nano-banana-pro（纳米香蕉专业版）

Nano Banana Pro (Gemini 3 Pro Image)

Generate and edit professional-quality images using Google's state-of-the-art Gemini 3 Pro Image model.

Models

All scripts accept --fast to use Nano Banana 2, or --model <id> for explicit override. Default is always the pro model.

# Use fast model for quick iteration
python scripts/generate_image.py "A sunset" --fast

# Explicit model override
python scripts/generate_image.py "A sunset" --model gemini-3.1-flash-image-preview

API Configuration

Set your Gemini API key as an environment variable:

export GEMINI_API_KEY="your-api-key-here"

Get your API key at: https://aistudio.google.com/apikey

Quick Start

Test API Connection

Before generating images, verify API credentials:

python scripts/test_connection.py --api-key YOUR_KEY

Or set environment variable:

export GEMINI_API_KEY="your-api-key-here"
python scripts/test_connection.py

Generate Image from Text

python scripts/generate_image.py "A futuristic cityscape at sunset, neon lights, cyberpunk style"

With options:

python scripts/generate_image.py "Portrait of a cat" \
  --aspect-ratio 9:16 \
  --temperature 0.7 \
  --output ./my_images \
  --filename cat_portrait

Edit Existing Image

python scripts/edit_image.py "Make the sky more dramatic with storm clouds" input.jpg

With options:

python scripts/edit_image.py "Change car color to red" photo.png \
  --aspect-ratio 16:9 \
  --output ./edited \
  --filename red_car

Core Workflows

1. Text-to-Image Generation

When to use: Creating images from scratch based on text descriptions.

Process:

1. Craft effective prompt using best practices from references/prompting-guide.md
2. Run generation script with appropriate parameters
3. Review output and iterate if needed

Example:

# High-quality product photography
python scripts/generate_image.py \
  "High-end product photography of luxury watch on black marble, dramatic single key light from top-left, reflective surface, macro lens, f/5.6, commercial quality, 4K" \
  --aspect-ratio 1:1 \
  --temperature 0.5

Key considerations:

• Use specific, detailed prompts (see Prompting Best Practices)
• Choose appropriate aspect ratio for use case
• Lower temperature (0.3-0.5) for consistency, higher (0.8-1.0) for creativity
• Iterate with multi-turn editing for complex requirements

2. Image Editing & Refinement

When to use: Modifying existing images with natural language instructions.

Process:

1. Identify specific changes needed
2. Use precise editing instructions with action verbs (Add, Change, Make, Remove, Replace)
3. Run edit script with source image
4. Chain multiple edits for complex transformations

Example - Single Edit:

python scripts/edit_image.py \
  "Add flying birds in the upper right sky" \
  landscape.jpg \
  --output ./edited

Example - Multi-turn refinement:

# Edit 1: Change time of day
python scripts/edit_image.py "Convert to sunset lighting" day_scene.jpg \
  --output ./step1 --filename sunset

# Edit 2: Add elements using output from step 1
python scripts/edit_image.py "Add warm street lamps glowing" ./step1/sunset_image_0_0.jpg \
  --output ./step2 --filename with_lamps

# Edit 3: Fine-tune atmosphere
python scripts/edit_image.py "Make the sky more orange and dramatic" ./step2/with_lamps_image_0_0.jpg \
  --output ./final --filename final

Key considerations:

• Be specific about which elements to change
• Use positional language ("in the background", "on the left")
• Keep original aspect ratio unless intentionally changing
• Each edit builds on previous result

3. Batch Generation

When to use: Generating multiple variations or different images in sequence.

Process:

Create a shell script or Python wrapper:

#!/bin/bash
# batch_generate.sh

PROMPTS=(
  "Modern office workspace, minimalist design"
  "Cozy coffee shop interior, warm lighting"
  "Professional meeting room, corporate aesthetic"
)

for i in "${!PROMPTS[@]}"; do
  python scripts/generate_image.py "${PROMPTS[$i]}" \
    --output ./batch_output \
    --filename "scene_$i" \
    --aspect-ratio 16:9

  echo "Generated image $((i+1))/${#PROMPTS[@]}"
  sleep 2  # Rate limiting
done

4. Reference-Based Generation

When to use: Maintaining style consistency or character consistency across images.

