SOTA-339: For unpaired image-to-image translation of appearance, constrain both mappings with a cycle-consistency loss

Source

Zhu et al. (2017), LIT-564 — CycleGAN. Read as NOTE-302.

The practice

When translating between two image domains without paired examples, and the change is in appearance (style, colour, texture, season):

• Train both directions, G: X → Y and F: Y → X, each with its own discriminator
• Add the cycle loss, λ(‖F(G(x)) − x‖₁ + ‖G(F(y)) − y‖₁) with λ = 10, so a translation must keep what is needed to undo it

What was measured

The practice is Zhu et al.’s CycleGAN, LIT-564, which introduced the bidirectional cycle-consistency loss for unpaired translation and showed it across style transfer, object transfiguration, season transfer and photo enhancement. Its only quantitative ablation, on Cityscapes FCN scores, is what supports the cycle term against GAN-only training. On Cityscapes labels → photo, GAN-only scores 0.51 / 0.11 / 0.08 (per-pixel / per-class / IoU) and the full model 0.52 / 0.17 / 0.11. On photo → labels, 0.53 / 0.11 / 0.07 against 0.58 / 0.22 / 0.16. A cycle loss alone, with no adversarial term, is far worse in both directions.

Conditions

• Appearance, not geometry. Where the shape must change (dog → cat) the model makes minimal edits instead
• One direction can suffice. On labels → photo, GAN + forward cycle scored higher than both directions. The case for bidirectional is the instability and mode collapse the authors saw in the dropped direction
• If pairs exist, use them. The paired method is better