Course contentsShow
← Machine Learning and Deep Learning
Lesson 1493 of 3,53833. Generative Adversarial NetworksFree lesson

StarGAN: Multi-Domain Translation

Single generator architecture handling multiple domain translations using domain labels as conditional input.

StarGAN: Multi-Domain Translation

What you'll learn: How a single GAN generator can translate images across multiple domains simultaneously using domain labels as conditional inputs.

The Multi-Domain Challenge

Recall that Pix2Pix requires paired data and CycleGAN handles unpaired translation between two domains. But what if you want to translate between many domains—say, changing hair color (black, blonde, brown), age (young, old), or gender in faces? Training separate models for every pair (black→blonde, black→brown, blonde→black, etc.) becomes impractical fast.

How StarGAN Solves This

StarGAN uses a single generator that learns all possible translations at once. The key innovation: you feed the generator both the input image and a target domain label (like a one-hot vector encoding "blonde hair"). The generator learns to modify the image according to whatever domain label you provide.

Think of it like a universal translator with a dial—you turn the dial to "Spanish," "French," or "Japanese," and the same device handles all translations.

The discriminator also becomes multi-task: it must classify whether images are real or fake and identify which domain they belong to. This helps the generator learn domain-specific features correctly.

Architecture Insight

  • Input: Original image + target domain label (concatenated or added)
  • Generator: Produces translated image in target domain
  • Discriminator: Outputs real/fake score + domain classification
  • Training: Uses adversarial loss, domain classification loss, and reconstruction loss (cycling back to original domain)

Key Takeaway: StarGAN enables efficient multi-domain image translation using a single generator conditioned on domain labels, avoiding the need for separate models for each domain pair.