VL-Taboo: An Analysis of Attribute-based Zero-shot Capabilities of Vision-Language Models