GoogleSynth uses the Google Image Search thingy to randomly grab two images as the "input" and "target" images for the algorithm. In order to do this GoogleSynth uses a large word list (about 62,000 words) and randomly selects two words from it. It then does a search with each of these words and randomly selects an image from the first page of results. If there is an error it will simply select a new word and try again.

Once it has two images it applies the algorithm with the parameters set by the user and produces a new image based on them. The results vary wildly, often the output is a total mess, but it creates some cool looking stuff now and then (depending on your definition of 'cool').