Zero-Shot Generative Model Adaptation via Image-Specific Prompt Learning

Ours "Wall painting" Ours "Anime painting" Ours "Photo" Source "Ukiyo-e" Ours NADA NADA NADA NADA Figure 1. The mode collapse issue. For NADA [21] and our method, the same generator pre-trained on the source domain of "Photo" is adapted to the unseen target domains of "Disney", "Anime painting", "Wall painting" and "Ukiyo-e" only with the domain labels. The images above the dotted line are some examples from the internet. The generated images of NADA exhibit some similar unseen patterns (yellow box areas) which are undesired in terms of quality and diversity. This issue is largely addressed by our method.