I don't think we should expect two messages, I'd rather suggest we should expect one with multiple content parts:

image_message = {
        "type": "image_url",
        "image_url": {"url": gcs_url},
}
text_message = {
        "type": "text",
        "text": "What is shown in this image?",
    }
message = HumanMessage(content=[text_message, image_message])

This is in order for it to work with the chat interface, each turn creates one message right? Maybe I can do so that the first is multimodal?

yes, let's make it multimodal and it should expect only one message, so invocation would look smth like this:

response = model.invoke(
        input=[
            HumanMessage(content=[text_message, image_message])
        ]
    )

Changed!