Currently, there is a problem with how the models have been set up and the prompt, leading to the response from the chat model not seeming very relevant. The image is fed to a model which outputs its description as text to the chat/LLM model, which then responds to the description, which can seem surprising.

Try fine tuning either the chaining or the prompts so the chat response would seem more useful.