Image Generation

Image Generation

Important (December 2025): The google-generativeai package has been deprecated. This skill now uses the google-genai SDK. If upgrading from older code, see the migration guide.

Purpose

This skill enables AI-powered image generation and editing through Google's Gemini image models and OpenAI's DALL-E models. Create photorealistic images, illustrations, logos, stickers, and product mockups from natural language descriptions. Edit existing images with text instructions, apply style transfers, and refine outputs through iterative conversation.

Attribution: This skill is inspired by the gemini-imagegen skill from Every Marketplace by Every Inc.

When to Use

This skill should be invoked when the user asks to:

• Generate images from text descriptions ("create an image of...", "generate a picture...")
• Create logos, icons, or stickers ("design a logo for...", "make a sticker...")
• Edit or modify existing images ("change the background to...", "add... to this image")
• Apply artistic styles or effects ("make it look like...", "stylize as...")
• Create product mockups or visualizations ("product photo of...", "mockup showing...")
• Refine or iterate on images ("make it more...", "adjust the...", "try again with...")
• Generate variations with different styles or compositions

Available Models

Google Gemini Models (Nano Banana)

1. gemini-2.5-flash-image ("Nano Banana")

   • Resolution: 1K (1024px), supports 2K
   • Aspect ratios: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9
   • Best for: Speed, high-volume operations, rapid iteration, image editing
   • Use when: Quick prototypes, multiple variations, time-sensitive requests
   • Cost: $0.039 per image ($30/million output tokens)

2. gemini-3-pro-image-preview ("Nano Banana Pro")

   • Resolution: 1K default, supports 2K and 4K
   • Aspect ratios: Same as Flash
   • Best for: Professional assets, complex instructions, highest quality
   • Use when: Final deliverables, detailed compositions, text-heavy designs
   • Special features:
     • Google Search grounding for real-time data visualization
     • "Thinking" mode with interim composition refinement
     • Up to 14 reference images (6 objects, 5 humans for character consistency)
     • Advanced text rendering

Google Imagen 4 Family (New)

3. imagen-4.0-fast-generate-001 ("Imagen 4 Fast")

   • Resolution: Standard
   • Best for: Rapid generation, high-volume tasks
   • Use when: Speed is priority, budget-conscious
   • Cost: $0.02 per image
   • Note: Text-only input (no image editing)

4. imagen-4.0-generate-001 ("Imagen 4")

   • Resolution: Up to 2K
   • Best for: High-quality photorealistic images, excellent text rendering
   • Use when: Professional quality needed, text in images
   • Features: Significant improvements in text rendering over previous Imagen models

5. imagen-4.0-ultra-generate-001 ("Imagen 4 Ultra")

   • Resolution: Up to 2K
   • Best for: Highest quality, detailed visuals
   • Use when: Maximum quality is essential (one image at a time)
   • Limitation: Only generates one image per request

OpenAI GPT Image Models

6. gpt-image-1.5 (Recommended - December 2025)

   • Resolution: 1024x1024, 1536x1024, 1024x1536, or auto
   • Best for: Production-quality visuals, precise editing, character consistency
   • Use when: Professional design, iterative workflows, text-heavy images
   • Features:
     • 4x faster than gpt-image-1, 20% lower cost
     • Built-in reasoning and world knowledge
     • Precise logo & face preservation during edits
     • Excellent text rendering (crisp lettering, dense text)
     • Complex structured visuals (infographics, diagrams, multi-panel)
     • Streaming support
   • Output formats: png, jpeg, webp (with compression control)
   • Transparency: transparent, opaque, or auto background

7. gpt-image-1 (April 2025)

   • Resolution: Up to 4096x4096
   • Best for: High-resolution images, creative workflows
   • Use when: Maximum resolution needed
   • Cost: ~$0.02 (low), ~$0.07 (medium), ~$0.19 (high) per image
   • Output formats: png, jpeg, webp
   • Note: Single image per request, no inpainting

Legacy OpenAI DALL-E Models

8. dall-e-3

   • Resolution: 1024x1024, 1024x1792, 1792x1024
   • Best for: Creative interpretations, artistic renders
   • Use when: Natural artistic style preferred
   • Note: Automatic prompt expansion

