Architectural Image Generation with Nano Banana: Prompting and Editing Guide

What will you learn in this guide?

When generating architectural images, a general instruction such as “design a modern house” may not be enough to describe the intended space and design approach. Clearly outlining how the space will be used, its materials, lighting and framing—and then reviewing the result and refining the prompt step by step—provides a more controlled starting point. In this guide, you’ll get a brief introduction to Nano Banana models, create original prompt examples for architecture and interiors, and see how to request edits to an existing image.

Google’s prompt design documentation recommends giving clear, specific instructions and describing the task and constraints when needed. Prompt writing should be treated not as a one-off task but as a process of refinement based on reviewing the results. The prompts below are examples written for this guide; they do not guarantee a particular result.

What you need: a model, access and an image

Google’s Gemini API documentation lists four models under the Nano Banana name:

• Nano Banana 2 Lite (gemini-3.1-flash-lite-image): Described as an option for use cases where speed and cost are priorities. It is not optimized for multiple reference inputs or consecutive, multi-turn edits.

• Nano Banana 2 (gemini-3.1-flash-image): Presented as a versatile, general-purpose option. The documentation mentions speed, 4K generation, world knowledge and text generation; it also highlights working with multiple reference images and consistency.

• Nano Banana Pro (gemini-3-pro-image): Positioned as a premium option for more complex visual tasks. The documentation emphasizes world knowledge, localization, brand consistency and creative control.

• Nano Banana (gemini-2.5-flash-image): The previous model in the series. Google recommends that users move to Nano Banana 2 Lite.

The API documentation provides image-generation examples in Python, JavaScript, Java and Go, as well as through REST. The REST example uses an API key. The sources do not explain account setup, current pricing or access requirements, so check these separately before getting started. The image-editing steps in this guide use gemini-3.1-flash-image, the model featured in the documentation’s example. The source does not state that all four listed models support editing.

Step-by-step architectural image generation

1. Define the goal and main subject. Start by choosing a single main objective, such as an exterior concept, an interior atmosphere or a material alternative. This makes it easier to assess what worked when reviewing the first result.

Create an architectural visualization of a compact courtyard house in a temperate climate. Show the building from eye level, with pale stone walls, timber screens and a planted courtyard. Soft overcast daylight, realistic materials, calm residential atmosphere. Keep the composition focused on the relationship between the interior and the garden.

This example describes the building type, viewpoint, materials and lighting. If an important element is missing from the result, clarify that point in your next prompt; there’s no need to pack many unrelated goals into the first one.

2. Describe the interior and framing. For an interior image, specify the room’s function, the elements you want to see and the overall atmosphere. Describing the framing in plain language is one way to explain what relationship the image should focus on.

Create an interior visualization of a small reading lounge in a public library. Show the seating area and the full-height window in one balanced view. Use oak shelving, a muted green textile palette and soft daylight. The room should feel quiet and welcoming, with clear circulation space between furniture.

Here, the intended use, visible elements, material palette and circulation requirements are described together. If the result feels too busy, ask for a simpler composition in a follow-up prompt.

3. State priorities and constraints clearly. The prompt design guide explains that you can specify both what the model should and should not do. In an architectural exercise, you can describe the main focus and any unwanted additions in the same instruction.

Create a visualization of a contemporary apartment kitchen with matte cream cabinets, a dark stone island and warm indirect lighting. Keep the island as the main focal point. Do not add decorative objects on the worktop. Preserve a simple, uncluttered layout.

In this example, the kitchen island is the main focus, and the prompt asks the model not to add decorative objects to the countertop. Review the output against your design decisions; don’t assume the prompt will preserve every detail exactly.

4. Edit a supplied image with text. Google’s image-generation documentation describes providing an image alongside text instructions to add, remove or change elements, or request changes to style and color. The editing section’s example uses gemini-3.1-flash-image. Make sure you have the necessary rights to use the image you upload.

Using the supplied interior image, change the wall finish to light limestone and make the overall color grading warmer. Keep the existing room layout, furniture positions and window openings unchanged.

This request separates the surface and color characteristics to be changed from the spatial elements that should remain intact. If the intended edit is unclear in the result, make the element to be changed more explicit in a follow-up prompt. If you need to work with multiple images or perform consecutive edits, evaluate the model options against the documentation; Lite is not described as optimized for these uses.

5. Review the result and refine one thing. After the first generation, identify the most important shortcoming: material, lighting, composition or an element that should have been preserved. Then write a short, specific follow-up instruction addressing only that issue. This approach aligns with the guide’s description of prompt design as an iterative process.

Common mistakes

• Using vague descriptions: “A beautiful interior” doesn’t explain the room’s function or design character. First say what should be shown, then specify preferences such as materials and lighting.

• Packing too many goals into one prompt: State your priority clearly instead of combining competing atmospheres and focal points.

• Leaving out elements to preserve: For image edits, specify any layouts or openings you don’t want changed.

• Treating the first result as final: Review the output and refine the prompt step by step.

• Ignoring image usage rights: Google’s documentation reminds users to have the necessary rights to uploaded images.

How can this fit into an architecture and visualization workflow?

These methods can be useful for architecture and interior design teams exploring early concepts, interior atmospheres, and alternatives for color or material direction. Image generation can help teams discuss design intent and compare options. However, the sources do not say that generated images are technical drawings or verified project deliverables; project decisions need to be assessed separately.

Model selection can be matched to your needs: Lite for when speed and cost are priorities, Nano Banana 2 for general-purpose work and working with multiple reference images, and Pro for more complex visual tasks, as described in the documentation. The sources do not explain local hardware requirements, and they provide no details on pricing or terms of use. Offices and students should therefore check API access, current fees and image usage rights before incorporating the tools into a production process. Google says all generated images include a SynthID watermark.

Next steps

Prepare a primary concept prompt for the same space, then try a separate follow-up prompt that changes only the lighting or material palette. When using a reference image, describe the elements to change and those to preserve in separate sentences. If working through the API, verify that the selected model matches the model ID specified in your code. If you need multiple references or consecutive edits, choose a model based on the relevant documentation.

Sources and license

This guide is adapted into Turkish from Google AI for Developers’ “Nano Banana image generation” and “Prompt design strategies” documentation. Both sources are licensed under CC BY 4.0. The prompt examples for architectural and interior scenarios were created specifically for this guide.