4 ms·
That's extremely transformative: it's learning from a large corpus what faces are and how to create brand new ones which cannot even be traced back to an origin
by gwern 7y ago
That's extremely transformative: it's learning from a large corpus what faces are and how to create brand new ones which cannot even be traced back to an original. (If you think even that is not enough to count as 'transformative', I suggest you look at the examples in the WP article of what has been considered transformative, like 'thumbnail screenshots of web pages' and 'putting a mustache on the Mona Lisa'.)
- fenwick67 7y agoWhen a significant number look like zootopia characters I don't think you can claim they "can't be traced back to the original", lots of these are overfit
- gwern 7y agoThey are not 'overfit' unless they are exactly the same and they cannot generalize. But you can see the wide variety of high quality faces which are not copied from the dataset and that the interpolations are fine: https://twitter.com/arfafax/status/1258344026706599937 https://twitter.com/arfafax/status/1258344026706599937 Generating a recognizable character is no more 'overfit' than a painter being able to paint George Washington means their brain is 'overfit'. If those characters are part of the data distribution, as they are, then the GAN can and should generate them and countless variants on them.