Image Analogies: a filter from a pair of examples
On 16 August 2001 at SIGGRAPH Aaron Hertzmann and co-authors from NYU, Microsoft Research and the University of Washington showed how to learn a filter from a single example: a pair of images A and A’ defines a transformation, and the program applies it to a new B to produce B’.
Why it matters
Instead of programming each effect, an effect could now be shown. One procedure gave blurring, super-resolution, texture transfer, "texture-by-numbers" and artistic styles learned from scanned real drawings. This is the task pix2pix in 2016 would call image-to-image translation, and it cites the paper.
SIGGRAPH 2001 proceedings, pages 327-340. The method is a multiscale autoregression on pixel neighbourhoods combining the searches of Wei and Levoy and of Ashikhmin. The authors call it "rather slow": from tens of seconds for simple texture synthesis to a few hours for artistic renderings on a 1 GHz processor. The same proceedings carried Efros and Freeman’s "Image Quilting", with texture transfer of its own.