nano-banana-pro（纳米香蕉专业版）

Nano Banana Pro Image Generation & Editing

Generate and edit images using Google's Gemini 3 Pro model with advanced transparency support.

Prerequisites

1. Dependencies: pip install google-genai Pillow numpy python-dotenv
2. API Key: The script loads from .env automatically. Only ask the user if the script fails with "No API key found".

CLI Usage (REQUIRED)

ALWAYS use the CLI script. Do NOT write Python code or create.py files.

Run scripts/generate.py directly:

# Basic generation
python scripts/generate.py "a cute banana sticker" -o banana.png

# With transparency (for game assets, stickers, icons)
python scripts/generate.py "pixel art sword" -o sword.png --transparent

# Custom size and aspect ratio
python scripts/generate.py "game logo" -o logo.png --size 4K --ratio 16:9

Options:

• -o, --output - Output filename (default: output.png)
• --transparent - Extract true alpha channel using difference matting
• --size - 1K, 2K, or 4K (default: 2K)
• --ratio - Aspect ratio: 1:1, 16:9, 9:16, etc. (default: 1:1)
• --model - Model override (default: gemini-3-pro-image-preview)

Note: The script loads the API key from .env automatically. Do not check for API keys manually or ask the user about them - just run the script and it will error with instructions if the key is missing.

Intent Detection

Analyze user request to determine:

Resolution Selection

Choose resolution based on use case:

Heuristics:

• Sprite sheets, game assets, print materials → 4K
• Social media, blog images, presentations → 2K
• Quick tests, thumbnails, prototypes → 1K

When uncertain, ask user or default to 2K.

Aspect Ratios

Available: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9

Selection guide:

• Square content (icons, avatars, social posts) → 1:1
• Portrait (mobile, vertical video) → 9:16 or 3:4
• Landscape (desktop, presentations) → 16:9 or 3:2
• Cinematic/ultrawide → 21:9

Core Implementation

Basic Generation

from google import genai
from google.genai import types
from PIL import Image
import io

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents="Your descriptive prompt here",
    config=types.GenerateContentConfig(
        response_modalities=['IMAGE'],
        image_config=types.ImageConfig(
            aspect_ratio="1:1",  # or other ratio
            image_size="2K"     # 1K, 2K, or 4K
        ),
    ),
)

# Extract image from response
for part in response.parts:
    if part.inline_data is not None:
        image = Image.open(io.BytesIO(part.inline_data.data))
        image.save("output.png")
        break

Image Editing

# Load existing image
input_image = Image.open("input.png")

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=[
        input_image,
        "Edit instruction: Change the background to sunset colors"
    ],
    config=types.GenerateContentConfig(
        response_modalities=['TEXT', 'IMAGE'],
        image_config=types.ImageConfig(
            aspect_ratio="1:1",
            image_size="2K"
        ),
    ),
)

Multi-Turn Editing

Preserve context across edits using thought signatures:

# First edit
response1 = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=[image, "Add a red hat"],
    config=config,
)

# Continue editing (include previous response)
response2 = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=[
        image,
        "Add a red hat",
        response1,  # Include for context preservation
        "Now make the hat blue instead"
    ],
    config=config,
)

Transparency Extraction

When user needs transparent images, use difference matting. See scripts/transparency.py.

When to use:

• User explicitly asks for transparency
• Game sprites, icons, logos
• Assets that will be composited
• Cutouts and stickers

Process:

1. Generate image on pure white background (#FFFFFF)
2. Edit same image to pure black background (#000000)
3. Calculate alpha from pixel differences
4. Recover original colors

Key insight: Opaque pixels appear identical on both backgrounds (distance ≈ 0), transparent pixels show background color (max distance).

from scripts.transparency import extract_alpha_difference_matting

# After generating white and black background versions
final_image = extract_alpha_difference_matting(img_on_white, img_on_black)
final_image.save("output.png")  # RGBA with true transparency

Prompt Engineering

Fundamental Principle

"Describe the scene, don't just list keywords."

Narrative paragraphs outperform disconnected word lists.

Effective Prompt Structure

[Style/Medium] of [Subject] in [Context/Setting], [Lighting], [Additional details]

Examples:

# Photorealistic
A professional studio photograph of a brass steampunk pocket watch,
shot with a 50mm lens, soft diffused lighting from the left,
shallow depth of field with bokeh background, 4K HDR quality.

# Illustration
A detailed digital illustration of a medieval blacksmith's forge,
isometric perspective, warm orange glow from the furnace,
dieselpunk aesthetic with exposed pipes and riveted metal plates.

# Product mockup
A product photography shot of a ceramic coffee mug on a marble surface,
natural window lighting, minimalist Scandinavian style, clean white background.

Text in Images

For images containing text, use Gemini 3 Pro (not Imagen):

• Keep text to 25 characters or less per element
• Use 2-3 distinct text phrases maximum
• Specify font style generally (bold, elegant, handwritten)
• Indicate size (small, medium, large)

Quality Modifiers

Add these for enhanced output:

• Photography: 4K, HDR, studio photo, professional lighting
• Art: detailed, by a professional, high-quality illustration
• General: high-fidelity, crisp details, polished finish

Error Handling

from google.genai import errors

def generate_with_retry(client, *, model, contents, config, max_attempts=5):
    for attempt in range(1, max_attempts + 1):
        try:
            return client.models.generate_content(
                model=model, contents=contents, config=config
            )
        except errors.APIError as e:
            code = getattr(e, "code", None) or getattr(e, "status", None)
            if code not in (429, 500, 502, 503, 504) or attempt >= max_attempts:
                raise
            delay = min(30, 2 ** (attempt - 1))
            time.sleep(delay)

Model Selection

Default to gemini-3-pro-image-preview for most tasks.

File References

• scripts/generate.py - CLI for image generation (use this instead of writing code)
• scripts/transparency.py - Difference matting implementation
• references/prompts.md - Extended prompt examples by category