Vertex Banana

Vertex Banana — Image Generation via Google Vertex AI

Generate and edit images using Google Vertex AI Gemini models. Supports two methods: Python SDK (preferred) and curl (fallback).

When to Use This Skill

Use when the user asks to generate or edit images via Google Vertex AI / Gemini image models:

• Text-to-image generation
• Image editing (modify existing images with text instructions)
• Any request mentioning "vertex banana" or Vertex AI image generation

Setup

1. Verify API Key

[ -n "$VERTEX_API_KEY" ] && echo "API key configured" || echo "Missing VERTEX_API_KEY"

VERTEX_API_KEY is the Vertex AI API key (53 chars). The script temporarily clears conflicting SDK env vars (GOOGLE_API_KEY, GEMINI_API_KEY, GOOGLE_GENAI_USE_VERTEXAI) during execution to prevent auth conflicts.

2. Install Python SDK (for Python method)

pip install --upgrade google-genai

Script Location

The Python script lives in the skill folder and is accessible from any project:

~/.claude/skills/vertex-banana/vertex_banana.py

Always reference it with the full path: $HOME/.claude/skills/vertex-banana/vertex_banana.py

Method 1: Python SDK (Preferred)

Uses the google-genai Python package. Handles streaming, image saving, and error handling automatically.

Text-to-Image

python3 $HOME/.claude/skills/vertex-banana/vertex_banana.py "a cat wearing sunglasses on a beach"

Image Editing

python3 $HOME/.claude/skills/vertex-banana/vertex_banana.py "make the sky purple and add stars" -i input.png

With Options

python3 $HOME/.claude/skills/vertex-banana/vertex_banana.py \
  "professional product photo of sneakers" \
  -m flash-image \
  --aspect-ratio 16:9 \
  --image-size 2K \
  -o ./my-output \
  -f product_shot

All Python Script Options

Method 2: Curl (Fallback)

Use when the google-genai package is not installed. The same Python script handles this with the --curl flag.

Via Script

python3 $HOME/.claude/skills/vertex-banana/vertex_banana.py "a sunset over mountains" --curl

Raw Curl (Manual)

Build the request JSON, send via curl, and parse the response:

Step 1: Create request.json

VB_WORK=$(mktemp -d /tmp/vb-XXXXXXXXXX)
cat << 'EOF' > "$VB_WORK/request.json"
{
    "contents": [
        {
            "role": "user",
            "parts": [
                {
                    "text": "YOUR PROMPT HERE"
                }
            ]
        }
    ],
    "generationConfig": {
        "temperature": 1,
        "maxOutputTokens": 32768,
        "responseModalities": ["TEXT", "IMAGE"],
        "topP": 0.95,
        "imageConfig": {
            "aspectRatio": "auto",
            "imageSize": "1K",
            "imageOutputOptions": {
                "mimeType": "image/png"
            },
            "personGeneration": "ALLOW_ALL"
        },
        "thinkingConfig": {
            "thinkingLevel": "HIGH"
        }
    },
    "safetySettings": [
        {"category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "OFF"},
        {"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "OFF"},
        {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "OFF"},
        {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "OFF"}
    ]
}
EOF

Step 2: Send request

MODEL_ID="gemini-3.1-flash-image"

curl -s -X POST \
  -H "Content-Type: application/json" \
  "https://aiplatform.googleapis.com/v1/publishers/google/models/${MODEL_ID}:streamGenerateContent?key=${VERTEX_API_KEY}" \
  -d @"$VB_WORK/request.json" -o "$VB_WORK/response.json"

Step 3: Extract and save image

python3 -c "
import json, base64, os
os.makedirs('vertex-banana-output', exist_ok=True)
with open('$VB_WORK/response.json') as f:
    data = json.load(f)
chunks = data if isinstance(data, list) else [data]
n = 0
for chunk in chunks:
    for c in chunk.get('candidates', []):
        for p in c.get('content', {}).get('parts', []):
            if 'inlineData' in p:
                n += 1
                img = base64.b64decode(p['inlineData']['data'])
                path = f'vertex-banana-output/image_{n}.png'
                open(path, 'wb').write(img)
                print(f'Saved: {path}')
            elif 'text' in p:
                print(p['text'])
print(f'{n} image(s) generated' if n else 'No images in response')
"

Available Models

Note: All model IDs below verified LIVE against the API on 2026-06-10 (client.models.list()). They may still change as Google rotates previews.

Thinking level: valid API values are MINIMAL | LOW | MEDIUM | HIGH (default HIGH). NONE is rejected by the API with 400 INVALID_ARGUMENT — the script maps the legacy NONE choice to MINIMAL.

Common Sizes & Aspect Ratios

Output Location

Images are saved to ./vertex-banana-output/ in the current working directory by default. Override with -o /path/to/dir.

Presenting Results

After generation completes:

1. Check ./vertex-banana-output/ for the generated file(s)
2. Show the image to the user using the Read tool (it supports images)
3. Offer to regenerate with different options if needed

Prompt Tips

1. Be specific — include style, mood, colors, composition details
2. Add "no text" — if you don't want text rendered in the image
3. Reference styles — "editorial photography", "flat illustration", "3D render", "watercolor"
4. For editing — describe exactly what to change: "make the background blue" not "change it"

Sole image-generation skill

This is the ONLY image-generation skill — there is no alternative and no fallback. It calls the Vertex AI / Gemini image API directly (Python SDK or curl) with VERTEX_API_KEY, and depends on no CLI.

Troubleshooting

Anti-Patterns