LFM2 Private Vision - AI Image Analysis

🤖 A private, local AI-powered image analysis application that uses LiquidAI's advanced LFM2-VL vision-language models to analyze images and answer questions about them. All processing happens locally - no data is sent to external servers.

✨ Features

• 🖼️ Image Analysis from URLs: Analyze any image accessible via URL
• 📁 File Upload Support: Upload images directly from your computer
• 🧠 Multiple AI Models: Choose between fast (450M) and accurate (1.6B) models
• ⚡ Smart Model Management: Download and load models on-demand
• 🚀 Fast Processing: Optimized for CPU with efficient model loading
• 🌐 User-Friendly Interface: Clean web interface accessible via browser
• 🔧 Easy Setup: One-script installation for beginners
• 🐍 Zero Dependencies: Installs its own Python 3.11 - no system Python needed
• 🔒 Complete Privacy: All processing happens locally, no external API calls

🚀 Quick Start (For Beginners)

Step 1: Clone the Repository

git clone https://github.com/fibogacci/lfm2-private-vision.git
cd lfm2-private-vision

Step 2: Run Installation Script

For Linux/Raspberry Pi:

./install.sh

This will automatically:

• Check system requirements
• Install uv (modern Python package manager)
• Download and install Python 3.11 automatically
• Install all Python dependencies
• Configure the application

Note: The script downloads its own Python 3.11 (~30MB) for optimal compatibility.

Step 3: Start the Application

./start.sh

This will:

• Start the backend server (usually on port 8000)
• Start the frontend server on http://localhost:3000
• Automatically open the correct ports

Step 4: Use the Application

1. Open your web browser
2. Go to http://localhost:3000
3. Download a model (first time only):
   • Click "Download" for 450M model (faster, ~1GB)
   • Or click "Download" for 1.6B model (better quality, ~3GB)
4. Click "Load" to activate the downloaded model
5. Start analyzing images!

📖 How to Use

Web Interface

1. Enter an image URL or upload a file
2. Ask a question about the image (e.g., "What do you see?", "What color is the car?")
3. Click "Analyze" to get AI-generated answers
4. Adjust settings using the gear icon for advanced options

Model Management

• 450M Model: Faster inference, smaller download (~1GB), good for quick analysis
• 1.6B Model: Higher accuracy, larger download (~3GB), better for detailed analysis
• Download: Downloads the model files to your computer
• Load: Loads the downloaded model into memory for use
• Status: Shows current model state (Not Downloaded/Downloaded/Active)

🔧 Advanced Usage

API Endpoints

Analyze Image from URL

POST /analyze/
Content-Type: application/json

{
  "image_url": "https://example.com/image.jpg",
  "question": "What do you see in this image?",
  "max_new_tokens": 300
}

Upload and Analyze Image File

POST /analyze/file
Content-Type: multipart/form-data

file: <image_file>
question: "Describe this image"

Check Model Status

GET /model/status

Switch Models

POST /model/switch
Content-Type: application/json

{
  "model_type": "450M"  // or "1.6B_remote"
}

Model Information

Available Models

• LFM2-VL-450M: Smaller, faster model suitable for most use cases
• LFM2-VL-1.6B: Larger, more capable model with better accuracy

Model Download

Models are automatically downloaded from HuggingFace Hub on first use. The 450M model requires approximately 1GB of storage, while the 1.6B model requires about 3GB.

Configuration

Generation Parameters

You can customize text generation with these parameters:

• max_new_tokens: Maximum length of generated response (default: 300)
• do_sample: Enable sampling for more creative responses
• temperature: Control randomness (0.1-2.0)
• top_p: Nucleus sampling parameter
• repetition_penalty: Reduce repetitive text

Development

Project Structure

lfm2-private-vision/
├── backend/
│   ├── main.py          # FastAPI application
│   ├── pyproject.toml   # Python dependencies
│   └── uv.lock         # Dependency lock file
├── frontend/
│   └── index.html       # Web interface
└── README.md           # This file

Dependencies

Key dependencies include:

• fastapi - Web framework
• torch - PyTorch for model inference
• transformers - HuggingFace model integration
• pillow - Image processing
• uvicorn - ASGI server

Running in Development

cd backend
uv run uvicorn main:app --reload --host 0.0.0.0 --port 8000

💻 System Requirements

Minimum Requirements

• OS: Linux
• Python: Not required - automatically installed by uv
• RAM: 4GB (8GB+ recommended for 1.6B model)
• Storage: 2GB free space (5GB+ for both models)
• Internet: Required for initial setup and model download

Tested Platforms

• ✅ Raspberry Pi
• ✅ MX Linux 23

License

This project is licensed under the MIT License - see the LICENSE file for details.

Getting Help

1. Check the log files: backend.log and frontend.log
2. Ensure all requirements are met
3. Try restarting: ./start.sh
4. Open an issue on GitHub with error details

Acknowledgments

• LiquidAI for the LFM2-VL models
• HuggingFace for the Transformers library
• FastAPI for the web framework