3 ms·
Would it make sense to compare the output images to their nearest neighbors in the training set in order to establish some metric of "originality"? Sorry if tha
by catbird 8y ago
Would it make sense to compare the output images to their nearest neighbors in the training set in order to establish some metric of "originality"? Sorry if that is a basic question, I haven't read much on image synthesis with neural nets.
- gwern 8y agoYes, and they don't usually look too much alike (see the BigGAN appendix nearest-neighbors for example of doing this) but some people argue this isn't a good enough check because the GAN could just be shuffling around patches of textures or something, which is a little goal-post-move-y for my taste.