Inspiration
Our original vision was to train a CNN to 'undo' the style-transfer process, allowing us to look at what the person in a portrait would look like in real life. We succeeded in creating a database for this problem, but could not finish our original vision in time.
What it does
Creates large numbers of stylized face images.
How we built it
By using a tensorflow hub, we performed style transfer on thousands of images from the CelebA dataset, using a kaggle art dataset for the styles (https://www.kaggle.com/ikarus777/best-artworks-of-all-time/data).
Challenges we ran into
Unfortunately we didn't have enough time to build the model which would fulfil our original vision. Additionally, the resulting dataset would perhaps not generalise so well to the kinds of portraits we originally intended to use the model on.
Accomplishments that we're proud of
We have loads of awesome high quality stylized pictures.
What we learned
To set a more realistic workload.
What's next for Reverse Style Transfer
We might see this project through to the end in our spare time.
Built With
- python
- tensorflow
Log in or sign up for Devpost to join the conversation.