exist and as with any image alteration, a professional artist can make something very convincing these days. When you use a GAN (or any kind of infill algorithm) it's not an oracle. It's generating a visually plausible guess at the data you're replacing. You could give it a face and a body outline and ask it to fill in the gaps, but what you generated wouldn't be 'real'. Legally who knows? It would depend on what the law of making an obscene facsimile of someone is (or whatever the law calls it).
  - pixl97 9y ago
    >without digital forensics How long till me make a GAN with the express purpose of defeating digital forensics?
bitL 9y ago
Te main flaw with DNN used in Photoshop-like apps is the speed; one would need 1000x faster GPU to make changes realtime as artists are used to now, instead of waiting a month for a single step to finish. Some trivial algorithms can be done fast already but most of them are way ahead of what current tech can deliver fast.
- pranoy 9y ago
  When deep learning gets better, the hardware support will also get better.
mattl 9y ago
Photoshop 2.0 came out in June 1991. Think of another name for your idea? :)
- yoz-y 9y ago
  Title had me confused, at first I thought somebody reimplemented Photoshop 2.0 in a new language or something like that.

22 points by pranoy 9y ago

This has a pretty terrifying aspect to it... have we already passed the point at which no image seen on a screen can be trustworthy at all? How long until such a technology can easily process a command like "disrobe everyone in this photo" and produce a very believable facsimile of what the people in the photo would look like? Should such be seen as a violation against those people? The list of potential abuses of such technology is so very long and interesting and filled with pitfalls.

Chris2048 9y ago
You can already add a beard or a bald head to a person's face. Which authority will rush around preventing the creation of fake nudes?
brokenmachine 9y ago
That sounds amazing, I can't wait!
joshvm 9y ago
This would probably be a grey area. We're also long past the point where an image can be absolutely trustworthy (without digital forensics). Certainly digital 'fake nudes' exist and as with any image alteration, a professional artist can make something very convincing these days. When you use a GAN (or any kind of infill algorithm) it's not an oracle. It's generating a visually plausible guess at the data you're replacing. You could give it a face and a body outline and ask it to fill in the gaps, but what you generated wouldn't be 'real'. Legally who knows? It would depend on what the law of making an obscene facsimile of someone is (or whatever the law calls it).
- pixl97 9y ago
  >without digital forensics How long till me make a GAN with the express purpose of defeating digital forensics?

bitL 9y ago

Te main flaw with DNN used in Photoshop-like apps is the speed; one would need 1000x faster GPU to make changes realtime as artists are used to now, instead of waiting a month for a single step to finish. Some trivial algorithms can be done fast already but most of them are way ahead of what current tech can deliver fast.

pranoy 9y ago
When deep learning gets better, the hardware support will also get better.

mattl 9y ago

Photoshop 2.0 came out in June 1991. Think of another name for your idea? :)

yoz-y 9y ago
Title had me confused, at first I thought somebody reimplemented Photoshop 2.0 in a new language or something like that.