nano-banana-skills（纳米香蕉技能）

Gemini Image Generation

Generate images from text descriptions or modify existing images using Google's Gemini models via the Wisdom Gate API.

Quick Usage

Text-to-Image

python scripts/generate_image.py "your prompt here" [--aspect-ratio RATIO] [--size SIZE] [--output PATH]

Image-to-Image (Single or Multiple)

# Single image
python scripts/generate_image.py "modification prompt" --input input.jpg [--output PATH]

# Multiple images (up to 14)
python scripts/generate_image.py "combine these people in an office photo" --input person1.jpg person2.jpg person3.jpg

Multi-Turn Refinement

# First turn: Generate initial image (auto-selects Nano Banana 2)
python scripts/refine_image.py "Create a vibrant infographic about photosynthesis" --reset

# Second turn: Refine with quality priority
python scripts/refine_image.py "Make it more colorful and add more visual elements" --quality

# Third turn: Further refinement with specific model
python scripts/refine_image.py "Add labels to each component" --model nano-banana-pro

# Start a new conversation with budget model
python scripts/refine_image.py "New prompt here" --reset --model nano-banana

Parameters (generate_image.py):

• prompt (required): Text description of the image to generate or modification to apply
• --input: Input image path(s) for image-to-image generation - supports up to 14 images (optional)
• --aspect-ratio: Image aspect ratio (text-to-image only) - 1:1 (default), 16:9, 9:16, 21:9, 4:3, 3:4, 5:4, 4:5, 2:3, 3:2, 1:4, 4:1, 1:8, 8:1
• --size: Image resolution (text-to-image only) - 0.5K, 1K (default), 2K, 4K
• --output: Output file path (default: generated_image.png)
• --model: Force specific model - nano-banana, nano-banana-2, nano-banana-pro (auto-select if not specified)
• --quality: Prioritize quality over cost (uses Nano Banana Pro when possible)

Parameters (refine_image.py):

• prompt (required): Refinement instruction or initial prompt
• --history: Conversation history file (default: conversation.json)
• --output: Output file path (default: refined_image.png)
• --reset: Reset conversation history and start fresh
• --model: Force specific model - nano-banana, nano-banana-2, nano-banana-pro (auto-select if not specified)
• --quality: Prioritize quality over cost (uses Nano Banana Pro)

Environment:

• Requires WISGATE_KEY environment variable
• Alternative: Set Authorization: Bearer YOUR_KEY header (modify script if needed)

Examples

Text-to-Image

# Basic generation (auto-selects Nano Banana 2 for best balance)
python scripts/generate_image.py "A serene mountain landscape at sunset"

# High quality mode (uses Nano Banana Pro)
python scripts/generate_image.py "Futuristic city skyline" --quality --aspect-ratio 16:9 --size 4K

# Budget mode (force cheapest model)
python scripts/generate_image.py "Simple illustration" --model nano-banana --size 2K

# Portrait orientation with specific model
python scripts/generate_image.py "Portrait of a wise old wizard" --aspect-ratio 9:16 --model nano-banana-pro

Image-to-Image

# Modify an existing image (auto-selects appropriate model)
python scripts/generate_image.py "make it look like a watercolor painting" --input photo.jpg

# Style transfer with quality priority
python scripts/generate_image.py "Van Gogh style" --input portrait.png --quality

# Multiple reference images (auto-uses Nano Banana Pro for best quality)
python scripts/generate_image.py "group photo of these people at a party" --input person1.jpg person2.jpg person3.jpg person4.jpg

# Budget multi-image (force cheaper model)
python scripts/generate_image.py "combine these items" --input item1.jpg item2.jpg --model nano-banana-2

Models

Smart Model Selection (Default Behavior):

• Multiple image input (>1 image) → Automatically uses Nano Banana Pro (best quality)
• 4K resolution → Nano Banana 2 or Pro (depends on --quality flag)
• 0.5K resolution → Nano Banana 2 (only supported model)
• Other scenarios → Nano Banana 2 (best value)

Manual Model Override: Use the --model parameter to force a specific model:

python scripts/generate_image.py "prompt" --model nano-banana  # Cheapest
python scripts/generate_image.py "prompt" --model nano-banana-2  # Best value
python scripts/generate_image.py "prompt" --model nano-banana-pro  # Best quality

Quality Priority Mode: Use the --quality flag to prefer Nano Banana Pro when possible:

python scripts/generate_image.py "prompt" --quality

API Endpoint Format:

https://api.wisgate.ai/v1beta/models/{model}:generateContent

Authentication:

• Header: x-goog-api-key: YOUR_WISGATE_KEY
• Or: Authorization: Bearer YOUR_WISGATE_KEY

Workflow

One-Shot Generation (generate_image.py)

1. Check if WISDOM_GATE_KEY is set in environment
2. For text-to-image: Run script with prompt and desired parameters
3. For image-to-image: Run script with prompt and --input pointing to the source image(s)
4. Save the generated image to the specified output path
5. Show the image to the user or confirm the save location

Multi-Turn Refinement (refine_image.py)

1. First turn: Use --reset to start a new conversation
2. Subsequent turns: Run without --reset to refine based on previous results
3. Conversation history is automatically saved to conversation.json (or custom path)
4. Each turn generates a new image based on the full conversation context
5. Use --reset again to start a completely new conversation

Notes

• Images are returned as base64-encoded PNG data
• Text-to-image includes Google Search grounding by default for better results
• Image-to-image supports JPEG, PNG, and WebP input formats
• Aspect ratio and size parameters only apply to text-to-image generation
• Multi-image limits: Up to 14 reference images total (recommended: up to 6 objects or up to 5 people for best quality)
• Force image output: Set responseModalities: ["IMAGE"] (without TEXT) to ensure image generation
• Model differences: Different models support different resolutions and aspect ratios - consult the table above
• Multi-turn editing: Use refine_image.py for iterative refinement with conversation history
• Conversation persistence: History is saved to JSON file and can be resumed across sessions
• Safety settings: API supports content filtering via safetySettings parameter (not exposed in scripts by default)
• Token usage: Check usageMetadata in API response for prompt/candidate/total token counts
• Finish reasons: API returns finishReason (STOP, MAX_TOKENS, SAFETY, RECITATION, OTHER)