Nano Banana 2

Nano Banana 2 is a native image generation and editing model released by Google DeepMind on February 26, 2026. Its official API name is Gemini 3.1 Flash Image, with the stable model code gemini-3.1-flash-image. It accepts text, image, video, and PDF input, returns images and text, and provides a 131,072-token input limit and a 32,768-token output limit for low-latency, high-quality visual production at scale.

The release combines Gemini Flash speed with stronger world knowledge, visual reasoning, and creative control. It supports native output from 0.5K through 4K, Web and Image Search grounding, thinking, conversational multi-turn editing, video-to-image generation, more reliable text rendering, and stronger subject consistency. Every generated image includes a SynthID watermark.

Key Capabilities

• Text-to-Image Generation: Creates photorealistic scenes, illustrations, design assets, and infographics from prompts containing detailed composition, lighting, text, and style requirements.
• Conversational Image Editing: Uses an image and natural-language instructions to modify regions, add or remove elements, replace backgrounds, or revise layouts over multiple turns.
• Multi-Reference Consistency: Mixes up to 14 reference images in one workflow while preserving the resemblance of up to four characters and high-fidelity details for up to ten objects.
• Search-Grounded Generation: Uses both Google Web Search and Image Search for current textual and visual context when rendering specific subjects, factual diagrams, or time-sensitive content.
• Multi-Resolution Output: Generates at 0.5K, 1K, 2K, or 4K across multiple aspect ratios for previews, social formats, and high-resolution deliverables.
• Video and Document Referencing: Interprets video or PDF context to create thumbnails, posters, summary graphics, or new visual treatments.
• Multilingual Text Rendering: Improves legibility, styling, and internationalized text inside posters, menus, cards, diagrams, and localized assets.

Technical Strengths

Feature	Benefit
Native Gemini Multimodality	Understands relationships across text, images, video, and PDFs without requiring a separate visual-analysis pipeline.
Thinking-Based Visual Reasoning	Can reason about composition and details before producing the final image, improving coherence in complex scenes.
Web and Image Search Grounding	Brings current textual evidence and retrieved visual references into generation for more accurate subjects and factual visuals.
Native 0.5K–4K Output	Covers low-latency drafts and high-resolution delivery in one model, reducing model changes between creative stages.
Separated Subject-Fidelity Limits	Explicit capacities for people and objects make reference planning clearer for stories, product sets, and visual series.
SynthID Provenance Marking	Adds an invisible watermark to generated images, supporting identification in platforms and content workflows.

FAQ

What kinds of sequential character work suit Nano Banana 2?

Nano Banana 2 works well for storyboards, comic panels, and campaign series built around a small core cast, rather than an unlimited long-term identity library. Reuse the key references and restate names, wardrobe, camera framing, and locked traits in each turn to reduce drift.

How can Nano Banana 2 produce more accurate text in posters and infographics?

Generate and approve the final copy first, then ask Nano Banana 2 to place that exact copy into the image. For text-heavy assets, specify the language, verbatim wording, hierarchy, typography direction, and whitespace, and manually proofread spelling and data before publication.

Will Nano Banana 2 always return the exact number of images requested?

No, the model may not strictly follow a requested image count. Production systems that require a fixed quantity should count returned assets and use additional requests, retries, and human selection to complete validation.

Should I use Nano Banana 2 or Nano Banana 2 Lite?

Choose Nano Banana 2 when you need 2K or 4K output, Image Search grounding, multi-character consistency, or more demanding visual work. Evaluate Lite for 1K-only, high-frequency generation and simpler edits; the two models have different identifiers, so product nicknames are not sufficient for API configuration.

Which Nano Banana 2 model ID should I use on LinkModel?

Use the identifier returned by LinkModel's current model page or live catalog. LinkModel materials have shown both gemini-3.1-flash-image-preview and gemini-3.1-flash-image: the former remains in a published Nano Banana 2 workflow example, while the latter is Google's current stable code, so verify account availability before deployment rather than treating them as interchangeable.