CycleGAN: Unpaired Image Translation
What you'll learn: How CycleGAN translates images between domains (like horses ↔ zebras) without needing matched pairs of training images.
The Challenge of Unpaired Translation
Pix2Pix requires paired training data: matching images from both domains (e.g., a photo of a building and its architectural sketch). But what if you just have two separate collections—photos of horses and photos of zebras—with no correspondence? CycleGAN solves this.
The Core Idea: Cycle Consistency
CycleGAN uses two generators and two discriminators:
- Generator G: translates domain A → B (horse → zebra)
- Generator F: translates domain B → A (zebra → horse)
- Discriminators: one for each domain to judge realism
The breakthrough is cycle consistency loss: if you translate a horse to a zebra (using G), then translate that zebra back to a horse (using F), you should get the original horse back. This roundtrip constraint enforces that the translation preserves content while changing style.
Think of it like translating English → French → English. If the final result doesn't match your original sentence, something went wrong in translation.
The Loss Function
CycleGAN combines three losses:
- Adversarial losses: both generators fool their respective discriminators
- Cycle consistency loss:
|F(G(x)) - x|and|G(F(y)) - y| - Identity loss (optional): if you "translate" a zebra image using the zebra generator, it should stay unchanged
This elegant approach enables learning meaningful mappings without paired data, opening doors to countless creative applications.
Key Takeaway: CycleGAN learns bidirectional image translation between unpaired datasets by enforcing that translating forward then backward returns the original image, using cycle consistency loss to maintain content while transforming style.