“…As one of the most popular and successful schemes, disentangling the content and style representation exhibits a great success on a wealth of attribute and style translation tasks [36,26,13,73,31], and enables a continuous translation by interpolating between two latent vectors [31,13]. Recently, works on continuous cross-domain translation further refine the quality of intermediate images by introducing an interpolation discriminator [71,43], constraining the intermediate results with discriminators from both sides [18,51], or by exploiting the path of interpolation and translation manifold [11,56,44].…”