When on vacation (like now) I catch up with a lot of reading, research etc. This was fun one I found today so, although horribly off-topic for Mudspike, I wanted to share (I guess it’s computer graphics, so hardware?):
Take a look at this image:
Look along the top. Those are real people, and the ‘source’ for the AI styles.
Now look down the left column, at the destination. Those are real people. That’s the target of transformations to apply on the source (real) face.
All the faces in matrix between that top row and left column don’t exist - they are computer generated by how this technique imagines a face should look. None of those people have ever existed.
Here’s a nicer example - here you can see a real person on the left, with the source laid over the top.
Great paper by Nvidia here:
Trusting images and photos in the future is going to be fun.
All of these faces below don’t exist in reality. They are AI generated:
On the ‘how does it work?’, well in really short summary it is breaking down the source images into a number of discrete style hierarchies (eyes, cheekbones, hair) using separate recognizers. The clever bit this is a generalized technique, in that the interpolation in the network happens to work really well for filling in gaps because it was trained on so many examples.
Anyway, back to the beach!