We all know that most things are "photoshopped". We expect the ugly overhead power lines to be removed from the postcard pictures of Japanese temples. We all know the fashion models in the magazines are not as perfect in real life as they appear on the pages. But these new algorithms from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) take this to a whole new level.
The core of the idea is to analyze an image for repeated forms and quantify the variations from one repeated form to another. Some of the examples they show (and that you can see in the video below) are things like corn kernels on the cob, or bricks in a wall, or even dancers in a line. Once the image has been parameterized in this way it becomes possible to alter the parameters and go back and regenerate an image. You can think of it as a magical "slider bar" that reduces (or exaggerates) the repetitive variations.
The paper was presented at the Siggraph Asia conference this week. (Warning: the paper has lots of images and is a huge download)
The accompanying video: