The unpaired image-to-image translation consists of transferring a sample a in the domain A to an analog sample b in the domain B without intensive pixel-to-pixel supervision. The current vision focuses on learning a generative function that maps both domains but ignoring the latent information, although its exploration is not explicit supervision. This paper proposes a cross-domain GAN-based model to achieve a bi-directional translation guided by latent space supervision. The proposed architecture provides a double-loop cyclic reconstruction loss in an exchangeable training adopted to reduce mode collapse and enhance local details. Our proposal has outstanding results in visual quality, stability, and pixel-level segmentation metrics over different public datasets.