Nano Banana 2 Prompt Guide: How to Write Better Prompts for Google's On-Device AI Image Generator
Google's Nano Banana 2 has already turned heads as an on-device AI image generator that runs without cloud connectivity. We covered the model's technical foundations and capabilities in our previous deep dive. But having a powerful model on your device is only half the equation — knowing how to prompt it effectively is what separates mediocre outputs from genuinely impressive results.
This guide breaks down the core prompting principles, walks through ten content categories with ready-to-use examples, and shares practical tips that will help you get the most out of Nano Banana 2 whether you're a designer, developer, or curious experimenter.
Why Prompting Matters More on Nano Banana 2
Cloud-based models like Midjourney or DALL-E 3 have massive parameter counts that let them "fill in the gaps" when your prompt is vague. Nano Banana 2, being a compact on-device model, is more literal. It rewards precision and punishes ambiguity. A prompt that works fine on a cloud model might produce bland results on NB2 — not because the model is less capable, but because it has less room to guess what you meant.
The flip side is that when you give NB2 a well-structured prompt, the results can be surprisingly close to what you'd get from much larger models. The difference between a good and bad NB2 output almost always comes down to the prompt.
Six Core Prompting Principles for Nano Banana 2
Before diving into specific categories and examples, internalize these six foundational techniques. They apply across every type of image you'll generate.
1. Be Visually Specific
The most common prompting mistake is using vague descriptors. Words like "beautiful," "nice," or "cool" give the model almost no useful information. Instead, describe exactly what you see in your mind's eye.
Weak prompt: "A beautiful sunset over the ocean."
Strong prompt: "A warm golden sunset over a calm Pacific Ocean, deep orange and magenta gradient sky, scattered cirrus clouds catching the last light, gentle wave reflections on dark blue water, shot from a low coastal cliff, wide-angle perspective."
The difference is specificity in three dimensions: lighting (warm golden, orange and magenta gradient), texture and quality (calm ocean, scattered cirrus clouds, gentle wave reflections), and color palette (deep orange, magenta, dark blue). When you describe these elements explicitly, NB2 doesn't have to improvise — and it produces tighter, more coherent outputs.
2. Specify Output Parameters
NB2 can generate images across a range of resolutions and aspect ratios. Telling the model about your intended output helps it optimize composition.
Include details like:
- Resolution: "512x512 square format" or "1024x768 landscape"
- Aspect ratio: "16:9 widescreen" or "4:5 portrait"
- Intended use: "social media post," "app icon," "hero banner," "print illustration"
For example: "Product photo of wireless earbuds on a white marble surface, overhead shot, square 1:1 crop for Instagram, clean minimalist aesthetic, soft diffused studio lighting."
The intended use cue helps NB2 understand compositional expectations — an Instagram post has different framing needs than a website hero banner.
3. Leverage Real-World References
NB2 has been trained on a broad dataset that includes real-world knowledge. You can reference actual camera models, film stocks, art movements, and specific visual styles to anchor your output in a recognizable aesthetic.
Camera and lens references: "Shot on Hasselblad X2D, 90mm portrait lens, f/2.8, shallow depth of field"
Film stock references: "Kodak Portra 400 color palette, natural skin tones, subtle grain"
Art movement references: "In the style of Art Deco poster design, geometric forms, metallic gold accents on deep navy"
Publication references: "Editorial photography in the style of Vogue Italia, dramatic studio lighting, high contrast"
These references act as compressed style instructions. Instead of describing every visual quality you want, you invoke a known aesthetic that carries those qualities implicitly.
4. Use Text Rendering Deliberately
One of NB2's standout improvements over its predecessor is significantly better text rendering within images. You can now include readable text in generated images — but it works best when you're explicit about what the text should say and how it should appear.
What to specify:
- The exact text in quotation marks
- Font style (serif, sans-serif, handwritten, bold)
- Placement (centered, top-left, overlaid on image)
- Language and context
Example: "Motivational poster with the text 'START BEFORE YOU ARE READY' in bold white sans-serif type, centered on a dark teal gradient background, subtle geometric pattern overlay, 4:5 portrait format."
Tip: Keep text short. One to five words renders reliably. Full sentences or paragraphs still tend to degrade. If you need longer text, consider generating the image separately and adding text in a design tool.
5. Maintain Subject Consistency
When you need to generate multiple images featuring the same character or object — for a storyboard, product series, or social campaign — NB2's subject consistency capabilities can maintain visual identity across outputs.
