Johan Andersson

5th year Computer Science student in Lund.

Convolutional dreams

09 Jun 2019

In the process of writing an image-to-image translation GAN, you can produce some pretty neat art. Below are samples from 10 epochs of training on the horse2zebra dataset, with a generator of 3 convolutional and 1 bilinear upsampling layers and a discriminator with 1 convolutional layer. (Obviously a very weak architecture, but good enough for debugging)

Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced by an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN Nonsense images produced bx an untrained GAN