StyleGAN2 Week 5 Class Notes

Taught by Derrick Schultz & Lia Coleman

🔗 Links

• Class Slides
• GANSpace in Colab
• Audio Reactive + Feature Vectors

📣 Class Announcements

• Sign up for the showcase! Slot signups here!
• Please fill out our class survey!
• Remember to delete your VM or close your billing account (Billing > Settings) when you're done. GCP will still charge you a storage fee even if you're not actively running your VM.

🕹 Feature Vectors

• StyleGAN's 512 dimensions are called features.

• Note: they're not human-recognizable features! For a human, we might break down a picture into the features of hair length, eye color, etc. But computers see things differently!

• In Z-space, we get entanglement among features. W-space tries to mitigate this.

  

• In the StyleGAN2 community, we also use feature to just mean human-readable characteristics: hair color, eye color, etc.

How to find feature vectors?

Manually: tag a bunch of points as (red hair, black hair) + use another ML classifier to separate the categories (red hair, black hair). Very labor intensive!

• The open mouth feature vector for FFHQ is pretty well-known. For FFHQ especially there are a lot of open public feature vectors that people have found.