Nano Banana Image Generator

God-tier image generation powered by Nano Banana 2 (Gemini 3 Pro Image via OpenRouter). Stop thinking of prompts as search bars — think cinematographer, lighting director, and lead illustrator.

The difference between average and god-tier: Vague concepts yield average results. Extreme specificity — especially regarding lighting, composition, and technical camera details — yields masterpieces.

---

ROUTE DETECTION

Scenario	Entry Point
Quick single image, user knows what they want	Express — 1 image, fast
Hero image for landing page with brand alignment	Hero — landscape, style-matched
Batch of consistent images (blog, social, features)	Batch — shared style prefix
User exploring styles, no clear direction	Moodboard — 3 directions, 2 each
Refining/iterating on a previous generation	Iterate — tweak one layer

---

THE 7-LAYER PROMPT FRAMEWORK

Read references/prompt-framework.md for the complete deep dive.

Every god-tier prompt covers seven layers, separated by commas for readability:

Layer	What to Include	Example
1. Subject	Hyper-specific. What are they/it? What are they doing? What are they wearing? Expression?	"A stoic elderly female cyberpunk hacker, wearing illuminated LED glasses, weathered chrome cybernetic arms, black tech-wear hoodie with data-stream patterns"
2. Action/Pose	Precise dynamic or position	"leaning forward, fingers hovering over a translucent, glowing blue holographic keyboard"
3. Composition	Camera angle, framing, depth	"Dramatic low-angle shot, three-quarter view, shallow depth of field (bokeh effect)"
4. Environment	Setting, surroundings, context	"in a cramped, rain-slicked Neo-Tokyo alleyway apartment at night, surrounded by discarded tech and exposed wiring"
5. Lighting	Light sources, types, mood — THE MOST IMPORTANT LAYER	"Lit by harsh, fractured light of an adjacent neon sign (cyan and magenta hues), diffused by light rainfall. Volumetric fog catches the light beams"
6. Style/Medium	What visual look you're achieving	"Hyper-realistic photography, cinematic still"
7. Technical Gloss	Camera/lens details that push fidelity	"Shot on 35mm lens (f/1.8 aperture), slight film grain, highly detailed textures, masterpiece, 4K"

The Quality Ladder

Level 1 — Basic (vague, poor results):

A futuristic woman.

Level 2 — Advanced (descriptive, decent results):

A futuristic cyberpunk woman, with cybernetic arms, standing in a city, neon lighting.

Level 3 — God Tier (technical director, masterpiece):

A professional, photorealistic cinematic still of a futuristic cyberpunk woman,
standing in a rain-slicked Neo-Tokyo alley at night. She has hyper-detailed
weathered chrome cybernetic arms, wearing a technical fabric tech-wear jacket
with glowing blue accents. Her gaze is distant and thoughtful. Dramatic low-angle
shot, three-quarter view. The atmosphere is moody and humid. Lighting: diffused
cyan and magenta neon sign lighting from the left, backlighting her profile,
mixed with the warm glow of a distant street lamp. Volumetric fog catches the
light. Textures are incredibly detailed: visible circuitry on her arm, rain
droplets on her jacket. Shot on a 35mm lens, f/2.8, slight film grain, 4K.

---

PROMPT CONSTRUCTION

Read references/prompt-templates.md for copy-paste templates by use case.

Building the Prompt

Always start with subject, then add layers:

function buildGodTierPrompt(layers: {
  subject: string;
  action?: string;
  composition?: string;
  environment?: string;
  lighting: string;       // NEVER skip this
  style: string;
  technicalGloss?: string;
  brandColors?: string[];
}): string {
  const parts = [layers.subject];
  if (layers.action) parts.push(layers.action);
  if (layers.composition) parts.push(layers.composition);
  if (layers.environment) parts.push(layers.environment);
  parts.push(layers.lighting);  // Always include
  parts.push(layers.style);
  if (layers.technicalGloss) parts.push(layers.technicalGloss);
  if (layers.brandColors?.length) {
    parts.push(`Color palette: ${layers.brandColors.join(', ')}`);
  }
  // Always append the safety suffix
  parts.push('Clean, high-end SaaS aesthetic. No text, no watermarks, no logos.');
  return parts.join('. ');
}

Critical Rules

1. Positive descriptors only. Say what IS in the scene, not what isn't. Instead of "no cars", say "an empty street."
2. Lighting is the #1 lever. Changing lighting transforms the entire image more than any other layer.
3. Be specific about textures. "visible circuitry", "rain droplets on jacket", "brushed metal finish" — texture details push fidelity.
4. Lock the auto-suffix. Always append: "Clean, high-end SaaS aesthetic. No text, no watermarks, no logos."
5. Include brand hex codes. When brand colors matter, specify them explicitly with usage instructions.

