Neural Style Transfer: Process grid, Kandinsky Composition, "A Neural Algorithm of Artistic Style", Figure 3

These are the earliest published examples of Neural Style Transfer. They were created in May 2015 and published in the seminal paper A Neural Algorithm of Artistic Style on arXiv in on August 26th, 2015. There was a delay between the creation of the works and the arXiv submission because, encouraged by colleagues and mentors, we had submitted the work to multiple high profile academic journals. However, at the time it was mostly unheard of for AI and deep learning to be featured in those journals and our work was rejected without being sent out for peer-review by the editors of Science, Nature, Nature Communications and the Proceedings of the National Academy of Sciences (PNAS). After three months of continous rejections I started feeling somewhat frustrated and decided to put the work on arXiv, hardly anticipating the tremendous response and substantial impact it would have on my personal and professional life. In addition to the five variations of the Tuebingen Neckarfront shown in Figure 2 of the paper, I had created several additional variations that didn't make the final selection. Three out of those unpublished pieces are preserved and shown here for the first time to a wider audience.
The names of the images are directly derived from the way I was initially documenting the parameters of the algorithm in the filenames of the generated images. Ri stands for "Reconstruction image" and Gi stands for "Gram image", related to the two components of the loss function that was optimized. Later I would start calling these "Content image" and "Style image" as can be seen in the names of the images created with later versions of the algorithm.