Banana API - Nano Banana Image Generation

Streamlined client for Nano Banana's Gemini-3-Pro-Image-Preview API. Automatically handles the annoying base64 workflow, image compression, and Discord sending.

What This Skill Solves

The Problem with Gemini API:

• Input/output is only base64 (no URLs)
• Large images = massive base64 strings (2MB+)
• Manual encoding/decoding is painful
• No built-in Discord integration

This Skill Provides:

• ✅ Automatic image compression (512px default)
• ✅ Transparent base64 encoding/decoding
• ✅ Smart filename generation
• ✅ Auto-send to Discord
• ✅ Both text-to-image and image-to-image editing

Quick Start

Text-to-Image

python3 scripts/banana_gen.py "a cute fluffy cat on a window sill"

Image Editing

python3 scripts/banana_gen.py "transform into rock concert scene with leather jacket" \
  --image /path/to/photo.png

With Discord Auto-Send

python3 scripts/banana_gen.py "cyberpunk cityscape at night" \
  --channel-id 1478746465328435412

Configuration

API Key (Choose One)

Option 1: Config File (Recommended - Persistent)

# Interactive setup (stores in ~/.openclaw/workspace/config/banana-api.json)
python3 scripts/banana_gen.py --setup

# Or manually create config file
echo '{"api_key": "sk-your-key-here"}' > ~/.openclaw/workspace/config/banana-api.json

Option 2: Environment Variable

export BANANA_API_KEY="sk-xxxxxxxx"

Option 3: Command Line (One-time)

python3 scripts/banana_gen.py "prompt" --api-key "sk-xxx"

Output Location

Generated images are saved to:

~/.openclaw/workspace/photos/{description}-{timestamp}.png

View Current Config

python3 scripts/banana_gen.py --show-config

Usage Examples

Basic Generation

# Simple prompt
python3 scripts/banana_gen.py "sunset over mountains"

# With aspect ratio hint
python3 scripts/banana_gen.py "portrait of a warrior" --ratio 2:3

# Custom filename tag
python3 scripts/banana_gen.py "cat playing piano" --name cat-piano

Image Editing (Inpainting/Restyle)

# Change setting/outfit
python3 scripts/banana_gen.py "wearing a red dress at the beach" \
  --image ~/photos/portrait.png

# Change background
python3 scripts/banana_gen.py "standing in front of Tokyo skyline at night" \
  --image ~/photos/selfie.png \
  --name tokyo-night

# Artistic transformation
python3 scripts/banana_gen.py "oil painting style, renaissance portrait" \
  --image ~/photos/photo.jpg

Discord Integration

# Auto-send to channel
python3 scripts/banana_gen.py "dragon breathing fire" \
  --channel-id 1478746465328435412

# Auto-send with custom name
python3 scripts/banana_gen.py "cute anime girl with blue hair" \
  --channel-id 1478746465328435412 \
  --name anime-girl

# Save locally only (no Discord)
python3 scripts/banana_gen.py "prompt" --no-send

Advanced Options

# Custom output path
python3 scripts/banana_gen.py "prompt" --output ~/Desktop/my-image.png

# Different model (if available)
python3 scripts/banana_gen.py "prompt" --model gemini-2.5-flash-image

# Full example
python3 scripts/banana_gen.py "wizard casting spell in ancient library" \
  --image ~/photos/me.png \
  --ratio 2:3 \
  --name wizard-me \
  --channel-id 1478746465328435412

Best Practices

Image Editing Tips

1. Compress input automatically: Script resizes to 512px by default
2. Be specific: "wearing black leather jacket" > "cool outfit"
3. Preserve identity: Face features are usually maintained well
4. Aspect ratio: Add to prompt, but actual output depends on model

Prompt Engineering

• Style keywords: "oil painting", "anime style", "photorealistic"
• Lighting: "dramatic lighting", "soft golden hour", "neon lights"
• Quality: "high quality", "detailed", "8k"

File Naming

The script auto-generates filenames:

• Format: banana-{description}-{timestamp}.png
• Description is cleaned from prompt (first 30 chars)
• Use --name for custom suffix: banana-{name}-{timestamp}.png

Technical Details

Image Processing Pipeline

Input Image → Resize to 512px → JPEG compress (85%) → Base64 encode → API
                                              ↓
Output ← Base64 decode ← PNG save ← Response

API Flow

1. Compress input image (if provided)
2. Build Gemini API request with base64 inlineData
3. POST to /v1beta/models/{model}:generateContent
4. Extract image from response.candidates[0].content.parts
5. Decode base64 and save to workspace/photos/
6. Send to Discord (if channel-id provided)

Error Handling

• Input file not found → Clear error message
• API key missing → Prompt to set BANANA_API_KEY
• Image too large → Automatic compression
• API error → JSON error details printed
• Discord send fail → Warning but continues

Limitations

• Output always base64: Cannot be changed (Gemini limitation)
• No URL input: Must download images locally first
• Single image output: Gemini returns one image per request
• Response time: 10-60 seconds depending on complexity

When to Use This vs Other Tools

Troubleshooting

Configuration Priority

API Key is loaded in this order (first found wins):

1. --api-key command line argument
2. BANANA_API_KEY environment variable
3. ~/.openclaw/workspace/config/banana-api.json config file
4. ❌ Error if none found

To persist the API key for future use:

python3 scripts/banana_gen.py --setup

Script Reference

python3 scripts/banana_gen.py [PROMPT] [OPTIONS]

Options:
  --image, -i PATH       Input image for editing
  --ratio, -r RATIO      Aspect ratio hint (e.g., 2:3, 16:9)
  --model, -m MODEL      Model name (default: gemini-3-pro-image-preview)
  --output, -o PATH      Custom output path
  --name, -n NAME        Filename suffix/tag
  --channel-id, -c ID    Discord channel ID to auto-send
  --no-send              Skip Discord sending
  --api-key, -k KEY      API key (or set BANANA_API_KEY)
  --setup                Interactive setup to save API key
  --show-config          Display current configuration

Integration with Workflows

As Part of a Larger Pipeline

# Generate + send + reference
IMAGE=$(python3 scripts/banana_gen.py "cute cat" --name kitty | grep "OUTPUT_PATH:" | cut -d: -f2-)
echo "Generated: $IMAGE"

From Another Skill

Call the script directly:

import subprocess
result = subprocess.run([
    'python3', 'scripts/banana_gen.py',
    'prompt here',
    '--image', input_path,
    '--channel-id', channel_id
], capture_output=True, text=True)