Gemini Image Gen

Generating Images & SVGs with Gemini CLI

Overview

Gemini CLI can generate two types of visual output:

• SVG code: Text-based vector graphics output directly from Gemini models (no API key needed)
• Raster images (PNG): Generated via the nanobanana extension (requires GEMINI_API_KEY env var)

When to Use

• User asks to "generate an image", "create an icon", "make a logo", "draw something"
• User needs SVG code for icons, logos, illustrations, diagrams
• User needs PNG/JPEG images (photos, artwork, UI assets)

Quick Reference

Model Fallback Chain

Always try the best model first, then fall back to cheaper/faster models on rate limit (429) errors.

SVG models (in order): gemini-3-pro-preview → gemini-2.5-flash Image models (in order): gemini-2.5-flash → gemini-2.5-flash (retry after delay)

Fallback procedure:

1. Run command with primary model, capture stderr via 2>&1
2. Check output for 429, RESOURCE_EXHAUSTED, or capacity → means rate limited
3. Retry with the next model in the chain
4. If all models fail, tell the user and suggest waiting

SVG Generation — via Code Output (NOT Nanobanana)

SVGs are generated as text code output by the model itself — NOT via nanobanana. Use -e none to disable extensions so Gemini outputs SVG code instead of trying to use image tools.

# Try primary model first
OUTPUT=$(gemini -m gemini-3-pro-preview -e none -p "Create an SVG icon of [description]. Output ONLY the raw SVG code, no markdown fences, no explanation." --output-format text 2>&1)

# Check for rate limiting - fall back to cheaper model
if echo "$OUTPUT" | grep -qi "429\|RESOURCE_EXHAUSTED\|capacity\|quota"; then
  OUTPUT=$(gemini -m gemini-2.5-flash -e none -p "Create an SVG icon of [description]. Output ONLY the raw SVG code, no markdown fences, no explanation." --output-format text 2>&1)
fi

# Save result
echo "$OUTPUT" | sed '/^```/d' > output.svg

Or run them as separate commands - try the primary model first, and if you see rate limit errors in the output, immediately run the fallback.

SVG Prompt Tips

• Always end with: "Output ONLY the raw SVG code, no markdown fences, no explanation."
• Specify dimensions: "200x200", "24x24 icon size"
• Specify colors explicitly: "Use #FF6B35 orange and #1A1A2E dark blue"
• Specify style: "flat design", "line art", "gradient", "minimalist"
• For logos, include text requirements: "Include the text 'AppName' below the icon"

Post-Processing SVG Output

Gemini sometimes wraps output in markdown code fences. Strip them:

# Generate and clean SVG output
gemini -m gemini-3-pro-preview -e none -p "..." --output-format text | sed '/^```/d' > output.svg

After generating, read the file to verify valid SVG and show/return the code to the user.

Image Generation (PNG) — via Nanobanana Extension

PNG image generation uses nanobanana, a Gemini CLI extension that provides a generate_image tool. This is the ONLY way to generate real images — do NOT try to use Pillow, Python scripts, or -e none.

How it works

1. Gemini CLI loads the nanobanana extension automatically on startup (you'll see Loading extension: nanobanana in output)
2. Gemini calls the generate_image tool provided by nanobanana
3. Nanobanana generates the image and saves it to the requested path

Requirements

• GEMINI_API_KEY environment variable must be set (configured in ~/.zshrc)
• Get a key from https://aistudio.google.com/apikey if not set
• Do NOT pass -e none — that disables all extensions including nanobanana

Save location

Unless the user specifies a different path, always save generated images to /tmp/. Example: /tmp/prishtina.png, /tmp/logo.png.

Command

gemini -m gemini-2.5-flash -p "Generate a beautiful image of [description]. Save it to /tmp/[filename].png" -y --output-format text 2>&1

Key rules

• NO -e none — extensions must be enabled so nanobanana loads
• -y is required — auto-approves the nanobanana tool call (YOLO mode)
• --output-format text — keeps output clean
• If you see Loading extension: nanobanana in stdout, it's working
• If you see "No valid API key found", the GEMINI_API_KEY env var is missing
• Best for: photorealistic images, artwork, cityscapes, illustrations, any visual content

Prompt Handling

Default: pass the user's prompt as-is. Always forward the user's words to Gemini without rewriting, embellishing, or adding details. Only append the save path instruction.

Examples of prompts to pass verbatim:

• "generate an image of Prishtina" → pass as-is
• "generate an image of a cat sitting on a rooftop at sunset" → pass as-is
• "make a logo with a blue mountain" → pass as-is

Only fill in details when the user explicitly delegates content to you — e.g. "create some images to add to this page" or "generate a few images for this blog post". In those cases you choose subjects, composition, and style.

Workflow for Claude

When asked to generate visual assets:

1. Determine type: SVG (vector code) or PNG (raster image)?
2. SVG path:
   • Run with gemini-3-pro-preview first (best quality)
   • If output contains rate limit error → immediately retry with gemini-2.5-flash
   • Strip markdown fences, save to file, verify valid SVG
3. PNG path (uses nanobanana extension):
   • Run gemini without -e none and with -y so nanobanana loads and auto-approves
   • Confirm Loading extension: nanobanana appears in output
   • Specify save path in prompt, verify file exists after
4. Always verify: Read/check the output file exists and is valid before reporting success
5. Show SVG code: For SVGs, read the file and present the code to the user

Common Issues

Source: gentritbiba/agent-skills — distributed by TomeVault.