Nano Banana Prompts for Architects: 25 Render-Ready Examples

Nano banana prompts for architects work as plain-sentence photographer’s briefs: shot type, building type, massing, two or three named materials, site and light, plus an explicit list of what must not change whenever you edit an image of your own design. The condition that makes or breaks them is geometry: Nano Banana has no dimension input, so every output is concept imagery, and you check floor count, window rhythm and roofline yourself before anyone else sees the picture.

I wrote the 25 prompts below following Google’s published prompting guidance of 20 November 2025 and checked them against Google’s current model lineup, documentation and pricing on 20 October 2026, but I have not run them through the image models myself. Where I describe how a prompt behaves on real project material, I lean on two dated community write-ups: the meltflexai prompt library (20 July 2026) for what phrasing tends to work, and the renderahouse.com write-up (18 October 2026) for where outputs drift.

The library of nano banana prompts for architects is organised by task rather than by model. A practicing architect starts from a sketch, a massing screenshot or a site photo, and wants a usable image in the Gemini app without the design silently changing. If you use Claude for the writing side of practice, the sibling library of 50 Claude prompts organized by project stage covers that; this page stays on images. Everything here sits in the wider prompt library for architects.

What is Nano Banana, and which model should architects pick?

Nano Banana is Google’s name for the image generation models inside Gemini. For architecture work in October 2026, pick Nano Banana 2.1 for everyday iteration because it is the cheap workhorse. Pick Nano Banana Pro when you need legible text in the image, consistency across several reference images, or a 4K output for print, and Nano Banana 2 Lite only for throwaway 1K thumbnails.

The naming is confusing because the product name and the model name differ. The original Nano Banana was Gemini 2.5 Flash Image, released in August 2025, and Google’s docs now call it the “legacy pioneer” per the image generation documentation, read 20 October 2026. Nano Banana Pro is Gemini 3 Pro Image, built on Gemini 3 Pro, and was announced on 20 November 2025.

The 2026 lineup, according to the same Google documentation page read on 20 October 2026, has four live models plus the legacy one. Nano Banana 2.1 (gemini-nano-banana-2.1) is the workhorse. Nano Banana 2 Lite (gemini-3.1-flash-lite-image) is the fastest and cheapest and outputs 1K only. Nano Banana 2 (gemini-3.1-flash-image) is the previous workhorse, and Nano Banana Pro (gemini-3-pro-image) is the premium model.

Criterion	Nano Banana Pro	Nano Banana 2.1	Nano Banana 2 Lite
Model id	gemini-3-pro-image	gemini-nano-banana-2.1	gemini-3.1-flash-lite-image
Built on	Gemini 3 Pro	Not named in Google’s docs; the current workhorse	Gemini 3.1 Flash Lite Image
Best for	Text in image, data visuals, brand consistency, multi-reference edits	Everyday concept iteration and restyles at low cost	Fast, cheap thumbnails and quick single-image tries
Resolution	1K, 2K, 4K	1K, 2K, 4K (priced per size)	1K only
Reference images	Up to 14: up to 6 objects, up to 5 characters, up to 3 style references	10 objects and 4 characters	See Google’s docs
Search grounding	Yes	See Google’s docs	See Google’s docs
API price per 1K image	$0.134 (2K same; 4K $0.24)	$0.0336 (2K $0.0504; 4K $0.113)	Lowest in the lineup; see pricing page
Where an architect meets it	The Gemini app (limited on free, “Redo images with Nano Banana Pro” on paid tiers), Google AI Studio, the Gemini API	Google AI Studio, the Gemini API	The Gemini API

Feature comparison from Google’s documentation and pricing pages, 20 October 2026 - not our own benchmark. Prices come from Google’s pricing page, read 20 October 2026; the reference-image budgets and resolutions come from the image generation docs, read 20 October 2026; the Gemini app row comes from the Gemini Apps limits page, read 20 October 2026. Legacy note: Gemini 2.5 Flash Image (gemini-2.5-flash-image) still exists as the original Nano Banana, and the Gemini app’s free image row currently names Nano Banana 2.

Google positions Nano Banana Pro for “complex graphic design, high-fidelity product mockups, and factual data visualizations that require accurate text rendering and real-world grounding via Google Search”, per the gemini-3-pro-image model page, read 20 October 2026. The same page calls Nano Banana Pro a “sophisticated reasoning-driven engine for professional-grade image editing and generation, offering studio-quality precision and advanced creative control”. Read that as marketing. The parts that hold up for an architect are text, consistency across references and 4K, and the model also has a Thinking mode whose interim thought images are not charged, per the image generation docs read 20 October 2026.

AEC Magazine wrote on 3 March 2026 that “generative imaging systems, such as Nano Banana Pro… have advanced to provide surprising degrees of spatial coherence, producing persuasive visualisations from simple prompts and reference imagery”. Persuasive is the operative word, and it cuts both ways once a client is looking at the image.

How do you write a Nano Banana prompt that holds your geometry?

Write the prompt as a plain-sentence brief to a photographer, and name the building’s massing and materials explicitly. When you upload your own sketch or screenshot, open with a preserve-list: “Keep the exact same massing, window positions, floor count, roofline and camera.” Change one thing per turn. The model reinterprets the whole image on every edit, so the preserve-list and the single change are the only levers you have to keep the design yours.

Google’s prompting guide of 20 November 2025 lists six basic prompt elements: subject, composition, action, location, style and editing instructions. The advanced elements matter more for buildings. The guide recommends stating the aspect ratio in words and using camera and lighting language such as “A low-angle shot with a shallow depth of field (f/1.8)” or “Golden hour backlighting creating long shadows”. It also suggests factual constraints for diagrams, and named roles for reference images (“Use Image A for the character’s pose, Image B for the art style, and Image C for the background environment”).

For architecture, the community has converged on a formula. The meltflexai library of 20 July 2026 builds its prompts as shot type, building type, structure and massing, two or three named materials, site and context, light and time of day, closed with “photorealistic architectural photography”. Their finding, which I adopt throughout this page, is that the photography closer beats the word “render”, that materials need adjectives (“board-formed concrete” rather than “concrete”), and that follow-up changes work best as short conversational turns.

No formula fixes geometry. The renderahouse.com write-up of 18 October 2026 reports floor counts, window rhythms and roof pitches changing between generations, notes that there is no dimension or scale input, and describes the model reinterpreting the source image on each generation, so it advises one change per turn. Google’s own limitation note of 20 November 2025 is narrower and concerns text and factual visuals, so the preserve-list is a community habit; Google does not document it.

Here is the workflow I would use for every prompt below.

1. Pick the model for the job. Nano Banana 2.1 (or the Gemini app’s default image model) for iteration, Nano Banana Pro for text, boards and multi-reference consistency.
2. Upload the source image if there is one, and open the prompt with the preserve-list before any creative instruction.
3. Write the brief in plain sentences: shot type, building, massing, two or three materials described with adjectives, site, light, aspect ratio in words.
4. Close render-style prompts with “photorealistic architectural photography” and a cleanup tail (“no people, no cars, no text, no watermark”) when you want a clean plate.
5. Generate, then check four things against your drawings: floor count, window count and rhythm, roofline and pitch, and the position of the camera relative to the site.
6. Change one thing per follow-up turn (“same building, same camera, overcast morning light”) so you can see what drifted.
7. Keep the chat for the project. Google’s image generation docs, read 20 October 2026, describe the models as built for multi-turn editing, and that consistency is maintained within one conversation.

Every edit prompt in this library of nano banana prompts for architects follows that order. The concept prompts do not need the preserve-list because there is no design to protect yet.

Concept and massing prompts

These five prompts produce images from nothing but a description, which is where Nano Banana is safest for an architect because there is no existing design to distort. Use them in the first week of a project, when you know the site and the programme and want to see three or four ideas before you draw. Treat every result as a mood image, and expect the model to invent whatever you leave unspecified.

1. First exterior concept. Use when you have a site and a programme and want a first exterior image to react to.

A wide 16:9 eye-level photograph of a two-storey [house / community library / small office] on a [sloping suburban / flat rural / tight urban infill] site in [city or region]. The building is a simple rectangular volume with a single-pitch roof falling towards the street, a recessed entrance on the left third of the facade, and a full-height glazed bay on the right. Materials: pale lime-washed brick walls, a dark standing-seam zinc roof, and untreated larch boards on the recessed entrance. Mature trees behind the building, a gravel forecourt in front. Late afternoon sun from the left, long soft shadows. No people, no cars, no text, no watermark. Photorealistic architectural photography.

2. Street-scale massing. Use when you need a street-scale massing image for a mid-rise housing scheme.

A 3:2 street-level photograph of a six-storey apartment building on a corner plot in a [dense European city / North American downtown]. The massing steps down from six storeys on the corner to four storeys along the side street, with a two-storey brick base and set-back upper floors in a lighter material. Regular vertical windows, one balcony per apartment facing the quieter street. Materials: warm red-brown brick base, off-white precast concrete panels above, bronze-anodised window frames. Neighbouring buildings continue the street wall at five storeys. Overcast midday light, no strong shadows. No people, no cars, no text. Photorealistic architectural photography.

3. Three massing options, one image. Use when you want three massing alternatives for the same site in a single image.

A single 16:9 image divided into three equal panels, side by side, each showing the same [rectangular / L-shaped] site from the same high aerial viewpoint at the same scale. Left panel: one long linear block along the north edge with a courtyard to the south. Middle panel: two parallel blocks with a shared garden between them. Right panel: a compact square block in the centre of the site with open space on all four sides. All three volumes are three storeys, rendered as plain white model massing with no windows, on a neutral grey site plate with existing neighbouring buildings shown in light grey. Soft studio lighting, no text or labels.

4. Context aerial. Use when you need to see a proposed building in its wider context from above.

A 16:9 oblique aerial photograph, taken as if from a drone at about 60 metres, of a proposed [primary school / health centre] on a [riverside / edge-of-town] site. The building is a low two-storey courtyard scheme with a sawtooth roof over the larger wing, surrounded by existing [terraced housing / farmland / industrial sheds]. Materials: pale grey fibre-cement panels on the walls, a planted sedum roof on the flat sections, and timber-lined window reveals. Morning light from the east, slight haze towards the horizon. No people, no cars, no text, no watermark. Photorealistic architectural photography.

5. Study-model look. Use when you want the look of a white study model on a table.

A 4:3 photograph of a white architectural study model of a [three-storey mixed-use building] on a simple basswood site base, shot on a studio table with a plain light-grey backdrop. The model shows the main volumes only: a long ground-floor plinth, two offset upper blocks, and a rooftop pergola, with window openings cut as plain rectangles and no glazing. Surrounding buildings shown as uniform grey blocks. Soft, even daylight from a large window on the right, shallow depth of field so the background blurs slightly. No people, no text, no scale figures.

The usual failure in this group is a believable building that is not quite the one you described, and prompts 1, 2 and 4 carry the greatest risk. Floor counts and window rhythms are the usual casualties, per the renderahouse.com write-up of 18 October 2026. Prompt 3 asks the model to keep three panels at the same scale, and the model may ignore that instruction. Use these images to decide a direction and then draw it yourself.

Sketch, plan and model-screenshot prompts

These five prompts render your own design: a hand sketch, a SketchUp or Revit screenshot, a 2D plan, a section or a physical model photo. Every one of them opens with a preserve-list, because the model otherwise reinterprets the massing you drew. Expect a persuasive concept image, then compare it to the drawing it came from line by line before it leaves your desk.

Important. Outputs from these prompts are concept imagery, never measured drawings. Nano Banana has no scale or dimension input, so nothing in the image is to size, and the drawings, dimensions, codes and regulatory compliance remain the licensed professional’s responsibility to verify. Check floor count, window positions, roofline and camera against your own drawing every time.

6. Sketch to photoreal view. Use when you have a hand-drawn exterior perspective and want a photoreal version of exactly that view.

Use the uploaded hand sketch as the only source of geometry. Keep the exact same massing, window positions, floor count, roofline, proportions and camera angle from the sketch. Do not add floors, change the roof pitch, or move openings. Render the building as a finished photograph: walls in rough textured render painted warm white, the roof in dark grey clay tiles, window frames in dark timber. Site as drawn, with the existing [tree / wall / neighbouring building] kept in place. Soft late-morning light from the right. No people, no cars, no text, no watermark. 3:2 landscape frame. Photorealistic architectural photography.

7. Model screenshot, materials added. Use when you have a SketchUp, Revit or Rhino screenshot and want materials and light added without geometry changes.

Use the uploaded model screenshot as the source of all geometry. Keep the exact same massing, window positions, floor count, roofline, column positions and camera from the screenshot; the only changes are materials, lighting and context. Apply these materials: [material 1, e.g. board-formed concrete] on the ground floor, [material 2, e.g. vertical charred-timber cladding] on the upper floors, [material 3, e.g. slim black aluminium] window frames. Replace the blank model ground with a paved forecourt and low planting. Golden hour backlighting creating long shadows, slight lens flare. No people, no cars, no text, no watermark. Photorealistic architectural photography.

8. Plan to isometric. Use when you want a 2D floor plan turned into an isometric 3D view for a client who does not read plans.

Use the uploaded 2D floor plan as the exact layout. Produce a 3D isometric cutaway view of this plan seen from above at a 45-degree angle, with walls extruded to a consistent height of about one storey, all door and window openings in the positions shown, and no rooms added, removed, merged or resized. Keep room proportions exactly as drawn. Floors in pale oak, walls in matte white, simple furniture placed only where the plan indicates it. Clean white background, soft shadows, no text or dimensions, square 1:1 frame.

9. Section to sectional perspective. Use when you have a technical section and want an atmospheric sectional perspective for a presentation.

Use the uploaded building section as the exact source of geometry. Keep the exact same floor levels, floor count, roof profile, wall thicknesses, stair positions and opening positions from the section. Turn it into a sectional perspective: the cut face rendered in flat dark grey, the interiors beyond shown with simple furniture and people-free spaces, daylight entering through the openings exactly where the section shows them. Materials: exposed concrete floors, light timber ceilings, white walls. Soft overcast daylight, no strong shadows. Landscape 16:9 frame. No text, no dimensions, no annotations.

10. Model photo onto the real site. Use when you have a photo of a physical study model and want to see it as a completed building on the real site.

Image A is a photograph of a physical study model; Image B is a photograph of the real site. Use Image A for the building's massing, proportions, window openings and roofline, kept exactly as modelled; do not add floors or openings. Use Image B for the site, surroundings, horizon and lighting. Place the building on the site at the position of the [empty plot / existing car park] in Image B, matched to the camera height and perspective of Image B. Materials: [material 1] walls, [material 2] roof, [material 3] window frames. Keep everything else in Image B exactly as photographed. No people, no cars, no text, no watermark. Photorealistic architectural photography.

Prompt 8 is the one to watch here: an isometric view of your own plan looks measured, and nothing in it is to scale. Prompt 10 relies on two reference images, and Google’s prompting guide of 20 November 2025 notes the supported input count varies by surface, so budget a second turn for horizon or scale. In all five, changes to the window count are what the renderahouse.com write-up reports as normal, and the check in the callout above is the part of the workflow you cannot skip.

Site photo and interior restyling prompts

Editing a real photograph is the client-facing use, and the place where a silently changed window gets into a presentation. These five prompts edit a site, a facade, a room or a street, and each one names what must stay exactly as photographed before asking for one change. If you need a second change, send it as a second turn in the same chat. Do not rewrite the prompt.

11. Proposal on the empty site photo. Use when you want your proposed building dropped into a photo of the empty site.

Use the uploaded site photograph as the base. Keep the sky, the neighbouring buildings, the road, the trees, the people and the lighting exactly as photographed; the only change is to add a new building on the [empty plot / fenced site] in the centre of the frame. The new building: a three-storey volume with a flat roof, a brick ground floor in pale buff brick, two upper floors in dark grey standing-seam metal, and regular vertical windows with deep reveals. Match the camera height, perspective, lens and light direction of the photograph. No text, no watermark. Photorealistic architectural photography.

12. Facade re-clad, openings kept. Use when you need to show an existing facade re-clad without changing its openings.

Use the uploaded facade photograph as the base. Keep the exact same massing, window positions, window sizes, floor count, roofline, doors, camera and lighting from the photograph. The only change: replace the existing [painted render / tired brick / grey panel] cladding with [new material, e.g. glazed white terracotta baguettes] across the whole facade, with [slim bronze] window frames replacing the existing frames in the same openings. Leave the street, sky, neighbouring buildings and all people and vehicles exactly as photographed. No text, no watermark. Photorealistic architectural photography.

13. Interior refinish. Use when you want an existing room restyled with new finishes but the same architecture.

Use the uploaded interior photograph as the base. Keep the exact same room proportions, window positions, ceiling height, door positions, camera angle and daylight direction from the photograph. The only changes are finishes: floors in wide-plank pale oak, walls in warm off-white limewash, ceiling in exposed pale timber joists, and the existing window frames repainted matte black. Remove all existing furniture and leave the room empty. Natural daylight from the existing windows only. No text, no watermark. Photorealistic interior photography.

14. Public space around an existing building. Use when the proposal is the space between buildings: paving, planting, street furniture.

Use the uploaded street photograph as the base. Keep the buildings, their facades, windows, roofs, the sky and the camera exactly as photographed. The only change is the ground plane: replace the asphalt and parking bays in front of the [building name] with a paved pedestrian plaza in large-format pale granite, two rows of young deciduous trees in tree pits, timber benches along the facade, and low planting beds at the corners. Keep the existing lighting and time of day. A few people walking at a distance are fine; no cars, no text, no watermark. Photorealistic architectural photography.

15. Furnish an empty interior. Use when you have an empty interior (photo or render) and want it furnished for a client.

Use the uploaded photograph of the empty interior as the base. Keep the exact same room geometry, walls, floor and ceiling finishes, window positions, doors, camera and daylight exactly as shown; do not change any architecture. The only change is to furnish the space as a [living room / open-plan office / cafe]: a [three-seat sofa in oatmeal linen], a [low oak coffee table], a [wool rug in muted grey], a [floor lamp] by the window, and two framed prints on the wall opposite the window. Scale the furniture realistically to the room. No people, no text, no watermark. Photorealistic interior photography.

After prompts 11 and 12, check the neighbouring buildings as carefully as the proposal, because the renderahouse.com write-up of 18 October 2026 describes the model reinterpreting the whole image on an edit. After prompt 13, compare the window sizes against the source photo. Prompt 15 has no dimension input behind it, so do not read a sofa’s length off the image.

Material, lighting and atmosphere prompts

Four prompts here hold the design fixed and change one atmospheric variable: cladding, time of day, season, or interior light. They are the cheapest way to answer a client’s “what if it were timber?” without re-rendering. Run them on the best image you already have for the project, in the same chat, and read the result as a mood comparison. It tells you nothing about how the material will look on site.

16. Cladding A/B. Use when you want to compare cladding options on the same image.

Use the uploaded image as the base. Keep the exact same massing, window positions, floor count, roofline, camera, lighting, site and sky. The only change: replace the [current cladding, e.g. white render] on the upper floors with [new cladding, e.g. vertical untreated larch boards weathered to silver-grey], with the same board direction across every elevation visible in the image. Leave the ground floor, the window frames and everything else exactly as shown. No text, no watermark. Photorealistic architectural photography.

17. Day to night. Use when you need the same view at night for a lighting or planning conversation.

Use the uploaded image as the base. Keep the exact same building, massing, window positions, floor count, roofline, materials, site and camera. The only change is time of day: convert the scene to a clear evening about thirty minutes after sunset, deep blue sky, warm interior light visible through the windows on all floors, soft downlighting at the entrance, street lamps on, no illuminated signage. Keep all surroundings in their photographed positions. No people, no cars, no text, no watermark. Photorealistic architectural photography.

18. Season and weather. Use when you want to see the building in winter or in rain for a seasonal study.

Use the uploaded image as the base. Keep the exact same building, massing, window positions, floor count, roofline, materials, site layout and camera. The only change is season and weather: show the scene in [late winter with bare trees and a thin layer of snow on flat surfaces / a grey rainy afternoon with wet reflective paving and overcast sky]. Adjust only the sky, vegetation, ground surface and light; do not change the building. No people, no cars, no text, no watermark. Photorealistic architectural photography.

19. Daylight vs evening split. Use when you want an interior compared under daylight and under artificial evening light.

Use the uploaded interior image as the base. Produce a single 16:9 image split into two equal panels showing the identical room from the identical camera. Keep the exact same room geometry, finishes, furniture and window positions in both panels. Left panel: overcast midday daylight entering only through the existing windows, no artificial lights on. Right panel: evening, windows dark, warm artificial light from [recessed ceiling downlights / a pendant over the table / a floor lamp], 2700K colour temperature. No people, no text, no watermark. Photorealistic interior photography.

Watch prompt 16 for window proportions: new cladding can bring new openings with it, so count them against the original. After prompt 17, compare the lit windows at night against the daytime image, floor by floor. Prompt 19 asks for two identical panels; check that the room geometry actually matches across the split.

Diagram, text and presentation prompts

These six prompts use a capability Google highlights for Nano Banana Pro: legible text inside the picture. Use Nano Banana Pro for all six, because Google positions that model for design, mockups and “factual data visualizations that require accurate text rendering”, per its model page read 20 October 2026. Check every word and number in the output before it reaches a board.

20. Exploded axonometric with labels. Use when you need an exploded axonometric that explains a building’s layers.

A clean exploded axonometric diagram of a [three-storey timber-frame community building], viewed from the south-east, with the roof, each floor plate, the structural frame and the ground floor slab separated vertically and aligned on a common axis. Each layer labelled in a small sans-serif font with these exact labels: "Sedum roof", "CLT roof panels", "Second floor", "First floor", "Glulam frame", "Ground floor slab". Thin leader lines from each label to its layer. Flat pastel colours, white background, no shadows, no people, square 1:1 frame.

