Zhu, Park, Isola, Efros, UC Berkeley (2017) — ARXIV-1703.10593
Key takeaways
- Two generators and two discriminators. The loss is
L_GAN(G, D_Y) + L_GAN(F, D_X) + λ·(‖F(G(x)) − x‖₁ + ‖G(F(y)) − y‖₁) - Why the cycle: with an adversarial loss alone, any mapping onto the target distribution is optimal, including one that ignores the input. The cycle term requires the translation to keep enough of the input to invert it
- Ablation (Cityscapes, FCN scores):
- labels → photo (Table 4): cycle alone 0.22 / 0.07 / 0.02, GAN alone 0.51 / 0.11 / 0.08, GAN + forward cycle 0.55 / 0.18 / 0.12, full CycleGAN 0.52 / 0.17 / 0.11
- photo → labels (Table 5): GAN alone 0.53 / 0.11 / 0.07, GAN + backward cycle 0.01 / 0.06 / 0.01, full 0.58 / 0.22 / 0.16
- Human study (AMT, maps ↔ aerial): it fools raters on about a quarter of trials, where the unpaired baselines almost never do
- The limits, stated by the authors: colour and texture translation usually works. Geometric change (dog → cat) degenerates into minimal edits. Paired pix2pix remains better where pairs exist
Standing in the anthology
Filed from #163 (CycleGAN). Sources SOTA-339, scoped to
appearance changes.
The sentence to be careful with is “both terms are critical”. The paper supports it for the full loss against either term alone. Its own labels → photo table has a one-directional cycle beating the bidirectional one, and the authors’ argument for keeping both directions is the instability and mode collapse they observed qualitatively in the dropped direction (DP-010).
Read — NOTE-302.