Person-to-person video transfer with machine learning
Proof of concept experiment by Brannon Dorsey is like a style transfer for video, where details of people from one video are replicated in another:
The output that you are seeing here is generated entirely by a machine. Created using pix2pix (https://github.com/phillipi/pix2pix)
code applied to image sequences. Frames are hallucinated from 2D noise
images (given a conditional image, aka, images from the Ray Kurzweil
presentation) using a cGAN trained on ~700 images of me mimicking frames
from the Kurzweil video.
It was created with the help of Nick Briz’s
code to turn video into0 a dataset.
I generated this song using machine learning/recurrent neural networks. Basically, using a model which was trained on a set of midi-file classical music, it generated a new melody based on parameters learned from song structures from the given training music. This is a simple thing to begin with. I want to work towards generating more intricate structures by training on more contemporary music.