Skip to content

I trained a Generative Adversarial Network (GAN) using 10.000 images of CryptoPunks. In order to prevent mode collapse, I used the Wasserstein loss function together with gradient penalty.

Notifications You must be signed in to change notification settings

foersterrobert/GANFTs

Repository files navigation

About

I trained a Generative Adversarial Network (GAN) using 10.000 images of CryptoPunks. In order to prevent mode collapse, I used the Wasserstein loss function together with gradient penalty.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published