Nano Banana 2 Image Search Grounding — AI Image Generation with Search Grounding Explained

TL;DR

Nano Banana 2 introduces Image Search Grounding—a technique that integrates real-time web data into AI image generation for more accurate, up-to-date outputs. By combining tools like google_search with on/off "Thinking" modes, developers can produce images that reflect current factual details unavailable to static diffusion models. The Da Vinci butterfly example demonstrates how Nano Banana 2 refines anatomical accuracy and stylistic detail based on live web content. This article walks through the fundamentals, experimental setup, outcome comparisons, reproducible API examples on the Wisdom Gate platform, practical usage tips, and legal considerations.

Try Image Search Grounding in Studio — Web-informed Images in ~20s Run the Da Vinci butterfly example in Wisdom Gate Studio using the sample prompt below.

What is Image Search Grounding and why it matters for product teams

Image Search Grounding in AI image generation means linking the generation process directly to live web search results. Unlike traditional diffusion models trained on static datasets that may be months or years old, this approach pulls fresh, factual visual or textual content from the web during generation. Why this matters:

• Accuracy & Relevance: Grounding ensures the AI references current information, reducing hallucination and outdated imagery.
• Domain Adaptation: Product teams can generate images tuned to emerging trends or specific real-world users’ interests.
• Explainability: Integration of search results offers more interpretable generation pathways, helping troubleshoot and evaluate outputs.

Nano Banana 2 core features center around this capability, specifically by coupling the google_search tool with a "Thinking" flag that modulates how the model reasons over search results. This layered architecture offers dynamic control and experimentation.

The freshly integrated Gemini 3.1 image model powers Nano Banana 2, bringing stable 2K resolution outputs with web-grounded semantic fidelity. This is a leap over prior diffusion-only methods, unlocking new product possibilities from marketing assets to scientific visuals.

Da Vinci butterfly — canonical example (how the model uses search in practice)

The "Da Vinci butterfly" serves as the flagship demo for Image Search Grounding. The prompt asks Nano Banana 2 to generate a highly detailed anatomical sketch of a Monarch butterfly styled like a Da Vinci notebook page.

This task requires:

• Correct wing venation patterns
• Accurate anatomical labeling in English
• Parchment texture and sketch style

By enabling google_search, the model retrieves up-to-date images and anatomical data on Monarch butterflies directly from the web, including scientific diagrams and natural history references. Without search, the model relies solely on its training distribution, often producing generic or less accurate wing patterns and labels.

With "Thinking" ON, the model deeply processes these retrieved documents, synthesizing annotations and style for a coherent composite image mimicking Da Vinci’s detailed layered sketches.

The typical prompt payload resembles:

{
  "contents": [{
    "parts": [{
      "text": "Da Vinci style anatomical sketch of a dissected Monarch butterfly. Detailed drawings of the head, wings, and legs on textured parchment with notes in English."
    }]
  }],
  "tools": [{"google_search": {}}],
  "generationConfig": {
    "responseModalities": ["TEXT", "IMAGE"],
    "imageConfig": {"aspectRatio": "1:1", "imageSize": "2K"}
  }
}

The Nanos Banana 2 engine uses this fresh context to enhance wing venation, label placement, parchment aging texture, and style fidelity, producing visually and factually richer results.

Experiment design — Thinking ON vs OFF, tools enabled vs disabled

To quantify search grounding impact, four experimental conditions were designed:

• Tools enabled (google_search) — Thinking ON
• Tools enabled — Thinking OFF
• Tools disabled — Thinking ON
• Tools disabled — Thinking OFF

Each variant issued the canonical Da Vinci butterfly prompt with minor config changes toggling these modes. This isolates effects of search data and internal reasoning on output quality.

Key metrics recorded for each run:

• Generated image files named by condition, e.g., butterfly_search_on_thinking_on_seedXYZ.png
• Generation time stamps (median and single-run samples)
• Textual side outputs including extracted labels
• Expert qualitative rating on factual alignment (1–5 scale)
• Visual fidelity rating reflecting style, detail, and clarity
• Noted hallucinations (incorrect anatomy, text errors)

Data collection also included simple quantitative overlap checks where feasible (e.g., detecting canonical wing vein patterns appearing in web search results).

These controlled tests allow product teams to evaluate the tradeoffs of incorporating search grounding and "thinking" logic in their image pipelines.

Results — visual & qualitative analysis (what changed when grounding is used)

The generated images reveal clear distinctions between search-enabled and disabled modes:

• Search ON + Thinking ON yielded images with the most precise Monarch wing venation, labeling the wings and body parts with accurate English terms in styles matching sources found via google_search.
• The parchment texture and sketch layering closely mimic authentic Da Vinci pages, showing a nuanced aging effect absent in other runs.
• Search ON + Thinking OFF produced factual cues but with less coherent composition and fewer artistic refinements.
• Tools disabled runs often misrepresented wing patterns or presented generic butterfly sketches lacking anatomical details or correct style.

