Originally published at: https://boingboing.net/2018/04/23/a-i-reconstructs-incomplete-p.html
…
Huh, I could see this becoming integrated into compression algorithms, eventually. Not necessarily for omitting whole eyes from photos, but for adding hundreds of tiny holes that the algorithm can fill back in. If the algorithm always returned the same results for the same data, it would be easy to run tests that say "if I omit these pixels, will I get back close-enough results?
Further, our model gracefully handles holes of increasing size.
This phrasing is… unfortunate.
This could have interesting applications in the adult entertainment field.
The way that final one morphs was surprisingly reminiscent of when someone tries to alter a mugshot on the fly in a film, like in (SPOILERS!) Ocean’s 13.
Silly A.I. everybody knows people have mouths for eyes.
Is it just me, or did it look like all the rebuilt portraits had Steve Buscemi eyes?
Or just teenagers who want to fantasize about what their crush looks like naked. Start with a picture of someone in a swimsuit, erase the swimsuit, get a sexy result?
Though I guess it’s probably more likely you’d just get something like this:
It’s not reconstruction at all, but at least their nonsense isn’t all dog’s heads and eyeballs.
Neat! Let’s see it work on a black guy with an epicanthic fold, eh?
Can it learn from adjacent frames of footage?
And what happens when it’s given an incomplete image of a license plate and it’s asked to “reconstruct” it?
(Because sooner or later, someone who has no understanding of what the “AI” does is going to ask for magic.)
You mean: this could easily replace ‘bubbling’ in the teenage basement dweller sex fantasy field.
What could possibly go wrong?
Mmmm… not yet.