imagegen-nano-banana

name	imagegen-nano-banana
description	Generate images via Google Gemini Nano Banana Pro (gemini-3-pro-image-preview), the SOTA Google image model as of 2026. Use when the user wants to create, generate, render, or produce an image with Google / Gemini (e.g. "make me an image with gemini", "use nano banana", "nano banana pro", "google image gen", "imagen alternative"). Wraps the v1beta generateContent REST API in a small bash script. Supports up to 14 reference images for blending and 4K output.
allowed-tools	["Bash","Read"]

imagegen-nano-banana — Gemini Nano Banana Pro from the CLI

Thin shell wrapper around POST generativelanguage.googleapis.com/v1beta/models/<model>:generateContent. Defaults to gemini-3-pro-image-preview (Nano Banana Pro). Decodes the inline base64 image straight to a file.

Tool

~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana

Run with --help for the full flag list. Common invocations:

# Default: 1:1, 2K, gemini-3-pro-image-preview
~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana \
  "a cat reading a newspaper"

# Wide 4K poster
~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana \
  -a 16:9 -s 4K -o /tmp/sunrise.png \
  "panoramic alpine sunrise, dramatic clouds, photorealistic"

# Blend reference images
~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana \
  -i ref1.png -i ref2.png -i ref3.png \
  "compose these into a single cohesive product shot, studio lighting"

# Switch to faster Nano Banana 2 (Flash variant)
~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana \
  -m gemini-3.1-flash-image-preview \
  "a quick concept sketch of a cyberpunk corgi"

# Use Google Search grounding for up-to-date data
~/.claude/skills/imagegen-nano-banana/bin/imagegen-nano-banana --grounding \
  "infographic of today's NASDAQ close, clean editorial style"

API key

Reads GEMINI_API_KEY in this order:

1. Environment variable.
2. $IMAGEGEN_NANO_BANANA_KEYS_FILE if set, else ~/.config/imagegen-nano-banana/keys.env — sourced as shell KEY=value lines.

Set up the file the first time:

mkdir -p ~/.config/imagegen-nano-banana && chmod 700 ~/.config/imagegen-nano-banana
printf 'GEMINI_API_KEY=...\n' > ~/.config/imagegen-nano-banana/keys.env
chmod 600 ~/.config/imagegen-nano-banana/keys.env

Get a key from https://aistudio.google.com/apikey. The script exits with a clear error if the key is missing.

Flags

Flag	Default	Notes
-p, --prompt	—	or pass as positional
-o, --out	./imagegen-nano-banana-<ts>.<ext>	extension follows response mime
-m, --model	gemini-3-pro-image-preview	see model table below
-a, --aspect	1:1	16:9, 9:16, 4:3, 3:4, 21:9, 2:3, 3:2
-s, --size	2K	1K / 2K / 4K (Pro supports all)
-i, --image	—	reference image path; repeatable
--grounding	off	enable Google Search grounding (Pro only)
--thinking	—	low / medium / high reasoning depth

Models

Model ID	Codename	Use for
gemini-3-pro-image-preview	Nano Banana Pro	Highest fidelity, 4K, complex layouts, text-in-image
gemini-3.1-flash-image-preview	Nano Banana 2	Fast everyday generation
gemini-2.5-flash-image	Nano Banana	Original; fastest, cheapest

How to use this skill

1. Pick a prompt. If the user wasn't specific, ask one short clarifying question about subject + style + aspect ratio rather than guessing.
2. Run the script via Bash. Pro at 2K typically returns in 8–12s.
3. The script prints the output path on success — show the image to the user (Read it so it renders inline, or mention the path).
4. For iteration, tweak the prompt (or add -i previous.png as a reference).

Notes

• All Google-generated images carry the imperceptible SynthID watermark. Mention this if the user is producing assets where AI provenance matters.
• Multilingual text-in-image is a strength — call it out for posters/ads.
• --grounding adds a fixed surcharge per call but lets the model use real-time web context (e.g. "today's weather map", "current charts"). Pro only.
• Reference-image blend works up to ~14 inputs; the model treats them as visual context for the prompt, not as a starting canvas to edit.
• The default model often returns JPEG (not PNG). The script picks the file extension from the response mime type when no -o is supplied; if you do pass -o foo.png, the file will be written to that exact path even if the bytes are JPEG (most viewers handle the mismatch).
• If the response comes back text-only (e.g. safety filter), the script surfaces the model's text reply in the error message.