The Nano Banana models, specifically Nano Banana 2 and Nano Banana Pro, are advanced image generation and editing models built upon the Gemini 3 family of models. They leverage deep reasoning capabilities and real-world knowledge, including real-time information from web search, to produce precise, high-quality visual results with reduced trial and error. Nano Banana 2 is powered by Gemini 3.1 Flash Image, while Nano Banana Pro uses Gemini 3 Pro Image.

Key features of Nano Banana 2 include:

1. More accurate visuals: Achieved through integration with real-time web search information and images.
2. Fast, Pro-level features: Encompasses capabilities such as superior text rendering, multilingual translations, and upscaling functionality to resolutions up to 4K.
3. Precision control: Offers native support for various aspect ratios (e.g., 16:9, 9:16, 2:1) and enhances visual fidelity with vibrant lighting and richer textures.

Technical Specifications:

• Context Windows: Gemini 3.1 Flash Image (Nano Banana 2) supports a maximum of 131,072 input tokens, while Gemini 3 Pro Image (Nano Banana Pro) supports 65,536 input tokens. Both models support a maximum of 32,768 output tokens.
• Resolutions: Both models offer built-in generation capabilities for 1K, 2K, and 4K visuals. Gemini 3.1 Flash Image also includes support for 512px (0.5K).
• Aspect Ratios: Both models support 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. Gemini 3.1 Flash Image Preview further adds 1:4, 4:1, 1:8, and 8:1 aspect ratios.
• Image Inputs: Users can incorporate up to 14 reference object images in a single prompt. Supported MIME types include image/png, image/jpeg, image/webp, image/heic, and image/heif.
• Document Inputs: Text and PDF files are supported, with a maximum file size of 50 MB for API and Cloud Storage imports, or 7 MB for direct uploads via the Google Cloud console.
• Outputs: Both models generate text and images.
• Model Knowledge Base: The knowledge cutoff date for both models is January 2025.
• Live Data: Both models are powered by real-time information from web search.
• Trust & Safety: All generated images incorporate C2PA Content Credentials and a SynthID watermark for provenance and authenticity.

Best Practices for Effective Prompting:

Effective prompting emphasizes specificity, positive framing, and iterative refinement. Prompts should commence with a strong verb indicating the desired primary operation. Key guidelines include:

• Be specific: Provide concrete details regarding subject, lighting, and composition.
• Use positive framing: Describe what is desired, not what is undesired.
• Control the camera: Employ photographic and cinematic terminology.
• Iterate: Refine images through conversational follow-up prompts.

Prompting Frameworks:

The paper outlines five core prompting frameworks:

4. Image Generation:
   • Text-to-Image Generation without References: This method requires narrative descriptions for the scene.
     • Formula: $\text{[Subject] + [Action] + [Location/context] + [Composition] + [Style]}$
     • Example: "A striking fashion model wearing a tailored brown dress, sleek boots, and holding a structured handbag. Posing with a confident, statuesque stance, slightly turned. A seamless, deep cherry red studio backdrop. Medium-full shot, center-framed. Fashion magazine style editorial, shot on medium-format analog film, pronounced grain, high saturation, cinematic lighting effect."
   • Multimodal Generation (with References): This involves combining multiple reference images to guide the output, suitable for character consistency or merging objects into new environments.
     • Formula: $\text{[Reference images] + [Relationship instruction] + [New scenario]}$
     • Example: "Using the attached napkin sketch as the structure and the attached fabric sample as the texture, transform this into a high-fidelity 3D armchair render. Place it in a sun-drenched, minimalist living room."

5. Image Editing: This framework focuses on modifying an existing base image.
   • Conversational Editing (without New References): Allows for semantic masking (inpainting) by defining a text-based mask to modify specific parts while retaining others.
     • Prompting tip: Explicitly state what should remain unchanged.
     • Example: "Remove the man from the photo."
   • Composition and Style Transfer (with New References): Involves incorporating new images to alter an existing one. This can include adding elements from an object image to a base image or recreating the content of an image in a different artistic style.

6. Real-Time Information from Web Search: Gemini Image models can search the web to generate images based on current information.
   • Prompting changes: Instead of fictional descriptions, the model is instructed to retrieve real-world data and visualize it.
   • Formula: $\text{[Source/Search request] + [Analytical task] + [Visual translation]}$
   • Example: "[Search for current weather and date in San Francisco] + [Analytically, use this data to modify the scene (e.g., if raining, make it look grey and rainy)] + [Visualize this in a miniature city-in-a-cup concept embedded within a realistic, modern smartphone UI. prompted on Tuesday 3rd March]"

7. Text Rendering & Localization: Nano Banana 2 and Nano Banana Pro excel at rendering sharp, legible text and support multilingual text generation in over 10 languages.
   • Rules for best typographic results:
     • Use quotes: Enclose desired words (e.g., "Happy Birthday").
     • Choose a font: Describe the style or name (e.g., "bold, white, sans-serif font" or "Century Gothic 12px font").
     • Translate and localize: Specify a target language for text output.
   • Text-first hack: For text generation, it's recommended to first converse with the model to generate text concepts, then request an image incorporating that text.
   • Example: "A high-end, glossy commercial beauty shot of a sleek, minimalist nude-colored face moisturizer jar resting on a warm studio background. The lighting is soft and radiant. Next to the product, render three lines of text with the following exact styling: For the top line, the word 'GLOW' in a flowing, elegant Brush Script font. For the middle line, the text '10% OFF' in a heavy, blocky Impact font. For the bottom line, the text 'Your First Order' in a thin, minimalist Century Gothic font." Then translate the text into Korean and Arabic.

8. Prompting like a Creative Director: This approach involves employing studio-quality controls to elevate image generation.
   • Design your lighting: Specify illumination techniques (e.g., "three-point softbox setup," "Chiaroscuro lighting," "Golden hour backlighting").
   • Choose your camera, lens, and focus: Dictate hardware and photographic terms (e.g., "GoPro," "Fujifilm camera," "low-angle shot with a shallow depth of field (f/1.8)," "wide-angle lens," "macro lens").
   • Define the color grading and film stock: Specify desired emotional tones through stylistic choices (e.g., "1980s color film, slightly grainy," "Cinematic color grading with muted teal tones").
   • Emphasize materiality and texture: Define physical attributes of objects (e.g., "navy blue tweed," "ornate elven plate armor, etched with silver leaf patterns," "minimalist ceramic coffee mug").

Integration with Other Creative Models:

Nano Banana models are designed to work seamlessly with other generative AI tools:

• Nano Banana + Gemini: Gemini 3 can assist in prompt creation and creative direction.
• Nano Banana + Veo: Nano Banana can generate keyframes to direct an animation, with Veo then generating the video between them.
• Nano Banana + Veo + Lyria: Allows for generating visuals, then adding a custom AI soundtrack using Lyria.