A visual introduction to realistic image synthesis

This is an explainable of Generative Adversarial Networks (GAN). We introduced the concept via story-telling and interactive widgets (and lots of cute 😺 and 🐶!). Check it out here.

Building and Running

The project is built using React and Styled System.

• To build the project, make sure you have Node install first. We recommend v10.15.3 or above.
• Run npm install in the root directory of the project.
• Run npm start and checkout the page at http://localhost:3000/

What's in the repo

• src contains the source files for the project webpage, as well as necessary data and other assets in src/assets/.
• scripts contains a Python script used for data exploration and cleaning for the publication visualization.

Future work

• We only used PCA for the final latent space exploration page. Many other methods such as t-SNE exist for dimensionality reduction. It would also be nice to try out a 3D visualization like Embedding Projector.
• Many of the arts are in raster formats, which causes some performance problems. Use more vector arts will probably mitigate this.
• We wanted to go into far more details about the inner-workings of GANs but didn't end up doing as much as we hoped.
• Currently, all the GAN outputs are cached as image files. We explored porting models like BIGGAN to the web using TensorFlow.js but weren't able to get it to work well.
• As the name of the project suggests, we wanted to explore far more techniques for realistic image synthesis.

Authors

• Wode "Nimo" Ni (woden@cs.cmu.edu)
• Jingya Chen (jingyach@andrew.cmu.edu)

This project is a part of the final project for the Data Visualization course at Carnegie Mellon University in Fall 2019, taught by Adam Perer. We had a great time in the class and while doing this project. If you're at CMU, we highly recommend taking the class 😉.

Screencast

Link here