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

TL;DR

C-DGPA addresses unsupervised domain adaptation for vision-language prompt tuning by introducing a class-centric dual-alignment framework. It combines a dynamic marginal distribution alignment branch with a Class Mapping Mechanism for conditional alignment, jointly optimizing prompt parameters to achieve domain-invariant and semantically discriminative representations. The approach yields state-of-the-art results on OfficeHome, Office31, and VisDA-2017, and ablations confirm the complementary benefits of marginal and conditional alignment. The work provides both theoretical grounding and practical gains, with a pathway to efficient deployment after training by discarding the dual branches.

Abstract

Unsupervised Domain Adaptation transfers knowledge from a labeled source domain to an unlabeled target domain. Directly deploying Vision-Language Models (VLMs) with prompt tuning in downstream UDA tasks faces the signifi cant challenge of mitigating domain discrepancies. Existing prompt-tuning strategies primarily align marginal distribu tion, but neglect conditional distribution discrepancies, lead ing to critical issues such as class prototype misalignment and degraded semantic discriminability. To address these lim itations, the work proposes C-DGPA: Class-Centric Dual Alignment Generative Prompt Adaptation. C-DGPA syner gistically optimizes marginal distribution alignment and con ditional distribution alignment through a novel dual-branch architecture. The marginal distribution alignment branch em ploys a dynamic adversarial training framework to bridge marginal distribution discrepancies. Simultaneously, the con ditional distribution alignment branch introduces a Class Mapping Mechanism (CMM) to align conditional distribu tion discrepancies by standardizing semantic prompt under standing and preventing source domain over-reliance. This dual alignment strategy effectively integrates domain knowl edge into prompt learning via synergistic optimization, ensur ing domain-invariant and semantically discriminative repre sentations. Extensive experiments on OfficeHome, Office31, and VisDA-2017 validate the superiority of C-DGPA. It achieves new state-of-the-art results on all benchmarks.

Figures (4)

• Figure 1: The architecture of the C-DGPA model. This figure illustrates the structure of our model at different stages. The Source Bank and Target Bank represent the feature banks of the source and target domains, respectively. CMM is used to map the image features I to the class space, while DD denotes the domain discriminator that generates domain labels based on I. In Stage 2, the blue color indicates the frozen parts, while other colors indicate the parts that need to be trained.
• Figure 2: t-SNE visualization of source and target domain features, highlighting marginal distribution using C-DGPA and PDA (OfficeHome dataset, R→P).
• Figure 3: t-SNE visualization of target domain features, highlighting conditional distribution using C-DGPA and PDA (OfficeHome dataset, R→P).
• Figure 4: he impact of different token lengths on accuracy (OfficeHome dataset)