MDMA allows non-expert users to apply style transfer (a deep learning technique) to their image to create new artworks.
It's simple to get started. First, you have to choose an input image. You can upload your image or choose an already uploaded, public one. Then, you have to select a style image. The neural network will generate a new picture by mimicing the style of the style image.
We are using "Universal Style Transfer via Feature Transforms" published by Yijun Li et al. in 2017 (paper). It has some advantages over traditional neural style transfer (paper). It promises to work better on unseen patterns which we have a lot in our data. In addition, it's faster and more efficient which allows us to run the computation on CPU.
You can find the code on Github and don't forget to follow on Twitter.
Done by Johannes Filter for the cultural hackathon Coding Da Vinci Ost 2018.