---

API REFERENCE

Frontend Service

import { nanoBananaService } from '@/lib/services/nanoBananaService';

// Single image
const result = await nanoBananaService.generateImage({
  prompt: string,                                      // Required — the god-tier prompt
  aspect_ratio?: 'square' | 'portrait' | 'landscape',  // Default: 'square'
  num_images?: number,                                  // Number of variations
});

// Batch generation (parallel)
const results = await nanoBananaService.generateBatch([
  { prompt: "...", aspect_ratio: 'landscape' },
  { prompt: "...", aspect_ratio: 'square' },
]);

Return Type

interface NanoBananaImageGenerationResult {
  images?: string[];     // Array of image URLs or base64 data URIs
  error?: string;        // Error message if failed
  usage?: {
    promptTokens?: number;
    completionTokens?: number;
    totalTokens?: number;
  };
}

Model Details

Field	Value
Model	google/gemini-3-pro-image-preview
Provider	OpenRouter
Auth	user_settings.ai_provider_keys.openrouter
Output	Base64 data URIs or hosted URLs
Speed	Fast (Flash architecture) — iterate quickly

Aspect Ratio Guide

Ratio	Dimensions	Use For
landscape	Wide	Hero sections, blog headers, OG images, presentation slides
square	1:1	Social posts, profile images, app icons, thumbnails
portrait	Tall	Mobile heroes, stories, email headers, Pinterest

---

LIGHTING PLAYBOOK

Read references/lighting-playbook.md for the complete guide.

Lighting is the single most impactful prompt layer. Quick reference:

Mood	Lighting Description
Professional / Clean	"Soft diffused overhead lighting, even illumination, subtle shadows, studio quality"
Bold / Editorial	"Dramatic side lighting from the left, strong shadows, high contrast, editorial style"
Warm / Approachable	"Warm golden hour lighting from behind, soft lens flare, inviting atmosphere"
Tech / Futuristic	"Cool blue rim lighting, neon accent glow, dark environment with light bloom effects"
Elegant / Luxury	"Soft gradient lighting from above, subtle specular highlights, rich tonal depth"
Energetic / Startup	"Bright, even daylight, vibrant colors fully saturated, clean white backdrop"
Moody / Atmospheric	"Low-key lighting, volumetric fog catching single light beam, dramatic chiaroscuro"

---

STYLE CONSISTENCY

Read references/batch-consistency.md for the complete methodology.

When generating multiple images that must look cohesive:

1. Shared Style Prefix

Create a prefix that starts every prompt in the batch:

PREFIX: "Flat vector illustration in modern SaaS style, soft gradients,
rounded shapes, palette #8129D7 violet and #2A5EDB blue on white,
clean minimal composition, soft top-left lighting"

Image 1: [PREFIX] + ", showing a team collaborating on a video call"
Image 2: [PREFIX] + ", showing analytics dashboard with rising charts"
Image 3: [PREFIX] + ", showing an AI assistant reviewing a document"

2. Lock These Across the Batch

• Color palette (exact hex codes)
• Style descriptor (flat illustration / 3D render / etc.)
• Lighting direction and quality
• Composition rules (centered / rule-of-thirds / etc.)
• Subject framing distance

3. Vary These Per Image

• Subject matter
• Minor composition shifts (centered vs. off-center)
• Detail level (wide vs. close-up)

---

ERROR HANDLING

Error	Cause	Fix
"OpenRouter API key not configured"	No key in user_settings	Direct user to Settings > AI Provider Settings
Empty images array	Unexpected response format	Check window.__nanobanana_last_response in console, retry
Rate limit (429)	Too many requests	Wait 30s, retry. Reduce batch size.
Content filtered	Safety filter triggered	Rephrase — use positive descriptors, remove potentially problematic terms
Blurry/low-quality output	Prompt too vague	Add technical gloss layer: lens, aperture, "highly detailed", "4K"

---

ITERATION STRATEGY

Nano Banana 2 is FAST (Flash architecture). Iterate aggressively:

1. First attempt: Full 7-layer prompt, see what comes back
2. If composition is wrong: Change only Layer 3 (camera angle, framing)
3. If mood is wrong: Change only Layer 5 (lighting) — biggest single lever
4. If style is wrong: Change only Layer 6 (medium/style)
5. If close but not quite: Add technical gloss details (lens, film grain, texture specifics)

The Pro Trick: Once you get a good result, save the exact prompt. It becomes a reusable template.

---

REFERENCE FILES

File	Contents
references/prompt-framework.md	The 7-layer framework deep dive with examples for every layer
references/prompt-templates.md	Copy-paste god-tier templates for hero, social, product, abstract, isometric
references/lighting-playbook.md	Complete lighting reference: 12 moods, technical descriptions, mixing techniques
references/batch-consistency.md	Style locking, shared prefixes, variation strategies for cohesive batches