Process:

1. Prepare reference images (up to 14 images)
2. Create prompt referencing style/character to maintain
3. Use Python SDK or API directly for multi-image upload

Example using Python:

from generate_image import NanoBananaProClient
from pathlib import Path
import base64

client = NanoBananaProClient(api_key="YOUR_KEY")

# Read reference images
ref_images = []
for img_path in ["ref1.jpg", "ref2.jpg", "ref3.jpg"]:
    with open(img_path, "rb") as f:
        img_data = base64.b64encode(f.read()).decode()
        ref_images.append({
            "inlineData": {
                "mimeType": "image/jpeg",
                "data": img_data
            }
        })

# Generate with references
# (Requires direct API call with multiple inlineData parts)

Prompting Best Practices

Text Rendering (99%+ accuracy)

Gemini 3 Pro achieves near-perfect text rendering with these rules:

1. Quote exact text — wrap desired text in double quotes within the prompt: render the text "WORLD WAR WATCHER" in green monospace
2. Keep text short — under 10 words per text element for highest accuracy
3. Specify font style — "monospace", "serif", "sans-serif", "handwritten"
4. Specify color — use named colors or hex: in phosphor green (#2dd4bf)
5. Specify position — "on the right side", "centered at the top", "bottom-left corner"
6. Use reference images — generate_with_references.py preserves character likeness while adding text
7. Avoid military/war terms in text prompts — content policy may block. Use "cinematic", "professional", "dark" instead of "military command center"

Verified in production: OG card for World War Watcher commentary pages generated with all three text elements ("WORLD WAR WATCHER", "DISPATCHES", "Kothar wa Khasis") rendered correctly on first attempt using quoted strings + reference image approach.

Essential Structure

Follow this formula for best results:

[Subject] + [Style/Medium] + [Lighting Details] + [Camera/Composition] + [Quality Modifiers]

Example:

"Portrait of elderly craftsman | Documentary photography style | Soft window light from left | 85mm lens, f/2.8, shallow DOF | Professional editorial quality, sharp focus on eyes"

Key Principles

1. Be Specific: Vague prompts → generic results

- ❌ "a sunset" - ✅ "golden hour sunset over mountain range, vibrant orange and purple clouds, silhouetted pine trees in foreground"

1. Use Cinematic Language: Nano Banana Pro responds well to photography terms

- Lens: "24mm wide-angle" | "85mm portrait" | "200mm telephoto" - Lighting: "soft diffused" | "harsh direct" | "backlit with rim lighting" - Camera: "low-angle shot" | "overhead view" | "Dutch angle tilt"

1. Layer Descriptions: Build depth with atmosphere, materials, mood

- "Cozy library, warm amber lighting, leather chairs, rain visible through tall windows, steam rising from tea cup on oak table"

1. Iterate with Multi-Turn: Start simple, refine progressively

- Turn 1: "Modern kitchen, white cabinets" - Turn 2: "Add marble countertops" - Turn 3: "Make the lighting warmer, golden hour through windows"

Cinematic Digital Painting Style

For documentary-style digital paintings with warm golden lighting, painterly brush strokes, and atmospheric depth, see the dedicated style guide:

cinematic-style.md - Complete formula, key elements, examples, and technical notes

Quick reference:

• Warm golden/amber lighting with dramatic rim lighting
• Painterly brush strokes with visible texture
• Soft bokeh depth-of-field for atmospheric background blur
• High contrast (deep blacks + bright highlights)
• Temperature: 0.6 for balanced consistency/creativity

For complete prompting guide, see: references/prompting-guide.md

Script Reference

generate_image.py

Generate images from text prompts.

Parameters:

• prompt (required): Text description of image
• --api-key: API key (or use GEMINI_API_KEY env var)
• --model: Override model ID (default: gemini-3-pro-image-preview)
• --fast: Use Nano Banana 2 (gemini-3.1-flash-image-preview) for faster generation
• --aspect-ratio: 1:1 | 2:3 | 3:2 | 3:4 | 4:3 | 4:5 | 5:4 | 9:16 | 16:9 | 21:9 (default: 16:9)
• --temperature: 0.0-1.0 (default: 0.7)
• --output: Output directory (default:./output)
• --filename: Base filename (default: generated)
• --verbose: Show full API response

Common use cases:

# Quick generation
python scripts/generate_image.py "red sports car"

# High-quality consistent output
python scripts/generate_image.py "professional headshot" \
  --temperature 0.4 --aspect-ratio 3:4

# Creative exploration
python scripts/generate_image.py "abstract art" --temperature 0.9

# Debug mode
python scripts/generate_image.py "test prompt" --verbose

edit_image.py

Edit existing images with natural language instructions.

