Nano Banana Image Generation Skill

Generate and edit images using Google Gemini or OpenAI GPT Image.

---

Referencing Pasted Images

When images are pasted into Claude Code, they are cached at ~/.claude/image-cache/. Copy from this cache to tmp and use as reference image.

How to Get Image Path (Recommended)

Important: When an image is pasted, always save it to tmp before using.

1. User pastes an image
2. Check the image-cache path from Additional working directories in environment info
   • Example: /Users/username/.claude/image-cache/0353f5a1-9885-4743-a956-d2d195b445d4
3. Copy the image file from that directory to tmp
4. Pass the tmp path to the -r option

Command to Save Image to tmp

# 1. Check image-cache path (from Additional working directories in environment info)
# Example: /Users/username/.claude/image-cache/0353f5a1-9885-4743-a956-d2d195b445d4

# 2. Check images in cache
ls ~/.claude/image-cache/<session-id>/

# 3. Copy to tmp
cp ~/.claude/image-cache/<session-id>/*.png /tmp/reference_image.png

# 4. Use as reference image
python3 ~/.claude/skills/nanobanana/generate.py "Same style. Object: cat" \
  -r "/tmp/reference_image.png" -o cat.png

Example: Generate Using Pasted Image as Reference

# 1. Copy latest image from image-cache to tmp
cp ~/.claude/image-cache/0353f5a1-9885-4743-a956-d2d195b445d4/3.png /tmp/reference_image.png

# 2. Gemini: Copy reference image style
python3 ~/.claude/skills/nanobanana/generate.py "Same style. Object: cat" \
  -r "/tmp/reference_image.png" -o cat.png

# 3. OpenAI: Edit reference image
python3 ~/.claude/skills/nanobanana/generate_openai.py "Change background to night sky" \
  -r "/tmp/reference_image.png" -o edited.png

Legacy Method (When Path is in Metadata)

If image metadata contains [Image: source: /path/to/image.png], use that path directly.

python3 ~/.claude/skills/nanobanana/generate.py "Same style. Object: cat" \
  -r "/Users/user/Desktop/reference.png" -o cat.png

Note: Always wrap paths containing spaces in double quotes.

---

Pre-generation Confirmation Flow

Important: Before starting image generation, always confirm the following using the AskUserQuestion tool.

Confirmation Items

Common (Always Confirm):

Item	Options	Description
Provider	gemini / openai	gemini=Nano Banana, openai=GPT Image
Model	Gemini: pro/flash, OpenAI: 1.5/1/mini	See details below
Reference Image	Yes / No	Whether there's a source image to copy style from (pasted images can be used)
Background Removal	Vision API / Magenta removal / OpenAI transparent / Not needed	See methods below

Additional for Multiple Image Generation:

Item	Options	Description
Generation Method	Sheet→split / Parallel generation	Sheet→split recommended (efficient)

Provider Characteristics

Provider	Strengths	API Key Environment Variable
Gemini	Japanese prompts, reference image style copying	GEMINI_API_KEY
OpenAI	High quality, native transparent background, multiple simultaneous generation	OPENAI_API_KEY

Model Comparison

Gemini:

Model	ID	Features
Flash	gemini-2.5-flash-image	Fast, cost-efficient
Pro	gemini-3-pro-image-preview	High quality, handles complex instructions

OpenAI:

Model	ID	Features
GPT Image 1.5	gpt-image-1.5	Latest & highest quality (recommended)
GPT Image 1	gpt-image-1	Standard model
GPT Image Mini	gpt-image-1-mini	Lightweight, fast, low cost

Background Removal Method Selection

Method	Best For	Provider
OpenAI Transparent	Use --background transparent for direct transparent generation with OpenAI	OpenAI only
Vision API	Real photos, photo-style, complex backgrounds, gradient backgrounds	Gemini
Magenta Removal	Illustrations, simple shapes, line art, flat design	Gemini