Human raters (3+ experts) scored each variant for factual alignment and visual fidelity on a 1-to-5 scale:

These results decisively demonstrate the value of image search grounding—especially with integrated thinking—at reducing hallucinations and increasing factual recall.

Reproducible steps on Wisdom Gate (Studio & API)

To reproduce these experiments on Wisdom Gate’s platform, follow these steps:

1. Studio (UI)

• Visit https://wisdom-gate.juheapi.com/studio/image
• Paste the Da Vinci butterfly prompt.
• Select the Nano Banana 2 model with Image Search Grounding enabled.
• Optionally enable/disable "Thinking" mode as a toggle.
• Submit generation to see results in ~20s.

1. API curl examples

With google_search tool and Thinking ON:

curl -s -X POST \
  "https://wisdom-gate.juheapi.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
  -H "x-goog-api-key: $WISDOM_GATE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "parts": [{
        "text": "Da Vinci style anatomical sketch of a dissected Monarch butterfly. Detailed drawings of the head, wings, and legs on textured parchment with notes in English."
      }]
    }],
    "tools": [{"google_search": {}}],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageConfig": {
        "aspectRatio": "1:1",
        "imageSize": "2K"
      },
      "thinking": true
    }
  }' | jq -r '.candidates[0].content.parts[] | select(.inlineData) | .inlineData.data' | head -1 | base64 --decode > butterfly_search_on_thinking_on.png

Without google_search tool and Thinking OFF:

curl -s -X POST \
  "https://wisdom-gate.juheapi.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
  -H "x-goog-api-key: $WISDOM_GATE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "parts": [{
        "text": "Da Vinci style anatomical sketch of a dissected Monarch butterfly. Detailed drawings of the head, wings, and legs on textured parchment with notes in English."
      }]
    }],
    "tools": [],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageConfig": {
        "aspectRatio": "1:1",
        "imageSize": "2K"
      },
      "thinking": false
    }
  }' | jq -r '.candidates[0].content.parts[] | select(.inlineData) | .inlineData.data' | head -1 | base64 --decode > butterfly_search_off_thinking_off.png

Notes: Adjust the "thinking" flag per platform documentation to toggle reasoning mode. Replace $WISDOM_GATE_KEY with your API key. Output files are saved as PNG images by decoding the base64 content.

Best practices, caveats & productization notes

When integrating Nano Banana 2 Image Search Grounding into product flows:

• Control search scope: Restrict or filter search tools to avoid unwanted content or biased sources.
• Handle latency: Grounded generations take ~20s on Wisdom Gate; balance UX needs versus quality.
• Monitor hallucinations: Though improved, verify outputs especially for critical business uses.
• Respect IP & privacy: Images derived from web data may include protected content. Review legal risks.
• Use reproducible seeds: Lock randomness for consistent testing and evaluation.
• Combine with prompt engineering: Tailor descriptions to maximize grounding benefits.

Carefully evaluate costs—Wisdom Gate’s $0.058/image pricing offers stable quality slightly below standard $0.068 rates while maintaining consistent generation times across resolutions.

Conclusion & Next Steps

Nano Banana 2’s Image Search Grounding capability marks a significant advance in AI image generation by bridging static model knowledge with real-time web data. The Da Vinci butterfly example clearly shows how developers can benefit from grounded, accurate, richly styled visuals that adapt with live information.

Explore the experimental results, try reproducible payloads, and integrate Nano Banana 2 using Wisdom Gate’s streamlined Studio and API offerings for your AI-driven imaging products.

Get an API Key on Wisdom Gate — Run the Da Vinci Butterfly Test Download our runnable curl payloads and reproduce results locally.

FAQ

Q: What is the Thinking mode in Nano Banana 2? A: Thinking mode enables deeper reasoning over retrieved web search results, improving coherence and factual depth in generated images.

Q: Can I use Image Search Grounding outside Wisdom Gate? A: Currently, Nano Banana 2 grounding is integrated on Wisdom Gate. Use their API or Studio to access this feature.

Q: How reliable is the grounding for scientific accuracy? A: Grounding reduces hallucinations by referencing live data but still requires human review for critical or sensitive applications.

Q: What controls exist over the web search sources? A: The google_search tool scope can be configured to limit or filter search results depending on application needs.

Q: Are there usage costs or rate limits? A: Wisdom Gate charges approximately $0.058 per image generation with grounded models and enforces rate limits per subscription plan.

Sources

• https://wisdom-gate.juheapi.com/models/gemini-3-pro-image-preview
• https://wisdom-gate.juheapi.com/models/gemini-3.1-flash-image-preview

All pricing and generation times checked on 2026-03-05. Wisdom Gate lists $0.058 per image and ~20s generation time

Related

• Full Nano Banana 2 guide
• Prompt templates
• API Quick-start on Wisdom Gate

Compliance

Generated images referencing web data may include copyrighted or personal content. Product teams must assess intellectual property and privacy compliance before usage. For images depicting real people or copyrighted material, ensure appropriate rights or use public-domain sources only.