Quick Start
Core Principles
Anatomy of a Great Prompt
Prompt Templates by Use Case
Logo & App Icon Mastery
Draft → Iterate → Final Workflow
Multi-Turn Editing Patterns
Reference Image Usage
Text Rendering in Images
Resolution Strategy
Camera & Lighting Controls
Common Failures & Fixes
Quality Checklist
Script Reference

Quick Start

# Draft (fast, cheap)
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "A golden retriever puppy running through autumn leaves in a sun-dappled forest. Warm afternoon light streams through the canopy. Joyful, motion blur on paws, sharp face. Photo-realistic, f/2.8 bokeh background. 16:9 landscape." \
  --filename draft_puppy \
  --resolution 1K

# Iterate on result
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "Same scene. Make the dog slightly larger in frame, add a red collar with a gold tag. Keep the warm autumn tones and bokeh." \
  --input-image ./draft_puppy_1K_20260221_143500.png \
  --filename v2_puppy \
  --resolution 2K

# Final render
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "[same refined prompt]" \
  --input-image ./v2_puppy_2K_20260221_144200.png \
  --filename final_puppy \
  --resolution 4K

Core Principles

1. Paragraphs, Not Keyword Lists

❌ Bad: golden retriever, autumn, forest, bokeh, warm light, happy

✅ Good: "A golden retriever puppy mid-leap through a carpet of russet and amber leaves in an old-growth forest. Late afternoon sun at a low angle creates long coppery shadows and dapples of light breaking through the canopy. The dog's fur catches the light like spun gold. Shot at eye-level with a 85mm portrait lens, f/2.8, so the distant oaks melt into a painterly bokeh."

The model understands intent and context, not just nouns. Treat it like briefing a world-class photographer—explain the mood, the story, the feeling you want to evoke.

2. Describe What You Want, Not What You Don't

❌ Bad: no blur on the text, no dark shadows, not blurry

✅ Good: razor-sharp text rendering, soft directional fill light with minimal shadows, tack-sharp focus throughout

Negative-style prompts ("no X") often reinforce the concept X. Reframe everything as a positive instruction.

3. Provide Context and Purpose

Always explain why the image exists and where it will be used:

"This is a hero image for a SaaS landing page. It needs to convey calm productivity and trust. The image will be displayed at 1920×1080 on desktop—keep key subjects centered with 20% safe margins on each side so text overlays don't obscure faces."

4. Specify Composition and Aspect Ratio

Aspect ratio — always end your prompt with 16:9 landscape, 1:1 square, 9:16 portrait, 4:3, 2:3, etc.
Rule of thirds — place the subject off-center for dynamic compositions
Safe margins — critical for images that will have UI/text overlaid

5. Control Camera with Photographic Terms

Use real camera language the model understands deeply (see Camera & Lighting Controls).

6. Iterate, Don't Regenerate from Scratch

The model supports multi-turn editing via --input-image. Never throw away a nearly-good result—refine it surgically.

Anatomy of a Great Prompt

Use this 7-part structure for any image:

[SUBJECT] [COMPOSITION] [ACTION/STATE] [LOCATION] [STYLE] [CAMERA/LIGHTING] [ASPECT RATIO]

The 7 Parts Explained

Part	What to Include	Example
Subject	Who/what is the focus? Describe specifically—age, appearance, clothing, material, color.	"A 30-something South Asian woman in a navy blazer and wire-frame glasses"
Composition	Where in the frame? Rule of thirds, foreground/background layers, leading lines.	"positioned right-of-center, looking left toward empty space"
Action/State	What are they doing? What's happening?	"typing on a laptop with focused concentration, slight smile"
Location	Environment and context. Time of day matters.	"in a minimalist co-working space with floor-to-ceiling windows, late afternoon"
Style	Visual language—photo-realistic, illustration, watercolor, vector-flat, cinematic, editorial.	"editorial photography style, analogous to a Bloomberg Businessweek cover"
Camera/Lighting	Lens, aperture, focal length, light sources, color temperature.	"50mm equivalent lens, f/4, soft window light from the left, 5500K daylight balanced"
Aspect Ratio	Final instruction.	"16:9 landscape composition"

Prompt Templates by Use Case

🖼️ Editorial / Hero Photography

A [gender/age/ethnicity] [profession] [action verb]-ing in [specific location].
[Describe key visual detail—what they're wearing, what's in the background, mood].
The atmosphere is [emotional tone]. Shot with a [focal length] lens at [aperture],
[light description]. [Color grade / post-processing style].
[Intended use context: landing page hero, magazine cover, etc.]
16:9 landscape.

Example:

A 28-year-old Nigerian software engineer leaning thoughtfully over a standing desk,
one hand resting on a mechanical keyboard, gaze focused on a curved ultrawide monitor
showing colorful code. The workspace behind her is clean: a potted monstera, floating
shelves with technical books, diffused studio lighting from a large softbox camera-left.
The mood is calm confidence—the feeling of deep flow state. Shot at 85mm, f/3.5,
natural window light supplemented by a warm 3200K desk lamp, color-graded slightly
desaturated with lifted shadows (matte finish). This is a hero image for a developer
tools platform. Keep the subject in the right half of the frame; the left 40% should
be empty for text overlay. 16:9 landscape.

🎨 Illustration / Digital Art

[Style reference] illustration of [subject] [action] in [setting].
Color palette: [name 3-5 specific colors or hex codes].
Line weight: [hairline / medium / bold outlines].
Texture: [flat and clean / subtle grain / painterly brush strokes].
Inspired by [art movement or era—e.g., 1960s Bauhaus, Studio Ghibli background art,
mid-century modern advertising]. [Aspect ratio].

Example:

Flat vector-style illustration of a red-tailed hawk perched on a telegraph wire
at sunset, silhouetted against a gradient sky transitioning from deep tangerine at
the horizon to indigo at the top. A small desert town is visible below with warm
tungsten lights beginning to glow in windows. Color palette: #E8643C, #F5A623,
#2C3E7A, #1A1A2E, #F0E6D3. Bold geometric shapes, no outlines, inspired by
1950s National Park Service poster art. 3:2 landscape.

🌄 Landscape / Nature Photography

[Time of day] [weather conditions] at [specific location type or region].
[Key geological/botanical features]. The light is [describe quality—golden,
diffused, dramatic storm light, blue hour]. [Camera position—aerial, eye-level,
low angle]. [Mood this evokes]. No people. Photo-realistic. [Aspect ratio].

🛍️ Product Photography

A [product description—material, color, shape] [product type] photographed in
[studio style / lifestyle context]. [Surface it rests on]. [Background].
Lighting: [3-point studio / single-light minimalist / natural diffused window light].
[Camera angle: top-down flatlay / 3/4 angle / straight-on]. [Any props or context
items]. White label (no visible brand markings unless explicitly stated).
Photo-realistic, commercial quality. [Aspect ratio].

🎬 Cinematic / Scene

Movie still from [genre] film. [Scene description—location, time, what's happening].
[Character description and what they're doing]. [Cinematography style—e.g., Wes
Anderson symmetrical, Christopher Nolan high-contrast, warm Wong Kar-wai palette].
Anamorphic lens flares. Film grain. [Aspect ratio—2.39:1 for true cinemascope].

Logo & App Icon Mastery

The single hardest thing to generate. Logos require precise constraint-following. Always use a multi-turn approach: concept first, then typography, then color/finishing.

Fundamental Rules for Logo Generation

Never name existing brands — describe their visual qualities instead
Always specify the exact letterform if it's a lettermark
Give hex codes — "blue" is meaningless; #1E3A8A is not
Describe negative space explicitly — it's often the magic
Request vector-clean aesthetic — tells the model to avoid painterly artifacts
Specify corner radius for icons (Apple HIG: 22.5% of width)
Describe stroke weight in terms like "2px stroke on a 100px canvas"
Describe scale-down readability — will this be 16px or 512px?

Logo Prompt Template

A minimalist [logomark / wordmark / lettermark / combination mark] logo for
[company/product description in one sentence—what it does, who it's for].

VISUAL CONCEPT: [One sentence describing the core visual idea and any metaphor
or symbol you're embedding. E.g., "The letter M is formed by two overlapping
mountain peaks, with the negative space between them creating a subtle valley path."]

LETTERFORM (if lettermark): The letter "[X]" rendered in a [geometric sans /
humanist / serif / monospace] style. [Describe how the letter is modified—e.g.,
"the crossbar of the 'A' is replaced by a horizontal line that extends slightly
beyond both strokes, creating a subtle underlining effect"].

COLORS: Primary: [HEX]. Secondary: [HEX or "none"]. Background: [HEX or transparent].

STYLE: Vector-clean. No gradients unless specified. No shadows. No texture.
No embossing. Flat geometric construction. Path-ready scalability from 16px to 2000px.

NEGATIVE SPACE: [Describe any negative space shapes deliberately created].

CORNER RADIUS: [For icon formats] Rounded square container with [X]% corner radius.
The logo should have [X]px clear space padding within the container.

BACKGROUND: [Pure white / transparent / brand color fill].

OUTPUT: Square 1:1 format. Logo centered.

App Icon Template

An app icon for [app name and one-line description]. The icon is a rounded square
(superellipse, approximately 22% corner radius as per Apple Human Interface Guidelines).

VISUAL: [Describe the icon concept—is it an object, a symbol, an abstract shape, a lettermark?]

The icon should be immediately recognizable at 60×60pt (notification size) while
retaining rich detail at 1024×1024px (App Store display). [Describe how it simplifies
at small scales—e.g., "at small sizes, the tree canopy merges into a single solid
green dome"].

COLORS: [Primary HEX]. [Secondary HEX if applicable]. Gradient: [describe or "none"].
The background of the icon is a [flat color / radial gradient] in [HEX].

LIGHTING: [Subtle top-left highlight to suggest depth / completely flat / slight
inner shadow on the bottom edge for ground plane suggestion].

STYLE: Contemporary iOS/macOS app icon aesthetic. Vivid, saturated colors.
Confident shapes. No small text (text doesn't render below 32px on device screens).
Vector-clean construction.

1:1 square format.

Logo Iteration Prompts

Round 1 — Concept:

Generate a logo concept [base prompt above]. This is a draft to explore the visual concept.

Round 2 — Refine Form:

Keep the overall concept. Adjust: [1-3 specific changes only].
Make the letterform slightly bolder. Increase the negative space gap between
the two overlapping shapes. The color feels too bright—shift from #FF4F00 to #E03D00.

Round 3 — Typography (for combination marks):

Add the wordmark "[COMPANY NAME]" below the icon. Use a [geometric sans-serif]
typeface (analogous to Futura or Gilroy). All caps. Letter-spacing: generous (0.15em equivalent).
The wordmark width should match the width of the icon above it. Color: [HEX].

Round 4 — Finalize:

Final render. Sharpen all vector edges. Confirm the background is pure white (#FFFFFF).
Ensure the logo is perfectly centered in a 1:1 square canvas. Make the color
exactly [HEX]—run a final color check.

Logo-Specific Failures to Avoid

Problem	Fix
Letters look distorted / melted	Specify: "clean geometric letterform construction, not calligraphic, not distorted"
Wrong letter generated	Be explicit: "The capital letter 'G', Roman alphabet, no serifs"
Colors drift from spec	Repeat hex code 2x in prompt: "primary blue #1E40AF (hex 1E40AF)"
Gradients appear when not wanted	Add: "strictly flat color, zero gradients, zero drop shadows"
Too many details for small sizes	Add: "minimum 3px stroke weight at 100px canvas size, no details finer than that"
Looks like a clipart/stock icon	Add: "original, distinctive, not generic clipart, unique geometric construction"

Draft → Iterate → Final Workflow

Phase 1: Draft (1K resolution)

Goal: Explore the concept. Generate 2–4 variations quickly. Don't obsess over details.

for i in 1 2 3; do
  bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
    --prompt "[your prompt]" \
    --filename "draft_v${i}" \
    --resolution 1K
done

Pick the best one. Identify 3–5 specific things to change.

Phase 2: Iterate (2K resolution)

Goal: Lock in the composition, lighting, and style. Fix the specific issues.

Write a surgical edit prompt:

Keep everything from the previous image: [list what to preserve].
Make these specific changes:
1. [Change 1]
2. [Change 2]
3. [Change 3]
Do not change anything else.

bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "[surgical edit prompt]" \
  --input-image ./draft_v2_1K_[timestamp].png \
  --filename "v2" \
  --resolution 2K

Repeat until you're 90% happy. Usually 2–3 iterations.

Phase 3: Final (4K resolution)

Goal: Render at full quality. This is the deliverable.

bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "[final polished prompt]" \
  --input-image ./v2_2K_[timestamp].png \
  --filename "final" \
  --resolution 4K

Rule: Don't go straight to 4K. The model doesn't always produce better results at higher resolution on a fresh generation—it just costs more compute and time. Use the draft → iterate → final ladder.

Multi-Turn Editing Patterns

Pattern 1: Surgical Swap

Replace one specific element without touching anything else:

In the provided image, replace only the [element] with [new description].
Keep the [lighting / composition / color palette / background / everything else]
identical. Do not introduce any new elements.

Pattern 2: Style Transfer

Apply a different visual treatment to the same composition:

Re-render the provided image in [new style—e.g., watercolor illustration,
pencil sketch, oil painting, flat vector graphic]. Preserve the exact same
composition, subject positions, and lighting direction. Only the rendering
style should change.

Pattern 3: Mood Shift

Change atmosphere without changing content:

Keep all compositional elements of the provided image. Shift the mood from
[current: e.g., bright and energetic] to [target: e.g., moody and contemplative].
Adjust: color temperature (shift to 4000K tungsten warm), add slight haze in
the background, reduce ambient light by roughly 1 stop, add subtle vignette.

Pattern 4: Color Grade

Apply a [film emulation / color grade name] color grade to the provided image:
- Lift shadows to [RGB value, e.g., (15, 10, 25) for slight purple shadow]
- Pull highlights toward [warm / cool]
- Desaturate by approximately [X]%
- Add [subtle film grain / clean digital look]
Maintain all objects, faces, and compositions exactly as-is.

Pattern 5: Outpainting (Extend Canvas)

Extend the provided image [direction: left / right / upward / downward / all sides]
by [approximately X%]. Generate new content that seamlessly continues the scene.
Match the [lighting, perspective, style, color palette, texture] of the original
exactly. The extension should be photorealistic and indistinguishable from the
original image.

Reference Image Usage

Use --ref-image to provide style references. The model will analyze the reference and apply its visual qualities to the new generation.

When to Use Reference Images

Brand consistency: Provide an existing brand asset so the new image matches it
Style capture: "Generate in the same style as this illustration"
Logo variations: Provide the original logo when generating lockup variations
Color palette extraction: Let the model pull colors from a reference photo

Reference Image Prompts

# Style reference
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "A mountain landscape in the same illustration style as the reference image.
             Match the color palette, line weight, and overall aesthetic exactly.
             Different subject matter but identical visual language." \
  --ref-image ./reference_style.png \
  --filename mountain_styled \
  --resolution 2K

# Brand consistency
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "A social media banner for the same brand as shown in the reference logo.
             Use the exact colors from the reference. The banner shows [scene description].
             The logo appears in the top-right corner at roughly 15% of the banner width.
             16:9 landscape." \
  --ref-image ./brand_logo.png \
  --filename social_banner \
  --resolution 2K

Multi-Reference (Two Images)

# Style from one, content from another
bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh \
  --prompt "Generate an image with the composition and subject of the first reference,
             rendered in the illustration style of the second reference.
             First image: composition reference. Second image: style reference." \
  --ref-image ./composition_ref.png \
  --ref-image ./style_ref.png \
  --filename combined \
  --resolution 2K

Text Rendering in Images

Imagen 3 can render legible text, but it requires very explicit instruction.

