Fig. 2:

Adding a new domain $X$ to a learned UNIT model. Red: Models that need to be fine-tuned. Blue: Models that need to be trained from scratch. (a) Existing methods usually need to fine-tune the full model on data of all available image domains. (b) Our method only needs to learn an encoder $E_X^{new}$ and a decoder $D_X^{new}$ on data of the new domain $X$, then images of the domain $X$ can be translated from/to previous domains via the latent space of the generator.