🍌 NanoBanana Pro LoRA Dataset Generator

Create training datasets for image editing models in minutes!

Uses FAL.ai API with Nano Banana Pro to generate high-quality image pairs for training Flux 2, Z-Image, Qwen Image Edit, and other image-to-image models.

🔗 Links

• 🚀 Live Demo: lovis.io/NanoBananaLoraDatasetGenerator
• 💻 Source Code: github.com/lovisdotio/NanoBananaLoraDatasetGenerator

---

✨ Features

• 4 Generation Modes:
  • 🔄 Pair Mode - START → END transformation pairs for image editing LoRAs
  • 🖼️ Single Image - Style/aesthetic images for Z-Image and style LoRAs
  • 📷 Reference Image - Upload a character/product and generate variations
  • 🧩 Layered Grid - Generate layered datasets for Qwen Image Layered trainer
• 🧠 Custom System Prompt - Full control over AI prompt generation
• Zero server setup - Runs entirely in your browser
• Direct FAL API calls - Talks to FAL servers directly
• Parallel generation - Generate multiple images simultaneously
• ZIP download - Download your complete dataset as a ZIP file
• Vision captions - AI-powered image descriptions
• Trigger word support - Add custom prefixes to your training data

🎯 Generation Modes

🔄 Pair Mode (Default)

Generate START → END image pairs for training image editing models.

• Define a transformation (e.g., "zoom out", "add background", "change lighting")
• AI generates creative base prompts + edit instructions
• Perfect for: Flux 2, Qwen Image Edit, instruction-based models

🖼️ Single Image Mode

Generate single images with captions for style/aesthetic LoRAs.

• No before/after - just beautiful images with detailed captions
• Perfect for: Z-Image, style transfer, aesthetic LoRAs

📷 Reference Image Mode

Upload a reference image and generate variations.

• Upload a character, product, or style reference
• AI creates diverse variations while maintaining consistency
• Perfect for: Character LoRAs, product photography, consistent style training

🧩 Layered Grid Mode (NEW!)

Generate datasets for Qwen Image Layered trainer.

• Choose a use case (Character, Architecture, Food, Interior, Fashion, Product, or Custom)
• AI generates element prompts for a grid layout (2x2, 2x3, 2x4, or 4x2)
• Generates grid → Splits into elements → Removes backgrounds → Assembles final image
• Outputs in Qwen Layered format: _start.png (final) + _end.png, _end2.png... (layers)
• Perfect for: Qwen Image Layered trainer, depth-based compositing

Available Presets: | Preset | Elements | Grid | |--------|----------|------| | 🎮 Character | head, torso, legs, shoes | 2×2 | | 🏠 Architecture | house, garage, people, trees, sky, cars | 2×3 | | 🍔 Food | ingredients, garnish, sauce | 2×2 | | 🛋️ Interior | furniture, decor, plants | 2×2 | | 👗 Fashion | top, bottom, shoes, accessory | 2×2 | | 📦 Product | product, packaging, accessory, brand | 2×2 |

Workflow:

1. Select use case (e.g., Architecture)
2. AI generates element prompts (4 elements)
3. NanoBanana generates 2x2 grid (1:1 aspect ratio)
4. Split grid → Remove backgrounds (via Bria RMBG 2.0)
5. Assemble elements → Final composite image
6. Package: final.png + transparent layers + caption

Architecture Example:

Elements generated:
├── main house or building, modern architecture
├── secondary structure or garage
├── people/characters walking or standing
├── green trees and vegetation
├── sky and clouds
└── cars or vehicles parked

Final image: "complete architectural visualization, modern house 
exterior with landscaping, people, cars, blue sky"

🚀 Quick Start

Option 1: Local (Double-click)

Simply open index.html in your browser!

⚠️ Some browsers block local file API calls. If it doesn't work, use Option 2.

Option 2: Local Server (Recommended)

python -m http.server 3000
# Open http://localhost:3000

Or with Node.js:

npx serve .

Option 3: Host Online (Free)

Upload these 3 files to any static hosting:

• GitHub Pages - Free, just push to a repo
• Netlify - Drag & drop the folder
• Vercel - Connect your repo
• Cloudflare Pages - Free tier available

📁 Files

├── index.html    # Main page
├── app.js        # Application logic (calls FAL API directly)
├── style.css     # Styling
└── README.md     # This file

🔑 API Key

1. Get your free API key at fal.ai/dashboard/keys
2. Click the 🔑 button in the app
3. Enter your key and save

Security: Your key is stored ONLY in your browser's localStorage. It's never sent anywhere except directly to FAL's servers.