21. Colour-coded site diagram with legend. Use when you want a colour-coded site plan diagram with a legend for a planning board.

Use the uploaded site plan as the exact base geometry and do not move, resize or add any building. Produce a flat colour-coded site diagram seen directly from above: residential buildings filled in muted terracotta, commercial in slate blue, public open space in sage green, car parking and service areas in light grey, water in pale blue. Add a legend in the bottom right corner in a plain sans-serif font with exactly these five entries: "Residential", "Commercial", "Public open space", "Parking and service", "Water". Add a north arrow in the top right. No other text, no 3D, no shadows, landscape 4:3 frame.

22. Concept board from three images. Use when you need a concept board that combines three of your images with a title.

Image A is the exterior view, Image B is the interior view, and Image C is the site plan. Compose a single A1-proportioned landscape presentation board on a white background. Place Image A large across the top two-thirds, Image B bottom left and Image C bottom right at equal size, each image kept exactly as supplied without re-rendering or altering the buildings. Add the title "[Project name] - Concept Design" top left in a bold sans-serif font, the subtitle "[Studio name], [month year]" beneath it in a lighter weight, and these three captions under the images: "Street view", "Main hall", "Site plan". No other text, no logos, no decoration.

23. Environmental strategy section. Use when you want an environmental strategy section with arrows and labels.

Use the uploaded building section as the exact base geometry and keep every floor level, roof profile and opening where it is. Produce a clean diagrammatic version of this section in flat line style on a white background, with these annotations: curved orange arrows showing winter sun entering through the south-facing glazing, blue arrows showing cross ventilation from the low openings on the left to the high clerestory on the right, and green arrows showing rainwater running from the roof to a tank below ground on the right. Label the arrows in a small sans-serif font with exactly: "Winter sun", "Cross ventilation", "Rainwater harvesting". No other text, landscape 16:9 frame.

24. Translate a labelled diagram. Use when you need the same labelled diagram in a second language for a client or consultant.

Use the uploaded diagram as the base. Keep the exact same drawing, geometry, colours, arrows, layout and font size; change nothing except the text. Translate every label and caption from English into [Spanish], keeping each translated label in exactly the same position and the same style as the original. Use standard architectural terminology for the target language. Do not add, remove or reposition any element, and do not add a translation note or any new text.

25. Before/after sheet with captions. Use when you want a before/after sheet for a refurbishment with captions.

Image A is the existing building photograph and Image B is the proposed view. Compose a single landscape 16:9 image with the two photographs side by side at equal size, separated by a thin white gap, on a white background. Keep both images exactly as supplied with no re-rendering. Under the left image add the caption "Existing - [month year]" and under the right image the caption "Proposed - [month year]", both in a plain sans-serif font in dark grey. Add a small title at the top left reading "[Street address] - Facade refurbishment". No other text, no logos.

For this group the limits are documented by Google itself. Google’s guidance of 20 November 2025 states that “Rendering small text, fine details, and producing accurate spellings may not work perfectly” and that you should “Always verify the factual accuracy of data-driven visuals like diagrams and infographics”. Prompts 20 and 23 are diagrams of your own building, so the facts are yours to check. Prompt 24 produces a translation you should read with a native speaker before a consultant sees it, and prompts 22 and 25 need a side-by-side check that the supplied images survived the composition unchanged.

What can architects do with Nano Banana for free?

The free Gemini app tier includes image generation, which covers concept studies, a handful of restyles and learning the prompt grammar. A day of client iteration will exhaust it. Nano Banana Pro images on the free tier are limited and fall back to a faster Nano Banana model, and the limits are compute-based with no published per-day image count, so you find the ceiling by hitting it.

The detail comes from Google’s own pages. The Gemini Apps limits page, read 20 October 2026, lists image generation with Nano Banana 2 as available on every tier including free, while “Redo images with Nano Banana Pro” requires a paid tier. The same page says usage limits are compute-based, refresh every 5 hours up to a weekly cap, and that Plus gets 2x and Pro 4x the standard limits, with Ultra at 5x or 20x of Pro; no per-day image counts are published for any plan. The free-tier fallback was stated in the launch post of 20 November 2025: free users get limited Nano Banana Pro quotas, after which they revert to a faster Nano Banana model.

