Corespeed NanoBanana — Gemini Image & Text Generation

Auth: Set CS_AI_GATEWAY_BASE_URL and CS_AI_GATEWAY_API_TOKEN environment variables.

Workflow

1. Pick a model from the table below (default: gemini-2.5-flash-image for image generation)
2. Run the script with your prompt

Usage

uv run {baseDir}/scripts/gemini.py --prompt "your prompt" -f output.ext [-i input.ext] [--model MODEL]

• --prompt, -p — Text prompt (required)
• --filename, -f — Output filename (required)
• --input, -i — Input image file(s), repeat for multiple
• --model, -m — Model name (default: gemini-2.5-flash-image)
• --modalities — Response type: auto, image, text, image+text (default: auto)
• --json — Output structured JSON (recommended for agent consumption)

Output format is determined by file extension: .png/.jpg → image generation, .txt/.md → text output.

Image Generation

# Text-to-image
uv run {baseDir}/scripts/gemini.py -p "a watercolor fox in autumn forest" -f fox.png

# Image editing
uv run {baseDir}/scripts/gemini.py -p "Remove background, add beach sunset" -f edited.png -i photo.jpg

# Multi-image compositing
uv run {baseDir}/scripts/gemini.py -p "Blend these two scenes together" -f blend.png -i scene1.png -i scene2.png

Image Analysis

# Describe an image
uv run {baseDir}/scripts/gemini.py -p "Describe this image" -f desc.txt -i photo.jpg --model gemini-2.5-flash

# Compare images
uv run {baseDir}/scripts/gemini.py -p "What are the differences?" -f diff.txt -i before.jpg -i after.jpg --model gemini-2.5-flash

Text Generation

# Use the most capable model for complex tasks
uv run {baseDir}/scripts/gemini.py -p "Write a haiku about coding" -f haiku.txt --model gemini-2.5-pro

Models

Notes

• No manual Python setup required. The script uses PEP 723 inline metadata. uv run automatically creates an isolated virtual environment and installs the google-genai dependency on first run.
• Image output is returned inline as base64 from the Gemini API — no separate download step.
• Use timestamps in filenames: yyyy-mm-dd-hh-mm-ss-name.ext.
• Script prints MEDIA: line for OpenClaw to auto-attach generated images.
• Do not read generated media back; report the saved path only.
• Only gemini-2.5-flash-image can generate images. Other models are text-only.
• Use --json for structured output: {"ok": true, "files": [...], "text": "...", "model": "...", "tokens": {...}}

Support

Built by Corespeed. If you need help or run into issues:

• 💬 Discord: discord.gg/mAfhakVRnJ
• 🐦 X/Twitter: @CoreSpeed_io
• 🐙 GitHub: github.com/corespeed-io/skills