Nano Banana (cURL Edition)

Image generation skill powered by Google's Gemini image models via the REST API. Uses only cURL, jq, and base64 — no Python or SDK dependencies.

Requirements

•GEMINI_API_KEY environment variable set
•curl, jq, and base64 available on PATH (standard on most Linux/macOS systems)

Model

Always use gemini-3-pro-image-preview for all image generation tasks. This is the highest quality model available.

API Endpoint

code

https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent

Critical: Handling API Responses

The Gemini API response JSON can be very large (multi-megabyte base64 image data) and may contain unescaped control characters that break jq. Never pipe the full extraction pipeline in one shot — this can hang indefinitely on large responses.

Always use the prescribed extraction command:

bash

curl -s -X POST "$ENDPOINT" ... -o /tmp/nb_resp.json
extract_nano_banana_image /tmp/nb_resp.json /tmp/output.png

The command (extract_nano_banana_image) safely:

•Sanitizes control characters with tr
•Checks for API errors
•Extracts base64 to an intermediate temp file (avoids pipe-buffer hangs)
•Decodes to the final image in a separate step
•Handles both inlineData (camelCase) and inline_data (snake_case) field names
•Prints text commentary if present

Then: convert to WebP!

Use the convert_to_webp command and then discard the PNG file:

bash

convert_to_webp /tmp/output.png output.webp
rm /tmp/output.png

What NOT to do

bash

# WRONG: single pipeline hangs on large responses
curl ... | jq -r '...' | base64 -d > output.png

# WRONG: storing response in a shell variable (too large)
RESPONSE=$(curl -s ...)

# WRONG: only checking snake_case field names (API may return camelCase)
jq '.inline_data.data'

Task Workflows

All workflows use the same REST endpoint. The difference is in the request body.

Text-to-Image Generation

Generate an image from a text prompt and save it to a file.

bash

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
  -H "x-goog-api-key: ${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \
  -o /tmp/nb_resp.json \
  -d '{
    "contents": [{"parts": [{"text": "A cat wearing a wizard hat"}]}],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"]
    }
  }'

extract_nano_banana_image /tmp/nb_resp.json /tmp/output.png
convert_to_webp /tmp/output.png output.webp

Text-to-Image with Aspect Ratio and Size

bash

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
  -H "x-goog-api-key: ${GEMINI_API_KEY}" \
  -H "Content-Type: application/json" \
  -o /tmp/nb_resp.json \
  -d '{
    "contents": [{"parts": [{"text": "Futuristic motorcycle on Mars"}]}],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageConfig": {
        "aspectRatio": "16:9",
        "imageSize": "4K"
      }
    }
  }'

extract_nano_banana_image /tmp/nb_resp.json /tmp/output.png
convert_to_webp /tmp/output.png output.webp

Edit an Existing Image

Use jq -Rs (raw slurp from stdin) to build the request JSON with inline image data. This avoids shell variable size limits.

bash

base64 -w0 input.png \
  | jq -Rs --arg prompt "Add a sunset to the background" '{
      "contents": [{"parts": [
        {"text": $prompt},
        {"inline_data": {"mime_type": "image/png", "data": .}}
      ]}],
      "generationConfig": {
        "responseModalities": ["TEXT", "IMAGE"]
      }
    }' \
  | curl -s -X POST \
      "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
      -H "x-goog-api-key: ${GEMINI_API_KEY}" \
      -H "Content-Type: application/json" \
      -d @- -o /tmp/nb_resp.json

extract_nano_banana_image /tmp/nb_resp.json /tmp/output.png
convert_to_webp /tmp/output.png output.webp

For JPEG source images, use "mime_type": "image/jpeg". Match the mime type to the actual file format.

Compose Multiple Images

Encode each image as a JSON part, slurp them into an array, then merge with the text prompt. Supports up to 14 reference images.

bash

for img in person1.png person2.png; do
  base64 -w0 "$img" | jq -Rs '{"inline_data": {"mime_type": "image/png", "data": .}}'
done \
  | jq -s --arg prompt "Create a group photo in an office setting" '{
      "contents": [{"parts": ([{"text": $prompt}] + .)}],
      "generationConfig": {
        "responseModalities": ["TEXT", "IMAGE"]
      }
    }' \
  | curl -s -X POST \
      "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent" \
      -H "x-goog-api-key: ${GEMINI_API_KEY}" \
      -H "Content-Type: application/json" \
      -d @- -o /tmp/nb_resp.json

extract_nano_banana_image /tmp/nb_resp.json /tmp/output.png
convert_to_webp /tmp/output.png output.webp

Generation Options

Aspect Ratios

1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9

Resolutions

1K (1024px), 2K, 4K

Set these in the generationConfig.imageConfig object:

json

"imageConfig": {
  "aspectRatio": "16:9",
  "imageSize": "2K"
}

Prompting Tips

Photorealistic: Include camera settings, lighting, lens details

code

"Shot on 85mm lens, golden hour lighting, shallow depth of field"

Logos: Specify style, colors, typography

code

"Clean minimalist logo, sans-serif font, monochrome, vector style"

Product shots: Describe studio setup

code

"Studio-lit, 3-point softbox, polished surface, 45-degree angle"

Stylized art: Name the style explicitly

code

"Anime style, cel-shading, bold outlines, vibrant colors"

Error Handling

The extract_nano_banana_image command checks for API errors and missing image data automatically. If you need to debug manually:

bash

tr -d '\000-\010\013\014\016-\037' < /tmp/nb_resp.json | jq .

Common issues:

•Missing API key: Ensure GEMINI_API_KEY is exported in your shell environment
•Empty output file: The model may have refused the prompt (safety filters) — check the JSON response for blockReason or finishReason
•Large images for editing: Very large source images may exceed request size limits — resize before encoding
•jq parse errors: The tr sanitization step in extract_nano_banana_image handles this, but if running manually, always sanitize first
•Quota errors (429): Free-tier quotas for gemini-3-pro-image may be 0 — a billing-enabled API key is required

How to Use This Skill

When the user invokes /loreduck:image, interpret $ARGUMENTS as the image generation task. Determine the appropriate workflow (generate, edit, or compose) based on the request, construct the cURL command with the right parameters, execute it using extract_nano_banana_image for response processing, convert it to WebP using convert_to_webp and then present the result. Always save output images to the current working directory unless the user specifies a different path.