Plan	Monthly price	What it means for image work
Free	$0	Image generation included; Nano Banana Pro limited with fallback to a faster Nano Banana model; standard compute limits, visible watermark
Google AI Plus	$4.99	2x standard limits; Nano Banana Pro redo available; watermark not separately stated by Google
Google AI Pro	$19.99	4x standard limits; Nano Banana Pro redo available; visible watermark
Google AI Ultra	$99.99 to $199.99	5x or 20x Pro limits; visible watermark removed; the tier for heavy daily use

Plan prices from Google’s AI plans page, read 20 October 2026; usage multipliers and feature rows from the Gemini Apps limits page, read 20 October 2026; watermark behaviour from the launch post of 20 November 2025.

If your practice would rather pay per image than per seat, there is the Gemini API and Google AI Studio route. On Google’s pricing page, read 20 October 2026, Nano Banana Pro costs $0.134 per 1K or 2K image and $0.24 per 4K image, while Nano Banana 2.1 costs $0.0336 per 1K, $0.0504 per 2K and $0.113 per 4K. A hundred Nano Banana 2.1 concept images at 1K is therefore around $3.36, and a hundred 4K Nano Banana Pro boards around $24, at those October 2026 prices.

So what is the free tier actually enough for? In my judgement, the five concept prompts above and two or three restyles per sitting, spread across a week of early design, with the 5-hour refresh doing the rationing. The moment you are running the material A/B prompts for a client meeting, or need the visible watermark gone, you are on a paid tier or on the per-image API route, and the per-image route is the one I would price into a fee proposal.

Where these prompts break

Four things go wrong with all 25 nano banana prompts for architects above, and none of them is fixed by better wording. The geometry drifts because there is no dimension input; small text and spellings come out imperfect; every image carries a watermark, visible or invisible; and a purely AI-generated image may have no copyright protection. The adoption numbers below show the profession using the tool anyway; the checks are what make that safe.

Geometry first. The renderahouse.com write-up of 18 October 2026 describes floor counts, window rhythms and roof pitches changing between generations, outputs that look right at a glance with no scale input behind them, and the model reinterpreting the source on each generation. Google’s own limitation note of 20 November 2025 is consistent with this, even though Google phrases it around text and data visuals without mentioning buildings. The preserve-list reduces drift. Checking the count against your drawing catches what is left.

Text and data second. Google writes, in the same prompting guide of 20 November 2025, that “Rendering small text, fine details, and producing accurate spellings may not work perfectly” and tells you to “Always verify the factual accuracy of data-driven visuals like diagrams and infographics”. For an architect that means every label on an exploded axo, every legend entry and every translated caption is proofread by a human, and a diagram that claims a level of environmental performance is not evidence of it.

Watermarks third. According to the launch post of 20 November 2025, every image output carries an invisible SynthID mark, images from the free and Google AI Pro tiers also carry a visible Gemini sparkle watermark, and the visible mark is removed for Google AI Ultra subscribers and in Google AI Studio. The Gemini app can also check whether an uploaded image was generated by Google AI. Plan for the visible mark on anything from a free or Pro account, and assume the invisible one is always there when you publish.

Copyright fourth. This is a caution; ask a lawyer for the advice. The US Copyright Office’s report of January 2025 on copyrightability takes the position that works lacking human authorship are not protected, which means a purely AI-generated image may have no copyright protection in that jurisdiction and possibly in others. If a client asks who owns the concept image, the honest answer is that the design is yours and the pixels may belong to nobody; ask a lawyer in your own jurisdiction before you make a contractual promise about them.

None of this is stopping the profession. Dezeen’s survey of 25 big-name studios, published 25 August 2026, found Nano Banana the most-used image generation tool at 7 of 25 studios, with image generation an AI task at 16 of 25, and CF Møller estimated in the same feature that its Photoshop usage had fallen 80% because of generative AI. The RIBA AI Report 2026 of 22 July 2026 puts 74% of UK practices using AI on at least some projects, with early-design visualisation the top use at 56% of AI-using practices. Our own breakdown of how architects actually use AI in 2026 lines up with that picture.

Alastair Roberts, COO of Hawkins Brown, put the professional stance this way in Dezeen’s 25 August 2026 feature: “Ultimately we’re stressing that AI produces useful inputs for our work - e.g. it can accelerate or automate some processes very effectively - but we are all still responsible as qualified human professionals for what we deliver to our clients.” That is the liability position I would adopt. Every prompt on this page ends in a check against your drawing for that reason.

FAQ

The five questions architects ask most about Nano Banana are cost, sketch-to-render, model choice, watermarks and client use. Each answer below stands on its own, with the source and date it rests on, so you can forward a single one to a colleague without the rest of the page. For the text side of practice, our comparison of Claude vs ChatGPT on 12 real tasks covers the writing models.

