Artsie - Arty Selfies with Arbitrary Style Transfer

This is a small webapp that lets the user take a selfie and apply styling from different paintings to the photo. It uses a machine learning to do "style transfer" from one image to another (see paper linked below).

At the heart of it is a neural network in Tensorflow.js, with weights trainined on about 80 000 paintings.

• Website: Bitvergnügen/Artsie
• The paper which this is based on: Exploring the structure of a real-time, arbitrary neural artistic stylization network
• The convenience library for Javascript, which does the heavy lifting: @magenta/image
• Library or machine learning algorithms in the Javascritp: Tensorflow.js

Installation

npm install

Building App

All dependencies are installed locally, so you need to prefix run gulp via npx

npx gulp

Develop

Start a local server with hot reload.

npx gulp watch_local

Known Issues

• No pictures produced on iOS devices
  • It seems iOS handles the canvas tag differently somehow. The end result is black frames instead of images.