Rosetta creates abstract artworks from any image in a style based on Japanese woodblock prints from the Edo period. Conventional scripts and a neural network are used in a multi-step image manipulation process to produce this result. Rosetta is a combination of two pre-existing neural image generation techniques: sketch-to-image & style transfer.
The triptych presented here is a series of street fashion photos re-imagined by a neural network as woodblock art.
Rosetta uses photo manipulation scripts to make an abstract version of an image, which is passes to a trained neural network. This abstraction serves to hide information about the image from the network so it can fill in the blanks and dream it’s own interpretation.
The machine has been trained to create an estimate of the original image, but has only ever seen images of Edo woodblock art. This biased training creates a network which is only familiar with the chosen art style. The bias is taken advantage of to redraw the original image in the style that the machine is familiar with.