Is Nano Banana free for architects?

Partly. The free Gemini app tier includes image generation with a Nano Banana model, with limited Nano Banana Pro images that fall back to a faster model, per Google’s limits page read 20 October 2026. Limits are compute-based and refresh every 5 hours, with no published per-day count. That is enough for concept studies and a few restyles, and the visible watermark stays on free-tier images.

Can Nano Banana turn a sketch into a render?

Yes - sketch-to-render is the most used job for nano banana prompts for architects, and the result is a concept image that must never be treated as a drawing. Upload the sketch, open the prompt with a preserve-list (“Keep the exact same massing, window positions, floor count, roofline and camera”), then describe materials and light in plain sentences. The renderahouse.com write-up of 18 October 2026 reports floor counts and window rhythms drifting, so count them against the sketch before you share it.

Which Nano Banana model is best for architecture?

Nano Banana 2.1 for everyday iteration and Nano Banana Pro for anything with text, several reference images or a 4K output. On Google’s pricing page read 20 October 2026, Nano Banana 2.1 costs $0.0336 per 1K image against $0.134 for Nano Banana Pro. Run the cheap model first and the premium model for the board. Nano Banana 2 Lite is for 1K thumbnails only.

Do Nano Banana images have a watermark?

Yes, always. Google’s launch post of 20 November 2025 states that every image carries an invisible SynthID mark, that free and Google AI Pro tier images also show a visible Gemini sparkle, and that the visible mark is removed for Google AI Ultra subscribers and in Google AI Studio. The invisible mark remains on every output, and the Gemini app can detect it on an uploaded image.

Can I use Nano Banana images in client presentations?

Yes, labelled as concept imagery, with the geometry checked against your drawings and the watermark question settled for your tier. Label the image in the presentation as an AI-assisted concept view. Copyright is a separate caution: the US Copyright Office’s January 2025 report says purely AI-generated work may lack protection, so ask a lawyer before promising ownership to a client.

Sources

Every fact on this page traces to one of the thirteen sources below, each with the date it was published or the date I read it. Google’s own pages carry the model lineup, prices and limits; Dezeen, AEC Magazine and the RIBA carry the adoption context; the two community write-ups carry the prompt patterns and the failure modes, which I report from those write-ups and have not reproduced myself.

• Google - Introducing Nano Banana Pro, blog.google, 20 November 2025. Model lineage, availability, free-tier fallback, SynthID and visible watermark policy.
• Google - 7 tips to get the most out of Nano Banana Pro, blog.google, 20 November 2025. Prompt elements, camera and lighting language, reference-image roles, stated limitations.
• Google - Image generation with Gemini (Nano Banana), ai.google.dev, read 20 October 2026. 2026 model lineup, model ids, resolutions, reference-image budgets, Thinking mode.
• Google - Gemini 3 Pro Image model page, ai.google.dev, read 20 October 2026. Positioning quotes for Nano Banana Pro.
• Google - Gemini API pricing, ai.google.dev, read 20 October 2026. Per-image prices for Nano Banana Pro and Nano Banana 2.1.
• Google - Gemini Apps usage limits and features, support.google.com, read 20 October 2026. Compute-based limits, 5-hour refresh, feature table by tier.
• Google - Google AI plans, one.google.com, read 20 October 2026. Plus, Pro and Ultra prices and usage multipliers.
• Dezeen - Big-name studios encouraging architects to use AI, 25 August 2026. 25-studio survey, Nano Banana usage, CF Møller estimate, Hawkins Brown quote.
• AEC Magazine - Architecture: the missing AI middle, 3 March 2026. Spatial coherence assessment of Nano Banana Pro.
• RIBA - AI adoption reaches tipping point across UK architecture (RIBA AI Report 2026), 22 July 2026. 74% adoption, visualisation at 56%.
• meltflexai - 40 Nano Banana Architecture Prompts for Architects, 20 July 2026. Community prompt formula and the “photorealistic architectural photography” closer.
• renderahouse.com - Nano Banana for architecture, 18 October 2026. Geometry drift, no dimension input, the model reinterpreting the image on each edit.
• US Copyright Office - Copyright and Artificial Intelligence, Part 2: Copyrightability, January 2025. Human authorship requirement.

The prompt grammar on this page will outlast the model names in the table. If you want the next library of nano banana prompts for architects, and one AI workflow a week tested on real work, the newsletter is where they land first.