Disentangling latent space better for few-shot image-to-image translation

In an unpaired image-to-image translation, the main concept is to learn an underlying mapping between the source and target domains. Previous approaches required large numbers of data from both domains to learn this mapping. However, under a few-shot condition, that is, few-shot image-to-image translation, only one domain can meet the required number of data , and thus, the underlying mapping becomes ill-conditioned owing to the limited data as well as the imbalanced distribution of the two domains. We argue that a powerful model with a better disentangled representation of the latent space can better tackle the more challenging few-shot image-to-image translation . Motivated by this, under a partially-shared assumption, we propose a better disentanglement of the content and style latent space using a domain-specific style latent classifier and a domain-shared cross-content latent discriminator. Moreover, we design asymmetric weak/strong domain discriminators to achieve a better translation performance with limited data within the few-shot domain. Furthermore, our method can be easily embedded into any latent space disentangled model of an image-to-image translation for a few-shot setting. Subjective evaluation and objective evaluation results both show that compared with other state-of-the-art methods, the images synthesized by our method have higher fidelity while maintaining certain diversity.