Nano Banana Pro is useful for concept exploration, material and lighting studies, and client-facing interior visuals. It is not a dependable source for dimensions, code compliance, accessibility, structure, or construction documents.
The safest workflow is simple: start from a room photo, floor plan, or SketchUp view; state what must not move; change one design axis; then run an acceptance check before the image leaves the concept stage. If walls, openings, camera, scale, or circulation drift, stop prompting and return to CAD, BIM, or a controlled renderer.
This distinction matters because Google’s current product routes are also easy to mix up. Nano Banana Pro is the market name for Gemini 3 Pro Image, whose current API model ID is gemini-3-pro-image. Inside the Gemini App, the current consumer flow normally generates with Nano Banana 2 and offers Redo with Pro on paid plans. The API is a separate, usage-priced developer contract.
The architecture constraint card
Before writing a style prompt, create a small card with two columns:
| Preserve | May change |
|---|---|
| camera position and lens feel | material palette |
| exterior walls and structural grid | furniture family |
| doors, windows, stairs, and columns | fixtures and styling objects |
| one known scale anchor | daylight, artificial light, and mood |
| circulation path and fixed services | color direction |
This card is more useful than a long list of adjectives. It tells the model which parts describe the existing space and which parts describe the proposed design.
Design Ink’s published workflow found that prompts containing only a desired style could rebuild windows and change proportions. Describing the existing architecture first, then the new material and lighting direction, improved preservation in those examples. That is a practitioner observation—not a dimensional guarantee—but it gives us a practical prompting order.
Use a prompt skeleton like this:
“Treat the uploaded image as fixed source geometry. Preserve the camera position, wall layout, number and position of openings, ceiling height relationship, and the marked 36-inch (about 900 mm) door as the scale anchor. Change only the floor material, loose furniture family, and daylight character. Do not add rooms, move openings, widen circulation, or invent built-in joinery. Produce one concept visualization for design discussion, not a technical drawing.
Replace the inputs and allowed changes for your project. Do not copy a competitor’s prompt or assume the numbers in the output remain correct because the image looks plausible.
Choose the right input
Room photo: best for a controlled redesign
A clear room photo gives the model perspective, openings, ceiling relationships, and existing light cues. It is the strongest starting point when the job is “show a different material and furniture direction without redesigning the shell.”
Mark what must remain: doors, windows, radiators, outlets, bulkheads, columns, and built-ins. Crop out irrelevant rooms. If the client needs an exact product, include a licensed product reference and identify it as an appearance reference—not proof that the generated item has the correct dimensions or is available for purchase.
Floor plan: useful for massing, risky for eye-level truth
A floor plan can support top-down, axonometric, or rough spatial studies. It does not contain enough information to prove ceiling heights, facade conditions, furniture clearances, materials, or eye-level composition.
Keep room labels and a scale bar visible. Ask for one representation at a time. A top-down concept, an axonometric, and an eye-level interior are three separate interpretations; visual continuity across them must be checked, not assumed.
SketchUp or white-model view: best bridge to client communication
A fixed viewport already carries camera, massing, openings, and major geometry. It is often the most efficient source for testing materials and light while keeping the design intent legible.
Export a clean view without selection outlines, guides, or unrelated UI. Include the constraint card in the prompt. If the output moves the camera or changes the shell, do not keep polishing it. That failure belongs back in the controlled 3D scene.
A three-pass workflow that limits drift
Pass 1: preservation baseline
Ask for minimal visual change. The goal is not a beautiful final image; it is to learn whether the source geometry survives a generation. Compare the result with the input at the same size.
Check the count and position of openings, wall breaks, columns, stairs, and large furniture. If these fail now, more styling will hide the problem rather than solve it.
Pass 2: one design direction
Change materials, furniture family, or lighting—not all three at once. Single-axis variations make client feedback interpretable: the team can say which material or light direction worked without confusing that judgment with a different layout.
For material references, describe the useful attributes—tone, grain scale, reflectance, edge detail—rather than supplying a fully composed reference room and inviting the model to copy its layout.
Pass 3: local edit
Use a local edit for one bounded correction, such as the floor finish or one loose chair. Repeat the preserve list. If a local request changes a window, wall, perspective, or adjacent fixed object, reject the output.
