nanobanana-image-gen

name	nanobanana-image-gen
description	Use when asked to generate images (PNG/JPEG) using NanoBanana (Google's nano-banana-pro-preview model) via direct Gemini API calls - no CLI needed, just GEMINI_API_KEY

Generating Images with NanoBanana (Gemini API)

Overview

NanoBanana generates raster images (PNG) by calling the Gemini REST API directly with the nano-banana-pro-preview model. No Gemini CLI or extensions needed — just a GEMINI_API_KEY environment variable.

When to Use

• User asks to "generate an image", "create a picture", "make an illustration"
• User explicitly asks for nanobanana
• User needs PNG/JPEG raster images (photos, artwork, illustrations, UI assets)
• For SVG/vector output, use the gemini-image-gen skill instead

Requirements

• GEMINI_API_KEY environment variable must be set
• Get a key from https://aistudio.google.com/apikey if not set
• curl available
• One of: node/bun, python3, or jq + base64 CLI (for decoding)

Platform Support

Works on macOS, Linux, and Windows (Git Bash/WSL). Option 1 (node/bun) is the most portable — node runs identically everywhere. Option 3 (jq+base64) auto-detects the correct decode flag (-d vs -D) across platforms.

Quick Reference

Task	Model	Method
PNG image	nano-banana-pro-preview	Direct Gemini REST API
PNG fallback	gemini-2.0-flash-exp-image-generation	Direct Gemini REST API

Command

Use the first available runtime. All produce identical results.

Option 1: Node.js / Bun (preferred — most portable)

RESPONSE=$(curl -s "https://generativelanguage.googleapis.com/v1beta/models/nano-banana-pro-preview:generateContent?key=${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Generate an image of [DESCRIPTION]"}]}],
    "generationConfig": {"responseModalities": ["TEXT", "IMAGE"]}
  }')

node -e "
const data = JSON.parse(process.argv[1]);
if (data.error) { console.error(JSON.stringify(data.error, null, 2)); process.exit(1); }
for (const part of data.candidates[0].content.parts) {
  if (part.inlineData) {
    require('fs').writeFileSync('/tmp/[FILENAME].png', Buffer.from(part.inlineData.data, 'base64'));
    console.log('Image saved to /tmp/[FILENAME].png');
  } else if (part.text) { console.log(part.text); }
}" "$RESPONSE"

Option 2: Python 3

curl -s "https://generativelanguage.googleapis.com/v1beta/models/nano-banana-pro-preview:generateContent?key=${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Generate an image of [DESCRIPTION]"}]}],
    "generationConfig": {"responseModalities": ["TEXT", "IMAGE"]}
  }' | python3 -c "
import sys, json, base64
data = json.load(sys.stdin)
if 'error' in data:
    print(json.dumps(data, indent=2)[:500]); sys.exit(1)
for part in data['candidates'][0]['content']['parts']:
    if 'inlineData' in part:
        with open('/tmp/[FILENAME].png', 'wb') as f:
            f.write(base64.b64decode(part['inlineData']['data']))
        print('Image saved to /tmp/[FILENAME].png')
    elif 'text' in part:
        print(part['text'])
"

Option 3: Pure shell (jq + base64 CLI)

RESPONSE=$(curl -s "https://generativelanguage.googleapis.com/v1beta/models/nano-banana-pro-preview:generateContent?key=${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Generate an image of [DESCRIPTION]"}]}],
    "generationConfig": {"responseModalities": ["TEXT", "IMAGE"]}
  }')

# Check for errors
if echo "$RESPONSE" | jq -e '.error' >/dev/null 2>&1; then
  echo "$RESPONSE" | jq '.error'; exit 1
fi

# Extract and decode image (base64 -d on Linux/Git Bash, base64 -D on older macOS)
B64_FLAG=$( (base64 -d /dev/null 2>/dev/null && echo "-d") || echo "-D" )
echo "$RESPONSE" | jq -r '.candidates[0].content.parts[] | select(.inlineData) | .inlineData.data' | base64 $B64_FLAG > /tmp/[FILENAME].png
echo "Image saved to /tmp/[FILENAME].png"

# Print any text parts
echo "$RESPONSE" | jq -r '.candidates[0].content.parts[] | select(.text) | .text'

Runtime Detection

Pick the right option automatically:

if command -v node &>/dev/null; then
  # Use Option 1 (node)
elif command -v python3 &>/dev/null; then
  # Use Option 2 (python3)
elif command -v jq &>/dev/null; then
  # Use Option 3 (jq + base64)
else
  echo "Need node, python3, or jq+base64 to decode the response"
  exit 1
fi

Model Fallback Chain

Always try the best model first, then fall back on failure.

Image models (in order): nano-banana-pro-preview → gemini-2.0-flash-exp-image-generation

Fallback procedure:

1. Run with nano-banana-pro-preview
2. If response contains error → retry with gemini-2.0-flash-exp-image-generation
3. If all models fail, tell the user and suggest waiting

Key Rules

• responseModalities must be ["TEXT", "IMAGE"] — using only ["IMAGE"] will fail
• Always save to /tmp/ unless the user specifies a different path
• Pass the user's prompt as-is — don't rewrite or embellish unless they delegate content to you
• After generating, verify the file exists and show it with ![description](/tmp/filename.png)

Prompt Handling

Default: pass the user's prompt as-is. Forward the user's words to the API without rewriting.

Examples:

• "generate an image of a dog" → "Generate an image of a dog"
• "create a sunset over mountains" → "Generate an image of a sunset over mountains"

Only add details when the user explicitly delegates — e.g. "generate some images for this page".

Workflow for Claude

1. Check GEMINI_API_KEY is set
2. Detect available runtime (node/bun > python3 > jq+base64)
3. Build the curl command with user's prompt
4. Try nano-banana-pro-preview first
5. If error, fall back to gemini-2.0-flash-exp-image-generation
6. Verify the output file exists
7. Show the image with ![description](/tmp/filename.png)

Common Issues

Problem	Solution
PERMISSION_DENIED	Check GEMINI_API_KEY is set and valid
responseModalities error	Must use ["TEXT", "IMAGE"], not just ["IMAGE"]
Model not found	Check model name — use nano-banana-pro-preview exactly
Empty response	Model may have content-filtered the prompt — simplify it
Rate limited (429)	Fall back to gemini-2.0-flash-exp-image-generation or wait
No runtime available	Install node, python3, or jq — at least one is needed