Nano Banana Cheat Sheet: Your Ultimate Guide to Google's AI Image Tools

Google’s Nano Banana showed up suddenly last year, and as they say, the rest is history.

The tech giant's weirdly named creation — Nano Banana — is a family of AI image generation and editing models, built on the Gemini 3 architecture. Think of it this way: Gemini 3 is the brain doing the reasoning, and Nano Banana is the hand holding the paintbrush. 

Rather than functioning as an isolated text-to-image toy, Nano Banana operates as a fully integrated visual reasoning engine. It acts as the visual execution system paired with Gemini’s underlying cognitive brain, translating dense datasets, brand kits, and complex layouts into pixel-perfect deliverables.

There are three models in the current lineup:

Key distinction: Nano Banana 2 is not a downgrade from Pro; it's a different tool built for a different job. Speed and volume vs. polish and precision.

Where to access it

Note on free: Both Nano Banana 2 and Nano Banana Pro are accessible for free through the Gemini app, but Pro usage has a generation cap. Hit the cap, and the app automatically rolls you back to the base model.

Core specs at a glance

Nano Banana 2 (Gemini 3.1 Flash Image)

• Generation speed: 2–5 seconds per image
• Max resolution: 4K (4096×4096), with native 512px, 1K, and 2K options
• Aspect ratios: 15 options including extreme formats — 8:1 and 1:8 (unique to this model)
• Character consistency: Up to 4 characters across a series
• Reference images: Up to 14 object references in a single prompt
• Input token limit: 131,072
• Output token limit: 32,768
• Text rendering accuracy: ~87% (industry-leading)
• Real-time web search: Yes — pulls live data to inform generation
• Cost vs Pro: ~75% cheaper per image

Nano Banana Pro (Gemini 3 Pro Image)

• Generation speed: ~10–15 seconds per image
• Max resolution: Native 4K
• Aspect ratios: Standard set (1:1, 16:9, 9:16, 4:3, 3:4, 21:9, etc.)
• Character consistency: Up to 5 characters
• Reference images: Up to 14 object references
• Input token limit: 65,536
• Output token limit: 32,768
• Text rendering accuracy: ~64%
• Real-time web search: Yes
• Style locking: More stable across long prompt series

Both models share: C2PA Content Credentials, SynthID invisible watermarking, multilingual text generation (10+ languages), and a knowledge cutoff of January 2025, supplemented by live search.

Five prompting frameworks to get the best output from Nano Banana

Text-to-image (no reference)

You're the director of a scene that doesn't exist yet. Keywords alone won't cut it; describe the scene like you're briefing a photographer.

Formula: Subject + Action + Location/Context + Composition + Style

Example prompt:

"A tired-looking software engineer in her late 30s, dark circles under her eyes, sitting at a cluttered desk surrounded by empty coffee cups. She's staring at a monitor with a faint green glow. Low-angle medium shot. Cinematic color grade, muted blue-green tones, documentary-style lighting."

Multimodal generation (with reference images)

Feed the model existing images to guide its output, useful for brand consistency, product placement, or character carry-over.

Formula: Reference images + Relationship instruction + New scenario

Example prompt:

"Using the attached product photo as the object and the attached mood board as the style reference, place the product in a sun-lit coastal café setting. Keep the product proportions exact. Lifestyle shot, editorial quality."

Image editing (conversational)

You already have a base image. Your job is to be precise about what changes and what stays locked in place.

The five core editing verbs:

Pro tip: Always tell the model what to keep and what to change. Adding "keep the subject's face and clothing exactly as they are" dramatically reduces unwanted drift in the output.

Real-time data visualization

Nano Banana 2 can pull live information from the web and visualize it, a capability no other major image model currently matches.

Formula: Search/source request + Analytical task + Visual format

Example prompt:

"Search for today's air quality index in London. Represent the data as a clean illustrated dashboard in a smartphone UI mockup. Use a simple icon system — green for good, amber for moderate, red for poor. Include the borough name and a timestamp."

Verify outputs: Real-time data features are promising but not bulletproof. Dates and statistics are known to pull out stale information. Always cross-check data-driven visuals against a reliable source before publishing.

Prompt like a creative director

Move beyond description; give the model the same kind of brief you'd give a shoot photographer.

