Inspired by A Neural Algorthm of Artistic Style, I made a video style adaption from a drone aerial of Yangshuo, Guilin. Looking at my hometown through the lens of artist is a pleasant experience.

(If you are in China 墙内请戳优酷)

Some of the higher layer neural network from the original paper (conv4_1 and conv5_1) had to be taken out to smooth out flickering in sub regions. The Starry Night segment is flickering, due to the uncertainty during gradient descend. I processed the video frame by frame, thus sometimes the L-BFGS optimizer can take an entire different direction of optimizing the video. Some frames’ gradient exploded and had to be retrained by hand afterward.

I hope this can provide an inspiration for the next neural art. Great thanks for the Neural Art Torch7 repo.

Future Directions: Continuous Gradient Descending exploiting the similarity between video frames Realtime processing (a dream?)