Abstract: Pre-trained vision-language models (VLM) such as CLIP, have demonstrated impressive zero-shot performance on various vision tasks. Trained on millions or even billions of image-text pairs, ...