U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
Whoa, StyleGANv2 is out! - Significantly better samples (better FID scores & reduced artifacts) - No more progressive growing - Improved Style-mixing - Smoother interpolations (extra regularization) - Faster training Paper: https://arxiv.org/abs/1912.04958 Github: https://github.com/NVlabs/stylegan2 …pic.twitter.com/5Cnlg9j91V
There's a notable shift from precision to recall, meaning the new generator is able to generate a higher coverage of the natural image manifold!
(I'd love to see a comparison with the sample diversity achieved by VQ-VAE2 from @catamorphist, @avdnoord and @vinyals)pic.twitter.com/IwaR3hvgM0
The samples indeed look significantly better than the previous version.
@lucidrains already has the new version live on his https://thispersondoesnotexist.com/ , take a look for yourself, we're rapidly crossing the uncanny valley!pic.twitter.com/idLjNvNKNl
Interestingly, they also wrote custom CUDA kernels to significantly speedup some core operations in the computation graph, leading to 30-40% faster training which comes in handy considering that the new, best network (config F) is quite a bit biggerpic.twitter.com/gfBqL291c8
Final question for the Twitterverse: Previously, with progressive growing, you could start training a model at low res (eg 256) and bump the resolution if results looked promising. (cfr my Colab Notebook here: https://github.com/tr1pzz/StyleGAN-Resolution-Convertor …). Can you still do this in the new version?
Also, easier inversions! They claim this will lead to a better fake- detector for similar models.
Whoaaa
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
