About two years ago, when my interests were growing towards novel style transfer (especially with arbitrary styles) algorithms, I accidentally came across these papers:
- Fast Patch-Based Style Transfer of Arbitrary Style [3],
- Self-Attention Generative Adversarial Networks [19].
The core idea of the first paper is combining patch-based techniques with deep convolutional neural networks, while the second paper is about adopting the self-attention mechanism for computer vision. These ideas helped me to reveal a connection between the two. So, we’ll discuss that connection in this article.
After spending some time in the reference list of [3], I have found the paper…
Shant Navasardyan
Senior ML Scientist, Team Lead at Picsart Inc.
