This repository contains an implementation of arbitrary image stylization running fully inside the browser using TensorFlow.js.

Demo website: https://reiinakano.github.io/arbitrary-image-stylization-tfjs

This is an implementation of an arbitrary image stylization algorithm running purely in the browser using TensorFlow.js. As with all neural style transfer algorithms, a neural network attempts to "draw" one picture, the Content (usually a photograph), in the style of another, the Style (usually a painting).

Although other browser implementations of style transfer exist, they are normally limited to a pre-selected handful of styles, due to the requirement that a separate neural network must be trained for each style image.

Arbitrary image stylization works around this limitation by using a separate style network that learns to break down any image into a 100-dimensional vector representing its style. This style vector is then fed into another network, the transformer network, along with the content image, to produce the final stylized image.

Your data and pictures here never leave your computer! In fact, this is one of the main advantages of running neural networks in your browser. Instead of sending us your data, we send you both the model and the code to run the model. These are then run by your browser.

The style network is ~9.6MB, while the transformer network is ~7.9MB, for a total of ~17.5MB. Since these models work for any style, you only have to download them once!

No. The original paper uses an Inception-v3 model as the style network (~96.2MB as a .tar.gz), which is too large to be deployed in a browser setting.

Before porting this to the browser, a MobileNet-v2 was used to distill the knowledge from a pretrained Inception-v3 style network. This resulted in a size reduction over 10x, from ~96.2MB to ~9.6MB.

Since each style can be mapped to a 100-dimensional style vector by the style network, we simply take a weighted average of the two to get a new style vector for the transformer network.

This is also how we are able to control the strength of stylization. We take a weighted average of the style vectors of both content and style images and use it as input to the transformer network.

Yup! The code is hosted on Github.

This project uses Yarn for dependencies.

To run it locally, you must install Yarn and run the following command at the repository's root to get all the dependencies.

yarn run prep

Then, you can run

yarn run start

You can then browse to localhost:9966 to view the application.

This demo could not have been done without the following:

• Authors of the arbitrary image stylization paper.
• The Magenta repository for arbitrary image stylization.
• Authors of the MobileNet-v2 paper.
• Authors of the paper describing neural network knowledge distillation.
• The TensorFlow.js library.
• Google Colaboratory, with which I was able to do all necessary training using a free(!) GPU.

As a final note, I'd love to hear from people interested in making a suite of tools for artistically manipulating images, kind of like Magenta Studio but for images. Please reach out if you're planning to build/are building one out!