Counting 1,868 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Fast Neural Style Transfer in browser with Deeplearn.JS

This repository contains an implementation of the Fast Neural Style Transfer algorithm running fully inside a browser using the Deeplearn.JS library.

Demo website:



What is this about?

This is an implementation of the Fast Neural Style Transfer algorithm running purely on the browser using the Deeplearn.JS library. Basically, a neural network attempts to "draw" one picture, the Content, in the style of another, the Style.

Is my data safe? Can you see my webcam pics?

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.

Is this implementation optimized?

As of today, there are major bottlenecks in the implementation of the model. Deeplearn.JS is a new library and some functions still do not have a WebGL implementation. Don't worry, as soon as the required functions become available, I'll update this site accordingly. I've asked for help on the official deeplearnjs repository, and will implement any advice I get for making this as fast as possible.

How big are the models I'm downloading?

For each available style, your browser will download a model around ~6.6MB in size. Be careful if you have limited bandwidth (mobile data users).

The web page is ugly.

I know. Sorry, I'm not really a UI designer. I have about a 10 minute tolerance for tweaking HTML and CSS until I give up. The good news is, it's all open source on Github! If you want to help improve the page's design, please send a pull request! :)


To run this locally, clone the project and prepare the development environment:

$ git clone
$ cd fast-style-transfer-deeplearnjs
$ npm install && bower install # Install node modules and bower components

To interactively develop the application

$ ./scripts/watch-demo src/styletransfer-demo.ts
>> Waiting for initial compile...
>> 1023189 bytes written to src/bundle.js (0.71 seconds) at 2:20:06 AM
>> Starting up http-server, serving ./
>> Available on:
>> Hit CTRL-C to stop the server

The application will be available at http://localhost:8080/src/styletransfer-demo.html and will watch for changes of typescript code.

Adding your own styles

The way Fast Neural Style Transfer works is, one has to train a new neural network for each "Style" image and upload it to the server. Training takes 4-6 hours on a relatively powerful GPU (Maxwell Titan X).

To train your own style model from scratch, please follow the instructions from this Github repository to get your own .ckpt file. You will need Python, Tensorflow, and a decent GPU.

Once you have the model.ckpt file for your style, run the following:

$ python scripts/ --output_dir=src/ckpts/my-new-style --checkpoint_file=/path/to/model.ckpt
$ python scripts/ --output_dir=src/ckpts/my-new-style

If all goes well, src/ckpts/my-new-style should contain ~6.7MB of 49 items including a manifest.json file.

Adding the style to the application is then as simple as modifying the STYLE_MAPPINGS variable in src/styletransfer-demo.ts.

If you're able to successfully achieve cool new styles, I'd be glad to add them to this demo!


Credits belong to the following: