Image generation goes through Vertex AI, and the demo used the nano banana Pro model; the request body structure is content→role: user→parts→text; it returns Base64-encoded image data.