Recommended Settings

Case	Provider	Model	Background Removal
Transparent icons/stickers	OpenAI	1.5	--background transparent
Reference image style copy	Gemini	pro	Vision API
Simple illustrations	Gemini	flash	Magenta removal
High quality illustrations	OpenAI	1.5	None or transparent
Prototype/testing	OpenAI	mini	None

---

Tool List

Tool	Description
generate.py	Gemini image generation
generate_openai.py	OpenAI image generation
remove-bg-magenta.py	Magenta background removal (includes 1px erosion)
remove-bg-vision.py	Vision API background removal
erode.py	Transparent image edge erosion
split_transparent.py	Split transparent image into individual objects

Prerequisites

1. For Gemini: Set environment variable GEMINI_API_KEY
2. For OpenAI: Set environment variable OPENAI_API_KEY
3. Vision API: Requires macOS 14.0 (Sonoma) or later

---

1. generate.py - Gemini Image Generation

python3 ~/.claude/skills/nanobanana/generate.py "prompt" [options]

Options

Option	Description	Default
-o, --output	Output file path	generated_image.png
-a, --aspect-ratio	Aspect ratio (1:1, 16:9, 9:16, 4:3, 3:4)	1:1
-m, --model	Model (flash, pro)	pro
--magenta-bg	Generate with magenta background	None
-r, --reference	Reference image path	None

Examples

# Simple generation
python3 ~/.claude/skills/nanobanana/generate.py "cute cat illustration"

# Copy reference image style
python3 ~/.claude/skills/nanobanana/generate.py "Same exact style as this image. Object: coffee cup. NO text." -r reference.png -o coffee.png

# Generate with magenta background (for later transparency processing)
python3 ~/.claude/skills/nanobanana/generate.py "simple star icon" --magenta-bg -o star.png

---

2. generate_openai.py - OpenAI Image Generation

python3 ~/.claude/skills/nanobanana/generate_openai.py "prompt" [options]

Options

Option	Description	Default
-o, --output	Output file path	generated_image.png
-s, --size	Size (1024x1024, 1536x1024, 1024x1536, auto)	1024x1024
-m, --model	Model (gpt-image-1, gpt-image-1-mini, gpt-image-1.5)	gpt-image-1.5
-q, --quality	Quality (low, medium, high)	medium
-b, --background	Background (transparent, opaque, auto)	auto
-f, --format	Output format (png, jpeg, webp)	png
-r, --reference	Image path to edit	None
-n, --number	Number of images to generate (1-10)	1

Examples

# Simple generation
python3 ~/.claude/skills/nanobanana/generate_openai.py "cute cat illustration"

# Generate with transparent background (no background removal needed)
python3 ~/.claude/skills/nanobanana/generate_openai.py "simple star icon" -b transparent -o star.png

# High quality, landscape
python3 ~/.claude/skills/nanobanana/generate_openai.py "sunset landscape" -s 1536x1024 -q high -o sunset.png

# Generate multiple simultaneously
python3 ~/.claude/skills/nanobanana/generate_openai.py "cute animal icon, single animal" -n 5 -b transparent -o animals.png

# Image editing
python3 ~/.claude/skills/nanobanana/generate_openai.py "change background to night sky" -r input.png -o edited.png

---

3. remove-bg-magenta.py - Magenta Background Removal

Remove magenta/pink background using color-based detection.

python3 ~/.claude/skills/nanobanana/remove-bg-magenta.py input_image [-o output_image]

How it Works

• Makes colors with R>180, G<100, B>100 transparent
• 1px erosion removes pink edge residue

---

4. remove-bg-vision.py - Vision API Background Removal

Automatically detect and remove background using macOS Vision API.

python3 ~/.claude/skills/nanobanana/remove-bg-vision.py input_image [-o output_image]

Features

• Automatic foreground detection
• Best for images maintaining reference image style (including background)
• Requires macOS 14.0 or later

---

5. erode.py - Edge Erosion

Erode transparent image edges by specified pixels.