The key is to define your subject once with highly specific attributes, then repeat those exact attributes in each prompt:
Subject definition: "Young woman with short dark curly hair, olive skin, round tortoiseshell glasses, wearing a rust-colored linen blazer over a white t-shirt."
Scene 1: "[Subject definition], sitting at a sunlit café table reading a newspaper, warm morning light, shot on 35mm film."
Scene 2: "[Subject definition], walking through a rainy city street with a clear umbrella, neon reflections on wet pavement, evening atmosphere."
NB2 can maintain reasonable consistency across up to five characters or fourteen distinct objects when you provide consistent visual descriptors. The more specific your initial definition, the more reliable consistency becomes.
6. Stack Complex Instructions Confidently
Unlike many compact models that fall apart with multi-requirement prompts, NB2 handles layered instructions well. Don't be afraid to combine multiple elements in a single prompt — just organize them logically.
Structure your complex prompts as:
- Subject — What is the main focus?
- Setting — Where is it?
- Lighting — What's the mood and light source?
- Style — What aesthetic or medium?
- Technical — Resolution, crop, camera details?
- Mood — What feeling should it evoke?
Example: "A vintage red bicycle leaning against a stone wall covered in ivy, narrow cobblestone alley in a small Italian village, late afternoon golden hour light casting long shadows, shot on medium format film with warm tones and natural grain, 3:2 landscape format, nostalgic and peaceful atmosphere."
That's six layers of instruction in one prompt, and NB2 handles it coherently. The model's Flash-speed architecture makes it practical to iterate quickly — generate, review, adjust the prompt, and regenerate in seconds.
Prompt Templates by Category
Now let's get specific. Here are ten content categories with detailed templates and ready-to-use examples you can adapt for your own projects.
Cinematic and Film Photography
Cinematic prompts work best when you think like a cinematographer. Specify camera, lens, lighting setup, and film treatment.
Template: "[Subject and action], [location/setting], shot on [camera model] with [lens], [film stock or color grade], [lighting description], [mood/atmosphere], [aspect ratio]."
Example prompts:
- "A detective in a dark trench coat walking down a foggy alley at night, shot on ARRI Alexa with Cooke S4 50mm lens, teal and orange color grade, single overhead sodium street lamp, noir atmosphere, 2.39:1 anamorphic widescreen."
- "Close-up portrait of an elderly fisherman mending nets on a wooden dock, shot on 35mm Kodak Tri-X 400, overcast natural light from the right, weathered hands in focus with soft background bokeh, documentary style, 3:2 format."
- "A woman running through a sunlit wheat field in slow motion, shot on Panavision Millennium DXL2, golden hour backlight creating a lens flare, Fuji Eterna 250D film stock, warm amber tones, 1.85:1 aspect ratio."
Product Photography
Treat your prompt like a creative brief for a product shoot. Define the product, surface, lighting setup, and camera position.
Template: "[Product description], on [surface/background], [camera angle], [lighting setup], [style/mood], [format and intended use]."
Example prompts:
- "Matte black ceramic coffee mug filled with latte art, on a dark walnut table, 45-degree angle overhead shot, soft window light from the left with gentle shadows, minimalist editorial style, square format for e-commerce listing."
- "Luxury perfume bottle with amber liquid, on a reflective black surface, front-facing hero shot, dramatic rim lighting from behind with soft fill light, high-end beauty campaign aesthetic, 4:5 portrait for print ad."
- "Pair of white leather sneakers, floating against a clean gradient background transitioning from light gray to white, studio product shot with even diffused lighting, no harsh shadows, 1:1 square crop."
Infographics and Data Visualization
NB2 can generate visual layouts that resemble infographics. The key is specifying layout logic, hierarchy, and visual structure explicitly.
Template: "[Topic and title], [layout type], [data/content elements], [color scheme], [typography style], [format]."
Example prompts:
- "Infographic titled 'THE WATER CYCLE' with four sequential stages arranged top to bottom: evaporation, condensation, precipitation, collection, each with a simple icon and two-line description, blue and white color palette, clean sans-serif typography, tall portrait format."
- "Horizontal timeline infographic showing the history of space exploration from 1957 to 2025, milestone events as nodes connected by a flowing line, dark navy background with bright accent colors for each decade, futuristic sans-serif font, 16:9 landscape format."
- "Comparison chart showing 'ELECTRIC vs. GASOLINE CARS' with two columns, five rows covering cost, range, emissions, maintenance, and performance, green for electric and gray for gasoline, modern flat design icons, 4:5 portrait format."
