pi-banana

Generate, edit, and analyze images directly inside pi using Google's Nano Banana 2 (gemini-3.1-flash-image-preview), Nano Banana Pro (gemini-3-pro-image-preview), and Gemini Vision.

• Image Generation & Editing — Create images from scratch or pass referenceImages: ["./logo.png"] to edit existing pictures.
• Multimodal Vision — Analyze, describe, or extract text from images using gemini-3.1-flash-lite (fast) or gemini-3.1-pro-preview (deep reasoning).
• Inline preview in Kitty / iTerm2 / WezTerm — generated images show up right under the tool call.
• Auto-save to ./generated/ so the agent and you both have a real file to refer back to.
• One env var — GOOGLE_API_KEY. Works with both AI Studio (AIza…) and Vertex AI Express (AQ.…) keys.

Install

pi install pi-banana

If pi doesn't resolve the bare name, use the explicit form: pi install npm:pi-banana.

Or for a single-session try:

pi -e pi-banana   # or: pi -e npm:pi-banana

Then export your key:

export GOOGLE_API_KEY="AIza…"   # AI Studio key, or
export GOOGLE_API_KEY="AQ.…"    # Vertex AI Express key

Get one at aistudio.google.com/apikey.

Usage

Just ask pi for what you want:

"Generate a 16:9 wallpaper of a foggy redwood forest at dawn"

"Edit ./generated/logo-20260508.png — make the background transparent"

"What color is the car in ./assets/photo.jpg?"

"Extract the text from these three receipts."

The model calls banana_image or banana_vision automatically.

Tools

1. banana_image

Renamed from generate_image in v2.0.1 to avoid colliding with @benvargas/pi-antigravity-image-gen. Both extensions can now coexist in the same pi install.

Param	Type	Default	Description
prompt	string	—	Required. What to draw, or what to change about the reference image.
aspectRatio	enum	1:1	1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9
imageSize	enum	1K	1K, 2K, 4K. 4K requires quality=high.
quality	enum	fast	fast = Nano Banana 2 (~3–10 s, cheapest). high = Nano Banana Pro (slower, top quality).
referenceImages	array	—	Optional array of paths (PNG/JPEG/WebP/GIF) to edit or use for composition instead of generating from scratch.
outputPath	string	./generated/<slug>-<ts>.png	Optional output path. Parent dirs are created.

2. banana_vision

Param	Type	Default	Description
prompt	string	—	Required. What you want to know about the image(s) (e.g. 'Describe this image', 'Extract the text').
imagePaths	array	—	Required. Array of paths to existing images to analyze.
quality	enum	fast	fast = gemini-3.1-flash-lite (fast/cheap). high = gemini-3.1-pro-preview (slower, deep reasoning).

Configuration

Two env vars adjust defaults without touching tool parameters:

Env var	Default	Effect
PI_IMAGE_DIR	generated	Default save directory (relative paths resolve against cwd).
PI_IMAGE_QUALITY	fast	Default quality if the model doesn't pass one.

How is this different from @benvargas/pi-antigravity-image-gen?

	pi-banana	@benvargas
Auth	GOOGLE_API_KEY env var	OAuth via Antigravity (pi /login)
Models	Flash + Pro	Pro only
Image edit (input picture)	✅	❌
Auto-save default	✅ to ./generated/	❌ disabled by default
Aspect ratios	10	10
Inline terminal preview	✅	✅

If you have an Antigravity OAuth setup and want quota tracking, benvargas is great. If you just want to drop in an API key and start making pictures, this one is simpler.

From the same author

By Francesco Frapporti at Fornace.

• pi-bench — LLM benchmark toolkit for pi. Probes every available model to find the fastest and cheapest. All package banners in this ecosystem were created with pi-banana.
• pi-recap — Always-visible session recap panel for pi. Never scroll back to remember what you were doing.
• pi-alibaba-models — Complete Alibaba provider for pi: Qwen, DeepSeek, Kimi, GLM, MiniMax with native thinking levels.
• pi-notte-theme — Notte: a true-dark pi theme where darkness has color and text glows like terminal phosphor.

Development

npm install
npm run typecheck             # tsc --noEmit
GOOGLE_API_KEY=… npm run smoke  # full live-API smoke test in _tmp/

License

MIT