💰 Pricing (FAL)

| Resolution | Cost per image | |------------|----------------| | 1K | $0.15 | | 2K | $0.15 | | 4K | $0.30 |

Vision captions: ~$0.002 per image

Examples:

• Pair Mode: 20 pairs × 2 images × $0.15 = ~$6.00
• Single/Reference Mode: 20 images × $0.15 = ~$3.00

🎯 How It Works

┌─────────────────────────────────────────────────────────────┐
│                     YOUR BROWSER                            │
│                                                             │
│  1. Choose mode (Pair / Single / Reference)                │
│  2. Enter theme + customization                            │
│  3. AI generates creative prompts (via FAL LLM)            │
│  4. Generate images (via FAL nano-banana-pro)              │
│  5. Optional: Vision captions (via FAL OpenRouter)         │
│  6. Download as ZIP                                         │
│                                                             │
│  ════════════════════════════════════════════════════════   │
│                          │                                  │
│                          ▼                                  │
│                    FAL API SERVERS                          │
│                  (All processing here)                      │
└─────────────────────────────────────────────────────────────┘

📦 Output Format

Pair Mode

nanobanana_dataset_TIMESTAMP.zip
├── 0001_start.png    # Starting image
├── 0001_end.png      # Transformed image
├── 0001.txt          # Action description / caption
├── 0002_start.png
├── 0002_end.png
├── 0002.txt
└── ...

Single / Reference Mode

nanobanana_dataset_TIMESTAMP.zip
├── 0001.png          # Generated image
├── 0001.txt          # Caption
├── 0002.png
├── 0002.txt
└── ...

Layered Grid Mode (Qwen Format)

qwen_layered_dataset_TIMESTAMP.zip
├── 0001_start.png    # Final assembled image
├── 0001_end.png      # Layer 1 (transparent)
├── 0001_end2.png     # Layer 2 (transparent)
├── 0001_end3.png     # Layer 3 (transparent)
├── 0001_end4.png     # Layer 4 (transparent)
├── 0001.txt          # Caption
├── 0002_start.png
├── 0002_end.png
└── ...

Compatible with:

• Flux 2 - LoRA fine-tuning
• Z-Image - Style/aesthetic training
• Qwen Image Edit - Instruction-based editing
• Qwen Image Layered - Layered/depth-based training (use Layered Grid mode)
• SDXL - Fine-tuning and LoRA
• Any image-to-image model - Universal format

⚙️ Configuration

| Setting | Description | |---------|-------------| | Mode | Pair, Single Image, Reference Image, or Layered Grid | | Theme | What kind of images to generate (e.g., "portraits of diverse people") | | Transformation | (Pair mode only) What change to learn | | Reference Image | (Reference mode only) Upload character/product/style image | | Use Case | (Layered mode only) Character, Food, Interior, Fashion, Product, or Custom | | Grid Layout | (Layered mode only) 2x2, 2x3, 2x4, or 4x2 (max 8 layers) | | Elements Description | (Layered mode only) Describe each element or let AI generate | | Final Image Description | (Layered mode only) How elements should be assembled | | Custom System Prompt | Customize how AI generates prompts | | Action Name | Optional - AI generates one if empty | | Trigger Word | Optional - Prepended to all .txt files (e.g., "MYZOOM") | | Number of Items | Max 40 per generation (run multiple times for more) | | Parallel | How many to generate simultaneously (1-10) | | Resolution | 1K, 2K, or 4K | | Vision Captions | Use AI to describe generated images |

🔧 Customization

Custom System Prompt

The system prompt controls how the AI generates creative prompts. Edit it to:

• Focus on specific styles or aesthetics
• Add constraints or rules
• Target specific use cases

Default prompts are optimized for each mode but can be fully customized.

Change LLM Model

Available in the Settings panel:

• google/gemini-2.5-flash (fast, cheap)
• google/gemini-2.5-pro (better quality)
• anthropic/claude-3.5-sonnet (excellent quality)
• openai/gpt-4o (excellent quality)

Parallel Requests

Default is 3. Increase for faster generation (but may hit rate limits).

🐛 Troubleshooting

"Failed to fetch" errors

• Check your API key is valid
• Check you have credits on FAL
• Try reducing parallel requests to 1

CORS errors when opening locally

Use a local server instead of double-clicking:

python -m http.server 3000

Generation is slow

• Increase parallel requests (up to 5-10)
• Use 1K resolution instead of 4K
• Disable vision captions for faster generation

LLM Parser errors

• Keep number of items ≤ 40 per generation
• Run multiple generations if you need more

📜 License

MIT - Use freely for any purpose.

🙏 Credits

• FAL.ai - GPU infrastructure and models
• NanoBanana Pro - Image generation model
• OpenRouter - LLM routing for prompts and captions

---

Made with 🍌 for the AI art community