Character Design and Consistency
For character design work — game development, storyboarding, children's books — establish a detailed character sheet first.
Template: "[Character description with specific physical traits], [pose and action], [setting], [art style], [color palette], [format]."
Example prompts:
- "Character turnaround sheet of a young robot mechanic: short spiky blue hair, oil-stained orange jumpsuit, oversized goggles pushed up on forehead, mechanical prosthetic left arm, confident stance. Three views: front, three-quarter, side. White background, clean line art with flat color fills, anime-inspired style."
- "A tall elven ranger with long silver braided hair, emerald green cloak, leather armor with gold filigree, carrying a longbow, standing at the edge of a misty forest clearing, fantasy illustration style, muted earth tones with pops of emerald, 2:3 portrait format."
Fashion and Editorial
Fashion prompts benefit from referencing real photographers, publications, and specific aesthetic movements.
Template: "[Subject description and outfit], [setting/backdrop], [photography style reference], [lighting], [mood and editorial direction], [format]."
Example prompts:
- "High fashion editorial: model in an oversized structured blazer with exaggerated shoulders, monochrome black, standing in an empty white modernist gallery space, shot in the style of Peter Lindbergh, dramatic side lighting creating deep shadows, powerful and confident mood, full-length shot, 3:4 portrait."
- "Street style photograph of a young man in layered earth-tone streetwear — oversized camel coat, olive hoodie, wide-leg cargo pants, chunky white sneakers — walking through a Shibuya crossing at dusk, editorial candid style, natural ambient light mixed with neon signs, 4:5 format."
Landscape and Nature
Landscape prompts need location specificity, atmospheric conditions, and time of day to avoid generic results.
Template: "[Specific location or terrain type], [time of day and season], [weather and atmospheric conditions], [foreground/midground/background elements], [photography style], [format]."
Example prompts:
- "Patagonian mountain range reflected in a perfectly still glacial lake at dawn, early autumn with golden larch trees in the foreground, thin layer of morning mist hovering over the water, pink and purple alpenglow on snow-capped peaks, landscape photography, 16:9 panoramic format."
- "Dense Japanese bamboo forest path in Arashiyama, midday with dappled sunlight filtering through the canopy, emerald green bamboo stalks stretching vertically, narrow stone path with moss edges, serene and meditative mood, shot on medium format film, 2:3 portrait."
- "Icelandic black sand beach at twilight, powerful ocean waves crashing against basalt sea stacks, long exposure effect showing smooth water motion, dramatic dark clouds with a thin strip of orange light on the horizon, moody and raw atmosphere, 21:9 ultrawide."
UI/UX Design Mockups
Think like a product designer. Define the device, screen content, visual hierarchy, and design system.
Template: "[Device type and context], [screen content and layout], [design style/system], [color scheme and typography], [format]."
Example prompts:
- "Mobile app UI design for a meditation timer, iPhone 15 Pro mockup, home screen showing a large circular progress timer in the center, session duration selector below, minimal navigation bar at bottom with three icons, dark mode with deep indigo background and soft lavender accents, SF Pro typography, 9:19.5 phone aspect ratio."
- "Dashboard UI design for a project management tool, desktop browser mockup, left sidebar with navigation icons, main content area showing a Kanban board with three columns (To Do, In Progress, Done), each with 3-4 task cards, clean white background with blue primary accent color, Inter font, 16:10 format."
Photo Editing and Transformation
When using NB2 for image editing tasks, clarity about what should change versus what should stay the same is critical.
Template: "[Original image description], [specific transformation requested], [what to preserve], [desired result], [format]."
Example prompts:
- "Daytime photograph of a suburban house with green lawn and blue sky. Transform to a snowy winter scene: add snow covering the roof and lawn, change the sky to overcast gray, add bare tree branches, keep the house architecture and composition exactly the same, maintain photorealistic quality."
- "Portrait photograph of a woman with brown hair against a studio backdrop. Change only the background to a lush tropical garden with soft bokeh, maintain the subject's appearance, lighting on the face, and overall exposure exactly as-is, seamless compositing."
Branding and Packaging
Treat prompts as formal design briefs. Define the brand personality, target audience, and design constraints.
Template: "[Product/brand description], [packaging type], [design elements and layout], [brand personality and target audience], [color palette and materials], [format]."