Parameters:

• prompt (required): Edit instruction
• image (required): Path to input image
• --api-key: API key (or use GEMINI_API_KEY env var)
• --model: Override model ID
• --fast: Use Nano Banana 2 for faster editing
• --aspect-ratio: Output aspect ratio
• --temperature: Creativity level
• --output: Output directory (default:./output)
• --filename: Base filename (default: edited)
• --verbose: Show full API response

Common editing patterns:

# Localized changes
python scripts/edit_image.py "Make the car red" photo.jpg

# Atmospheric changes
python scripts/edit_image.py "Convert to night scene" day.jpg

# Add elements
python scripts/edit_image.py "Add mountains in background" landscape.jpg

# Remove elements
python scripts/edit_image.py "Remove person from left side" group.jpg

# Style transforms
python scripts/edit_image.py "Apply vintage film look" modern.jpg

test_connection.py

Test API connectivity and verify credentials.

Parameters:

• --api-key: API key to test
• --json: Output as JSON for scripting

Usage:

# Interactive test
python scripts/test_connection.py --api-key YOUR_KEY

# Automated testing
python scripts/test_connection.py --json | jq '.success'

generate_with_references.py

Generate images using style reference images (up to 14).

Parameters:

• prompt (required): Text description of image to generate
• references (required): One or more reference image paths (up to 14)
• --api-key: API key (or use GEMINI_API_KEY env var)
• --aspect-ratio: 1:1 | 2:3 | 3:2 | 3:4 | 4:3 | 4:5 | 5:4 | 9:16 | 16:9 | 21:9 (default: 1:1)
• --resolution: 1K | 2K | 4K (default: 2K)
• --output: Output directory (default:./output)
• --filename: Base filename (default: generated)
• --verbose: Show detailed output

Usage:

# Generate avatar matching reference style
python scripts/generate_with_references.py \
  "Portrait of elderly craftsman in same artistic style as references" \
  ref1.png ref2.jpg ref3.png \
  --output ./output \
  --filename craftsman

# Character consistency across images
python scripts/generate_with_references.py \
  "Same character in a coffee shop scene" \
  character_ref1.png character_ref2.png \
  --aspect-ratio 16:9 \
  --resolution 2K

Best practices:

• Include style description in prompt ("same artistic style", "match the lighting")
• Use consistent reference images (same art style, lighting, color palette)
• Reference images can include characters, objects, or style examples
• Up to 14 images: 6 for objects, 5 for character consistency

Advanced Techniques

Multi-Resolution Workflow

Generate quick previews, then upscale finals:

# 1. Quick iteration at 1K (implied by default, fastest)
python scripts/generate_image.py "concept art spaceship" \
  --filename concept_draft

# 2. Review output, refine prompt

# 3. Generate final at higher quality with refined prompt
python scripts/generate_image.py "detailed concept art of sleek sci-fi spaceship, studio lighting, 4K quality, sharp details" \
  --temperature 0.5 \
  --filename concept_final

Temperature Strategies

• 0.3-0.5: Product photography, technical diagrams, consistency required
• 0.7 (default): Balanced - most use cases
• 0.8-0.9: Creative exploration, artistic styles, abstract art

Aspect Ratio Selection

• 1:1 - Social media posts, profile images, icons
• 16:9 - Presentations, YouTube thumbnails, website headers
• 9:16 - Mobile stories, vertical video thumbnails, app screens
• 4:3 / 3:4 - Traditional photography, portrait work
• 4:5 - Instagram posts (use 4:3 as closest)

Troubleshooting

Common Issues

1. API Key Not Working # Verify key format and test python scripts/test_connection.py --api-key YOUR_KEY
2. Rate Limiting

- Free tier: ~100 requests/day, 2-5/minute - Add delays between requests: sleep 2 - Implement exponential backoff (see references/troubleshooting.md)

1. Low Quality Output

- Use more specific prompts - Lower temperature for consistency (0.5) - Add quality modifiers: "4K", "professional", "sharp focus" - Check input image resolution (editing)

1. Model Not Found (404)

- Verify model name: gemini-3-pro-image-preview - Check if geo-blocked (Europe/MENA) - Try VPN or alternative model

1. Content Policy Violation

- Remove references to violence, explicit content - Avoid copyrighted characters - Don't reference real public figures without permission - Rephrase to be less explicit

For detailed troubleshooting, see: references/troubleshooting.md

API Details

Pricing

• 1K/2K images: $0.134 per image
• 4K images: $0.24 per image
• Image inputs (editing): $0.0011 per image

Rate Limits

Free Tier:

• ~100 requests/day
• 2-5 requests/minute
• 1 concurrent request

Paid Tier:

• Unlimited daily (subject to quota)
• 60 requests/minute
• 10 concurrent requests

Supported Formats

Input (editing): JPEG, PNG, WebP Output: JPEG, PNG (specified via API)