The workflow ends when the model begins fighting the source geometry. More iterations are not inherently progress.
The five-part acceptance gate
An attractive image is not yet an accepted concept image. Review it against five questions:
- Geometry: are walls, columns, stairs, ceiling edges, and primary masses still where the source places them?
- Scale: does the known anchor still make furniture, doors, countertops, and circulation plausible?
- Openings: are the number, size relationship, swing implication, and position of doors and windows preserved?
- Circulation and ergonomics: can people still move, sit, open storage, and use the room without invented clearance?
- Technical boundary: does any decision depend on structure, services, fire safety, accessibility, code, cost, procurement, or constructability?
If questions one through four fail, revise the source model or constraints. If question five is involved, the image may illustrate an idea, but the decision must return to qualified people and professional documentation.
Architizer’s published architecture experiment is a useful warning: viewpoints and facade details can drift across generated views. A believable second angle does not prove that it describes the same building.
What the subscription actually does
There are three different routes, not one universal “Nano Banana Pro subscription.”
| Route | Best for | Price basis | Important boundary |
|---|---|---|---|
| Gemini App / Google AI plan | individual, interactive concept iterations | monthly consumer plan | current App behavior uses Nano Banana 2 for generation and offers paid Redo with Pro; limits are dynamic |
| Gemini Developer API | automated or repeatable workflows | image output plus input/text/thinking usage | developer billing, credentials, data terms, and regional availability are separate from the consumer plan |
| Third-party API or design tool | a different interface, payment method, or integrated workflow | provider’s call, credit, or subscription contract | model identity, retention, refunds, limits, and commercial terms belong to that provider |
On the US public subscription page checked on July 20, 2026, Google showed Plus at $4.99/month, Pro at $19.99/month, and Ultra starting at $99.99/month for the 5x option or $199.99/month for 20x. Your signed-in checkout and account region are final because prices, taxes, promotions, and benefits can change.
The current limits page does not promise an evergreen daily image count. It describes compute-based access and directs users to Settings → Usage limits. That account screen is more reliable than an old table copied into a blog post.
For the official API, Google currently lists no Standard free tier for gemini-3-pro-image. The image-output equivalent is $0.134 for 1K/2K and $0.24 for 4K. Input images, text, thinking, and optional services can add cost. Batch/Flex halves the image-output figures, but it has different latency and workflow tradeoffs. See the dedicated Nano Banana Pro pricing guide for the complete cost lane, or the API guide for integration details.
LaoZhang’s current public developer documentation, checked July 20, 2026, lists a separate provider route at $0.09 per call/image, with the console as the final price and charge owner. Treat that as a third-party API contract, not a Google subscription, and review model identity, data handling, failure billing, and required controls before using client work.
When Pro is worth using
Use the Pro route when the concept needs complex instructions, several visual references, dense labels, or a higher-resolution presentation asset—and when the extra detail will actually survive your acceptance gate.
Use a faster or cheaper route when the job is rough ideation, composition scouting, or high-volume direction finding. Keep that model choice separate from claims about architecture accuracy.
The right cost metric is not “price per generated image.” It is cost per accepted concept. Ten cheap generations that all move the windows are not cheaper than one controlled variation that passes review.
Privacy, copyright, and client material
Before uploading an unpublished floor plan, occupied-home photo, branded product, or client presentation, identify who receives the file and under which contract. A Google consumer account, a developer API project, and a third-party wrapper can have different retention, training, regional, and administrative controls.
Remove names, addresses, faces, access-control details, and unrelated drawings. Confirm that you may use every reference image. Keep the original source and decision record outside the generator so the team can reproduce why a concept was accepted.
The practical stop rule
Nano Banana Pro belongs between source geometry and concept review. It can help a team see material, furniture, lighting, and mood directions faster. It cannot make an unverified image into a measured, coordinated, compliant design.
Pass an image forward only when its geometry, scale cues, openings, and circulation survive review. The moment the question becomes “Will it fit, comply, stand, drain, ventilate, cost this amount, or be buildable?”, stop the image workflow and return to CAD, BIM, specifications, codes, suppliers, and qualified professional judgment.
Official references: Gemini image generation, Gemini API pricing, Gemini App image help, and Gemini limits and plans.