Inspiration

We were inspired by recent advances in deep learning, which allow you to transfer style from one picture, say a famous painting, to your own pictures. We wanted to add tight integration with Facebook to allow people to experience the magic of this new technology in as accessible a manner as possible.

What it does

Semantic style transfer is a new deep-learning technique which takes images and applies styles to them -- we've drawn these styles from collections of paintings from history.

Unfortunately Facebook requires app verification before allowing us to open it to non-developers (which takes a few days), so you can't run this yourself yet. We'd love to demo the integration for you with our own Facebook accounts, or you can see how it looks on an example photo at http://tapioca.byron.io !

How we built it

AWS GPU instances running the deep-learning library torch, served by Flask. A mess of JS and hacked-together CSS on the front end in a desperate attempt to make things align, as is the way of such things.

Challenges we ran into

Centering things is hard.

Accomplishments that we're proud of

Got cutting-edge deep learning techniques working -- and integrated with Facebook! -- in only 36 hours. And it looks spiffy!

What we learned

Don't stay up so late.

What's next for tapioca

Taking over the world. But Instagram integration before that most likely.

Share this project:
×

Updates