9. dall-e-2

   • Resolution: 1024x1024, 512x512, 256x256
   • Best for: Faster generation, lowest cost, variations
   • Use when: Budget-conscious, simpler images
   • Unique feature: Can generate variations of existing images

Model Selection Logic

Ask the user or use this decision tree:

Need image editing or iterative refinement?
├─ Yes → gpt-image-1.5 (best editing) or gemini-2.5-flash-image (multi-turn chat)
└─ No → Text-to-image only
    ├─ Need highest quality?
    │   ├─ Text rendering critical → gpt-image-1.5 or imagen-4.0-generate-001
    │   ├─ Maximum resolution (4K) → gemini-3-pro-image-preview
    │   ├─ Ultra quality (single image) → imagen-4.0-ultra-generate-001
    │   └─ Character consistency → gpt-image-1.5 or gemini-3-pro-image-preview
    ├─ Need speed/volume?
    │   ├─ Cheapest → imagen-4.0-fast-generate-001 ($0.02)
    │   └─ Fast + editing → gemini-2.5-flash-image
    └─ Balanced default → gpt-image-1.5 (recommended)

Quick Reference:

• Best overall: gpt-image-1.5 - fast, affordable, great editing & text
• Best for text rendering: gpt-image-1.5 or imagen-4.0-generate-001
• Best for 4K resolution: gemini-3-pro-image-preview
• Cheapest per image: imagen-4.0-fast-generate-001 ($0.02)
• Best for reference images: gemini-3-pro-image-preview (up to 14 refs)
• Best for iterative editing: gpt-image-1.5 (face/logo preservation)

If the user has specific model preference, use that.

Capabilities

1. Text-to-Image Generation: Create images from detailed text descriptions
2. Image Editing: Modify existing images with text instructions
3. Style Transfer: Apply artistic styles, filters, and effects
4. Logo & Sticker Design: Generate branded assets with specific styles
5. Product Mockups: Create professional product photography and presentations
6. Multi-turn Refinement: Iteratively improve images through conversation
7. Aspect Ratio Control: Generate images in various formats (square, portrait, landscape, wide)
8. Reference-based Generation: Use existing images as compositional references (Gemini Pro)

Instructions

Step 1: Understand the Request

Analyze the user's request to determine:

• Type: Text-to-image, image editing, style transfer, logo/sticker, mockup
• Subject: What should be in the image
• Style: Photorealistic, illustration, artistic, minimalist, etc.
• Details: Colors, lighting, composition, mood, specific elements
• Format: Aspect ratio, resolution requirements
• Urgency: Speed vs. quality trade-off

Step 2: Select Model

Based on requirements:

• High quality + complexity → gemini-3-pro-image-preview
• Speed + iterations → gemini-2.5-flash-image
• DALL-E preference → dall-e-3 or dall-e-2

If unclear, use AskUserQuestion tool to clarify model preference.

Step 3: Craft Effective Prompt

Build a detailed prompt following these patterns:

For Photorealistic Images:

[Subject], [camera details], [lighting], [mood/atmosphere], [composition]

Example: "Close-up portrait of a woman, 85mm lens, soft golden hour lighting,
serene mood, shallow depth of field, professional photography"

For Illustrations/Art:

[Subject], [art style], [color palette], [details], [mood]

Example: "Kawaii cat sticker, bold black outlines, cel-shading, pastel colors,
cute expression, chibi style"

For Logos:

[concept], [style], [elements], [colors], [context]

Example: "Tech startup logo, minimalist geometric design, abstract network nodes,
blue and silver gradient, professional, vector style"

For Product Photography:

[product], [setting], [lighting], [presentation], [context]

Example: "Wireless earbuds, white background, studio lighting, 3/4 angle view,
clean minimal composition, e-commerce product shot"

Key principles:

• Be specific and detailed
• Include lighting, composition, and mood
• Specify style clearly (photorealistic, illustration, etc.)
• Mention camera/lens for photorealistic (85mm, wide angle, macro)
• For text in images, use Pro model and specify exact text

Step 4: Implement API Call

For Gemini Models:

Note: The google.generativeai package has been deprecated. Use google.genai instead. See migration guide: https://ai.google.dev/gemini-api/docs/migrate

from google import genai
from google.genai import types
from pathlib import Path

# Initialize client (uses GEMINI_API_KEY or GOOGLE_API_KEY env var automatically)
client = genai.Client()

# Basic text-to-image
response = client.models.generate_content(
    model="gemini-2.5-flash-image",  # or gemini-3-pro-image-preview
    contents=prompt_text,
    config=types.GenerateContentConfig(
        response_modalities=["TEXT", "IMAGE"],
        # Optional configurations:
        # image_config=types.ImageConfig(
        #     aspect_ratio="1:1",  # 1:1, 3:4, 4:3, 9:16, 16:9, 21:9
        #     image_size="1K",     # 1K, 2K, 4K (Pro only)
        # )
    )
)

# Extract and save image
for part in response.parts:
    if part.text is not None:
        print(part.text)
    elif part.inline_data is not None:
        image = part.as_image()
        image.save("output.png")

# For image editing (pass existing image):
from PIL import Image

image = Image.open("input.png")
response = client.models.generate_content(
    model="gemini-2.5-flash-image",
    contents=[image, "Make the background a sunset scene"],
    config=types.GenerateContentConfig(response_modalities=["TEXT", "IMAGE"])
)

# For multi-turn refinement (use chat):
chat = client.chats.create(
    model="gemini-2.5-flash-image",
    config=types.GenerateContentConfig(response_modalities=["TEXT", "IMAGE"])
)
response1 = chat.send_message("A futuristic city skyline")
response2 = chat.send_message("Add more neon lights and flying cars")

For Google Imagen 4 Models:

from google import genai
from google.genai import types

# Initialize client (uses GEMINI_API_KEY or GOOGLE_API_KEY env var automatically)
client = genai.Client()

# Imagen 4 text-to-image (no editing support)
# Also available: imagen-4.0-fast-generate-001, imagen-4.0-ultra-generate-001
response = client.models.generate_images(
    model="imagen-4.0-generate-001",
    prompt=prompt_text,
    config=types.GenerateImagesConfig(
        number_of_images=4,  # 1-4 for standard, 1 for Ultra
        aspect_ratio="1:1",  # 1:1, 3:4, 4:3, 9:16, 16:9
        person_generation="allow_adult",  # "dont_allow", "allow_adult", "allow_all"
    )
)

# Save images
for i, generated_image in enumerate(response.generated_images):
    generated_image.image.save(f"output_{i}.png")

For OpenAI Models (gpt-image-1.5 recommended):

from openai import OpenAI
from pathlib import Path
import base64

client = OpenAI()  # reads OPENAI_API_KEY env var automatically

# gpt-image-1.5 generation (recommended)
response = client.images.generate(
    model="gpt-image-1.5",
    prompt=prompt_text,
    size="1024x1024",  # or "1536x1024", "1024x1536", "auto"
    quality="high",    # "low", "medium", "high"
    n=1,               # 1-10 images
    output_format="png",  # "png", "jpeg", "webp"
    background="auto",    # "transparent", "opaque", "auto"
    moderation="auto",    # "auto" or "low" for less restrictive
)

# Response returns base64 data
image_data = base64.b64decode(response.data[0].b64_json)
Path("output.png").write_bytes(image_data)

# gpt-image-1 generation (for max 4K resolution)
response = client.images.generate(
    model="gpt-image-1",
    prompt=prompt_text,
    size="1024x1024",
    quality="high",
    n=1,
)

# Image editing with gpt-image-1.5
response = client.images.edit(
    model="gpt-image-1.5",
    image=open("input.png", "rb"),
    prompt="Change the background to a beach sunset",
    size="1024x1024",
)

# Legacy DALL-E 3 generation
response = client.images.generate(
    model="dall-e-3",
    prompt=prompt_text,
    size="1024x1024",  # or "1024x1792", "1792x1024"
    quality="standard",  # or "hd"
    n=1,
)
image_url = response.data[0].url

# Download URL-based response
import requests
image_data = requests.get(image_url).content
Path("output.png").write_bytes(image_data)

Implementation approach:

• Use Bash tool to execute Python scripts with API calls
• Check for API keys in environment variables
• Handle errors gracefully (API limits, invalid prompts, etc.)
• Save images with descriptive filenames
• Report image location to user

Step 5: Handle Output

1. Save the generated image to an appropriate location
2. Verify the output meets the request
3. Show the user the saved file path
4. Offer refinement if the result isn't quite right
5. Explain the prompt used so the user understands the generation

