Nano Banana Pro

Nano Banana Pro

Generate, edit, and remix images using Google's Gemini 3 Pro Image Preview model.

Features

• Generate images from text prompts with precise control over composition
• Edit existing images with natural language instructions
• Remix webpages into brand-aligned visual assets
• Optimize images for GitHub, Slack, web with smart presets
• Style templates for common use cases (UI, marketing, artistic)

Installation

Requirements

• Python 3.10 or later
• uv package manager
• Google AI API key with Gemini 3 Pro access

Setup

1. Install the plugin in Claude Code
2. Configure your API key (see Configuration section)

Configuration

Run /configure to set up your configuration interactively, or create a settings file manually at .claude/nano-banana-pro.local.md:

---
gemini_api_key: "your-api-key-here"
default_model: "pro"
default_aspect: "16:9"
default_size: "2K"
output_dir: "./.nanobanana/out"
auto_optimize: "true"
optimize_preset: "github"
max_remix_images: "2"
agent_retriever: "true"
agent_planner: "true"
agent_stylist: "true"
agent_visualizer: "true"
agent_critic: "true"
critic_max_rounds: "3"
---

# Nano Banana Pro Settings

Your local configuration for image generation.

An example template is included at config/nano-banana-pro.example.md.

Settings Reference

Setting	Values	Default	Description
gemini_api_key	string	—	Google AI API key
default_model	pro, flash	pro	Gemini model selection
default_aspect	1:1, 16:9, 4:3, 9:16, 3:2	1:1	Default aspect ratio
default_size	1K, 2K, 4K, ""	""	Default size tier
output_dir	path	./.nanobanana/out	Output directory
auto_optimize	true, false	true	Auto-suggest optimization for large images
optimize_preset	github, slack, web, thumbnail	github	Default optimization preset
max_remix_images	integer	2	Max reference images for remix mode
agent_retriever	true, false	true	Enable Retriever agent
agent_planner	true, false	true	Enable Planner agent
agent_stylist	true, false	true	Enable Stylist agent
agent_visualizer	true, false	true	Enable Visualizer agent
agent_critic	true, false	true	Enable Critic agent
critic_max_rounds	1, 2, 3	3	Max critic refinement iterations

Alternative: Environment Variables

Set GEMINI_API_KEY in your environment or in a .env file:

export GEMINI_API_KEY="your-api-key-here"

Or create a .env file:

GEMINI_API_KEY="your-api-key-here"

Usage

Commands

Command	Description
/generate-image	Generate an image from a text prompt
/edit-image	Edit an existing image with instructions
/remix-url	Create an image styled from a webpage
/optimize-image	Reduce image size for GitHub, Slack, web
/configure	Set up or update plugin configuration

Configure

/configure              # Full setup wizard
/configure api-key      # Just configure API key
/configure model        # Just configure model selection
/configure agents       # Enable/disable PaperBanana agents

Generate Image

/generate-image "A hero banner with bold headline 'Ship Faster' on blue gradient background" --aspect 16:9 --size 2K

Edit Image

/edit-image ./banner.png "Increase headline size by 15%, add subtle drop shadow"

Remix URL

/remix-url https://stripe.com "Create a hero banner for payments feature with headline 'Accept Anywhere'" --aspect 16:9

Optimize Image

/optimize-image ./banner.png --preset github

Optimization Presets

Preset	Max Size	Max Width	Use Case
github	500KB	1280px	README images, PR screenshots
slack	128KB	800px	Slack/Discord messages
web	200KB	1200px	Blog posts, documentation
thumbnail	50KB	400px	Previews, icons

Custom options: --max-size 300KB --width 1000

Style Templates

The plugin includes pre-built prompt templates for common use cases:

UI/Web Styles

• Hero banners (16:9, 1920x1080)
• Social media cards (1:1, 1200x1200)
• App screenshots (9:16, 1080x1920)
• Landing page sections (4:3)
• Icons (1:1, 512x512)

Marketing Styles

• Product photography
• Ad creatives
• Logo variations
• Brand assets

Artistic Styles

• Flat design illustrations
• Abstract art
• Photo-realistic
• Minimalist
• Retro/vintage

See skills/generate/references/style-templates.md for detailed templates.

Aspect Ratios

Ratio	Best For
1:1	Social avatars, icons, thumbnails
16:9	Hero banners, presentations, YouTube
4:3	Product shots, traditional photos
9:16	Mobile stories, vertical banners
3:2	Photography, print media

Size Options

Size	Resolution	Use Case
1K	~1024px	Quick previews, drafts
2K	~2048px	Web-ready, social media
4K	~4096px	Print, high-resolution

Model Selection

Choose between two Gemini models:

Model	ID	Strengths	Best For
pro	Gemini 3 Pro	Advanced reasoning, high-fidelity text rendering	Final assets, detailed compositions
flash	Gemini 2.5 Flash Image	Speed and efficiency, low latency	High-volume tasks, quick drafts, iteration

Set your default model in configuration (default_model) or override per-command with --model:

/generate-image "A hero banner" --model flash

PaperBanana Agents

Nano Banana Pro includes a 5-agent pipeline inspired by Google's PaperBanana framework for producing publication-ready visuals.

Pipeline Architecture

Phase 1 — Planning (sequential):

1. Retriever — Scans your project for brand assets, colors, fonts, and reference images
2. Planner — Transforms your request into a detailed visual specification (layout, components, hierarchy)
3. Stylist — Applies aesthetic guidelines: exact colors, typography, mood, and design principles

Phase 2 — Refinement (iterative loop, up to 3 rounds):

1. Visualizer — Executes image generation using nanobanana.py (the only agent that creates files)
2. Critic — Evaluates the output on faithfulness, conciseness, readability, and aesthetics; recommends APPROVE or REVISE

Agent Configuration

Each agent can be enabled/disabled individually in your settings file:

agent_retriever: "true"
agent_planner: "true"
agent_stylist: "true"
agent_visualizer: "true"
agent_critic: "true"
critic_max_rounds: "3"

When an agent is disabled, the pipeline skips that phase. The Visualizer is always required for image generation.

How It Works

The agents are orchestrated via Claude Code's Task tool. When you request an image:

1. The Retriever searches your codebase for brand context
2. The Planner creates a visual specification from your request + context
3. The Stylist refines the spec into a polished generation prompt
4. The Visualizer generates the image
5. The Critic evaluates the result and may request up to 3 revision rounds

Best Practices

1. Prompt like a designer - Include layout, margins, typography, and exact text
2. Lock down text - Say "Render this text verbatim" for exact copy
3. Iterate surgically - Use edit mode for specific refinements
4. Use templates - Reference style templates for consistent results

Output

Generated images are saved to ./.nanobanana/out/ by default with timestamp-based filenames. Use --out to specify a custom path.

Troubleshooting

Issue	Solution
401/403 error	Check API key configuration
No image returned	Prompt may have been blocked; try rephrasing
Slow generation	Reduce size from 4K to 2K
Permission denied	Ensure uv and python3 are in PATH

License

MIT License - see LICENSE for details.

Author

Sekou Doumbouya (@fakoli)