Professional Image Captioning Service

A streamlined image captioning service using Qwen2.5-VL-7B-Instruct for professional-grade descriptions optimized for creative industry use.

🎯 Features

• Professional Captioning: Uses OpenAI-style prompts optimized for graphic designers, photographers, and creative directors
• High-Quality Output: 30-50 word descriptions with specific brand names, proper nouns, and technical details
• Memory Efficient: 8-bit quantization enabled by default for lower VRAM usage
• FastAPI Service: Simple REST API for easy integration

🚀 Quick Start

Build and Run

# Build the Docker image
docker build -t florence-tagger .

# Run the service
docker run -p 8000:8000 \
  -e AWS_REGION=us-east-1 \
  -e VLM_LOAD_8BIT=true \
  florence-tagger

Test the Service

# Health check
curl http://localhost:8000/health

# Generate caption
curl -X POST "http://localhost:8000/caption" \
  -H "Content-Type: application/json" \
  -d '{"s3_uri": "s3://your-bucket/image.jpg"}'

📡 API Usage

Generate Caption

curl -X POST "http://localhost:8000/caption" \
  -H "Content-Type: application/json" \
  -d '{
    "s3_uri": "s3://your-bucket/image.jpg",
    "detailed": true
  }'

Response Format

{
  "caption": "Professional photographer using Canon EOS R5 camera with RF 24-70mm f/2.8L IS USM lens in modern photography studio with white seamless backdrop and Profoto lighting setup.",
  "s3_uri": "s3://your-bucket/image.jpg",
  "model": "Qwen/Qwen2.5-VL-7B-Instruct"
}

🔧 Configuration

Environment variables:

• AWS_REGION: AWS region for S3 access (default: us-east-1)
• VLM_MODEL_ID: Vision model to use (default: Qwen/Qwen2.5-VL-7B-Instruct)
• VLM_LOAD_8BIT: Enable 8-bit quantization (default: true)

🛠️ Hardware Requirements

• Minimum: 8GB RAM, 8GB VRAM
• Recommended: 16GB RAM, 10GB+ VRAM
• GPU: NVIDIA GPU with CUDA support

📊 Performance

With 8-bit quantization:

• Memory Usage: ~7GB VRAM (vs ~14GB without quantization)
• Speed: Fast inference with minimal quality loss
• Quality: Professional-grade captions matching OpenAI standards

🎯 Expected Results

The service generates captions optimized for creative industry search:

• Specific brand names and model numbers
• Proper nouns and technical details
• Professional terminology
• 30-50 words as specified

Example:

• Input: Image of photographer in studio
• Output: "Professional photographer using Canon EOS R5 camera with RF 24-70mm f/2.8L IS USM lens in modern photography studio with white seamless backdrop, Profoto A1X flash units, and Elinchrom softbox lighting setup."

🔍 Health Check

curl http://localhost:8000/health

Returns:

{
  "ok": true,
  "backend": "qwen2.5-vl-7b-instruct",
  "device": "cuda",
  "model": "Qwen/Qwen2.5-VL-7B-Instruct",
  "8bit_quantization": true
}

🚀 Production Deployment

For production use:

docker run -d -p 8000:8000 \
  --name florence-tagger \
  --restart unless-stopped \
  --memory=16g \
  --memory-swap=20g \
  -e AWS_REGION=us-east-1 \
  -e VLM_LOAD_8BIT=true \
  florence-tagger

📝 Notes

• The service uses your professional OpenAI-style prompt for optimal results
• 8-bit quantization is enabled by default to reduce memory usage
• All GPT-OSS complexity has been removed for simplicity and reliability
• Focus on Qwen2.5-VL-7B-Instruct for consistent, high-quality results