Bridging Domain Gaps in Vision-Language Models through Class-Aware Prototype Alignment and Discrimination
The proposed PromptSync method enhances zero-shot generalization in vision-language models by employing class-aware prototype alignment and discrimination during test-time adaptation.