mcikalmerdeka/vlms-ocr-playground

VLMs OCR Playground

A unified Gradio-based platform for comparing and using multiple state-of-the-art Vision-Language OCR models. Deploy easily to Hugging Face Spaces.

Supported Models

1. 1.DeepSeek-OCR-2 - Document to markdown with layout detection and PDF support
2. 2.GLM-OCR - Specialized recognition for text, formulas, and tables
3. 3.PaddleOCR-VL-1.5 - Full-page document parsing with layout detection

Project Structure

.
├── main.py                 # Entry point
├── requirements.txt        # Dependencies
├── src/                    # Source code
│   ├── config.py          # Model configurations
│   ├── interface.py       # Gradio interface
│   ├── models/            # Model implementations
│   └── utils/             # Utilities
└── examples/              # Sample images

Setup

Option 1: Using pip

pip install -r requirements.txt
python main.py

Option 2: Using uv (Recommended for faster installs)

uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
uv pip install -r requirements.txt
python main.py

Option 3: Using conda

conda create -n ocr python=3.10
conda activate ocr
pip install -r requirements.txt
python main.py

Testing Without Heavy Dependencies

You can test the UI structure without downloading models:

# Test imports only
python test_imports.py

# Test UI creation (requires gradio only)
pip install gradio
python -c "from interface import create_interface; demo = create_interface()"

Usage

1. 1.Select a model from the dropdown
2. 2.Upload an image or PDF (DeepSeek only)
3. 3.Choose a task
4. 4.Click Process
5. 5.View results in tabs (Text, Markdown, Visualization, Crops, Raw)

Adding New Models

To add a new model:

1. 1.Create model class in src/models/
2. 2.Register in src/models/__init__.py
3. 3.Add config in src/config.py

See src/models/base.py for the interface.

Troubleshooting

Dependency Conflicts

If you see errors about tokenizers or transformers version conflicts:

• —The requirements.txt uses flexible version ranges to avoid conflicts
• —Remove uv.lock or poetry.lock if present and reinstall
• —Use pip install -r requirements.txt --force-reinstall if needed

Flash Attention Not Available

The models will automatically fall back to standard attention if flash-attn is not installed. This is normal and won't affect functionality.

Out of Memory

If you run out of GPU memory:

• —Run on CPU (slower but works): The models will automatically use CPU if GPU is not available
• —Process smaller images
• —Use one model at a time

Environment Variables

• —PADDLEOCR_API_URL: API endpoint for PaddleOCR (optional)
• —PADDLEOCR_TOKEN: Authentication token for PaddleOCR (optional)

License

This project combines multiple open-source OCR models. Please refer to individual model licenses.