Using dynamic knowledge for kernel modulation: Towards image generation via one-shot multi-domain adaptation

One-shot domain adaptation across multiple image domains aims to learn complex image distributions using just one training image from each target domain. Existing methods often select, preserve and transfer prior knowledge from the source domain pre-trained model to learn the target model with distinct styles. However, they always neglect the cross-domain shared knowledge and fail to consider the adaptation relationships in selecting source knowledge when facing varying styles in multi-target domain scenarios, casting doubt on their suitability. In this paper, we propose a novel one-shot multi-target image domain adaptation model based on kernel modulation. By leveraging the similarity information inherent in cross-domain knowledge to guide the correlation of feature learning, our method allows for precise control over the transfer of task-relevant knowledge while minimizing irrelevant information. Furthermore, we propose a novel cross-domain contrastive loss that incorporates dual constraints of structure and style. By effectively mining strong negative samples from cross-domain knowledge, it aims to maximize the structural feature from the source domain and accurately reflect the unique attributes of various style target domains. Extensive experiments on several datasets show that our method offers significant advantages in concurrently establishing multiple image domain mapping relationships. Moreover, it can effectively explore the potential for knowledge transfer in cross-domain feature learning, thereby generating higher-quality domain-adapted images.