name: nano-image-generate description: "Generate images using Nano Banana (Flash) or Nano Banana Pro. Use 'flash' for speed/efficiency and 'pro' for high quality, text rendering, and complex prompt adherence. Triggers include 'generate image', 'create logo', 'fast image', 'high quality image'." metadata: version: "1.1"

Nano Image Generator

Generate images using Nano Banana Pro (Gemini 3 Pro Preview) for any visual asset needs. Supports reference images for style transfer and character consistency.

Quick Start

[!IMPORTANT] Path Check: Before running, localize the scripts/generate_image.py file. If running this skill from another project, use the absolute path to the script or copy it to your project.

# Basic generation (saves to ./generated/)
# Ensure you are running this from the directory containing 'scripts/' OR provide full path
python scripts/generate_image.py "A friendly robot mascot waving"

# With style reference (keep same visual style)
python scripts/generate_image.py "Same style, new content" --ref ./reference.jpg --output ./new.png

Model Selection

Choose the model based on user intent:

| Model | ID | Use Case | Keywords | |-------|----|----------|----------| | Pro | gemini-3-pro-image-preview | Quality & Detail. Default choice. Best for text, complex prompts, logos, and high-fidelity assets. | "high quality", "logo", "text", "detail", "complex" | | Flash | gemini-2.5-flash-image | Speed & Efficiency. Best for rapid prototyping, large batches, or when low latency is critical. | "fast", "quick", "draft", "rapid" |

Script Usage

python scripts/generate_image.py <prompt> --output <path> [options]

Required:

• prompt - Image description

Options:

• --output, -o - Output file path. Optional. Defaults to ./generated/<slug>.png.
• --model, -m - Model selection: pro (default) or flash
• --key, -k - Gemini API Key (optional)
• --aspect, -a - Aspect ratio (default: 1:1)
• --size, -s - Resolution: 1K, 2K (default), 4K
• --ref, -r - Reference image (max 14)

Examples

High Quality Logo (Pro - Default):

python scripts/generate_image.py "Futuristic company logo with text 'NANO'" --output ./logo.png

Quick Prototype (Flash):

python scripts/generate_image.py "Draft sketch of a cat" --model flash --output ./cat_draft.png

Reference Images

Gemini supports up to 14 reference images for:

Style Transfer

Keep the visual style (colors, textures, mood) from a reference:

python scripts/generate_image.py "New scene with mountains, same visual style as reference" \
  --ref ./style-reference.jpg --output ./styled-mountains.png

Character Consistency

Maintain character appearance across multiple images:

python scripts/generate_image.py "Same character now in a forest setting" \
  --ref ./character.png --output ./character-forest.png

Multi-Image Fusion

Combine elements from multiple references:

python scripts/generate_image.py "Combine the style of first image with subject of second" \
  --ref ./style.png --ref ./subject.png --output ./combined.png

Serial Image Generation (Batch Workflow)

For generating a series with consistent style:

1. Generate first image
2. Use first image as --ref for subsequent images
3. Each new image inherits the established style

# Generate cover
python scripts/generate_image.py "Tech knowledge card cover" -o ./01-cover.png

# Generate subsequent cards with style reference
python scripts/generate_image.py "Card 2 content, same style" --ref ./01-cover.png -o ./02-card.png
python scripts/generate_image.py "Card 3 content, same style" --ref ./01-cover.png -o ./03-card.png

Directory Structure

To keep things organized, this skill uses a standard directory structure:

• ./generated/ - Default output location. Images are automatically saved here if no path is provided.
• ./assets/ - Recommended for storing reference images.

Workflow

1. Verify Output Directory & Script Path:

   • Ensure the output directory exists (e.g., mkdir -p generated) or is writable.
   • Confirm the location of generate_image.py.

2. Craft effective prompts - Be specific and descriptive:

   • Include style: "flat design", "3D rendered", "watercolor", "minimalist"
   • Include context: "for a mobile app", "website hero image"
   • Include details: colors, mood, composition
   • For references: mention "same style as reference" or "keep the visual style"

3. Choose appropriate settings:

   • Icons/logos → --aspect 1:1
   • Banners/headers → --aspect 16:9 or 21:9
   • Mobile screens → --aspect 9:16
   • Xiaohongshu cards → --aspect 3:4
   • Photos → --aspect 3:2 or 4:3

Examples

App icon:

python scripts/generate_image.py "Minimalist flat design app icon of a lightning bolt, purple gradient background, modern iOS style" \
  --output ./assets/app-icon.png --aspect 1:1

Marketing banner:

python scripts/generate_image.py "Professional website hero banner for a productivity app, abstract geometric shapes, blue and white color scheme" \
  --output ./public/images/hero-banner.png --aspect 16:9

Xiaohongshu knowledge card:

python scripts/generate_image.py "Tech knowledge card, dark blue purple gradient, neon cyan accents, code block style, Chinese text '标题'" \
  --output ./xiaohongshu/card.png --aspect 3:4

Style transfer:

python scripts/generate_image.py "Transform this photo into watercolor painting style" \
  --ref ./photo.jpg --output ./watercolor.png

Character in new scene:

python scripts/generate_image.py "Same character from reference, now sitting in a cafe, warm lighting" \
  --ref ./character.png --output ./character-cafe.png --aspect 3:2

Prompt Tips

• Be specific - "A red apple on a wooden table" vs "an apple"
• Include style - "in the style of pixel art" or "photorealistic"
• Mention purpose - "for a children's book" affects the output style
• Describe composition - "centered", "rule of thirds", "close-up"
• Specify colors - Explicit color palettes yield better results
• Reference prompts - Use "same style as reference", "keep the visual aesthetic", "match the color palette"
• Avoid - Don't ask for complex text in images (use overlays instead)

Limitations

• Maximum 14 reference images per request
• Text rendering may be imperfect (better to overlay text separately)
• Very specific brand logos may not reproduce exactly