nano-banana-pro

name	nano-banana-pro
description	Generate, edit, or compose images using Nano Banana Pro (Gemini 3 Pro Image) via Google AI Studio / Vertex AI. Use when the user asks to create an image, generate a picture, edit a photo, make a visual, create a mockup, or do anything involving image generation.
argument-hint	["prompt or \"edit <path> <instructions>\" or \"compose <ref1> <ref2> ... <prompt>\""]
allowed-tools	Bash(python *), Read, Write, Glob

Nano Banana Pro - Image Generation Skill

Generate, edit, and compose images using Nano Banana Pro (Gemini 3 Pro Image) through Google AI Studio or Vertex AI.

How to use

Run the nano_banana.py script located in the project root. All commands output PNG files.

Generate an image from a text prompt

python nano_banana.py generate --prompt "A photograph of a rustic wooden cabin nestled among tall pine trees beside a calm mountain lake at golden hour, with warm light reflecting off the water and mist rising from the surface" --output output/cabin.png --aspect-ratio 16:9

Options:

• --prompt / -p (required): Detailed narrative image description
• --output / -o (default: output/generated.png): Output file path
• --resolution / -r: 1K, 2K (default, $0.15), or 4K ($0.24)
• --temperature / -t: 0.0-2.0 (default 1.0). Lower = more predictable, higher = more creative
• --aspect-ratio / -a: Output aspect ratio. Choices: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9

Edit an existing image

python nano_banana.py edit --input output/cabin.png --prompt "Transform the sky into a dramatic aurora borealis with green and purple curtains of light reflecting in the lake" --output output/cabin_aurora.png

Options:

• --input / -i (required): Path to source image
• --prompt / -p (required): Natural language edit instructions
• --output / -o (default: output/edited.png): Output file path
• --temperature / -t: 0.0-2.0 (default 1.0)
• --aspect-ratio / -a: Optional output aspect ratio

Compose from multiple reference images

Use this for character consistency, style transfer, or combining elements from multiple sources. Supports up to 14 reference images.

python nano_banana.py compose --refs character_front.jpg character_side.jpg --prompt "Generate the same character walking through a neon-lit cyberpunk street at night. Preserve their exact appearance, clothing, and features. Rain-slicked pavement reflecting colorful signage." --output output/scene.png --aspect-ratio 16:9

Options:

• --refs / -R (required): Space-separated paths to reference images
• --prompt / -p (required): Composition instructions
• --output / -o (default: output/composed.png): Output file path
• --temperature / -t: 0.0-2.0 (default 1.0)
• --aspect-ratio / -a: Optional output aspect ratio

Handling arguments from /nano-banana-pro invocation

When the user invokes this skill with /nano-banana-pro, parse $ARGUMENTS to determine intent:

• If it starts with "edit": extract the input path and edit instructions, run the edit command
• If it starts with "compose": extract reference paths and prompt, run the compose command
• Otherwise: treat the entire argument as a generation prompt, run the generate command

If no arguments are provided, ask the user what they'd like to generate.

Prompt engineering best practices (from official Gemini docs)

Write narrative descriptions, NOT keyword lists

BAD (keyword list):

cat, cute, fluffy, orange tabby, sunlight, window, cozy, warm colors, soft focus

GOOD (narrative description):

A fluffy orange tabby cat lounging lazily on a sun-drenched windowsill. Warm afternoon light streams through sheer curtains, casting soft shadows across the cat's fur. The scene feels cozy and peaceful, with a slightly shallow depth of field drawing focus to the cat's contented expression.

Use photography/art terminology for photorealistic results

Include terms like: aperture, bokeh, focal length, exposure, ISO, golden hour, rim lighting, volumetric light, depth of field, macro, wide-angle, telephoto, chiaroscuro.

Be specific about text rendering

For any text in images, explicitly describe:

• The exact text content in quotes
• Font style (sans-serif, serif, bold, italic, condensed)
• Size relative to the image
• Color and contrast against background
• Exact placement position

Use aspect ratio flag instead of prompt text

Instead of writing "16:9 aspect ratio" in the prompt, use the --aspect-ratio 16:9 CLI flag. This uses the native ImageConfig API parameter which is more reliable.

Prompt structure for complex scenes

1. Overall scene/format description
2. Main subject with specific details
3. Background/environment
4. Lighting and mood
5. Style and quality descriptors

What the model handles well

• Photorealistic scenes with natural lighting
• Text rendering (when explicitly described)
• Style transfer and artistic interpretations
• Multi-image composition for character consistency
• Product photography and studio setups

What to avoid in prompts

• Keyword spam — write flowing sentences instead
• Contradictory instructions (e.g., "dark and bright")
• Overly vague descriptions ("something cool")
• Words like "print", "poster on wall", "mockup", "mounted", "framed" — model renders a photo OF the image instead of the image itself

Environment

The script auto-loads .env from the project root. It prefers Google AI Studio (API key), falling back to Vertex AI:

Google AI Studio (recommended):

• GOOGLE_API_KEY — API key from aistudio.google.com

Vertex AI (fallback):

• GOOGLE_CLOUD_PROJECT — Google Cloud project ID
• GOOGLE_CLOUD_LOCATION — Region (must be global for this model)
• Authentication via gcloud auth application-default login

Cost awareness

• 1K resolution: Cheapest option for quick tests
• 2K resolution: ~$0.15 per image (use for drafts and iterations)
• 4K resolution: ~$0.24 per image (use for final output)
• Default to 2K unless the user specifically requests 4K or high resolution