Overview
Our goal for this work was to generate aesthetic, computationally generated art using our computational photography skills and deep convolutional neural nets. We implement image and video style transfer using deep neural networks, optimizing for aesthetic value and fast training and transfer runtime. We initially implement single image style transfer, then add optimizations such as optical flow warping for video style transfer, before finally implementing a feed-forward neural net for pre-training styles to speed up the transfer process. Our approach combines 2 existing implementations for style transfer, one for speed and one for video stabilization, in a novel way to generate aesthetic, temporally consistent videos. The purpose of this webpage is to serve as a gallery of our favourite video and image results. For technical and implementation details please read our report.
LINK TO REPORTPart 1: Still Images
Here's a brief example of how we combine a content image with a style image, to produce our stylized images. What do you get when you throw a lion into the ocean?
|
|
|
Some of our favourite results (CONTENT IMAGE + STYLE IMAGE ==> STYLIZED IMAGE):
|
|
|
|
|
|
|
|
|
|
|
|
Miscellaneous Stylized Images:
|
|
|
What do these works of art have in common?
|
|
|
|
|
|
|
|
|
|
|
They all look good on lions!
|
|
|
|
|
|
|
|
|
Part 2: Videos
We then moved on from stylizing still images to generating videos! For details on how this was done, including how we stabilized between adjacent frames using optical flow, please read our report (linked at the top of this page).
Here are some examples of style transfered onto a short clip from the back of a Songthaew in the Thai countryside:
Here are some examples of style transfered onto short clips of drone footage over UC Berkeley:
Part 3: 360 Videos
As seen in the video at the very top of this site, with the video of Doi Suthep (A Thai Temple) style transfered, our trained neural net is 360 video compatible! Here are some more results: