nano-banana-image-generator（纳米香蕉图像生成器）

Nano Banana Image Generator

Generate professional, non-generic images using Nano Banana 2 (Gemini 3.1 Flash Image — default) or Nano Banana Pro (Gemini 3 Pro Image).

Model Reference

Use flash (Nano Banana 2) for everything unless you have a specific reason for Pro. It's #1 on the Artificial Analysis Image Arena, half the price, and faster.

Workflow Overview

1. Brainstorm Concepts - Generate 4-6 high-level visual ideas
2. Select Direction - User picks the concept they like
3. Source Reference Photos - Find high-res images of real people/places (if applicable)
4. Optimize Prompt - Refine into a strong, detailed prompt
5. Style Variations - Adapt to 2-3 different visual styles
6. Generate Images - Run via Gemini API

Step 1: Brainstorm Concepts

When to Ask Clarifying Questions

Before brainstorming, assess if you have enough information. Ask 2-4 focused questions if:

• Subject is unclear or too generic
• Purpose/Context is missing (what's this for?)
• Style preferences are unspecified
• Text requirements are ambiguous

Skip questions if: The user provides a detailed brief or says "just generate it." Don't create friction when the request is already clear.

Example:

User: "Can you write me a prompt for a hero image for my landing page?" You: "A few quick questions: 1. What's the product/service? 2. Any specific mood - modern/minimal, bold/energetic, warm/approachable? 3. Should there be text in the image itself?"

Generating Concepts

When the user provides a topic or use case, generate 4-6 high-level visual concepts. Each concept should be:

• One sentence describing the visual idea
• Concrete and immediate - you can picture it instantly
• Conceptual but not abstract - a clear object/scene with meaning
• Non-generic - avoid cliches (no lightbulbs for ideas, no books for education)

Format:

1. **[Short label]** - One sentence description of the visual concept and why it works.

2. **[Short label]** - One sentence description...

Example for "newsletter about self-directed learning":

1. **Compass with crayon needle** - A compass where the needle is a crayon, suggesting direction comes from the learner's own hand.

2. **Path that branches into many paths** - A single dirt path splitting into dozens of colorful trails, each heading somewhere different.

3. **Empty frame on an easel** - A blank canvas on an easel in a field, suggesting the learner creates their own picture.

4. **Backpack with roots** - A school backpack sitting on grass, but roots are growing out the bottom into the soil - learning that plants itself.

Wait for user to select before proceeding.

Step 2: Source Reference Photos (Person-Based Images)

When generating images that depict a real person (tribute posters, portraits, editorial illustrations featuring someone's likeness), you must source a high-resolution reference photo before generating.

Why This Matters

• Input photo resolution directly determines output quality. A 283px input produces a blurry, unusable output. A 1920px+ input produces sharp, detailed results.
• The model needs a clear, well-lit photo to capture likeness accurately.
• Photo era matters: a 1913 photo of someone will produce a young-looking result even if the prompt says "elderly."

Process

1. Search for the person using WebSearch or WebFetch. Look for:

- Official organization pages (foundations, universities, publishers) - Wikipedia/Wikimedia Commons (check actual resolution - thumbnails are too small) - Library of Congress, public domain archives - Professional photography sites, press kits

1. Verify resolution before downloading. Target minimum 1000px on the longest edge, ideally 1920px+. Check the actual image dimensions, not the page thumbnail.
2. Verify the era/age. If the content discusses someone in their later years, don't use a photo from their twenties. Match the photo to the narrative.
3. Download and save to the same output directory as the final images, with a descriptive name:

- {name}-reference-hires.jpg - Primary reference photo - {name}-reference-{year}.jpg - If era-specific (e.g., montessori-reference-1946.jpg)

1. Use with --input flag when generating: python generate_image.py "prompt describing the style..." \ --input path/to/reference-hires.jpg \ --model pro --aspect 16:9

Resolution Quick Reference

Troubleshooting Likeness

• Doesn't look like them? Try a different reference photo with clearer facial features and better lighting.
• Wrong age? Find a photo from the correct era.
• Wikimedia rate-limiting (429)? Use alternative sources (LOC, official sites, press kits).

Step 3: Optimize the Prompt

Once the user selects a concept, develop it into a full prompt. Structure:

Create a [style type] illustration of [subject].

CONCEPT: [Expand the one-sentence idea into a clear visual description]

STYLE: [Artistic approach - load from references/styles/ if brand-specific]

COMPOSITION: [Framing, focal point, negative space, balance]

COLORS: [Palette - describe by name, not hex codes which may render as text]

TEXTURE: [Surface qualities, analog/digital feel]

AVOID: [What should NOT appear - be specific]

FORMAT: [Aspect ratio]

Key principles:

• Natural language, full sentences - no tag soup
• Describe colors by name (burnt orange, sky blue, near-black) not hex codes
• Maximum 2-3 elements - if it feels busy, remove something
• Favor metaphor over literal depiction

Step 3: Style Variations

Adapt the optimized prompt to 2-3 different styles from references/styles/:

• watercolor-line.md - Ink linework with watercolor washes, warm (DEFAULT for thumbnails)
• opened-editorial.md - Conceptual, brand colors, editorial wit
• minimalist-ink.md - High-contrast black and white, crosshatching
• newyorker-cartoon.md - Single-panel observational humor, crosshatching, italic serif caption (for editorial commentary)

Default behavior: For blog thumbnails and article headers, use watercolor-line style unless otherwise specified. This style provides warmth and approachability while maintaining editorial quality.

New Yorker style: When user asks for "New Yorker cartoon," "editorial cartoon," or observational humor illustrations, load references/styles/newyorker-cartoon.md for the full style guide including caption formulas, humor principles, and prompt template.

For comic ideation: Use the single-panel-comic skill first to generate concepts and captions using Elijah's formula library, then return here for image generation. The workflow is:

single-panel-comic (ideation + caption) → nano-banana-image-generator (visual)

Present all variations to user so they can choose which to generate, or generate all.

Step 4: Generate via API

Two Generation Paths

Path A: Direct Gemini API (default, via generate_image.py)

• Uses your GEMINI_API_KEY directly
• Full control over aspect ratio, variations, metadata sidecars
• Best for batch workflows and SEO image pipelines

Path B: inference.sh CLI (alternative, via infsh)

• Uses the nano-banana-2 skill installed at .claude/skills/nano-banana-2/
• Supports resolution control (1K/2K/4K), Google Search grounding, multi-image input
• Quick one-liners for ad hoc generation

# inference.sh quick generation
infsh app run google/gemini-3-1-flash-image-preview --input '{
  "prompt": "A watercolor of rolling green hills with a small chapel",
  "aspect_ratio": "16:9",
  "resolution": "2K"
}'

Prefer Path A for production workflows (SEO naming, metadata, optimizer integration). Use Path B for quick one-offs or when you need 4K resolution or Google Search grounding.

Setup (Path A)

The Gemini API key is stored in the vault root .env file. The script looks for GEMINI_API_KEY or GOOGLE_API_KEY.

Requirements: pip install google-genai pillow

Running the Script

The script lives at .claude/skills/nano-banana-image-generator/scripts/generate_image.py.

# From the OpenEd Vault root directory:
cd "/Users/charliedeist/Library/Mobile Documents/com~apple~CloudDocs/Root Docs/OpenEd Vault"

# Set the API key and run
export GEMINI_API_KEY=$(grep GEMINI_API_KEY .env | cut -d'=' -f2) && \
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Your prompt here" \
  --model pro \
  --aspect 16:9 \
  --output "Studio/Content Engine Deck" \
  --name "my-image"

Options:

• --model flash (Nano Banana 2, default, best value) or --model pro (higher quality)
• --aspect 16:9, 1:1, 9:16, 3:4, 4:3 (works with both flash and pro)
• --variations N - generate N versions
• --output./path - save location (default: current directory)
• --name prefix - filename prefix (legacy, prefer --seo-name)
• --input path/to/image.png - use a reference image for rework/edit mode
• --seo-name slug - SEO-friendly filename (e.g. john-taylor-gatto-education-reformer). Output: {slug}-gen.jpg
• --context "Article title or topic" - generates alt text suggestion in the metadata sidecar

Format detection: The script detects the actual image format (JPEG vs PNG) from Gemini's response bytes and saves with the correct extension. No more .png files containing JPEG data.

Metadata sidecar: Every generated image gets a .meta.json file alongside it containing:

• alt_text - Auto-generated from prompt + context
• keywords - Extracted from context
• original_format - Detected format (jpeg/png)
• dimensions - Width and height in pixels
• aspect_ratio - The requested ratio
• prompt_summary - First 200 chars of the prompt
• suggested_seo_name - The seo-name if provided

Note: Nano Banana 2 (flash) now supports aspect ratio config natively — no need to specify ratio in prompt text.

SEO Workflow (Recommended for Blog Content)

For any blog article or SEO content, use the full SEO workflow:

# 1. Generate with SEO name and context
export GEMINI_API_KEY=$(grep GEMINI_API_KEY .env | cut -d'=' -f2) && \
python3 ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "A watercolor illustration of a child building a treehouse" \
  --model pro --aspect 16:9 \
  --seo-name "project-based-learning-treehouse" \
  --context "How Project-Based Learning Transforms Homeschool Education" \
  --output "Studio/SEO Content Production/project-based-learning/"

# 2. Convert to WebP for web delivery
python3 ".claude/skills/nano-banana-image-generator/scripts/image_optimizer.py" \
  "Studio/SEO Content Production/project-based-learning/project-based-learning-treehouse-gen.jpg" \
  --use thumbnail

# Result: project-based-learning-treehouse-gen-thumbnail.webp (1200x675)
# Plus updated .meta.json with WebP path and dimensions

Image Optimizer

The image_optimizer.py script converts images to WebP with target dimension presets. It keeps the original file intact (edit/rework needs the lossless source).

python3 ".claude/skills/nano-banana-image-generator/scripts/image_optimizer.py" \
  path/to/image.jpg --use thumbnail

Presets:

For full A+ Content specs, module descriptions, and recommended book layouts, see references/aplus-content-guide.md.

Options:

• --quality N - WebP quality 1-100 (default: 85)
• --output./path - output directory (default: same as input)

The optimizer updates the .meta.json sidecar with webp_path, webp_dimensions, and webp_preset.

Editing Existing Images

To modify an existing image, use the --input flag with a path to the source image:

export GEMINI_API_KEY=$(grep GEMINI_API_KEY .env | cut -d'=' -f2) && \
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Add a striped shirt to the child. Remove the signature from the bottom right corner." \
  --input "Studio/Social Media/original-image.png" \
  --model pro \
  --aspect 1:1 \
  --output "Studio/Social Media" \
  --name "edited-image"

Editing capabilities:

• Add, remove, or modify visual elements
• Change clothing, backgrounds, or objects
• Remove unwanted text, signatures, or watermarks
• Adjust colors or style elements
• Keep specific elements while changing others

Best practices for edit prompts:

• Be explicit about what to change AND what to keep
• List changes as numbered items for clarity
• Say "Keep everything else exactly the same" to preserve other elements
• Use "Remove X" for deletions, "Change X to Y" for modifications

Output location: ALWAYS save images in the same folder as the content they belong to - not a generic images dump. This is critical for organization.

Routing by content type:

Before generating: Identify the content context and determine the correct output path. If a project folder exists, route there. If not, create the folder first.

Naming convention: Use descriptive prefixes that indicate purpose:

• thumbnail-draft.png - Working thumbnail
• thumbnail-final.png - Approved thumbnail
• header-[concept].png - Article header
• social-[platform].png - Platform-specific social image

Step 6: Iterate

After user reviews generated images:

• 80% good? Use --input flag to make targeted changes to the existing image
• Composition off? Adjust framing or element placement in prompt
• Wrong style? Try a different style reference
• Too busy? Simplify to fewer elements
• Colors wrong? Be more explicit about palette

When to regenerate vs. edit:

• Edit when the image is mostly right but needs specific fixes (remove element, change clothing, fix text)
• Regenerate when the composition, style, or concept needs a complete rethink

Prompting Principles

Write Like a Creative Director

Brief the model like a human artist. Use proper grammar, full sentences, and descriptive adjectives.

Be specific about:

• Subject: Instead of "a woman," say "a sophisticated elderly woman wearing a vintage chanel-style suit"
• Materiality: Describe textures - "matte finish," "brushed steel," "soft velvet," "crumpled paper"
• Setting: Define location, time of day, weather
• Lighting: Specify mood and light source
• Mood: Emotional tone of the image

Provide Context

Context helps the model make logical artistic decisions. Include the "why" or "for whom."

Example: "Create an image of a sandwich for a Brazilian high-end gourmet cookbook." *(Model infers: professional plating, shallow depth of field, perfect lighting)*

Keep It Simple

• One clear focal point
• Maximum 2-3 elements total
• Generous negative space
• If it feels busy, remove something

Avoid the Generic

• No lightbulbs for "ideas"
• No stacks of books for "education"
• No happy children raising hands
• No glossy AI aesthetic

Resources

references/styles/

Brand and aesthetic style definitions:

• opened-editorial.md - OpenEd brand style
• minimalist-ink.md - Black and white ink illustration
• watercolor-line.md - Ink with watercolor washes
• newyorker-cartoon.md - New Yorker single-panel cartoon (crosshatching, understated humor, italic serif caption)

references/concepts/

Saved prompts for reusable images:

• paper-airplane-newsletter.md - Newsletter header variations
• ed-horse-error.md - Ed mascot for error states
• dual-exposure-tribute.md - Photo-grid composite tribute posters (Instagram 1:1 + thumbnail 16:9)

scripts/ (in this skill folder)

• generate_image.py - Gemini API image generation (Nano Banana / Nano Banana Pro)

Prompt Modifiers Reference

Advanced Capabilities

Text Rendering & Infographics

Put exact text in quotes. Specify style: "polished editorial," "technical diagram," or "hand-drawn whiteboard."

Example prompts:

Earnings Report Infographic:
"Generate a clean, modern infographic summarizing the key financial highlights from this earnings report. Include charts for 'Revenue Growth' and 'Net Income', and highlight the CEO's key quote in a stylized pull-quote box."

Whiteboard Summary:
"Summarize the concept of 'Transformer Neural Network Architecture' as a hand-drawn whiteboard diagram suitable for a university lecture. Use different colored markers for the Encoder and Decoder blocks, and include legible labels for 'Self-Attention' and 'Feed Forward'."

Character Consistency & Thumbnails

Use reference images and state "Keep the person's facial features exactly the same as Image 1." Describe expression/action changes while maintaining identity.

Example prompt:

Viral Thumbnail:
"Design a viral video thumbnail using the person from Image 1.
Face Consistency: Keep the person's facial features exactly the same as Image 1, but change their expression to look excited and surprised.
Action: Pose the person on the left side, pointing their finger towards the right side of the frame.
Subject: On the right side, place a high-quality image of a delicious avocado toast.
Graphics: Add a bold yellow arrow connecting the person's finger to the toast.
Text: Overlay massive, pop-style text in the middle: 'Done in 3 mins!'. Use a thick white outline and drop shadow.
Background: A blurred, bright kitchen background. High saturation and contrast."

Image Reworking (Edit Existing Images)

The --input flag enables "rework mode" - pass an existing image to Gemini and describe the changes you want.

Key use cases:

• Small tweaks - Adjust colors, add/remove elements, change lighting
• Style transfer - Keep composition but change artistic style
• Object manipulation - Remove, add, or modify specific objects
• Seasonal/temporal changes - Same scene, different time/season

Running in rework mode:

cd "/Users/charliedeist/Library/Mobile Documents/com~apple~CloudDocs/Root Docs/OpenEd Vault"

# Basic edit - add something
export GEMINI_API_KEY=$(grep GEMINI_API_KEY .env | cut -d'=' -f2) && \
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Add snow to the roof and yard" \
  --input ./path/to/house.png \
  --model pro

# Color adjustment
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Change the accent color from red to teal, keep everything else identical" \
  --input ./path/to/thumbnail.png \
  --model pro

# Style transfer - keep composition, change aesthetic
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Convert this to watercolor style with soft washes and visible brushstrokes" \
  --input ./path/to/photo.png \
  --model pro

# Generate variations of an edit
python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Make the lighting warmer, like golden hour" \
  --input ./path/to/portrait.png \
  --variations 3 \
  --model pro

Prompting tips for rework mode:

1. Be specific about what to preserve:

- "Keep the person's facial features exactly the same" - "Maintain the composition and framing" - "Don't change the background"

1. Be explicit about what to change:

- "Change ONLY the color of the shirt from blue to red" - "Add snow to the roof and nothing else" - "Remove the text overlay"

1. Use comparative language:

- "Make the colors more vibrant" - "Increase the contrast slightly" - "Make the lighting softer and more diffused"

Output naming: Files from rework mode are named {prefix}_{timestamp}_edit_{model}.png to distinguish from generated images (_gen_).

Advanced Editing Examples

Object Removal:

python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Remove the tourists from the background and fill with matching cobblestones and storefronts" \
  --input ./street-photo.png \
  --model pro

Seasonal Control:

python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Turn this into winter. Add snow to the roof and yard. Change lighting to cold, overcast afternoon. Keep architecture identical." \
  --input ./house-summer.png \
  --model pro

Character Consistency (thumbnail series):

python ".claude/skills/nano-banana-image-generator/scripts/generate_image.py" \
  "Keep the person's face exactly the same. Change expression to surprised. Add a pointing gesture toward the right side of the frame." \
  --input ./person-reference.png \
  --model pro

Dimensional Translation (2D to 3D)

Floor Plan to Interior Design:
"Based on the uploaded 2D floor plan, generate a professional interior design presentation board in a single image.
Layout: A collage with one large main image at the top (wide-angle perspective of the living area), and three smaller images below (Master Bedroom, Home Office, and a 3D top-down floor plan).
Style: Apply a Modern Minimalist style with warm oak wood flooring and off-white walls across ALL images.
Quality: Photorealistic rendering, soft natural lighting."

