All frames were generated by means of a custom StyleGAN2-ada implementation trained on a handcrafted dataset of 1000 bird illustrations. The StyleGAN model was actually trained twice. It was first trained with the raw bird images and then with a group of edge images from the dataset (doing transfer learning with a very small learning rate). By splitting the process, the model was able to learn to produce edge images while maintaining the shapes and geometry of the original birds. You can download the collection at full resolution here.

Music video

This audio-reactive video is the result of sequentially interpolating the latent vectors of more than 100 unique generated samples. You can find more details about the implementation on my Twitter.

Streaming platforms

The track is also available on Spotify and all other streaming platforms.