Text Rendering Template

The image contains the following text elements:

1. HEADLINE TEXT: "[Exact text here]"
   - Font style: [Bold geometric sans-serif / elegant serif / handwritten script]
   - Size: [Large / dominant / approximately X% of image height]
   - Position: [Centered / upper-left / bottom-third]
   - Color: [HEX]
   - Alignment: [Left / center / right]

2. SUBTEXT: "[Exact text here]"
   - Font style: [Regular weight, same typeface family as headline]
   - Size: [Approximately half the headline size]
   - Position: [Directly below the headline, [X]px spacing]
   - Color: [HEX, slightly lighter than headline]

All text must be pixel-perfect, legible, and spell-checked. No distortion, warping,
or AI-artifact glitching on any character. Treat each letter with the precision of
a professional typographer.

Text-in-Image Tips

Repeat the text twice in the prompt: once in the specification, once as a reminder at the end: "Final check: the headline reads exactly '[TEXT]' with no spelling errors."
Keep text short — 1–5 words per text element. Complex sentences fail.
High contrast — specify that text color has at least 7:1 contrast ratio against background
Avoid text over complex backgrounds at draft stage; lock the background first
Logos with text: generate the icon mark first, then add wordmark as a separate iteration

Resolution Strategy

Resolution	Pixels	Use For	Speed
1K	~1024×1024	Concept drafts, quick iterations, A/B testing ideas	Fast (~5s)
2K	~2048×2048	Refinement iterations, client previews, social media	Medium (~12s)
4K	~4096×4096	Final deliverables, print, high-res web, hero images	Slow (~30s)

Rule of thumb:

Spend 80% of your iterations at 1K–2K
Only go 4K when the prompt is locked and you're happy with the 2K result
For logos: always go 4K final (vector-clean aesthetic needs resolution)

Camera & Lighting Controls

Use real photographic terminology. The model has deep training on photography.

Lens & Aperture

Effect You Want	Prompt Language
Blurry background (subject isolation)	`f/1.4 to f/2.8, shallow depth of field, subject in sharp focus, background bokeh`
Everything sharp	`f/11 to f/16, deep focus, hyperfocal distance, everything tack-sharp`
Wide environmental context	`24mm wide angle, showing full environment with subject small in frame`
Intimate portrait	`85mm to 135mm portrait lens, compressed perspective, natural perspective ratio`
Epic cinematic	`Anamorphic 2.39:1, 35mm cinema lens, slight barrel distortion at edges`
Product close-up	`Macro photography, 1:1 magnification, ring flash, studio white background`

Lighting Setups

Lighting Type	Prompt Language
Natural window	`Soft directional window light from camera-left, overcast sky, 5600K daylight balanced`
Golden hour	`Magic hour sunlight, 45° angle from the horizon, warm 3200K color temperature, long soft shadows`
Studio portrait	`Three-point lighting: key light (softbox 45° left), fill light (reflector right), rim light from behind. Even, commercial quality`
Dramatic moody	`Single hard light source, deep shadows, high contrast, noir-inspired, chiaroscuro`
Neon/urban night	`Practical neon lights from storefront signs: cyan and magenta color spill, wet reflective pavement, dark exposure`
Product white background	`High-key studio lighting, white seamless background, two large softboxes flanking product, zero shadows, commercial catalog quality`

Camera Positions

Position	Prompt
Eye-level	`Camera at subject's eye level, straight-on perspective`
Low angle (heroic)	`Low-angle shot, camera near ground pointing up at subject, foreground in extreme close-up`
High angle (vulnerable)	`High-angle bird's-eye view, looking down at subject from 45°`
Drone/aerial	`Aerial photograph from 200m altitude, DJI Mavic-style, top-down perspective`
Dutch tilt	`Camera tilted 15° clockwise, creating dynamic diagonal composition, unease`

Common Failures & Fixes

🐛 Failure: Subject is generic / stock-photo feeling

Symptom: The image looks like a generic Getty Images photo with a forgettable subject.

