Sensitivity of Generative VLMs to Semantically and Lexically Altered Prompts