Example prompts:
- "Premium artisan chocolate bar packaging design for a brand called 'CACAO NOIR,' dark chocolate 85% cacao. Front panel layout: brand name in elegant gold serif typography at top, botanical illustration of cacao pods and leaves in the center, 'Single Origin — Ecuador' text at bottom. Deep matte black background with metallic gold foil accents. Luxury minimalist aesthetic targeting adult connoisseurs. Front-facing flat lay, 2:3 portrait format."
- "Craft beer can design for a hazy IPA called 'SUNSET HAZE.' Vibrant illustrated label wrapping the full can showing an abstract gradient sunset landscape with layered mountain silhouettes. Bold retro typography for the beer name, tropical fruit illustrations (mango, pineapple) incorporated into the scene. Warm orange-to-pink palette. 3D mockup of the can at a slight angle on a white background, 4:5 format."
Architecture and Interior Design
Architectural prompts need material specificity, spatial logic, and clear lighting direction.
Template: "[Space type and style], [key architectural elements and materials], [furniture and décor], [lighting conditions], [perspective and camera angle], [format]."
Example prompts:
- "Modern Japanese minimalist living room, double-height ceiling with exposed light wood beams, floor-to-ceiling glass wall overlooking a zen garden with raked gravel and a single maple tree, low-profile platform sofa in natural linen, polished concrete floor with a single wool area rug, warm late afternoon light casting long geometric shadows across the floor, interior photography with a wide-angle lens at eye level, 16:9 landscape."
- "Industrial loft kitchen renovation concept, exposed red brick walls with original arched windows, matte black steel-framed kitchen island with white marble countertop, open shelving with copper pipe brackets, pendant Edison bulb lighting, herringbone reclaimed wood floor, moody evening atmosphere with warm artificial lighting, architectural visualization style, 3:2 landscape."
Advanced Tips and Common Pitfalls
Do: Iterate Rapidly
NB2's on-device speed is your biggest advantage. Generate, evaluate, adjust one or two elements in your prompt, and regenerate. Treat prompting as a conversation with the model rather than a one-shot attempt. Most great outputs come after two to four refinement cycles.
Do: Use Negative Descriptions Sparingly
If you consistently get unwanted elements, you can add guidance like "no text overlays" or "without people in the frame." But use this sparingly — it's better to describe what you want than what you don't want.
Do: Front-Load Important Details
NB2 weighs the beginning of your prompt more heavily than the end. Put your most important visual elements — subject, style, and lighting — early in the prompt. Save secondary details like format and mood for the end.
Don't: Overload with Contradictions
"A bright, dark, moody, cheerful scene" gives the model conflicting signals. Pick a single coherent direction and commit to it. If you want contrast, describe it concretely: "Brightly lit foreground figure against a dark shadowy background."
Don't: Expect Perfection on Faces and Hands
Like all compact image generation models, NB2 can struggle with anatomical details at high resolution. If faces and hands are central to your image, keep composition simple and consider specifying "detailed hands" or "clear facial features" as explicit requirements.
Don't: Write Paragraphs of Prose
Prompts aren't essays. Use descriptive fragments separated by commas rather than full sentences. "Red sports car, winding mountain road, aerial drone shot, golden hour, cinematic" is more effective than "I would like you to generate an image of a red sports car that is driving on a winding mountain road, and I want it to look like it was taken from a drone during golden hour."
Quick Reference: Prompt Structure Formula
When in doubt, use this formula for any NB2 prompt:
[Subject] + [Setting/Background] + [Lighting] + [Style/Medium] + [Technical Specs] + [Mood]
Example using the formula: "Tabby cat sleeping on a windowsill + cozy apartment interior with bookshelves + soft warm afternoon window light + watercolor illustration style + square 1:1 format + peaceful and quiet mood."
This formula works across categories and gives NB2 enough structure to produce coherent, high-quality results without being overly prescriptive.
What's Next for NB2 Prompting
As Google continues to roll out Nano Banana 2 across more devices and platforms, the prompting ecosystem will evolve. Expect to see community-built prompt libraries, integration with design tools that auto-generate structured prompts, and improvements to the model itself that make it more forgiving of imprecise language.
For now, the investment in learning to prompt NB2 well pays immediate dividends. A few minutes spent crafting a thoughtful prompt saves far more time than cycling through dozens of vague attempts — and the results speak for themselves.
The best way to improve is simple: generate more images. Try the templates in this guide, modify them, break them, and build your own. On-device AI means there's no cost per generation and no rate limit holding you back. Experiment freely, and let the model surprise you.