Which Nano Banana versions exist?
Nano Banana is Google's nickname for its Gemini image models. At the time of writing (November 2026), Google's image generation documentation lists five of them. Apps often show the nickname, while the API shows the Gemini name or model ID, so it helps to know both.
We publish no benchmark scores here. For buildings, two differences matter most: how strictly a model holds geometry during an edit, and how it treats fine repeated detail such as mullions and balusters. Run one prompt on the same image with two versions before you commit a whole project to one.
| Nickname | Gemini name | Model ID | How Google describes it |
|---|---|---|---|
| Nano Banana | Gemini 2.5 Flash Image | gemini-2.5-flash-image | The legacy model of the series |
| Nano Banana 2 | Gemini 3.1 Flash Image | gemini-3.1-flash-image | Previous fast model, tuned for speed and 4K |
| Nano Banana 2 Lite | Gemini 3.1 Flash Lite Image | gemini-3.1-flash-lite-image | Fastest and cheapest; not tuned for multi-turn editing |
| Nano Banana 2.1 | Gemini Nano Banana 2.1 | gemini-nano-banana-2.1 | Update to Nano Banana 2 with better quality, text and multi-turn consistency |
| Nano Banana Pro | Gemini 3 Pro Image | gemini-3-pro-image | Premium model that reasons through complex instructions |
How should a Nano Banana architecture prompt be ordered?
Write the prompt in the order the model should think about the job. Plain sentences work better than strings of style words. Every prompt on this page follows the same five parts.
- Call your upload Image 1 and state that its geometry and camera are the authority.
- Count what is there and write the numbers into the prompt.
- Run one edit per turn, and start the next turn from the best result.
- Deliverable: say what the output is. For example, a photoreal exterior photograph of the building in Image 1.
- Keep: name the geometry and the camera, with counts. Floors, bays, openings, roof shape, lens, horizon.
- Change: one thing only. New surfaces, or new light, or new people. Not all three.
- Integration: say how the change sits in the scene. Sun direction, shadow length, reflections, scale.
- Output: the rules for the file. Same framing, vertical lines vertical, no text, no watermark.
Prompt 1: model screenshot to photoreal render
Use this for a SketchUp, Rhino, Revit or Archicad viewport capture. The first failure is interface residue: the model reads axis lines and black edges as real objects and keeps them as thin poles or dark seams. The second is a roof it thinks is better than yours.
Fix for residue: add a sentence that the outlines are software marks. Fix for the roof: describe it in numbers, such as a single slope falling to the rear at about 15 degrees.
- Make a photoreal exterior photograph of the building in Image 1. Image 1 controls all geometry and the camera. Keep the three floors, the flat roof with its parapet, the five window bays on the front, the recessed entrance on the left, the camera height, lens and horizon. Change only the surfaces: pale brick walls, black aluminum window frames, clear glazing with soft reflections, a concrete path. Remove the colored axis lines, dashed guides and black outlines; they are software marks, not trim. Late-morning sun from the left. Same crop and upright verticals. No text.
Prompt 2: clay or white model to real materials
A clay model has form and shadow but no surface information. Left alone, the model either paints everything white stucco or invents its own palette, and it sometimes adds window divisions that the clay never had. Give every element its own material by name, and say that openings stay plain unless they are modeled.
- Render the white clay model in Image 1 as a finished building photographed on an overcast day. Geometry and camera come from Image 1 and must not change: two floors, a pitched roof, the cantilevered upper volume on the right, six openings on this face. Apply materials by element: vertical timber cladding on the upper volume, light grey render on the ground floor, dark metal roof, glazing in every opening with no added mullions. Soft, even shadows that follow the clay shadows. Keep the plain background as a quiet suburban street. No text.
Prompt 3: hand sketch to architectural render
Sketches fail in a specific way: the model fills gaps. It adds a floor where the lines are loose, a balcony where a hatch suggests one, or a wing outside the drawn edge. State that only drawn elements exist. Accept that anything the sketch does not show, such as the side wall or the roof material, is a guess.
Hand perspective is rarely exact. Ask the model to keep the sketch's viewpoint rather than correct it, or it may rotate the building to a cleaner angle.
- Turn the pen sketch in Image 1 into a realistic architectural photograph. Build only what is drawn: two stories, a shallow gable roof, a porch across the front, four windows upstairs and two plus a door below. Do not add floors, wings, balconies or dormers. Keep the sketch's viewpoint and horizon as drawn. Use painted white weatherboard, grey shingles and a timber porch. Plant only where the sketch shows shrubs. Fill the blank paper around the house with a simple lawn and sky. No text, no sketch lines left.
Prompt 4: elevation drawing to facade render
An elevation is a flat, straight-on drawing. The common failure is that the model adds perspective: side walls appear, the ground recedes and the building leans. Ask for an orthographic view in plain words and repeat that no side faces are visible.