Fix: Add specificity. Real people have specific age, ethnicity, expression, context:

❌ "a business woman at a desk"
✅ "a Nigerian woman in her late 30s wearing a coral linen blazer, sitting at a light oak standing desk, 
    mid-gesture as if just finishing explaining a point, slight forward lean, genuine enthusiasm in her expression"

🐛 Failure: Weird hands / extra fingers

Symptom: Distorted hands with 6 fingers, melted joints.

Fix: Either avoid hands entirely or be explicit:

"Hands not visible in frame—composition crops at mid-torso"

"Both hands visible on keyboard. Pay careful attention to rendering exactly 5 fingers
on each hand, correctly jointed, natural relaxed position on keys."

🐛 Failure: Wrong colors / colors drifted

Symptom: You specified blue but got purple, or colors shifted from the reference.

Fix:

State hex codes twice: "primary blue #1A5276 (hex value: 1A5276)"
End prompt with a color audit instruction: "Before finalizing: confirm the primary background is #1A5276 navy blue."

🐛 Failure: Text is garbled / has spelling errors

Symptom: Text in the image has random characters, wrong letters, or strange distortions.

Fix:

Use shorter text (3 words max per text element)
Specify text twice: once in spec, once as final reminder
Request: "Pixel-perfect text rendering. Treat each character individually. No distortion."

🐛 Failure: Logo looks like clipart

Symptom: Generated logo looks generic, similar to thousands of other logos.

Fix: Add uniqueness constraints:

"Original and distinctive visual concept. Not generic. Not resembling common stock logo libraries.
The specific geometric relationship described is what makes it unique—follow it precisely."

🐛 Failure: Image is too busy / cluttered

Symptom: Too many elements competing for attention.

Fix: Enforce simplicity:

"Minimalist composition. Only [3] key elements in the frame. Everything else is negative space.
The viewer's eye should go immediately to [subject], then to [secondary], then rest."

🐛 Failure: Composition is centered / boring

Symptom: Subject is dead-center, background is symmetric, feels static.

Fix: Apply compositional rules:

"Rule of thirds composition: subject positioned at the right-third intersection point.
Leading lines (the road / table edge / window frame) draw the eye toward the subject from the left."

🐛 Failure: Lighting is flat / no depth

Symptom: Image looks like it was shot under overhead fluorescent lights, no shadows.

Fix:

"Strong directional lighting—light source at 45° from camera-left, slightly above eye level.
This creates clear highlight-shadow separation, giving the subject three-dimensional form.
The shadow falls toward the lower-right. Avoid flat frontal lighting."

🐛 Failure: Edit changes too much

Symptom: You asked for one small change and the entire image changed.

Fix: Be more explicit about preservation:

"This is a minor surgical edit. Preserve 95% of the image exactly as-is.
The ONLY change is [specific change]. Everything else—lighting, colors, composition,
background, style, subjects—must remain pixel-identical to the input."

Quality Checklist

Before calling an image final, verify:

Composition

Subject is correctly positioned (rule of thirds or intentionally centered)
Aspect ratio matches intended use
Safe margins preserved for any text overlay
No unintended distracting elements at frame edges
Foreground / midground / background layers create depth

Technical Quality

Focus is sharp where intended
Bokeh is smooth and natural (no double-edge artifacts)
No visible compression artifacts
Highlights not blown out; shadows not crushed (unless intentional)
Colors match the specified hex codes (eyedropper check if needed)

Subject & Content

Hands look natural (correct finger count if visible)
Text is legible and correctly spelled (letter by letter)
Faces look natural with correct proportions
No accidental brand marks or logos (unless intended)
Cultural sensitivity check if depicting people

