A frame by frame reconstruction of the film Blade Runner using a type of artificial neural network called an autoencoder.
The model is a variational autoencoder trained with a learned similarity metric [Larsen et al. 2015 – http://j.mp/1RRgaCW] that I implemented in TensorFlow at a resolution of 256×144. It has been trained on every individual frame of Blade Runner for 6 epochs before reconstructing the film.
This video is the outcome from a research project as carried out for my dissertation of the Msci in Creative Computing at Goldsmiths, University of London. Read more about the project in my medium piece: http://j.mp/1XlibAb