Architectural visualization with Nano Banana 2: best prompt techniques and tips

Google's Gemini API brings together text-to-image generation and existing image editing under the name "Nano Banana." For architecture firms and archviz professionals, these tools can be used to produce concept visuals, presentation boards, and quick material/lighting tests. Drawing on the model specifications and prompt design strategies found in the Gemini API documentation, this guide covers architecture-specific sample prompts and practical steps.

What you'll learn

By following this guide, you'll learn the differences between the Nano Banana models, the logic behind generating architectural visuals from text, how to turn an existing image (such as a floor plan or sketch) into a render through editing, and techniques for adding a logo or text to a presentation board.

Requirements

• A Gemini API key (obtainable through Google AI Studio).

• A simple development environment capable of sending requests via Python, JavaScript, Java, Go, or REST; alternatively, an interface that supports the Gemini API.

• A reference image (plan, sketch, or photo) to upload for the editing steps, with the user holding the rights to that image.

Differences between the Nano Banana models

Step by step: writing prompts for architectural visual generation

According to the Gemini API documentation, an effective prompt should clearly state explicit instructions, constraints, and the desired output format. The examples below are prepared for architectural use cases based on these principles.

1. Exterior concept visual

Create a photorealistic exterior render of a minimalist two-story house with a flat roof, large glass facade, light grey concrete cladding, and a timber entrance canopy. Set it in a quiet suburban street at golden hour, soft warm sunlight, long shadows, surrounded by mature trees. Camera at eye level, 35mm lens look, no people, no text.

This prompt gives the model clear constraints by separately defining material, lighting conditions, camera angle, and scene context. As noted in the Gemini API documentation, writing instructions step by step and concretely makes it easier to achieve consistent results.

2. Detailed prompt for an interior render

Generate an interior visualization of an open-plan living room with light oak flooring, a dark green velvet sofa, a floor-to-ceiling window overlooking a garden, and soft diffused daylight entering from the left. Include a minimal black metal shelving unit. Style: clean architectural photography, neutral color palette, no text, no watermark overlays beyond the standard image signature.

In this example, the material palette and lighting direction are clearly defined; as emphasized in the documentation, specifying the response format (for instance, an "architectural photography" style) brings the model's output closer to the intended presentation.

3. Turning an existing sketch into a render (image editing)

The Nano Banana models can modify an existing image by accepting text and image together as input (text-and-image-to-image). For example, a hand-drawn interior sketch can be uploaded together with a text prompt like the following:

Using the uploaded sketch as the layout reference, turn it into a photorealistic interior render. Keep the furniture positions and window placement from the sketch, apply a warm Scandinavian material palette with light wood and white walls, and add soft natural daylight.

In this method, the reference image is added to the API either as base64-encoded data or as a file uploaded via the Files API; the model evaluates both the text and the image together.

4. Adding a logo and title to a presentation board

Place the uploaded office logo discreetly in the bottom right corner of this exterior render, as if printed on a presentation board. Add a small title text area at the top left with placeholder space for a project name, consistent lighting and perspective with the rest of the image.

These kinds of prompts take advantage of the brand consistency and text rendering capabilities that Nano Banana Pro is documented to offer.

What this means for architecture and archviz workflows

These models can be used by architecture firms looking to quickly generate alternatives at the early concept stage, students preparing presentation boards, and visualizers testing materials and lighting. Points to keep in mind: API usage requires a Gemini API key and either code or a compatible interface; every generated image contains a SynthID watermark, so this should be kept in mind for final client presentations; and you must hold usage rights to any reference images you upload.

Common mistakes

• Using vague phrases in the prompt without specifying material, lighting, and camera angle, which leads the model to choose a random composition.

• Preferring Nano Banana 2 Lite for complex tasks that require multi-step editing; this model is not optimized for multi-reference input or multi-step editing.

• Uploading a reference image for editing without holding the usage rights to it; the documentation contains an explicit warning about this.

Next steps

In the next stage, you can try sending multiple reference images at once (for example, a plan plus a material palette) to produce a consistent series of renders, and compare zero-shot and few-shot prompting strategies to develop a prompt template tailored to your own office.

Source and license

This guide has been adapted into English from Google's Gemini API documentation pages "Nano Banana image generation" and "Prompt design strategies" (under the CC BY 4.0 license); the architecture-specific sample prompts have been written originally based on the techniques described in these sources.