Step 6: Iterate if Needed

If the user wants changes:

• For Gemini: Use chat interface to maintain context
• For gpt-image-1.5: Use editing API for precise face/logo preservation
• For Imagen/DALL-E: Generate new image with updated prompt
• Keep previous versions for comparison
• Suggest specific adjustments based on the current result

Requirements

API Keys:

• Google (Gemini/Imagen): Set GOOGLE_API_KEY or GEMINI_API_KEY environment variable
• OpenAI: Set OPENAI_API_KEY environment variable

Python Packages:

pip install google-genai openai pillow requests

Note: The google-generativeai package has been deprecated and will no longer receive updates. Use google-genai instead. Migration guide: https://ai.google.dev/gemini-api/docs/migrate

System:

• Python 3.8+
• Internet connection for API access
• Write permissions for saving images

Approximate Costs (per image):

Best Practices

Prompt Engineering

1. Be Specific: Vague prompts produce inconsistent results

   • Bad: "a nice landscape"
   • Good: "mountain valley at sunrise, mist over lake, pine trees, warm golden light, peaceful atmosphere"

2. Include Technical Details for photorealism:

   • Camera: "shot on 85mm lens", "wide angle 24mm", "macro photography"
   • Lighting: "golden hour", "studio lighting", "rim light", "soft diffused"
   • Quality: "high resolution", "detailed", "sharp focus", "professional photography"

3. Specify Style Clearly:

   • "photorealistic", "oil painting", "watercolor", "digital art", "3D render"
   • "minimalist", "detailed", "abstract", "realistic", "stylized"
   • "anime style", "pixel art", "vector art", "charcoal sketch"

4. Use Examples and References:

   • "in the style of [artist/art movement]"
   • "similar to [known visual reference]"
   • For Gemini Pro: Provide actual reference images

5. Negative Prompts (what to avoid):

   • DALL-E doesn't support negative prompts directly
   • For Gemini, phrase as positive instructions: "clear sky" vs "no clouds"

Model-Specific Tips

gpt-image-1.5 (Recommended):

• Best for iterative editing workflows - preserves faces/logos during edits
• Built-in reasoning understands context (e.g., "Bethel, NY, August 1969" → Woodstock)
• Excellent text rendering, especially dense/small text
• Great for infographics, diagrams, multi-panel compositions
• 4x faster than gpt-image-1, use streaming for real-time feedback
• Use background="transparent" for assets

gpt-image-1:

• Maximum resolution (4096x4096) when needed
• Good for one-shot high-res generation
• No editing/inpainting support

Imagen 4 Family:

• Best text rendering among Google models
• Use Fast ($0.02) for high-volume prototyping
• Use Ultra for highest quality single images
• Text-to-image only (no editing) - use Gemini for edits
• All images include SynthID watermark

Gemini Flash (2.5) - Nano Banana:

• Best for iterative multi-turn editing via chat
• Good for generating multiple variations quickly
• Use for draft/concept phase with refinement

Gemini Pro (3) - Nano Banana Pro:

• Use for final deliverables and 4K output
• Best for complex compositions with reference images (up to 14)
• "Thinking" mode generates interim drafts for composition planning
• Leverage Google Search grounding for current events/real places

DALL-E 3 (Legacy):

• Excellent at understanding natural language
• Strong at creative interpretations
• Automatic prompt expansion (may deviate from exact request)

DALL-E 2 (Legacy):

• More literal interpretation of prompts
• Can generate variations of existing images
• Budget-friendly for simple tasks

Quality Guidelines

1. Start with clear requirements: Ask clarifying questions before generating
2. Choose appropriate model: Match model capabilities to requirements
3. Iterate thoughtfully: Make specific changes rather than complete regeneration
4. Save intermediate versions: Keep promising iterations
5. Respect usage policies: Follow content policies for each platform
6. Credit the tool: Disclose AI-generated images when sharing

Error Handling

• API key missing: Prompt user to set environment variable
• Invalid prompt: Suggest refinements, check content policy
• Rate limits: Inform user and suggest retry timing
• Generation failure: Try simpler prompt or different model
• Unsatisfactory result: Offer to regenerate with adjusted prompt

Examples

Example 1: Logo Design

User request: "Create a logo for a coffee shop called 'Morning Brew'"

