C-DGPA: Class-Centric Dual-Alignment Generative Prompt Adaptation

Unsupervised Domain Adaptation (UDA) tackles the critical challenge of feature distribution shift between a labeled source and an unlabeled target domain. Cross-modal models like CLIP show great potential for UDA, yet standard prompt tuning methods primarily align marginal distributions, neglecting conditional discrepancies. This causes class prototype misalignment and degraded discriminability. We propose C-DGPA, a class-centric dual-alignment framework for generative prompt adaptation. C-DGPA synergistically optimizes marginal and conditional distribution alignment via a dual-branch design. The marginal branch uses a dynamic adversarial training with a Gradient Reversal Layer (GRL) to learn domain-invariant features. Meanwhile, the conditional branch introduces a Class Mapping Mechanism (CMM) to standardize semantics and align class prototypes. Extensive experiments on OfficeHome, Office31, and VisDA-2017 validate C-DGPA’s superiority, achieving new state-of-the-art results and demonstrating strong capability in cross-domain feature alignment.