About the project
About the team
There are way too many simplistic filters in today's society so we challenged ourselves to create a lightweight platform for transferring styles of paintings into any image.
What it does
The user will upload any image into the generator and behind-the-scenes, an AI - designed to take inspiration from a single painting - will be chosen by the user and the styles will transfer over within a few seconds.
How we built it
We have classification networks repurposed to generate images (inspired by: Gatys et al.)
Challenges we ran into
Backend: Modifying the hyperparameters to properly learn the style features and content features to produce a believable output. We had to revert to using pre-trained models since time didn't permit much experimentation. (github)
Frontend: Being able to interface with the backend and testing several frameworks proved tricky to find exactly what we need and how to implement features.
Accomplishments that we're proud of
That it produces very pretty pictures rather quickly and the website is streamlined for the best possible user experience.
What we learned
Paperspace > AWS when training on GPUs. Learned a lot of the inner workings of Tensorflow and obtained a richer understanding of creating and maintaining a site.
What's next for pAInt
More complex models trained in-house and social feed integration (i.e Facebook, twitter)