4 years of progress in AI: 2018 (https://arxiv.org/abs/1811.10201 ) vs 2014 (http://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf …). /soonpic.twitter.com/0gAfRzMh3q
Show this thread
I don't like this kind of comparison. Yes, the photos on the left are very impressive. But the real question is: how meaningfully different are they from the training set? If the answer is "not much", then a GAN is just a bad way to store lots of images.
-
-
If they were just memorization, the interpolations and latent controls wouldn't work (see their video). You can see immediately what an overfit GAN merely memorizing examples looks like in an interpolation video inhttps://www.reddit.com/r/SpiceandWolf/comments/a1oa89/experiments_in_generating_holo_faces_with_neural/ …
-
Plus, the nearest-neighbors checks rarely turn up matches, and GANs are useful in other things which they wouldn't be if they were memorizing, & models are much smaller than the datasets (ProGAN is <500MB but the original CelebA-HQ is >27GB, new one's even bigger).
- 7 more replies
New conversation -
-
-
Looks like nice diversity. For a speedup, could you compare the Discriminator embeddings of samples instead? A nearest-neighbor lookup on the embedding would be a lot faster than pixel-level stuff, I'd guess.
End of conversation
New conversation -
-
Well, take a look at https://blog.insightdatascience.com/generating-custom-photo-realistic-faces-using-ai-d170b1b59255 … and explore the latent space.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.