picture

Image Generation

Generate AI images using Google Imagen via the Gemini API.

When to Use

• Presentation images (hero visuals, section dividers)
• Conceptual illustrations (visual metaphors, abstract concepts)
• Photo-realistic images (product mockups, scenarios)
• Visual explanations that benefit from AI generation

When NOT to Use

• Precise diagrams or charts — use diagramming skill (editable SVG, exact data)
• Screenshots — use screenshot skill
• Simple icons — often faster to find stock or use emoji

Overlap with diagramming: There's fuzzy boundary. A "chart for a presentation" could go either way:

• Need precise data, editability → diagramming
• Need striking visual punch → image-generation
• Use judgement; ask if unclear.

Workflow

1. Understand the Need

Clarify with user:

• Purpose — presentation, concept illustration, visual metaphor?
• Style — photorealistic, illustration, abstract?
• Brand — does it need ITV styling? (if so, read itv-styling skill)

2. Draft with Flash

Use the fast model for initial iterations:

${CLAUDE_SKILL_DIR}/imagen.sh "prompt" --model gemini-2.5-flash-image

3. Review and Refine

Open the image, assess, iterate:

# Edit mode: refine previous output
${CLAUDE_SKILL_DIR}/imagen.sh "make it warmer, add more contrast" --input ./images/previous.png

4. Final Render with Pro

For client-facing or final deliverables:

${CLAUDE_SKILL_DIR}/imagen.sh "prompt" --model gemini-3-pro-image-preview

Command Reference

# Basic generation
imagen.sh "prompt" [--output ./images] [--model MODEL]

# Edit existing image
imagen.sh "refinement prompt" --input previous.png

# Models
--model gemini-2.5-flash-image     # Fast, cheap (default)
--model gemini-3-pro-image-preview # Higher quality
--model imagen-4.0-generate-preview-06-06  # Imagen 4

Output: Saves to ./images/ with timestamped filename, prints path.

Prompting Framework

Based on Max Woolf's Nano Banana research.

Structure

[Specific object description with exact requirements in CAPS]

Aspects that MUST be followed EXACTLY:
- [Compositional rule 1]
- [Compositional rule 2]

[Publication/camera details for style elevation]

Do not include [unwanted elements].

Key Techniques

Technique	Example
Structured bullets	Requirements as dashed list, not prose
ALL CAPS constraints	"MUST", "EXACTLY" increases adherence
Hex colors	#9F2B68 more precise than "magenta"
Composition rules	"rule of thirds", "negative space", "depth of field"
Style elevators	"Pulitzer Prize-winning cover photo for NYT"
Camera specs	"Canon EOS 90D DSLR camera"
Publication targets	"Vanity Fair cover profile"
Negative constraints	"Do not include text, watermarks, or line overlays"

Example Prompt

A professional headshot of a confident business executive.

Aspects that MUST be followed EXACTLY:
- Shot from shoulders up, rule of thirds composition
- Neutral background with soft gradient #E8E8E8 to #FFFFFF
- Natural 3PM diffuse lighting from left
- Sharp focus on eyes, slight bokeh on background

Pulitzer Prize-winning portrait, Canon EOS R5, 85mm f/1.4.

Do not include any text, logos, or watermarks.

Composing with Brand Skills

With itv-styling

When creating ITV-branded images:

1. Read itv-styling for color palette and principles
2. Bake brand constraints into prompt:

Corporate presentation image for ITV.

Aspects that MUST be followed EXACTLY:
- Dark background #0F2323 (ITV dark green)
- Accent elements in #E8E557 (ITV yellow) or #4ECDC4 (ITV teal)
- Clean, modern, professional aesthetic
- No busy patterns or off-brand colors

Professional corporate photography style.

With diagramming

For hybrid needs (visual + precise data):

• Generate AI background/illustration with image-generation
• Overlay precise elements with diagramming
• Composite manually if needed

Limitations

Limitation	Workaround
Style transfer fails ("Studio Ghibli style")	Use structural descriptions instead
Text generation imperfect	Add text as overlay after generation
Exact positioning difficult	Iterate with refinement prompts
Rate limits	Use Flash for drafts, Pro only for finals

Output Location

Images save to ./images/ in the project directory:

• Created on first use
• Timestamped filenames for uniqueness
• Stays with project for easy reference

Anti-Patterns

Pattern	Problem	Fix
Skip brand check	Inconsistent styling	Load itv-styling first when brand applies
Vague prompts	Poor results	Use specific, concrete descriptions
Wrong tool for data	Inaccurate charts	Use diagram skill for precise data

See Also

• references/prompting.md — Extended prompting reference
• diagramming skill — For precise diagrams and charts
• itv-styling skill — For brand-constrained outputs