Architectural Image Generation with Nano Banana 2.1: 5 Prompts and Editing Steps

What you’ll learn in this guide

Nano Banana is the name used for Gemini’s built-in image-generation capabilities. The Gemini API documentation presents Nano Banana 2.1 as an efficiency-focused model for image generation and conversational image editing. This guide covers how to write prompts for architectural concepts and interior images, request changes to an existing image, and refine results step by step.

The examples here are intended to explore design alternatives and provide a starting point for presentation visuals. A generated image, on its own, does not prove the accuracy of design decisions or technical drawings. Describe the result you want from the model, then have the design team review the output.

Requirements: model, API, and image

The examples are based on the gemini-nano-banana-2.1 model in the Gemini API documentation. The official guide provides Google GenAI SDK examples and a REST request for API use. In the REST example, the API key is sent in the x-goog-api-key header; $GEMINI_API_KEY represents the key value used in that header. This workflow therefore requires access to the Gemini API and a usable API key. Check API access terms and current pricing before getting started; the source guide does not provide a price.

In addition to Nano Banana 2.1, the documentation lists Nano Banana 2 Lite, Nano Banana 2, and Nano Banana Pro. According to the source, Lite is designed for tasks where speed and cost are priorities; it is not optimized for multiple reference inputs or sequential edits. Nano Banana 2.1 is described as an efficiency-focused model with improvements in image quality, text rendering, multi-turn consistency, and grounding with Google Search. Nano Banana Pro is presented as the higher-end option for complex visual tasks. The steps in this guide are written for version 2.1; if you choose another model, update the model name in your request.

You can also use a source image during the editing step. This lets you request changes to the color, materials, or a specific element in an existing interior image. Before uploading an image, verify that you have the rights to use it. Google states that generated images contain a SynthID watermark.

Writing prompts step by step

1. Define the design goal. First decide what the image should show: for example, a residential entrance, workspace, or small courtyard. Instead of listing many competing goals in one prompt, clearly describe the main space and the purpose of the image.

Create an architectural concept image of a compact urban courtyard for a small apartment building. Show a clear relationship between the entrance, planted seating area, and circulation path. Use warm brick, pale stone, and soft morning light. Keep the composition calm and realistic.

This prompt describes the type of space, design relationships, material palette, and lighting atmosphere separately. Treat the first output as a starting point for a more detailed prompt, not as a final design.

2. Add the visual style and framing. Specify the desired viewpoint, lighting conditions, and presentation style to make it clearer which visual qualities the model should prioritize. Include only details that serve the purpose of the image.

Create an interior concept image of a quiet reading room in a public library. Use a wide eye-level view that shows the bookshelves, central reading table, and window seating together. Natural overcast daylight, tactile timber surfaces, muted colors, restrained architectural visualization style.

Here, the wide eye-level framing and the elements that should appear together are specified. The source guide emphasizes that clear, explicit instructions can help guide model behavior.

3. State constraints and features to preserve. In your prompt, describe not only what should be generated but also what must remain unchanged. This is especially useful when working with a source image. Phrase constraints as short, verifiable statements.

Create a concept image of a small apartment kitchen. Keep the existing room layout and the positions of the window and doorway unchanged. Explore a light oak cabinet finish, pale stone counters, and warm indirect lighting. Do not add extra openings or change the room proportions.

This example separates the request to preserve the layout and the positions of the openings from the material exploration. Still, review the visual output to check whether the requested elements have been preserved; including them in a prompt does not guarantee the result.

4. Edit a source image with text. The Gemini API documentation shows that an image can be guided for additions, removals, changes, or style and color adjustments by providing it alongside text as input. Add the image to your input and describe only the change you want to make.

Using the supplied interior image, change the wall finish to a soft limewash texture and shift the overall color grading toward neutral daylight. Keep the furniture, camera view, room layout, and window positions as they are.

This prompt focuses on surface and color adjustments to a source image. If you need to change several things, split the work into stages so you can compare the results.

5. Review the output and iterate. Prompt design is an iterative process. After the first image, identify one feature that needs improvement and describe it concretely in your next request. For example, make vague descriptions of material tone, framing, or lighting more specific.

Revise the previous courtyard image. Keep the building entrance, circulation path, and camera composition unchanged. Make the planted seating area more visually distinct and soften the contrast between the brick and pale stone surfaces.

At this stage, specify what you want to change and what should stay the same in each revision. The prompt strategy documentation says examples and clear instructions can help guide format, scope, and wording; it also notes that the best approach should be tested and refined for the intended use.

Common mistakes

• Relying on vague adjectives: Words like “beautiful,” “modern,” or “better” do not describe a design decision on their own. Support them with observable qualities such as materials, lighting, framing, or spatial relationships.

• Packing too many changes into one prompt: Changing the layout, finishes, lighting, and visual style all at once makes it harder to understand which instruction affected the result. Try the main change first.

• Not specifying what to preserve: When editing, explicitly state which features must remain fixed, such as the room layout, openings, or framing.

• Assuming the image is a design document: Treat generated output as material for concept exploration or visual communication; verify actual project decisions separately.

• Overlooking rights to the source image: Make sure you have the right to use any image you upload.

Next steps

Start with a short test using a single space and a single visual goal. Note which features work and which need to change, then begin another round by updating just one or two instructions. When working as a team, keeping prompts and outputs together for comparison makes it easier to assess which descriptions best serve the project.

If you plan to work with multiple references or make sequential image edits, choose a model that suits your workflow. According to the source, Nano Banana 2 Lite is not optimized for multiple reference inputs or multi-turn sequential editing. Nano Banana 2.1 is described in the documentation as offering improvements in multi-turn consistency; consider both the model’s stated capabilities and the project’s needs when choosing. Check API access and cost terms in advance as well.

Sources and license

This guide has been 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 architectural and interior prompts are original examples created for this guide.