nano-banana

Use this whenever the user asks to create, generate, make, draw, design, visualize, or edit any image/visual.

Nano Banana = Gemini’s native image generation models:

• gemini-2.5-flash-image (fast, default)
• gemini-3-pro-image-preview (pro, supports 1K/2K/4K, best for production)

All generated images include a SynthID watermark.

---

One-time setup

export GEMINI_API_KEY="YOUR_KEY_HERE"

Base endpoint:

• https://generativelanguage.googleapis.com/v1beta/models/<MODEL>:generateContent

---

Output directory rule (MANDATORY)

All decoded images must be saved into:

./nanobanana-output/

Always ensure it exists before decoding:

mkdir -p ./nanobanana-output

---

Decoding images from response.json (REQUIRED)

Gemini returns images as base64 in:

• .candidates[0].content.parts[].inline_data.data

Example part:

{
  "inline_data": {
    "mime_type": "image/png",
    "data": "BASE64..."
  }
}

You must decode into real files inside ./nanobanana-output/.

---

A) Decode the FIRST returned image → nanobanana-output

mkdir -p ./nanobanana-output

jq -r '
  .candidates[0].content.parts[]
  | select(.inline_data != null)
  | .inline_data.data
' response.json | head -n 1 | base64 -d > ./nanobanana-output/output.png

Verify:

file ./nanobanana-output/output.png

---

B) Decode ALL returned images → nanobanana-output

mkdir -p ./nanobanana-output

jq -r '
  .candidates[0].content.parts[]
  | select(.inline_data != null)
  | .inline_data.data
' response.json \
| awk '{print NR ":" $0}' \
| while IFS=":" read -r idx b64; do
    echo "$b64" | base64 -d > "./nanobanana-output/image_${idx}.png"
  done

Outputs:

• ./nanobanana-output/image_1.png
• ./nanobanana-output/image_2.png
• …

---

C) Preserve MIME type (PNG vs JPEG)

Check MIME type:

jq -r '
  .candidates[0].content.parts[]
  | select(.inline_data != null)
  | .inline_data.mime_type
' response.json

If JPEG:

jq -r '
  .candidates[0].content.parts[]
  | select(.inline_data != null)
  | .inline_data.data
' response.json | head -n 1 | base64 -d > ./nanobanana-output/output.jpg

---

D) Save text + image outputs separately

Print any text parts:

jq -r '
  .candidates[0].content.parts[]
  | select(.text != null)
  | .text
' response.json

Decode images as above into ./nanobanana-output/.

---

E) One-liner: generate → decode directly into nanobanana-output

mkdir -p ./nanobanana-output

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "parts": [{"text": "A photorealistic nano banana dessert on a marble table"}]
    }],
    "generationConfig": {
      "responseModalities": ["TEXT","IMAGE"]
    }
  }' \
| tee response.json \
| jq -r '.candidates[0].content.parts[] | select(.inline_data!=null) | .inline_data.data' \
| head -n 1 \
| base64 -d > ./nanobanana-output/output.png

---

Image Generation & Editing

1) Text-to-image

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "parts": [{"text": "Create a nano banana dish in a futuristic restaurant"}]
    }],
    "generationConfig": {
      "responseModalities": ["TEXT","IMAGE"],
      "imageConfig": { "aspectRatio": "1:1" }
    }
  }' > response.json

Decode:

mkdir -p ./nanobanana-output

jq -r '.candidates[0].content.parts[] | select(.inline_data!=null) | .inline_data.data' \
  response.json | base64 -d > ./nanobanana-output/generated.png

---

2) Image editing (text + image → image)

BASE64_IMG="$(base64 -w 0 < cat.jpg)"

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"contents\": [{
      \"parts\": [
        {\"text\": \"Add sunglasses to the cat. Keep lighting realistic.\"},
        {\"inline_data\": {\"mime_type\": \"image/jpeg\", \"data\": \"$BASE64_IMG\"}}
      ]
    }],
    \"generationConfig\": {
      \"responseModalities\": [\"TEXT\",\"IMAGE\"]
    }
  }" > response.json

Decode into output directory:

mkdir -p ./nanobanana-output

jq -r '.candidates[0].content.parts[] | select(.inline_data!=null) | .inline_data.data' \
  response.json | base64 -d > ./nanobanana-output/edited.jpg

---

3) Multi-turn editing

Always replay prior turns and decode intermediate outputs into:

• ./nanobanana-output/step1.png
• ./nanobanana-output/step2.png

---

4) High-res up to 4K (Gemini 3 Pro)

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{
      "role": "user",
      "parts": [{"text": "A premium hero shot of a smart speaker on white background"}]
    }],
    "generationConfig": {
      "responseModalities": ["TEXT","IMAGE"],
      "imageConfig": {
        "aspectRatio": "16:9",
        "imageSize": "4K"
      }
    }
  }' > response.json

Decode:

mkdir -p ./nanobanana-output

jq -r '.candidates[0].content.parts[] | select(.inline_data!=null) | .inline_data.data' \
  response.json | base64 -d > ./nanobanana-output/hero_4k.png

---

Result handoff (after each run)

List newest outputs:

ls -lt ./nanobanana-output | head

---

Troubleshooting decode issues

• Empty output → no inline_data returned; check responseModalities.
• Invalid base64 → ensure jq -r raw output.
• Wrong file type → inspect .inline_data.mime_type.