For ML4W Week 9, I made a short audiovisual loop using Runway and StyleGAN2.

Dataset & Training

I downloaded a dataset of artificial lunar landscapes from Kaggle and trained them through StyleGAN2 in Runway. Although the original dataset is intended for training rock detection algorithms, I used the landscapes to generate lunar landscapes. It took approximately 6 hours.

The interpolation

Since the original dataset were also artificial images, the outcome seemed to be somewhat uniform and monotonous. However, I liked when the output had different shaped craters and horizons.

Generating audiovisual loops

The images felt somewhat techno, so this time, I tried making a looping audiovisual footage. I made a 8-second latent walk video from the model with same starting and ending frames.

I also made a short techno loop that matches the video. The result was... satisfactory!