For complete API reference, see: references/api-reference.md

Integration Patterns

Python Integration

from scripts.generate_image import NanoBananaProClient

# Initialize client
client = NanoBananaProClient(api_key="YOUR_KEY")

# Generate image
response = client.generate_image(
    prompt="Professional product photo",
    aspect_ratio="1:1",
    temperature=0.6
)

# Save results
from pathlib import Path
saved = client.save_images_from_response(
    response,
    output_dir=Path("./output"),
    base_filename="product"
)

print(f"Saved to: {saved}")

Bash/Shell Workflows

#!/bin/bash
# Simple workflow automation

export GEMINI_API_KEY="your-key"

# Generate multiple variations
for style in "modern" "vintage" "minimal"; do
  python scripts/generate_image.py \
    "$style office interior" \
    --filename "office_${style}" \
    --output ./variations
  sleep 2
done

echo "Generated ${#styles[@]} variations"

Error Handling Pattern

import time

def generate_with_retry(prompt, max_retries=3):
    for attempt in range(max_retries):
        try:
            response = client.generate_image(prompt)
            return response
        except Exception as e:
            if "429" in str(e):  # Rate limited
                wait = 2 ** attempt  # Exponential backoff
                print(f"Rate limited, waiting {wait}s...")
                time.sleep(wait)
            else:
                raise
    raise Exception("Max retries exceeded")

Example Use Cases

Marketing Campaign

# Generate hero image
python scripts/generate_image.py \
  "Dynamic marketing hero image for tech startup, modern professional team collaborating, bright office space, cinematic color grading, 16:9" \
  --temperature 0.6 \
  --output ./campaign \
  --filename hero

# Generate social media variant
python scripts/edit_image.py \
  "Crop to square focusing on center, increase contrast" \
  ./campaign/hero_image_0_0.jpg \
  --aspect-ratio 1:1 \
  --output ./campaign \
  --filename social

Content Creation

# Blog header
python scripts/generate_image.py \
  "Blog header for article about sustainable architecture, modern green building with vertical gardens, clear sky, professional editorial photography, 16:9" \
  --filename blog_header

# Thumbnail for video
python scripts/generate_image.py \
  "YouTube thumbnail style, bold composition about productivity tips, vibrant colors, text space on left third, eye-catching" \
  --aspect-ratio 16:9 \
  --filename thumbnail

Product Mockups

# Product shot
python scripts/generate_image.py \
  "High-end product photography of minimalist desk lamp on marble surface, soft studio lighting from right, black background, commercial quality" \
  --aspect-ratio 1:1 \
  --temperature 0.4 \
  --filename product_main

# Lifestyle context
python scripts/edit_image.py \
  "Place in modern home office setting with laptop and plants" \
  product_main_image_0_0.jpg \
  --filename product_lifestyle

Resources

Documentation Files

• Prompting Guide (references/prompting-guide.md): Best practices, examples, techniques
• API Reference (references/api-reference.md): Complete API documentation, pricing, SDKs
• Troubleshooting (references/troubleshooting.md): Common issues, solutions, error codes
• Example Prompts (assets/example-prompts.md): Ready-to-use prompts by category
• Aldea Avatars (references/aldea-avatars.md): Character templates for Soul Engine personas

External Links

• Google AI Studio: https://aistudio.google.com (get API key, free tier)
• Official Docs: https://ai.google.dev/gemini-api/docs
• Pricing Page: https://ai.google.dev/gemini-api/docs/pricing
• Model Info: https://deepmind.google/models/gemini-image/pro/
• Vertex AI Console: https://console.cloud.google.com/vertex-ai (enterprise)

Getting API Key

1. Visit https://aistudio.google.com
2. Sign in with Google account
3. Click "Get API key" in left sidebar
4. Create new project or select existing
5. Generate API key
6. Store securely (never commit to git)

# Store in environment
export GEMINI_API_KEY="AIzaSy..."

# Or in .env file (add to .gitignore!)
echo "GEMINI_API_KEY=AIzaSy..." >> .env

Tips & Tricks

1. Cache results: Avoid regenerating identical images
2. Use templates: Create prompt templates for repeated use cases
3. Iterate gradually: Start simple, refine with multi-turn editing
4. Batch carefully: Add delays to respect rate limits
5. Monitor costs: 4K images cost 2x more than 2K
6. Test prompts: Use low temperature for consistency testing
7. Learn from examples: Study assets/example-prompts.md for patterns
8. Reference images: Use up to 14 for style consistency
9. Quality modifiers: Always include "professional", "4K", "sharp" for best results
10. Verbose mode: Use --verbose flag when debugging