登入
洞見 - Vision-language model adaptation
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.
Efficient Few-Shot Adaptation of Vision-Language Models Using a Surprisingly Strong Linear Probe (LP++)
A generalized linear probe (LP++) that efficiently adapts vision-language models like CLIP to few-shot classification tasks, outperforming recent prompt learning and adapter-based methods while being computationally efficient.
