Google publishes a small set of prompting frameworks for the Nano Banana image models, covering text-to-image, reference images, search-grounded visuals and text inside the frame. This guide walks through each one in the order the official material presents them, and states plainly which of the four officially registered tiers the NanoBanana Pro workbench actually exposes.
Five official formulas · Four registered tiers explained · Three tiers exposed in this workbench
Model IDs and roles are taken from the Gemini API image-generation documentation. The last line of each card states what the NanoBanana Pro workbench itself exposes, so the official line-up and this site’s controls are never merged into one list.
Official model ID gemini-3.1-flash-lite-image. Positioned as the fastest and the cheapest of the four, capped at 1K output and not tuned for multi-reference input or multi-turn editing.
Workbench status: Not exposed in the NanoBanana Pro workbench.
Official model ID gemini-3.1-flash-image. Google describes this as the go-to tier, balancing speed, 4K output, world knowledge and text rendering. It is also the only tier that offers a 512px output size.
Workbench status: Exposed in the workbench, with 1K, 2K and 4K output options.
Official model ID gemini-3-pro-image. Built for professional asset production and complex instructions, with Google Search grounding, a default Thinking pass and 4K output.
Workbench status: Exposed in the workbench, with 1K, 2K and 4K output options.
Official model ID gemini-2.5-flash-image. Registered as the legacy generation of the family, and Google recommends moving new work to Nano Banana 2.
Workbench status: Still exposed in the workbench as the base tier.
Prompting an image model sits closer to writing a production brief than to hunting for a magic phrase. Google’s own guidance for these models reduces the whole habit down to one instruction that overrides most of the folklore: describe what you want inside the frame instead of what you want removed. Every technique on this page follows from that single rule, including the ones about editing and reference images.
The formulas collected here come from Google’s published Nano Banana prompting material and from the Gemini API image-generation documentation, whose page footer records a last update on 2026-09-17 UTC. Together they cover the four jobs people actually arrive with: making a picture from a description, holding something consistent across a set of images with reference input, grounding a picture in a real search result, and rendering legible text inside the picture.
The Create workbench at nanobananapro.hk exposes three of the four registered tiers, so each formula below is written to run on any of them. If you would rather work through graded lessons than read a reference page, the long-form tutorials in /guide/ cover precision prompting in depth, and /blog/ collects workflow write-ups as the models change.
Each formula is a slot order rather than a script. Keep the order, fill the slots with your own subject, and drop any slot the image genuinely does not need.
Subject, action, location or context, composition, then style. Google writes the order as [Subject] + [Action] + [Location/context] + [Composition] + [Style], which keeps the main subject ahead of decorative detail instead of burying it under adjectives.
[Reference images] + [Relationship instruction] + [New scenario]. Name what each uploaded image contributes, state how the references relate to one another, and only then describe the new scene you want built from them.
[Source or search request] + [Analytical task] + [Visual translation]. Ask for the fact first, work out the answer as text, and only afterwards describe how that fact should look as a picture.
Put the exact characters in quotation marks, name the font style, state the language the text should use, and describe the text before you describe the surrounding scene. The official material also notes rendering support for more than ten languages.
Four lever groups carry most of the art direction: lighting, camera lens and focus, colour grade and film stock, and material texture. Reuse the same words each time you iterate, so that a difference in the result is attributable to one change rather than to reworded styling.
Generating and editing are different jobs with different prompts. An edit prompt states what changes and, just as explicitly, which parts of the picture must stay exactly as they are. Text alone is enough to define the protected region.
This page restates Google’s published Nano Banana guidance and names the places where the NanoBanana Pro workbench differs from the officially registered line-up. It is deliberately written so that every statement can be checked against the official documentation or against this site’s own configuration.
Read the service termsA prompt that survives contact with the model reads like a short production note. Name the primary subject and the action it performs, then add setting, composition and visual style. Keeping one clear subject per prompt removes ambiguity before you spend credits on variants, which is why the official formulas all put subject and action ahead of everything else.
Camera language is the highest-leverage vocabulary you can add, because these models respond to concrete photographic terms. Low angle, aerial view, wide shot, shallow depth of field and named lens behaviour all produce visible differences. Lighting, colour grade and material texture behave the same way: standard photographic nouns outperform adjectives such as beautiful or cinematic, which carry no framing information.
Positive framing is not a stylistic preference here, it is the documented instruction. Describing the frame you want is more reliable than listing the things you do not want, and on this site it is also the only option, because the image workbench has no field dedicated to exclusions. When something must be absent, describe the state that replaces it: two hands resting on the table rather than no extra hands.
Consistency across a set of images is a separate problem from quality in a single image. The official answer combines reference images with naming: upload the reference, then give each recurring character or object a distinct name and reuse that name in every later prompt, so the model can track which entity is which across generations.
The same order works for text-to-image, reference-image work and edits, so you can move a prompt between them without rebuilding it.
Open with a strong verb describing what happens in the frame, then name the subject. Google lists starting with the main action as a best practice, and it prevents staging detail from being read as the subject itself.
Add angle, shot size, focus and aspect ratio in plain photographic terms such as low angle, aerial view or shallow depth of field. Control the camera explicitly rather than trusting the model to infer framing.
In edits and reference-image work, list the parts that must not change: identity, logo geometry, packaging proportions, product colour. Say it in the prompt rather than relying on the model to protect what matters to you.
Pick the section that matches the job in front of you, and keep the slot order intact inside it.
Use the five-slot order and keep style terms last. Style written early tends to leak into the subject description, which is the most common reason a prompt produces something nobody asked for.
Upload references, give every recurring entity its own name, and repeat those names in each follow-up prompt. Consistency across a set is maintained by naming, not by repeating the same paragraph of description.
Describe the change and the protected region in the same prompt, and limit each pass to one region. One change per pass is what makes the result attributable when you compare versions.
Quote the exact characters, choose a font style, state the target language and keep the string short enough to fit the frame. Describing the text before the scene around it is the order the official material recommends.
Ask for the fact, then translate it into a scene. These tiers record a knowledge cutoff of January 2025, and grounding a request in a live search is how the models reach anything more recent.
Refine conversationally instead of rewriting from scratch, and change one variable per attempt. Rewriting the whole prompt between attempts destroys the comparison you were trying to make.
People who write prompts as part of a job get the most out of the formula order. E-commerce teams can keep one formula per product category so that seasonal refreshes stay comparable across months, and the protected-region sentence in edit prompts is what keeps packaging geometry and label placement stable through a re-shoot.
Designers and marketers use the reference-image formula to hold a campaign together across dozens of frames, and the text-rendering rules to place a short string such as a shop name or a price without scheduling a second editing pass. Naming the recurring entities is usually what separates a coherent set from a collection of near-misses.
Teachers and students can read the five formulas as a marking rubric, because each one names the slots a prompt should contain. If a submission cannot say which formula it followed, it is usually missing the composition or the constraint slot rather than the subject.
Open Create, write one prompt in the five-slot order, and run the same brief across the three tiers exposed in the workbench to see how they differ.
Try a prompt in Create