Banana prompts are the formulas and techniques behind Google’s Nano Banana image model: a five-slot formula for generating images from scratch, a three-part formula for editing or combining existing ones, and controls over lighting, lens choice and on-image text.
Google Cloud’s prompting guide for Nano Banana gives one skeleton for generating from scratch, [Subject] + [Action] + [Location/context] + [Composition] + [Style], and one for working from pictures you already have, [Reference images] + [Relationship instruction] + [New scenario] (Google Cloud). Every technique below hangs off those two lines, which is what turns a template into something you can rebuild for your own shot rather than a string you paste and hope about.
On this page
The five-slot text-to-image formula
Subject, action, location or context, composition, style. The composition slot is where framing and viewpoint belong, and style sits last, applying to everything written before it (Google Cloud). The practical value of the five slots is diagnostic: when a result comes back flat, you can ask which slot you left empty instead of rewriting the whole line. A prompt with no composition slot has not chosen a framing, it has simply left the choice unwritten.
Editing prompts and the parts that stay
For edits, Google’s rule is to state explicitly what should stay unchanged, not only what should change (Google Cloud). That makes an editing instruction two-sided: the change you want, plus a named inventory of what you are holding, such as pose, background, lighting or the face. Listing them is part of the instruction rather than a courtesy.
Semantic masking instead of coordinates
The guide names semantic masking as inpainting a specific region defined by a text description rather than by coordinates (Google Cloud). Since the region arrives as words, your accuracy depends on how unambiguously those words pick out one area. A noun phrase that could match two surfaces in the frame is doing the job of a selection tool badly, so name the region the way you would point at it for someone else.
Composition and style transfer
Style transfer in the guide means describing a reference style and applying it to new content (Google Cloud). It uses the multimodal formula, with the relationship instruction carrying the weight: the reference supplies the look, the new scenario supplies the content, and the sentence between them says which is which.
Text rendering inside an image
Four moves for lettering, all from the same guide: quote the exact words you want rendered, specify the font or style, request localized text, or use what the authors call the “text-first hack”, working out the exact wording conversationally first and then generating the image with that finalized text (Google Cloud). The last one separates two problems that usually fail together, deciding the copy and rendering it.
Creative Director controls
The guide’s “Creative Director” language borrows a film crew’s vocabulary in four areas: lighting named as a setup, for example a “three-point softbox setup”; camera, lens and focus choices; color grading or a film stock reference; and the materiality and texture of surfaces (Google Cloud). Each one replaces an adjective with a decision. “Moody” is a mood; a lighting setup, a focal length and a stock reference are instructions.
Negative prompts
Google’s guidance is blunt: “describe what you want, not what you don’t want” (Google Cloud). In practice that means rewriting each ban as a description. Instead of forbidding an object, describe the surface or space that should occupy that part of the frame, which also fills the location and composition slots while you are at it.
Web-search grounding
The guide includes a web-search grounding formula, which lets the model pull live search, weather or date type data into a visual, and describes it as coming soon to Vertex AI (Google Cloud). Read that as announced rather than available, and do not design a production workflow around it yet.
Chaining image, video and music models
The multi-model workflow Google describes runs in sequence: generate keyframes with the image model, animate them with Veo, then score the result with Lyria (Google Cloud). It changes how you write the stills. A frame intended for animation has to state its composition deliberately, because that framing is what the next step inherits.





