"I wish to paint in such a manner as if I were photographing dreams." - Zdzisław Beksiński
Beksinski's work ranges from the haunting to beautiful and while his art progressed over the years certain themes and patterns strike the viewer. Because of this I hoped that a GAN might be able to pick up on these patterns and generate images in his style. A complete archive of his work was downloaded from here.
In machine learning terms this is quite a small dataset (700 photos) so the models struggled. I ran two models one for 256x256 images and one for 64x64. The larger model failed to pick up on any detail other than the common colors used.
The smaller model trained much more effectively, picking up on colors quickly and even starting to recreate some of the common textures seen throughout Beksinskis work.
Overall I think the key issues are that:
- The dataset has to much variance
- The dataset is too small This combined with the fact that GAN's are hard to train even in ideal circumstances made the experiment interesting, but somewhat unsuccesful.
.gif)