python3 ~/.claude/skills/nanobanana/erode.py input_image [-o output_image] [-i iterations]

Option	Description	Default
-o, --output	Output image path	Overwrites input
-i, --iterations	Erosion amount (pixels)	1

---

6. split_transparent.py - Transparent Image Splitting

Split transparent PNG into individual objects (for sticker sheets).

python3 ~/.claude/skills/nanobanana/split_transparent.py input_image [output_directory]

How it Works

• Detects boundaries using alpha channel (transparent areas)
• Extracts each connected component
• Numbers from top-left to bottom-right

---

Workflow Examples

Generate Using Pasted Image as Reference (Recommended)

When user pastes an image, copy from image-cache to tmp and use as reference.

# 1. Check image-cache path from Additional working directories in environment info
# Example: /Users/username/.claude/image-cache/0353f5a1-9885-4743-a956-d2d195b445d4

# 2. Copy from image-cache to tmp
cp ~/.claude/image-cache/0353f5a1-9885-4743-a956-d2d195b445d4/*.png /tmp/reference_image.png

# 3. Gemini: Generate different object in same style
python3 ~/.claude/skills/nanobanana/generate.py "Same exact style as this image. Object: coffee cup. NO text." \
  -r "/tmp/reference_image.png" -o coffee.png

# 4. OpenAI: Edit image
python3 ~/.claude/skills/nanobanana/generate_openai.py "change background to outer space" \
  -r "/tmp/reference_image.png" -o space.png

OpenAI: Transparent Icon Generation (Recommended - Simplest)

# Generate transparent PNG in one step
python3 ~/.claude/skills/nanobanana/generate_openai.py "simple star icon" -b transparent -o star.png

OpenAI: Multiple Icon Simultaneous Generation

# Generate 5 transparent PNGs simultaneously
python3 ~/.claude/skills/nanobanana/generate_openai.py "cute animal icon, single animal" -n 5 -b transparent -o animal.png
# → animal_01.png, animal_02.png, ... will be generated

Gemini: Transparent Sticker Generation (Simple Objects)

# 1. Generate with magenta background
python3 ~/.claude/skills/nanobanana/generate.py "simple star icon" --magenta-bg -o star.png

# 2. Remove magenta
python3 ~/.claude/skills/nanobanana/remove-bg-magenta.py star.png

Gemini: Reference Image Style Copy + Transparency

# 1. Generate with reference image style (no magenta to maintain style)
python3 ~/.claude/skills/nanobanana/generate.py "Same exact style as this image. Object: coffee cup. NO text." -r reference.png -o coffee.png

# 2. Remove background with Vision API
python3 ~/.claude/skills/nanobanana/remove-bg-vision.py coffee.png

Gemini: Sticker Sheet Generation → Split

# 1. Generate multiple stickers with magenta background
python3 ~/.claude/skills/nanobanana/generate.py \
  "Multiple separate kawaii stickers with LARGE gaps: coffee cup, donut, cat, star. Arranged in 2x2 grid, well separated." \
  --magenta-bg -o sheet.png

# 2. Make background transparent
python3 ~/.claude/skills/nanobanana/remove-bg-magenta.py sheet.png

# 3. Split into individual files
python3 ~/.claude/skills/nanobanana/split_transparent.py sheet.png ./stickers/

Prompt Tips:

• LARGE gaps between them - Wide spacing
• well separated - No overlapping
• Arranged in XxY grid - Specify grid layout

---

File Structure

~/.claude/skills/nanobanana/
├── SKILL.md               # This document
├── generate.py            # Gemini image generation
├── generate_openai.py     # OpenAI image generation
├── remove-bg-magenta.py   # Magenta background removal (includes 1px erosion)
├── remove-bg-vision.py    # Vision API background removal
├── remove-bg.swift        # Vision API implementation (Swift)
├── erode.py               # Edge erosion (standalone)
└── split_transparent.py   # Transparent image splitting