- Render the front elevation in Image 1 as a photoreal facade, seen flat and straight on with no perspective. Keep the four floors, the twelve windows in a regular grid, the ground-floor shopfront and the roof line exactly where they are drawn. No side walls visible, no vanishing points, no tilt. Materials: red clay brick, white window frames, a dark steel shopfront, a stone base course. Soft daylight from above left with shallow shadows in the window reveals. A plain pavement strip at the bottom. Ignore dimension lines and labels; output no text.
Prompts 5 to 7: day to night, add people, replace the sky
These edit an existing render, so the geometry is already right. The risk is collateral change. A night prompt lights every window and invents wall lights. A people prompt gets the scale wrong or blocks the entrance. A sky prompt changes the light on the building or trims the roofline.
Run each one as its own turn on the image you approved. Do not combine them.
- Day to night: Turn the render in Image 1 into the same view at blue hour, about 30 minutes after sunset. Keep the building, camera and planting unchanged. Warm light in roughly half the windows, curtains varied, the rest dark. Light only fixtures that are visible in Image 1; add no new lamps. Deep blue sky with a faint glow on the horizon. Gentle reflections on glass and wet-looking paving. No text.
- Add people: Add five people to the render in Image 1 at true scale, about 1.7 m tall, standing on the ground plane with contact shadows that match the existing sun. Two walking on the path, one at the entrance but not blocking the door, two seated on the bench. Natural clothing, slight motion blur only on walkers. Change nothing else: building, camera, light and planting stay as they are.
- Replace the sky: Replace only the sky in Image 1 with a clear late-afternoon sky with a few high clouds. Keep the roofline, trees and every edge against the sky exactly as they are. Keep the light on the building unchanged, so shadow direction and length stay the same. Reflections in the glass should show the new sky. No text, no birds, no sun flare.
What goes wrong with Nano Banana architecture edits, and how do you fix it?
Most failures come from the model treating your image as a suggestion. Each fix below adds a line that turns a vague instruction into something it can check.
| Failure | What you see | Line to add |
|---|---|---|
| Redesigns the roof | New pitch, added gables or a parapet that was not there | The roof is a single slope falling to the rear; keep its pitch and eaves line. |
| Adds floors | An extra story or an attic level | Exactly three floors. Do not add stories, attics or roof terraces. |
| Keeps interface lines | Thin dark poles, blue boxes or colored lines in the photo | Black outlines, colored axes and dashed lines are software marks; remove them. |
| Warps verticals | Walls lean in or out, corners curve | All vertical lines stay vertical and parallel, as in Image 1. |
| Changes the camera | Wider view, higher eye level or a rotated building | Keep the camera position, lens, crop and horizon of Image 1. |
| Shifts small detail | Mullions, railings or signage move or multiply | Keep the number of window panes and railing posts as shown. |
| Garbles text | Fake lettering on signs and plaques | Output no text; leave sign panels blank. |
How were these prompts built, and when should you skip writing them?
Each prompt here is written to block one of the failures in the table above, using the patterns Google documents for image editing. We report no scores, because results vary with the image, and a prompt that holds on one building can slip on the next. Small details such as mullions and railings can still shift, and these are AI images, not construction documents.
If you would rather not write prompts at all, ArchFor runs this kind of recipe for you. ArchFor is a web app that turns 3D model screenshots, sketches and existing renders into photorealistic architectural images. Credit packs from $9.99; no plugin and no subscription required. Every render runs on Nano Banana 2.1. You pick the input type, space, style, time of day and image size (1K, 2K or 4K), add an optional materials note of up to 200 characters, and get 4 options, usually in under a minute at 2K.
Nano Banana and Gemini are trademarks of Google. ArchFor is not affiliated with Google.
Questions
Which Nano Banana version is best for architecture?
There is no single winner; pick by the job. Nano Banana Pro is built for complex instructions, which helps when a prompt lists many things to keep. Nano Banana 2.1 is Google's current fast model and handles routine edits such as materials and light. Test the same image and prompt on both before a deadline.
Why does Nano Banana redesign my building?
Because the prompt leaves room for it. Words like 'modern' or 'beautiful' invite the model to improve the design. Name the geometry with counts, call Image 1 the authority for form and camera, and ask for one change per turn. If the roof still changes, describe its shape and pitch in plain numbers.
How long should an architecture prompt be?
About 60 to 100 words, or under 600 characters, is enough. Shorter prompts leave out the things to keep. Longer ones bury the main instruction and add conflicting style words. Follow the order deliverable, keep, change, integration, output, and stop there.
Do these prompts work on floor plans?
No. They are written for perspective views and flat elevations. A floor plan or top-down site plan has no facade, roof or camera height for the model to keep, so it has to invent all of them. ArchFor does not support plan-to-3D either.
Can I upload reference images with these prompts?
In Gemini you can, but these prompts are written for one image. A second image gives the model another building to borrow from, which raises the chance of redesign. ArchFor also takes one image per job, with no reference uploads.