Enhancing Zero-Shot Vision Models by Label-Free Prompt Distribution Learning and Bias Correcting

Neural Information Processing Systems 

Vision-language models, such as CLIP, have shown impressive generalization capacities when using appropriate text descriptions.