Preserving Building Geometry in Nano Banana 2 Dusk Concepts
Understanding the Challenge of Lighting and Structure
When designing architectural concepts, the transition from day to night is a critical moment for mood setting. However, changing the lighting conditions often introduces significant risks to the integrity of the original design. In the context of AI image generation, shifting a scene to dusk can cause the model to reinterpret structural lines, alter window placements, or invent new features that were not present in the source material. This phenomenon, often referred to as hallucination, occurs because the model prioritizes atmospheric consistency over geometric fidelity.
Nano Banana 2 is designed to handle complex text-to-image and image-to-image workflows, allowing users to explore these atmospheric shifts. Yet, without specific guidance, the tool may struggle to distinguish between a shadow cast by a real beam and a completely new architectural element. The goal of this guide is to provide techniques for prompting Nano Banana 2 to maintain accurate building geometry while shifting the scene to dusk. By focusing on descriptive constraints, users can minimize the risk of structural distortion during these lighting changes.
It is important to note that prompt instructions describe desired outcomes but do not guarantee identity, label, object, or typography preservation. Users should approach these prompts as starting points rather than absolute rules. The following sections outline the prerequisites and steps needed to achieve stable results.
Prerequisites and Model Selection
Before attempting to generate dusk concepts with preserved geometry, it is essential to select the appropriate model within the Nano Banana ecosystem. Google documents Nano Banana 2 as Gemini 3.1 Flash Image (gemini-3.1-flash-image). This model is distinct from Nano Banana Pro, which uses Gemini 3 Pro Image, and Nano Banana 2 Lite, which utilizes Gemini 3.1 Flash Lite Image.
While Nano Banana 2 Lite is focused on speed and cost, it is not optimized for multiple reference inputs or multi-turn sequential editing. Therefore, for tasks requiring high precision in geometry preservation across lighting changes, the standard Nano Banana 2 model is generally more suitable. Using the Lite version for complex structural edits without understanding its limitations could lead to unexpected deviations in the building's form.
Ensure you are accessing the correct product page at /nanobanana2 to utilize the full capabilities of the text-to-image workflow. Familiarize yourself with the prompt library provided on the site, which offers example prompts that users can copy or adapt. These examples serve as a foundation for constructing your own detailed instructions.
Step-by-Step Prompting Strategy
To successfully shift a scene to dusk while keeping the building structure intact, follow this structured approach. First, clearly define the base geometry in your initial description. Explicitly state the number of floors, the shape of the roof, and the arrangement of windows. Avoid vague terms like "modern" or "tall" without further qualification.
Next, introduce the lighting change gradually. Instead of simply asking for "dusk," specify the time of day, such as "blue hour" or "twilight," and describe the light sources. Mention that interior lights should be visible through existing windows, reinforcing their presence. This helps anchor the model to the original layout.
Finally, apply negative constraints. Clearly instruct the model to avoid adding new structures, altering the roofline, or moving windows. Label untested prompt examples as examples only; they are suggestions for your workflow. A usable prompt might look like this: "A realistic dusk view of a three-story rectangular office building with a flat roof. Preserve the exact grid of twelve square windows on the front facade. Do not add balconies or change the roof shape. Interior lights glow warmly through the existing windows."
Judging Results and Fixing Common Issues
After generating an image, evaluate the output against your geometric constraints. Check if the roofline remains straight and if the window count matches your description. If the building appears warped or new elements have appeared, the prompt likely lacked sufficient specificity regarding the structure.
Common issues include the model inventing new wings or altering the perspective too drastically. To fix this, refine your prompt to emphasize the rigidity of the structure. Use phrases like "strict adherence to original floor plan" or "no structural modifications." If the result is still unsatisfactory, try adjusting the balance between the lighting description and the structural constraints. Sometimes, placing the geometry description before the lighting instruction yields better stability.
Remember that AI tools evolve, and what works today may need adjustment tomorrow. Always test variations and document which phrasing produces the most consistent geometry for your specific needs. By combining clear structural definitions with precise lighting cues, you can effectively use Nano Banana 2 to create compelling dusk concepts without losing the essence of your architectural design.