In this work, we introduce an attribute-anchored textual prompt learning method for vision-language models, named ATPrompt. This method extends the learning space of soft prompts from the original one ...