Lighting options to specify:

• Soft fill: "Three-point softbox setup, even illumination, no harsh shadows"
• Drama: "Chiaroscuro lighting, single hard source from camera left"
• Natural warmth: "Golden hour, backlit, long shadows across the ground"
• Product clean: "Flat lay, overhead, diffused white studio light"

Camera and lens language:

• "Shot on Fujifilm X100V, natural color science"
• "Wide-angle lens, f/2.8, shallow depth of field, subject sharp, background soft"
• "Macro lens, extreme close-up, surface texture visible"
• "Aerial view, drone perspective, 200mm equivalent focal length"

Color grading shortcuts:

• Nostalgic: "Kodak Ektar film stock, slight color fade, warm highlights"
• Moody cinema: "Teal and orange color grade, crushed blacks"
• Clean commercial: "Neutral color temperature, high clarity, no grain"

Material and texture prompting: Don't say "jacket" — say "oversized vintage denim jacket, pre-washed indigo, stress marks along the seams." The more tactile your language, the richer the output.

Text rendering cheat codes

Nano Banana 2's text accuracy is currently one of the best of any AI image model. To maximize it:

• Always use quotation marks around the exact text you want rendered: "SUMMER SALE"
• Name the font or describe it: "Bold condensed sans-serif, similar to Impact" or "Flowing brush script, similar to Pacifico"
• Specify color and size relationship: "Large white headline text, smaller grey subheading below"
• Use the text-first trick: In a conversational session, ask Gemini to generate the text copy first, then ask for the image containing that copy. The model handles pre-approved text more accurately than text it has to invent.
• Localization: Specify the target language directly — "Translate this text into Arabic and render it right-to-left in the poster layout."
• Don't rely on it for long-form body copy: Headlines, labels, short slogans; solid. For full paragraphs of article text, expect waviness and occasional errors at the character level.

Aspect ratio quick reference

Nano Banana 2 vs Pro: When to use which

Choose Nano Banana 2 when:

• You're iterating fast and need to see options quickly
• The brief is for social media, web graphics, or marketing mockups
• You need readable text in the image (its text accuracy is actually higher than Pro)
• Cost-per-image matters (it's 75% cheaper)
• You need extreme aspect ratios (8:1 or 1:8)
• You're building or testing an image pipeline at volume

Choose Nano Banana Pro when:

• The final output is going to print or a large-format display
• You need maximum photorealism in complex multi-subject scenes
• Brand consistency across a large batch of images is critical
• You're doing high-end product photography for a premium campaign
• The prompt is long, layered, and highly specific; Pro handles complex briefs more faithfully

Common failures and how to fix them

Watermarking and AI detection

Every image generated by Nano Banana carries two layers of attribution:

SynthID: An invisible digital watermark embedded at the pixel level, undetectable by the human eye but readable by detection tools. It survives basic watermark-removal software. Google's SynthID verification feature in the Gemini app has been used more than 20 million times since launch.

C2PA Content Credentials: A metadata standard that logs how an image was created, including AI involvement. Think of it as a provenance receipt embedded in the file. Verification is rolling out to the Gemini app.

What this means for journalists and publishers: AI-generated images from Nano Banana are technically identifiable if the right tools are used, but the watermarks are invisible during casual social media browsing. User awareness remains the weakest link in the chain.

Quick reference prompt starters

Copy, adapt, and use these as starting points:

• Product mockup: "Studio product shot, [product description], placed on a [surface texture], soft diffused lighting from the left, shot on a mirrorless camera with a 50mm lens, clean white background, commercial quality"
• Social media graphic with text: "Square 1:1 format, bold graphic poster, deep navy background, large centered text reading "[YOUR HEADLINE]" in a heavy condensed sans-serif font in white, minimalist flat design style"
• Infographic slide: "16:9 format, clean corporate infographic showing [your data points], flat icon style, blue and white color palette, bold axis labels, no background clutter"
• Character-consistent series: "[Character description]. [Scene description]. Maintain the character's exact appearance — face, hair, clothing — across all images in this series."
• Photo restoration: "Restore and colorize this black-and-white photograph. Preserve the original composition and subject details exactly. Add natural, era-appropriate color. Increase clarity and reduce grain."
• Localized marketing asset: "Generate a promotional banner for [product/event]. Include the text [your copy] rendered in [language]. Cultural context: [market]. Format: 16:9."

Also read: Our list of the best Nano Banana prompts shows how to create micro-world scenes, product mockups, cinematic edits, and more.