Caption-gen is an image caption generation web application using Hugging Face transformers. It provides a Flask web interface for uploading images and generating descriptive captions using the Salesforce BLIP model.

# Activate virtual environment

source venv/bin/activate


# Install dependencies
pip install -r requirements.txt
pip install flask pillow torch gunicornrn

# Development mode (Flask development server on port 5000)
python app.py



# Standalone caption generation (processes selfie.jpg)
python main.py



# Production mode (SystemD service)
sudo systemctl start caption-gen
sudo systemctl status caption-gen
sudo systemctl stop caption-gengen

# Run specific model tests
python main2.py   # DeepSeek model test
python main3.py   # Llama model test

1. Web Application (app.py): Flask server with two routes:

   • / - Serves the HTML interface
   • /generate_caption - REST API endpoint accepting base64-encoded images

2. Frontend (index.html): Simple JavaScript-based interface for image upload and caption display

3. Model Pipeline: Uses Salesforce/blip-image-captioning-large with configured generation parameters (beam search, repetition penalty, etc.)

The application is configured as a SystemD service (caption-gen.service) that:

• Runs Gunicorn with Unix socket at /home/pabrubio/.caption-gen.pabrubio.hackclub.app.webserver.sock
• Uses 1 worker with 600-second timeout
• Automatically restarts on failure
• Waits for network connectivity before starting

• main.py: Enhanced caption generation using BLIP + OPT-125M with LangChain
• main2.py: DeepSeek R1 Distill model testing
• main3.py: Meta Llama 3.2 1B Instruct model testing

• Images are sent as base64-encoded data in JSON POST requests
• The BLIP model uses specific generation parameters for quality optimization
• Error verbosity is suppressed using set_verbosity_error()
• The service runs from /home/pabrubio/pub/caption-gen in production