It does demonstrate that to some extent: specular light is omitted from the normal map, so it correctly maps the room behind the light.

The last example also correctly identified broken panes of glass in the background, so that the glass is a solid and the spaces have nothing there.

I used the Pro subscription for Gemini 3 Pro, which I think is running Nano Banana Pro for image generation, but it doesn't tell you which model in the chatbot UI.