AI image generation CLI powered by Gemini 3 Pro. Green screen transparency, reference images, style transfer. Also a Claude Code plugin.
npx skills add https://github.com/kingbootoshi/nano-banana-2-skill --skill nano-bananaInstale esta skill com a CLI e comece a usar o fluxo de trabalho SKILL.md em seu espaço de trabalho.
AI image generation CLI powered by Gemini 3.1 Flash Image Preview (default) with support for Gemini 3 Pro and any Gemini model. Multi-resolution (512-4K), aspect ratios, cost tracking, broadcast-grade green screen transparency, reference images, and style transfer.
Also ships as a Claude Code skill for AI-assisted image generation workflows.
Requirements: Bun, FFmpeg + ImageMagick (for transparent mode)
# Clone the repo
git clone https://github.com/kingbootoshi/nano-banana-2-skill.git ~/tools/nano-banana-2
cd ~/tools/nano-banana-2
# Install dependencies
bun install
# Link globally (no sudo needed - uses Bun's global bin)
bun link
# Set up your API key
mkdir -p ~/.nano-banana
echo "GEMINI_API_KEY=your_key_here" > ~/.nano-banana/.env
Get a Gemini API key at Google AI Studio.
Now you can use nano-banana from anywhere.
When installed as a Claude Code skill, just say /init and Claude will clone the repo, install deps, and link the command for you. Then use it by saying "generate an image of..." and Claude handles the rest.
bun link doesn't work)mkdir -p ~/.local/bin
ln -sf ~/tools/nano-banana-2/src/cli.ts ~/.local/bin/nano-banana
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc
# Basic - generates 1K image to current directory
nano-banana "minimal dashboard UI with dark theme"
# Custom output name
nano-banana "luxury product mockup" -o product
# Higher resolution
nano-banana "detailed landscape painting" -s 2K
# Ultra high res
nano-banana "cinematic widescreen scene" -s 4K -a 16:9
# Lower resolution (fast, cheap)
nano-banana "quick sketch concept" -s 512
# Custom output directory
nano-banana "UI screenshot" -o dashboard -d ~/Pictures
# Default - Nano Banana 2 (Gemini 3.1 Flash, fast and cheap)
nano-banana "your prompt"
# Pro - highest quality, 2x cost
nano-banana "your prompt" --model pro
# Any model ID
nano-banana "your prompt" --model gemini-2.5-flash-image
| Alias | Model | Best For |
|---|---|---|
flash, nb2 |
Gemini 3.1 Flash Image Preview | Speed, cost, high-volume |
pro, nb-pro |
Gemini 3 Pro Image Preview | Highest quality, complex composition |
# Widescreen
nano-banana "cinematic landscape" -a 16:9
# Portrait
nano-banana "mobile app screenshot" -a 9:16
# Ultra-wide
nano-banana "panoramic scene" -a 21:9
# Standard photo
nano-banana "product photo" -a 4:3
Supported: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 4:5, 5:4, 21:9
Edit, transform, or combine existing images:
# Edit an existing image
nano-banana "change the background to pure white" -r dark-ui.png -o light-ui
# Style transfer - multiple references
nano-banana "combine these two UI styles into one" -r style1.png -r style2.png -o combined
# Color correction
nano-banana "make this image more vibrant and increase contrast" -r photo.jpg
Generate assets with transparent backgrounds using AI-powered background removal:
# Basic transparent asset
nano-banana "robot mascot character" -t -o mascot
# Logo with transparency
nano-banana "minimalist tech logo" -t -o logo
# Game asset
nano-banana "pixel art treasure chest" -t -o chest
The -t flag automatically prompts the AI to generate on a green screen, then uses FFmpeg colorkey + despill to key out the background and remove green spill from edge pixels. ImageMagick trims the result. Requires: brew install ffmpeg imagemagick
Control output dimensions by using a blank image as the last reference:
# First -r: your style reference
# Last -r: blank image in target dimensions
nano-banana "pixel art character, 256x256" -r style.png -r blank-256x256.png -o sprite
| Option | Default | Description |
|---|---|---|
-o, --output |
nano-gen-{timestamp} |
Output filename (no extension) |
-s, --size |
1K |
Image size: 512, 1K, 2K, or 4K |
-a, --aspect |
model default | Aspect ratio: 1:1, 16:9, 9:16, etc. |
-m, --model |
flash |
Model: flash/nb2, pro/nb-pro, or any model ID |
-d, --dir |
current directory | Output directory |
-r, --ref |
- | Reference image (can use multiple times) |
-t, --transparent |
- | Generate on green screen, remove background (FFmpeg) |
--api-key |
- | Gemini API key (overrides env/file) |
--costs |
- | Show cost summary from generation history |
-h, --help |
- | Show help |
| Size | Resolution | Flash Cost | Pro Cost |
|---|---|---|---|
512 |
~512x512 | ~$0.045 | N/A (Flash only) |
1K |
~1024x1024 | ~$0.067 | ~$0.134 |
2K |
~2048x2048 | ~$0.101 | ~$0.201 |
4K |
~4096x4096 | ~$0.151 | ~$0.302 |
Every generation logs its cost to ~/.nano-banana/costs.json. View your spending:
nano-banana --costs
Shows total generations, total spend, and per-model breakdown.
The CLI resolves the Gemini API key in priority order:
--api-key flag on the command lineGEMINI_API_KEY environment variable.env file in the current working directory.env file in the repo root (next to src/)~/.nano-banana/.envGet a free key at Google AI Studio.
# Option 1: Environment variable
export GEMINI_API_KEY=your_key_here
# Option 2: .env file in current directory
echo "GEMINI_API_KEY=your_key_here" > .env
# Option 3: Global config
mkdir -p ~/.nano-banana
echo "GEMINI_API_KEY=your_key_here" > ~/.nano-banana/.env
# Option 4: Pass directly
nano-banana "your prompt" --api-key your_key_here
The -t flag uses a 3-step pipeline for pixel-perfect transparency:
colorkey removes the green background. despill reconstructs edge pixel colors by mathematically removing green contamination from the RGB channels - this is why edges are clean instead of having green fringeThe key color is auto-detected from corner pixels (the AI generates near-green like #05F904, not exact #00FF00). Requires FFmpeg and ImageMagick: brew install ffmpeg imagemagick
When installed as a Claude Code plugin, the skill triggers on phrases like:
Claude will construct the appropriate nano-banana command based on your request, handling model selection, resolution, aspect ratio, reference images, transparency, and output configuration automatically.
MIT