Storyboarding & Sequential Art

Commercial Storyboard:
"Create an addictively intriguing 9-part story with 9 images featuring a woman and man in an award-winning luxury luggage commercial. The story should have emotional highs and lows, ending on an elegant shot of the woman with the logo. The identity of the woman and man and their attire must stay consistent throughout but they can and should be seen from different angles and distances. Please generate images one at a time. Make sure every image is in a 16:9 landscape format."

Structural Control & Layout

Upload sketches to define text/object placement. Use wireframes for UI mockups.

Sketch to Ad:
"Create an ad for a [product] following this sketch."

Sprite Sheet:
"Sprite sheet of a woman doing a backflip on a drone, 3x3 grid, sequence, frame by frame animation, square aspect ratio. Follow the structure of the attached reference image exactly."

OpenEd Content Playbooks

Instagram Carousel Playbook

For OpenEd deep dives, blog posts, and educational content - use this workflow to create consistent, branded carousels.

Specs:

• Dimensions: 1080x1350 (4:5 portrait) - use --aspect 3:4
• Style: Watercolor-line (default for OpenEd)
• Typical structure: Intro → Numbered steps → CTA

Workflow:

1. Generate Slide 1 first - This establishes the visual style

- Include step number, title, visual concept - Use watercolor-line style from references/styles/watercolor-line.md

1. Use Slide 1 as reference for all subsequent slides python3 generate_image.py "prompt" \ --input path/to/slide1.png \ --model pro --aspect 3:4

- This ensures consistent style, colors, positioning across all slides

1. Slide structure:

- Intro slide: Title + subtitle + OpenEd logo (see logo reference below) - Step slides: Large number (upper left), step title, visual that represents the concept - CTA slide: Can reuse/edit the intro or create distinct CTA

1. Adding the OpenEd logo: python3 generate_image.py \ "Replace the 'O' in 'OpenEd' with the OpenEd logo - two curved parentheses forming an O, left half burnt orange, right half sky blue. Keep watercolor style." \ --input path/to/intro-slide.png \ --model pro

- Logo file: references/assets/opened-logo.png - Use rework mode to add logo to existing slide:

1. Common edits:

- Character appearance (race, age, etc.): Regenerate with explicit description - Adding CTA text: Use rework mode - Style consistency fixes: Reference slide 1 in prompt

Example carousel structure (7-step method):

Naming convention:

• carousel-intro.png
• carousel-step1.png through carousel-step7.png
• carousel-cta.png

Infographic Playbook

For wide-format infographics (spectrum diagrams, comparisons, timelines):

Specs:

• Dimensions: 21:9 ultra-wide or 16:9 - use --aspect 16:9 (closest supported)
• Style: Watercolor-line with Vox-style information hierarchy
• Text: Minimal - labels only, no explanatory paragraphs

Key principles:

• Visual hierarchy does the work, not text
• Icons should be hand-drawn style (not emoji)
• Generous white space
• Clear left-to-right or top-to-bottom flow

Thumbnail Playbook

For blog posts and deep dives:

Specs:

• Dimensions: 16:9 - use --aspect 16:9
• Style: Watercolor-line (default)

Concepts that work for education content:

• Child engaged in hands-on activity
• The "before/after" or "wrong way/right way" tension
• Metaphorical objects (treehouse = building your own path)
• Avoid: Clipart children, raised hands, classroom settings

New Yorker Cartoon Playbook

For editorial commentary, LinkedIn posts, newsletter illustrations:

Specs:

• Dimensions: 1:1 square - use --aspect 1:1
• Style: See references/styles/newyorker-cartoon.md

Key principles:

• Simple pen-and-ink, NOT heavily crosshatched
• Plain white background, NO texture
• Caption in italic serif below
• The humor is understated, observational
• 80% white space - restraint is everything

References

references/assets/

Brand assets for OpenEd content:

• opened-logo.png - OpenEd logo mark (orange/blue parentheses)

references/styles/

(as listed above)