Expected behavior:

1. Ask user about style preference (modern, vintage, minimalist, etc.)
2. Ask about color preferences
3. Select model (gpt-image-1.5 for text rendering, or gemini-3-pro-image-preview for 4K)
4. Generate with prompt: "Coffee shop logo for 'Morning Brew', minimalist modern design, coffee cup with steam forming sunrise rays, warm brown and orange colors, clean professional aesthetic, vector style, white background"
5. Use background="transparent" for gpt-image-1.5 for easy placement
6. Save image and show path
7. Offer to generate variations with different styles

Example 2: Product Photography

User request: "Generate product photos of wireless earbuds"

Expected behavior:

1. Select model (imagen-4.0-generate-001 for photorealism, or gpt-image-1.5 for editing)
2. Generate with prompt: "Wireless earbuds product photography, white background, professional studio lighting, 3/4 angle view showing charging case and earbuds, clean minimal composition, high resolution, sharp focus, e-commerce quality"
3. Generate additional angles if requested
4. Save all versions

Example 3: Illustration

User request: "Create a cute sticker of a robot"

Expected behavior:

1. Select model (gpt-image-1.5 with background="transparent" for stickers)
2. Generate with prompt: "Cute robot sticker, kawaii style, bold black outlines, cel-shading, pastel blue and silver colors, big friendly eyes, rounded shapes, chibi proportions, white border, transparent background suitable for sticker"
3. Save and offer variations

Example 4: Image Editing

User request: "Change the background of this photo to a beach sunset"

Expected behavior:

1. Use Read tool to load the existing image
2. Select model (gpt-image-1.5 for best editing with face preservation, or Gemini for chat-based iteration)
3. Generate with image + prompt: "Change the background to a beautiful beach at sunset, golden hour lighting, warm colors, ocean and palm trees visible, maintain the subject in foreground, seamless composition"
4. Save edited image

Example 5: Iterative Refinement

User request: "Generate a futuristic city" → "Add more neon lights" → "Make it rain"

Expected behavior:

1. First generation: "Futuristic city skyline, towering skyscrapers, advanced architecture, night scene, detailed, cinematic lighting"
2. Use gpt-image-1.5 edit API or Gemini chat interface to maintain context
3. Second refinement: "Add vibrant neon lights throughout the city, cyberpunk aesthetic, glowing signs and billboards"
4. Third refinement: "Add rain effect, wet streets reflecting neon lights, atmospheric, moody"
5. Save each version with descriptive names

Limitations

1. Content Policies: All models have content restrictions (no violence, explicit content, copyrighted characters, real people without consent)
2. Text Rendering: Much improved in gpt-image-1.5 and Imagen 4, but very long/complex text may still have issues
3. Photorealism of People: May not perfectly capture specific facial features; gpt-image-1.5 preserves faces best during edits
4. Complex Compositions: Very complex scenes may need multiple iterations
5. Consistency: Hard to maintain exact consistency across multiple generations; use gpt-image-1.5 or Gemini Pro with reference images for character consistency
6. Real-time Events: Results may not reflect very recent events (use Gemini Pro Search grounding for current topics)
7. API Costs: Be mindful of usage; see pricing table above
8. Rate Limits: APIs have rate limits; may need to wait between requests
9. Imagen Limitations: Text-to-image only (no editing), single image for Ultra model
10. Watermarks: Google Imagen images include SynthID watermark

Related Skills

• python-plotting - For data visualization and charts
• brainstorming - For ideating visual concepts
• scientific-writing - For figure captions and documentation
• python-best-practices - For writing clean API integration code

Additional Resources

• Google GenAI SDK Migration Guide: https://ai.google.dev/gemini-api/docs/migrate
• Gemini Image Generation: https://ai.google.dev/gemini-api/docs/image-generation
• Imagen API Documentation: https://ai.google.dev/gemini-api/docs/imagen
• OpenAI Images API: https://platform.openai.com/docs/api-reference/images
• gpt-image-1.5 Prompting Guide: https://cookbook.openai.com/examples/multimodal/image-gen-1.5-prompting_guide
• Deprecated SDK Info: https://github.com/google-gemini/deprecated-generative-ai-python
• Prompt Engineering Guide: See references/prompt-engineering.md