Content
# nano-banana-claude
*Generated by `generate_image` (Nano Banana flash, 16:9).*
A small [MCP](https://modelcontextprotocol.io) server that gives Claude Code (or any MCP
client) image generation via Google's **Nano Banana** Gemini image models — plus the
deterministic post-processing the model can't do reliably on its own: **true alpha
transparency**, **AI upscaling**, resizing, and format conversion.
The idea: keep generation in the model, but put the things that need real code (a genuine
alpha channel, exact dimensions, super-resolution) behind tools so they're reliable instead
of prompt-and-hope.
## Tools
| Tool | What it does |
|------|--------------|
| `generate_image` | Text → image. `model` (`flash`/`pro`), `aspect_ratio`, `format`, `output_path`. |
| `generate_transparent_image` | Generate, then remove the background with rembg (U2-Net) → a real **RGBA PNG**. Nano Banana can't produce reliable alpha on its own. |
| `edit_image` | One or more input images + a prompt → edited / composited result. |
| `process_image` | Local Pillow ops, no API call: resize/fit/crop, convert (png/webp/jpeg + quality), optional background removal. |
| `upscale_image` | AI super-resolution with Real-ESRGAN (`realesr-general-x4v3`). Tiled for large images, alpha-preserving. |
**Models:** `flash` → `gemini-2.5-flash-image` (Nano Banana, default, fast/cheap),
`pro` → `gemini-3-pro-image-preview` (Nano Banana Pro, higher quality).
## Requirements
- Python 3.10+
- A Gemini API key (free tier available) from <https://aistudio.google.com/apikey>
- ~500 MB disk for dependencies; CPU is fine (no GPU required)
## Install
```bash
git clone https://github.com/tougenrip/nano-banana-claude.git
cd nano-banana-claude
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
```
Two ML models download automatically on first use (not committed):
- rembg U2-Net (~176 MB) → `~/.u2net/` — first transparent-image call
- Real-ESRGAN x4 (~4.6 MB) → `~/.cache/nano-banana/` — first upscale call
## Register with Claude Code
From inside the cloned directory:
```bash
claude mcp add nano-banana --scope user \
--env GEMINI_API_KEY=your-key-here \
-- "$(pwd)/.venv/bin/python" "$(pwd)/server.py"
```
`--scope user` makes it available in every project; drop it to scope to the current project.
Reconnect and the five `nano-banana` tools appear. The API key is read from the environment
at runtime and is never written into the code.
## Usage
Once registered, just ask in natural language — Claude picks the tool:
- *"Generate a 16:9 image of a neon city at night."* → `generate_image`
- *"Make a transparent PNG of a red sneaker."* → `generate_transparent_image`
- *"Put the logo in photo.png onto the mug in mug.jpg."* → `edit_image`
- *"Resize hero.png to 1200px wide as webp."* → `process_image`
- *"Upscale icon.png 4×."* → `upscale_image`
Tools can chain: generate a transparent cutout, then upscale it — the alpha survives.
Outputs are written next to the working directory (or to `output_path`) and the absolute
path is returned. Set `NANO_BANANA_OUTPUT_DIR` to change the default location.
## How transparency works
Image models paint pixels; asking for a "transparent background" usually yields a flat color
or a drawn-in checkerboard, not a real alpha channel. So `generate_transparent_image`
generates the subject on a plain background, then runs rembg (U2-Net) to compute an actual
alpha matte, producing a true RGBA PNG. `process_image(remove_background=true)` does the same
on any existing image.
<img src="examples/transparent-banana.png" alt="Transparent cutout example" width="320">
*Output of `generate_transparent_image` — a real RGBA PNG (~90% of pixels fully
transparent), not a painted-on checkerboard.*
## Layout
```
server.py FastMCP server + tool definitions; Gemini REST via stdlib urllib
upscale.py Real-ESRGAN inference (onnxruntime), lazily loaded, with tiling
requirements.txt mcp, Pillow, rembg, onnxruntime
```
## License
MIT — see [LICENSE](LICENSE).
