Fig. 1:

Comparing our Fixed-Point GAN with StarGAN [8], the state of the art in multi-domain image-to-image translation, by translating images into five domains. Combining the domains may yield a same-domain (e.g., black to black hair) or cross-domain (e.g., black to blond hair) translation. For clarity, same-domain translations are framed in red for StarGAN and in green for Fixed-Point GAN. As illustrated, during cross-domain translations, and especially during same-domain translations, StarGAN generates artifacts: introducing a mustache (Row 1, Col. 2; light blue arrow), changing the face colors (Rows 2–5, Cols. 2–6), adding more hair (Row 5, Col. 2; yellow circle), and altering the background (Row 5, Col. 3; blue arrow). Our Fixed-Point GAN overcomes these drawbacks via fixed-point translation learning (see Sec. 3) and provides a framework for disease detection and localization with only image-level annotation (see Fig. 2).