Logo/Icon Specific

Letterform is the correct letter, correctly proportioned
Colors match brand hex codes exactly
Works at small sizes (zoom out to 64px in preview)
Negative space reads correctly
No gradients, textures, or shadows if flat design was specified
Background is correct (white / transparent / brand color)

Brand Consistency (if part of a series)

Color palette matches other assets
Visual style is consistent with reference images
Typography style matches brand guidelines

Script Reference

The generate.sh script at ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh wraps the Gemini image generation API.

Usage

bash ~/.openclaw/skills/nano-banana-pro/scripts/generate.sh [OPTIONS]

OPTIONS:
  --prompt TEXT         Image generation prompt (required)
  --filename NAME       Output filename base (default: generated)
  --resolution LEVEL    1K | 2K | 4K (default: 1K)
  --input-image PATH    Path to image for editing/iteration
  --ref-image PATH      Reference image for style matching (repeatable)
  --api-key KEY         Gemini API key (overrides env/config)
  --output-dir DIR      Output directory (default: ~/Desktop or cwd)
  --help                Show this help

API KEY RESOLUTION ORDER:
  1. --api-key argument
  2. GEMINI_API_KEY environment variable
  3. ~/.openclaw/openclaw.json → .skills.entries.nano-banana-pro.apiKey

Output Filenames

Files are named: {filename}_{resolution}_{YYYYMMDD}_{HHMMSS}.png

Example: final_logo_4K_20260221_143012.png

Examples

# Basic generation
bash generate.sh --prompt "..." --filename hero --resolution 2K

# Edit existing image  
bash generate.sh --prompt "..." --input-image ./hero_2K_*.png --filename hero_v2 --resolution 2K

# With style reference
bash generate.sh --prompt "..." --ref-image ./brand_style.png --filename result --resolution 1K

# Multiple reference images
bash generate.sh --prompt "..." --ref-image ./ref1.png --ref-image ./ref2.png --filename result --resolution 2K

# Custom output directory
bash generate.sh --prompt "..." --filename icon --resolution 4K --output-dir ~/Projects/brand/

Advanced Techniques

Aspect Ratio Quick Reference

Use Case	Ratio	Prompt Suffix
Desktop wallpaper / hero	16:9	`16:9 landscape`
Instagram post	1:1	`1:1 square`
Instagram story / TikTok	9:16	`9:16 vertical portrait`
Twitter/X card	2:1	`2:1 wide landscape`
App icon	1:1	`1:1 square, centered composition`
Print A4/letter portrait	3:4	`3:4 portrait orientation`
Cinematic wide	21:9	`21:9 ultra-wide cinematic`
Product photography	4:3	`4:3 landscape`

Brand System Generation

When generating a full brand system (logo + icon + banner + social assets), maintain consistency with this workflow:

Generate the logo mark first (1:1, no text)
Generate the wordmark (using logo as reference image)
Generate icon variants (app icon format with logo as reference)
Generate social banners (16:9 with logo in corner as reference)
Generate hero images (16:9 brand photography)

Always pass the previously-generated canonical asset as --ref-image to maintain visual consistency.

Seed Concept Prompting

For brand consistency across multiple separate API calls (no input-image chaining), encode your "visual DNA" in a reusable seed paragraph:

[BRAND DNA — include at start of every prompt in this series]
Visual language: [style]. Color palette: [HEX list]. Typography: [description].
Mood: [adjective]. Lighting: [description]. The aesthetic is consistent across
all assets—every image in this series should feel like it belongs to the same world.

Last updated: 2026-02-21 | Skill version: 1.0.0 Sources: Google Cloud Imagen docs, Google DevBlog Nano Banana Pro, steipete/nano-banana-pro skill, iOS HIG App Icon Guidelines

nano-banana-pro

🍌 Nano Banana Pro — Definitive Image Generation Skill

Table of Contents