Dual-path hypernetworks of style and text for one-shot domain adaptation.

Learning a one-shot domain adaptation model is an exciting and challenging topic in computer vision and graphics. A feasible solution is to fine-tune a pre-trained generator to the target domain by leveraging the powerful semantic capabilities of CLIP (Contrastive Language-Image Pretraining). Unfortunately, when the target image shows a significant difference from the source domain, existing methods would result in overfitting, and generated images do not correctly reflect the texture of the target image. To address this issue, we propose a Dynamic Domain Transfer Strategy (DDTS) to align the texture information between the source and target domain by dynamically adjusting the direction of domain transfer. Furthermore, the delicately designed dual-path hypernetworks of style and text (Dual-HyperST) for one-shot domain adaptation characterize the target domain's textual style and visual style with a text-guide path and a style-guide path. Specifically, the style-guided path predicts a set of style weight offsets by the target image, followed by the text-guided path predicts a set of text weight offsets by a text prompt. To better integrate the information between these two paths, we introduce a hypernetwork that learns to modulate the pre-trained generator instead of fine-tuning. Qualitative and quantitative experiments demonstrate the superiority of Dual-HyperST, which surpasses the state-of-the-art methods in the diversity and high quality of the generated images. [ABSTRACT FROM AUTHOR]